You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by tq...@apache.org on 2025/02/17 16:20:36 UTC
(tvm) 01/01: Cleanup relay related legacy components
This is an automated email from the ASF dual-hosted git repository.
tqchen pushed a commit to branch refactor
in repository https://gitbox.apache.org/repos/asf/tvm.git
commit 37cb59046af135022a2a487b9bc21ed1fafb5fee
Author: tqchen <ti...@gmail.com>
AuthorDate: Mon Feb 17 11:00:19 2025 -0500
Cleanup relay related legacy components
---
CMakeLists.txt | 2 -
NEWS.md | 2772 --------------------
cmake/config.cmake | 3 -
cmake/modules/LibInfo.cmake | 1 -
cmake/modules/contrib/ArmComputeLib.cmake | 2 -
cmake/modules/contrib/ONNX.cmake | 22 -
conftest.py | 17 +-
docs/arch/index.rst | 4 +-
docs/arch/pass_infra.rst | 126 +-
include/tvm/ir/affine_type.h | 150 --
include/tvm/ir/expr.h | 74 +-
include/tvm/ir/function.h | 8 +-
include/tvm/ir/memory_pools.h | 361 ---
include/tvm/ir/module.h | 38 +-
include/tvm/ir/op.h | 12 +-
include/tvm/ir/tensor_type.h | 127 -
include/tvm/ir/type_functor.h | 6 +-
include/tvm/node/structural_equal.h | 6 +-
include/tvm/relax/dataflow_pattern.h | 2 +-
include/tvm/relax/expr.h | 4 +-
include/tvm/relax/type.h | 67 +-
include/tvm/runtime/logging.h | 4 +-
include/tvm/target/generic_func.h | 171 --
include/tvm/target/virtual_device.h | 6 +-
include/tvm/tir/expr.h | 4 +-
python/setup.py | 2 +-
python/tvm/contrib/cutlass/build.py | 2 +-
python/tvm/contrib/hexagon/session.py | 186 +-
python/tvm/contrib/msc/core/transform/transform.py | 4 +-
python/tvm/contrib/target/__init__.py | 16 -
python/tvm/contrib/target/coreml.py | 233 --
python/tvm/ir/__init__.py | 4 +-
python/tvm/ir/affine_type.py | 79 -
python/tvm/ir/attrs.py | 2 +-
python/tvm/ir/base.py | 6 +-
python/tvm/ir/expr.py | 20 +-
python/tvm/ir/function.py | 4 +-
python/tvm/ir/instrument.py | 5 +-
python/tvm/ir/json_compact.py | 144 -
python/tvm/ir/module.py | 6 +-
python/tvm/ir/op.py | 4 +-
python/tvm/ir/tensor_type.py | 61 -
python/tvm/ir/transform.py | 16 +-
python/tvm/ir/type.py | 4 +-
.../meta_schedule/testing/custom_builder_runner.py | 173 --
.../testing/dataset_collect_models.py | 85 -
.../meta_schedule/testing/dataset_extract_tasks.py | 104 -
.../testing/dataset_sample_candidates.py | 187 --
python/tvm/meta_schedule/testing/tune_onnx.py | 196 --
python/tvm/meta_schedule/testing/tune_relay.py | 213 --
python/tvm/meta_schedule/testing/tune_utils.py | 125 -
python/tvm/relax/__init__.py | 2 +-
python/tvm/relax/analysis/analysis.py | 2 +-
python/tvm/relax/expr.py | 2 +-
python/tvm/relax/frontend/onnx/onnx_frontend.py | 2 +-
python/tvm/relax/frontend/torch/fx_translator.py | 1 -
python/tvm/relax/op/qdq.py | 4 +-
python/tvm/relax/testing/ast_printer.py | 4 +-
python/tvm/relax/ty.py | 8 +-
python/tvm/testing/utils.py | 23 +-
python/tvm/tir/analysis/analysis.py | 28 -
python/tvm/topi/signal.py | 16 +-
python/tvm/topi/sparse_reshape.py | 10 +-
python/tvm/topi/testing/one_hot.py | 2 +-
python/tvm/topi/transform.py | 26 +-
src/contrib/msc/core/ir/graph_builder.cc | 12 +-
src/contrib/msc/core/ir/graph_builder.h | 12 +-
src/contrib/msc/core/transform/layout_utils.h | 2 +-
src/contrib/msc/core/transform/rewrite_utils.h | 2 +-
src/contrib/msc/core/transform/set_expr_name.cc | 32 +-
src/contrib/msc/core/utils.h | 2 +-
src/ir/affine_type.cc | 82 -
src/ir/diagnostic.cc | 2 -
src/ir/function.cc | 5 -
src/ir/memory_pools.cc | 242 --
src/ir/module.cc | 7 +-
src/ir/source_map.cc | 2 -
src/ir/tensor_type.cc | 66 -
src/ir/transform.cc | 7 -
src/ir/type_functor.cc | 7 -
src/relax/analysis/graph_partitioner.cc | 20 -
src/relax/analysis/graph_partitioner.h | 6 +-
src/relax/analysis/struct_info_analysis.cc | 4 +-
src/relax/ir/dataflow_matcher.cc | 2 +-
src/relax/ir/expr.cc | 2 +-
src/relax/ir/type.cc | 16 +-
src/relax/transform/fold_constant.cc | 2 +-
src/relax/transform/fuse_tir.cc | 4 +-
src/runtime/contrib/bnns/bnns_wrp.h | 2 +-
src/runtime/contrib/cblas/gemm_common.h | 2 +-
src/runtime/contrib/onnx/onnx_module.cc | 80 -
src/runtime/contrib/tensorrt/tensorrt_calibrator.h | 2 +-
src/runtime/contrib/tensorrt/tensorrt_ops.h | 2 +-
src/script/printer/ir/ir.cc | 14 +-
src/script/printer/relax/type.cc | 6 +-
src/support/libinfo.cc | 5 -
src/target/generic_func.cc | 182 --
src/te/operation/create_primfunc.h | 2 -
src/tir/ir/expr.cc | 4 +-
tests/lint/check_file_type.py | 2 -
.../python/contrib/test_hexagon/infrastructure.py | 14 -
.../contrib/test_hexagon/test_relax_integration.py | 130 +-
tests/python/ir/test_ir_type.py | 11 +-
tests/python/nightly/test_nnapi/infrastructure.py | 2 +-
.../relax/test_analysis_struct_info_analysis.py | 10 +-
.../test_analysis_suggest_layout_transforms.py | 2 +-
tests/python/relax/test_ast_printer.py | 8 +-
tests/python/relax/test_blockbuilder_core.py | 2 +-
tests/python/relax/test_dataflow_pattern.py | 4 +-
tests/python/relax/test_expr.py | 6 +-
tests/python/relax/test_struct_info.py | 4 +-
.../test_transform_merge_composite_functions.py | 1 -
tests/python/relax/test_transform_normalize.py | 2 +-
tests/scripts/task_mypy.sh | 10 -
114 files changed, 249 insertions(+), 6798 deletions(-)
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 1edb821086..24504047d8 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -108,7 +108,6 @@ tvm_option(USE_TFLITE "Build with tflite support" OFF)
tvm_option(USE_TENSORFLOW_PATH "TensorFlow root path when use TFLite" none)
tvm_option(USE_COREML "Build with coreml support" OFF)
tvm_option(USE_BNNS "Build with BNNS support" OFF)
-tvm_option(USE_TARGET_ONNX "Build with ONNX Codegen support" OFF)
tvm_option(USE_ARM_COMPUTE_LIB "Build with Arm Compute Library" OFF)
tvm_option(USE_ARM_COMPUTE_LIB_GRAPH_EXECUTOR "Build with Arm Compute Library graph executor" OFF)
tvm_option(USE_TENSORRT_CODEGEN "Build with TensorRT Codegen support" OFF)
@@ -494,7 +493,6 @@ include(cmake/modules/contrib/LibTorch.cmake)
include(cmake/modules/contrib/TFLite.cmake)
include(cmake/modules/contrib/CoreML.cmake)
include(cmake/modules/contrib/BNNS.cmake)
-include(cmake/modules/contrib/ONNX.cmake)
include(cmake/modules/contrib/ArmComputeLib.cmake)
include(cmake/modules/contrib/TensorRT.cmake)
include(cmake/modules/contrib/NNAPI.cmake)
diff --git a/NEWS.md b/NEWS.md
deleted file mode 100644
index 2b575f7aa2..0000000000
--- a/NEWS.md
+++ /dev/null
@@ -1,2772 +0,0 @@
-<!--- Licensed to the Apache Software Foundation (ASF) under one -->
-<!--- or more contributor license agreements. See the NOTICE file -->
-<!--- distributed with this work for additional information -->
-<!--- regarding copyright ownership. The ASF licenses this file -->
-<!--- to you under the Apache License, Version 2.0 (the -->
-<!--- "License"); you may not use this file except in compliance -->
-<!--- with the License. You may obtain a copy of the License at -->
-
-<!--- http://www.apache.org/licenses/LICENSE-2.0 -->
-
-<!--- Unless required by applicable law or agreed to in writing, -->
-<!--- software distributed under the License is distributed on an -->
-<!--- "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY -->
-<!--- KIND, either express or implied. See the License for the -->
-<!--- specific language governing permissions and limitations -->
-<!--- under the License. -->
-
-TVM Change Log
-==============
-
- - [On-going version](#on-going-version)
- - [0.8](#08)
- - [Accepted RFCs](#accepted-rfcs)
- - [Features and Improvements](#features-and-improvements)
- - [TE, TIR, TVMScript](#te-tir-tvmscript)
- - [AutoTVM, AutoScheduler, Meta Schedule](#autotvm-autoscheduler-meta-schedule)
- - [Operator Coverage](#operator-coverage)
- - [Training](#training)
- - [Relay](#relay)
- - [MicroTVM, AOT, Graph Executor and VM](#microtvm-aot-graph-executor-and-vm)
- - [Arithmetic Analysis](#arithmetic-analysis)
- - [Frontends](#frontends)
- - [Codegen Backends and Runtime](#codegen-backends-and-runtime)
- - [BYOC Integration with Vendor Libraries: TensorRT, ACL, VitisAI](#byoc-integration-with-vendor-libraries-tensorrt-acl-vitisai)
- - [TVMC](#tvmc)
- - [Rust Binding](#rust-binding)
- - [Misc](#misc)
- - [0.7](#07)
- - [New Features](#new-features)
- - [Automatic Scheduling (Experimental)](#automatic-scheduling-experimental)
- - [BYOC](#byoc)
- - [Operator Coverage](#operator-coverage-1)
- - [Quantization](#quantization)
- - [Relay](#relay-1)
- - [Runtime and Backend](#runtime-and-backend)
- - [Rust Support](#rust-support)
- - [TIR](#tir)
- - [TE](#te)
- - [TVMC(Experimental)](#tvmcexperimental)
- - [Feature Improvement](#feature-improvement)
- - [Accelerator and Microcontroller Support](#accelerator-and-microcontroller-support)
- - [Arithmetic Analysis](#arithmetic-analysis-1)
- - [AutoTVM and Graph Tuner](#autotvm-and-graph-tuner)
- - [BYOC](#byoc-1)
- - [Codegen](#codegen)
- - [Dynamism Support](#dynamism-support)
- - [Frontend and User Interface](#frontend-and-user-interface)
- - [Relay](#relay-2)
- - [Operator Coverage](#operator-coverage-2)
- - [Runtime and Backend](#runtime-and-backend-1)
- - [Quantization](#quantization-1)
- - [TE](#te-1)
- - [TIR](#tir-1)
- - [Performance Improvements](#performance-improvements)
- - [Documentation](#documentation)
- - [Bug Fixes](#bug-fixes)
- - [API Changes](#api-changes)
- - [Deprecation](#deprecation)
- - [0.6](#06)
- - [Relay in Production](#relay-in-production)
- - [Relay Virtual Machine](#relay-virtual-machine)
- - [Training](#training-1)
- - [Quantization](#quantization-2)
- - [Accelerator and Microcontroller Support](#accelerator-and-microcontroller-support-1)
- - [Rust Support](#rust-support-1)
- - [Operator Support](#operator-support)
- - [Frontend and User Interface](#frontend-and-user-interface-1)
- - [Runtime and Backend Support](#runtime-and-backend-support)
- - [Language and Architecture](#language-and-architecture)
- - [Symbolic shape enhancement](#symbolic-shape-enhancement)
- - [Language and Architecture](#language-and-architecture-1)
- - [Arithmetic Analysis](#arithmetic-analysis-2)
- - [Runtime and Backend Support](#runtime-and-backend-support-1)
- - [Frontend and User Interface](#frontend-and-user-interface-2)
- - [AutoTVM](#autotvm)
- - [Performance Improvements](#performance-improvements-1)
- - [Documentation](#documentation-1)
- - [Build and Test](#build-and-test)
- - [Bug Fixes](#bug-fixes-1)
- - [Known Issues](#known-issues)
- - [Depreciations](#depreciations)
- - [0.5](#05)
- - [0.4](#04)
- - [0.3](#03)
- - [0.2](#02)
- - [0.1](#01)
- - [Initial version](#initial-version)
-
-This file records the changes in TVM library in reverse chronological order.
-
-## On-going version
-
-Refer to the Roadmap issue for complete list on on-going version features.
-If you check in something that is not reflected in Roadmap issue, please reply
-to that issue so it can get added.
-
-## 0.8
-
-Apache TVM v0.8 brings several major exciting experimental features, including:
-- PaddlePaddle frontend
-- TVMScript: round-trippable python-based syntax for TIR
-- TorchScript integration
-- TensorIR scheduling language
-- TensorRT and CUTLASS integration via BYOC
-- Int4 TensorCore support in AutoTVM
-- MicroTVM Project API and Zephyr, Arduino support
-- AOT executor
-- Robost Windows support
-- Affine analysis infra: iter-affine-map
-- Improved Vulkan backend
-- CUDA graph support in TVM runtime
-
-Besides, The community has been working together to refactor and evolve the existing infrastructure, including but not limited to:
-- Relay compilation engine
-- Relay pattern language
-- CI and build process
-- Refactoring documentation and tutorials
-- Stablizing AutoScheduler
-- Stablizing TVMC command line driver interface
-- Stablizing target system
-- Frontend coverage, quantization, dynamic shape, training
-
-Full changelog: https://gist.github.com/junrushao1994/c669905dbc41edc2e691316df49d8562.
-
-### Accepted RFCs
-
-The community has adopted a [formal RFC process](https://github.com/apache/tvm-rfcs). Below is a list of the formal RFCs accepted by the community since then:
-- [RFC-0005] Meta schedule (AutoTIR)
-- [RFC-0006] Automatic mixed-precision pass and support
-- [RFC-0007] Parametrized unit tests
-- [RFC-0008] MicroTVM Project API
-- [RFC-0009] Unified static memory planner
-- [RFC-0010] Target-registered compiler flow customisation
-- [RFC-0011] Arm® Ethos-U integration
-- [RFC-0014] Pipeline executor
-- [RFC-0015] Use CMSIS-NN with TVM
-- [RFC-0019] Add PaddlePaddle frontend
-- [RFC-0020] Extend metadata in project option
-- [RFC-0022] TIR non-scalar constants
-- [RFC-0023] Adding annotation field to `tir.allocate` nodes
-- [RFC-0025] PyTorchTVM
-- [RFC-0027] Formalize TVM documentation organization
-- [RFC-0028] Command line composition from internal registry
-- [RFC-0029] Migrating target attributes to IRModule
-- [RFC-0030] Command line configuration files
-- [RFC-0031] C Device API
-- [RFC-0036] TVMScript namespace
-- [RFC-0041] Update TVMScript block syntax
-
-### Features and Improvements
-#### TE, TIR, TVMScript
-
-- TVMScript parser and printer [#7630](https://github.com/apache/tvm/pull/7630) [#9115](https://github.com/apache/tvm/pull/9115) [#9286](https://github.com/apache/tvm/pull/9286)
-- Scheduleable TIR (S-TIR) infrastructure, analysis and lowering passes [#7553](https://github.com/apache/tvm/pull/7553) [#7765](https://github.com/apache/tvm/pull/7765) [#7847](https://github.com/apache/tvm/pull/7847) [#8114](https://github.com/apache/tvm/pull/8114) [#8121](https://github.com/apache/tvm/pull/8121) [#7873](https://github.com/apache/tvm/pull/7873) [#7923](https://github.com/apache/tvm/pull/7923) [#7962](https://github.com/apache/tvm/pull/7962) [#7848](https://github.com/a [...]
-- S-TIR schedule primitives: `compute-inline`, `reverse-compute-inline`, `fuse`, `split`, `rfactor`, `storage-align`, `vectorize`, `unroll`, `bind`, `reorder`, `cache-read`, `cache-write`, `compute-at`, `reverse-compute-at`, `decompose-reduction` [#8170](https://github.com/apache/tvm/pull/8170) [#8467](https://github.com/apache/tvm/pull/8467) [#8544](https://github.com/apache/tvm/pull/8544) [#8693](https://github.com/apache/tvm/pull/8693) [#8716](https://github.com/apache/tvm/pull/8716) [...]
-- While loop in TIR [#7425](https://github.com/apache/tvm/pull/7425) [#9004](https://github.com/apache/tvm/pull/9004)
-- Metaprogramming in S-TIR via `specialize` [#8354](https://github.com/apache/tvm/pull/8354)
-- Support Return value in TIR [#7084](https://github.com/apache/tvm/pull/7084) [#7932](https://github.com/apache/tvm/pull/7932)
-- Storage scope support in `PointerType` [#8017](https://github.com/apache/tvm/pull/8017) [#8366](https://github.com/apache/tvm/pull/8366) [#8463](https://github.com/apache/tvm/pull/8463)
-- Creation of S-TIR via TE compute [#7987](https://github.com/apache/tvm/pull/7987)
-
-#### AutoTVM, AutoScheduler, Meta Schedule
-
-- PopenPoolExecutor is used to replace python native library to provide better multiprocessing support as well as enable auto-tuning in Jupyter notebooks for AutoTVM and AutoScheduler [#6959](https://github.com/apache/tvm/pull/6959) [#8492](https://github.com/apache/tvm/pull/8492) [#8913](https://github.com/apache/tvm/pull/8913) [#8820](https://github.com/apache/tvm/pull/8820) [#8851](https://github.com/apache/tvm/pull/8851)
-- AutoScheduler improvement and stabilization: task scheduler, layout rewrite, early stopping, dispatching [#6945](https://github.com/apache/tvm/pull/6945) [#6750](https://github.com/apache/tvm/pull/6750) [#6987](https://github.com/apache/tvm/pull/6987) [#7156](https://github.com/apache/tvm/pull/7156) [#8862](https://github.com/apache/tvm/pull/8862) [#8995](https://github.com/apache/tvm/pull/8995) [#7571](https://github.com/apache/tvm/pull/7571) [#7376](https://github.com/apache/tvm/pull [...]
-- AutoScheduler support for sparse workloads [#7313](https://github.com/apache/tvm/pull/7313) [#7635](https://github.com/apache/tvm/pull/7635) [#8065](https://github.com/apache/tvm/pull/8065)
-- AutoScheduler support for Vulkan, ROCm, Mali [#7626](https://github.com/apache/tvm/pull/7626) [#7038](https://github.com/apache/tvm/pull/7038) [#7132](https://github.com/apache/tvm/pull/7132)
-- AutoTVM support for int4 TensorCore [#7831](https://github.com/apache/tvm/pull/7831) [#8402](https://github.com/apache/tvm/pull/8402)
-- Meta Schedule core infrastructure, builder runner and database [#8615](https://github.com/apache/tvm/pull/8615) [#8623](https://github.com/apache/tvm/pull/8623) [#8642](https://github.com/apache/tvm/pull/8642) [#8817](https://github.com/apache/tvm/pull/8817) [#9079](https://github.com/apache/tvm/pull/9079) [#9132](https://github.com/apache/tvm/pull/9132) [#9154](https://github.com/apache/tvm/pull/9154) [#9053](https://github.com/apache/tvm/pull/9053) [#9059](https://github.com/apache/t [...]
-
-#### Operator Coverage
-- Operators for Int-8 vision transformer on GPU [#7814](https://github.com/apache/tvm/pull/7814)
-- Optimizing NMS and ROI-related kernel on GPU [#7257](https://github.com/apache/tvm/pull/7257) [#7172](https://github.com/apache/tvm/pull/7172) [#7136](https://github.com/apache/tvm/pull/7136) [#7796](https://github.com/apache/tvm/pull/7796) [#7463](https://github.com/apache/tvm/pull/7463) [#6516](https://github.com/apache/tvm/pull/6516) [#7440](https://github.com/apache/tvm/pull/7440) [#7666](https://github.com/apache/tvm/pull/7666) [#8174](https://github.com/apache/tvm/pull/8174)
-- Support and optimize sparse operators [#8605](https://github.com/apache/tvm/pull/8605) [#7477](https://github.com/apache/tvm/pull/7477) [#7435](https://github.com/apache/tvm/pull/7435) [#6889](https://github.com/apache/tvm/pull/6889) [#6580](https://github.com/apache/tvm/pull/6580) [#8437](https://github.com/apache/tvm/pull/8437)
-- Sort-related operators and optimization [#9184](https://github.com/apache/tvm/pull/9184) [#7669](https://github.com/apache/tvm/pull/7669) [#8672](https://github.com/apache/tvm/pull/8672) [#7611](https://github.com/apache/tvm/pull/7611) [#7195](https://github.com/apache/tvm/pull/7195) [#7056](https://github.com/apache/tvm/pull/7056) [#6978](https://github.com/apache/tvm/pull/6978)
-- Support for einsum operator [#6370](https://github.com/apache/tvm/pull/6370)
-- Matmul, dense operators and their optimization [#8921](https://github.com/apache/tvm/pull/8921) [#8527](https://github.com/apache/tvm/pull/8527) [#8234](https://github.com/apache/tvm/pull/8234) [#8250](https://github.com/apache/tvm/pull/8250) [#6616](https://github.com/apache/tvm/pull/6616) [#8229](https://github.com/apache/tvm/pull/8229) [#8401](https://github.com/apache/tvm/pull/8401) [#7404](https://github.com/apache/tvm/pull/7404) [#8669](https://github.com/apache/tvm/pull/8669)
-- Convolution and pooling operators and their optimization [#8620](https://github.com/apache/tvm/pull/8620) [#8936](https://github.com/apache/tvm/pull/8936) [#8584](https://github.com/apache/tvm/pull/8584) [#7075](https://github.com/apache/tvm/pull/7075) [#7142](https://github.com/apache/tvm/pull/7142) [#7515](https://github.com/apache/tvm/pull/7515) [#6999](https://github.com/apache/tvm/pull/6999) [#6899](https://github.com/apache/tvm/pull/6899) [#6840](https://github.com/apache/tvm/pul [...]
-- Scatter and gather operators and their optimization [#8479](https://github.com/apache/tvm/pull/8479) [#7600](https://github.com/apache/tvm/pull/7600) [#7044](https://github.com/apache/tvm/pull/7044) [#7464](https://github.com/apache/tvm/pull/7464) [#7233](https://github.com/apache/tvm/pull/7233) [#6533](https://github.com/apache/tvm/pull/6533) [#6856](https://github.com/apache/tvm/pull/6856) [#6854](https://github.com/apache/tvm/pull/6854) [#7927](https://github.com/apache/tvm/pull/792 [...]
-- Prefix scan, cumsum and cumprod [#7722](https://github.com/apache/tvm/pull/7722) [#7303](https://github.com/apache/tvm/pull/7303) [#7314](https://github.com/apache/tvm/pull/7314) [#7334](https://github.com/apache/tvm/pull/7334) [#7123](https://github.com/apache/tvm/pull/7123) [#6868](https://github.com/apache/tvm/pull/6868)
-- Dynamic shape and shape functions [#7414](https://github.com/apache/tvm/pull/7414) [#6979](https://github.com/apache/tvm/pull/6979) [#6912](https://github.com/apache/tvm/pull/6912) [#6898](https://github.com/apache/tvm/pull/6898) [#6373](https://github.com/apache/tvm/pull/6373) [#8068](https://github.com/apache/tvm/pull/8068) [#7490](https://github.com/apache/tvm/pull/7490) [#7487](https://github.com/apache/tvm/pull/7487)
-- Miscellaneous improvement. Operators including: reshape, resize, pad, PRNG, transpose, where, softmax, concat, nll_loss, space_to_batch_nd, batch_to_space_nd, slice_like; Libraries including thrust, cuDNN, cuBLAS, MIOpen; Improving schedules for generic reduction and softmax. [#8592](https://github.com/apache/tvm/pull/8592) [#7375](https://github.com/apache/tvm/pull/7375) [#7287](https://github.com/apache/tvm/pull/7287) [#7184](https://github.com/apache/tvm/pull/7184) [#7131](https://g [...]
-
-#### Training
-
-- Relay AutoDiff [#7677](https://github.com/apache/tvm/pull/7677) [#8318](https://github.com/apache/tvm/pull/8318)
-- TE AutoDiff [#7321](https://github.com/apache/tvm/pull/7321)
-- Gradient operators [#7685](https://github.com/apache/tvm/pull/7685) [#7340](https://github.com/apache/tvm/pull/7340) [#6767](https://github.com/apache/tvm/pull/6767) [#8307](https://github.com/apache/tvm/pull/8307) [#7357](https://github.com/apache/tvm/pull/7357) [#6827](https://github.com/apache/tvm/pull/6827)
-
-#### Relay
-
-- Pattern language and mixed-mode visitor: matching more IR constructs, fuzzy matching; converting more passes to non-recursive. [#8843](https://github.com/apache/tvm/pull/8843) [#7754](https://github.com/apache/tvm/pull/7754) [#7355](https://github.com/apache/tvm/pull/7355) [#7332](https://github.com/apache/tvm/pull/7332) [#7282](https://github.com/apache/tvm/pull/7282) [#7151](https://github.com/apache/tvm/pull/7151) [#7120](https://github.com/apache/tvm/pull/7120) [#6958](https://git [...]
-- Improving or adding passes including ExtractOperators, SimplifyExpr, DynamicToStatic, DefuseOps, ConvertLayout, FoldConstant. Added a set of utilities that allows a model to be run efficiently on TensorCores [#9253](https://github.com/apache/tvm/pull/9253) [#9245](https://github.com/apache/tvm/pull/9245) [#8996](https://github.com/apache/tvm/pull/8996) [#7827](https://github.com/apache/tvm/pull/7827) [#9034](https://github.com/apache/tvm/pull/9034) [#7807](https://github.com/apache/tvm [...]
-- TECompiler and refactoring of compilation workflow [#9103](https://github.com/apache/tvm/pull/9103) [#8974](https://github.com/apache/tvm/pull/8974) [#8886](https://github.com/apache/tvm/pull/8886) [#8802](https://github.com/apache/tvm/pull/8802) [#8501](https://github.com/apache/tvm/pull/8501) [#8526](https://github.com/apache/tvm/pull/8526) [#8486](https://github.com/apache/tvm/pull/8486) [#8597](https://github.com/apache/tvm/pull/8597) [#7518](https://github.com/apache/tvm/pull/7518 [...]
-- Quantization and automatic-mixed precision [#8883](https://github.com/apache/tvm/pull/8883) [#8810](https://github.com/apache/tvm/pull/8810) [#8644](https://github.com/apache/tvm/pull/8644) [#7613](https://github.com/apache/tvm/pull/7613) [#8069](https://github.com/apache/tvm/pull/8069) [#8341](https://github.com/apache/tvm/pull/8341) [#8126](https://github.com/apache/tvm/pull/8126) [#8460](https://github.com/apache/tvm/pull/8460)
-- Parser, printer and diagnostic [#7347](https://github.com/apache/tvm/pull/7347) [#6274](https://github.com/apache/tvm/pull/6274) [#6692](https://github.com/apache/tvm/pull/6692) [#8352](https://github.com/apache/tvm/pull/8352) [#8000](https://github.com/apache/tvm/pull/8000)
-
-#### MicroTVM, AOT, Graph Executor and VM
-
-- Pipeline Executor [#8702](https://github.com/apache/tvm/pull/8702) [#9108](https://github.com/apache/tvm/pull/9108)
-- CUDA graph integration in graph executor [#7616](https://github.com/apache/tvm/pull/7616)
-- Enable add `set_output_zero_copy` in graph executor [#8497](https://github.com/apache/tvm/pull/8497)
-- VM: memory allocation improvement, shape function improvement and misc [#7746](https://github.com/apache/tvm/pull/7746) [#7451](https://github.com/apache/tvm/pull/7451) [#7413](https://github.com/apache/tvm/pull/7413) [#7210](https://github.com/apache/tvm/pull/7210) [#8040](https://github.com/apache/tvm/pull/8040) [#6938](https://github.com/apache/tvm/pull/6938) [#8661](https://github.com/apache/tvm/pull/8661) [#7676](https://github.com/apache/tvm/pull/7676) [#8285](https://github.com/ [...]
-- AOT compilation and execution [#8697](https://github.com/apache/tvm/pull/8697) [#7785](https://github.com/apache/tvm/pull/7785) [#8014](https://github.com/apache/tvm/pull/8014) [#8023](https://github.com/apache/tvm/pull/8023) [#8096](https://github.com/apache/tvm/pull/8096) [#8075](https://github.com/apache/tvm/pull/8075)
-- Project API infrastructure: [#8380](https://github.com/apache/tvm/pull/8380) [#8963](https://github.com/apache/tvm/pull/8963) [#8708](https://github.com/apache/tvm/pull/8708) [#8019](https://github.com/apache/tvm/pull/8019)
-- MicroTVM, Zephyr, Arduino RVM, AutoTVM support [#9320](https://github.com/apache/tvm/pull/9320) [#8941](https://github.com/apache/tvm/pull/8941) [#7804](https://github.com/apache/tvm/pull/7804) [#7786](https://github.com/apache/tvm/pull/7786) [#7449](https://github.com/apache/tvm/pull/7449) [#7891](https://github.com/apache/tvm/pull/7891) [#7915](https://github.com/apache/tvm/pull/7915) [#8055](https://github.com/apache/tvm/pull/8055) [#8037](https://github.com/apache/tvm/pull/8037) [# [...]
-- The pure C runtime (CRT) [#7398](https://github.com/apache/tvm/pull/7398) [#7333](https://github.com/apache/tvm/pull/7333) [#7095](https://github.com/apache/tvm/pull/7095) [#7225](https://github.com/apache/tvm/pull/7225)
-- Model library format [#8270](https://github.com/apache/tvm/pull/8270) [#8072](https://github.com/apache/tvm/pull/8072) [#7938](https://github.com/apache/tvm/pull/7938)
-
-#### Arithmetic Analysis
-
-- Tighter bounds and more simplification on cast [#6771](https://github.com/apache/tvm/pull/6771) [#7045](https://github.com/apache/tvm/pull/7045)
-- Introducing iterator (quasi-) affine map detection [#6667](https://github.com/apache/tvm/pull/6667) [#7752](https://github.com/apache/tvm/pull/7752) [#7759](https://github.com/apache/tvm/pull/7759)
-- Inverse of iterator affine map [#8384](https://github.com/apache/tvm/pull/8384) [#8427](https://github.com/apache/tvm/pull/8427)
-- Subspace division in iterator affine map [#7760](https://github.com/apache/tvm/pull/7760)
-
-#### Frontends
-
-- PaddlePaddle initial support [#8645](https://github.com/apache/tvm/pull/8645) [#9124](https://github.com/apache/tvm/pull/9124) [#9126](https://github.com/apache/tvm/pull/9126) [#9295](https://github.com/apache/tvm/pull/9295) [#9370](https://github.com/apache/tvm/pull/9370) [#9236](https://github.com/apache/tvm/pull/9236) [#9283](https://github.com/apache/tvm/pull/9283)
-- ONNX support, including better handling of control flow, coverage of more operators, better dynamic shape support, more tests. [#9265](https://github.com/apache/tvm/pull/9265) [#9178](https://github.com/apache/tvm/pull/9178) [#9146](https://github.com/apache/tvm/pull/9146) [#8894](https://github.com/apache/tvm/pull/8894) [#8966](https://github.com/apache/tvm/pull/8966) [#8967](https://github.com/apache/tvm/pull/8967) [#7818](https://github.com/apache/tvm/pull/7818) [#9000](https://gith [...]
-- TensorFlow and TFLite, including more operators, better TensorArray support and quantization [#9404](https://github.com/apache/tvm/pull/9404) [#9256](https://github.com/apache/tvm/pull/9256) [#8689](https://github.com/apache/tvm/pull/8689) [#7789](https://github.com/apache/tvm/pull/7789) [#7736](https://github.com/apache/tvm/pull/7736) [#8763](https://github.com/apache/tvm/pull/8763) [#8647](https://github.com/apache/tvm/pull/8647) [#8648](https://github.com/apache/tvm/pull/8648) [#855 [...]
-- PyTorch: more operators including activations, inplace operators, RNNs, NMS [#9371](https://github.com/apache/tvm/pull/9371) [#9204](https://github.com/apache/tvm/pull/9204) [#9185](https://github.com/apache/tvm/pull/9185) [#9135](https://github.com/apache/tvm/pull/9135) [#9133](https://github.com/apache/tvm/pull/9133) [#9015](https://github.com/apache/tvm/pull/9015) [#8839](https://github.com/apache/tvm/pull/8839) [#8718](https://github.com/apache/tvm/pull/8718) [#8699](https://github [...]
-- MXNet support. More operators and NLP model coverage in GluonNLP [#7568](https://github.com/apache/tvm/pull/7568) [#7409](https://github.com/apache/tvm/pull/7409) [#7209](https://github.com/apache/tvm/pull/7209) [#7191](https://github.com/apache/tvm/pull/7191) [#7062](https://github.com/apache/tvm/pull/7062) [#6561](https://github.com/apache/tvm/pull/6561) [#6699](https://github.com/apache/tvm/pull/6699)
-- Misc: CoreML, Keras, DarkNet, etc. [#7667](https://github.com/apache/tvm/pull/7667) [#6676](https://github.com/apache/tvm/pull/6676) [#6651](https://github.com/apache/tvm/pull/6651) [#6963](https://github.com/apache/tvm/pull/6963) [#7949](https://github.com/apache/tvm/pull/7949) [#7035](https://github.com/apache/tvm/pull/7035) [#7446](https://github.com/apache/tvm/pull/7446) [#8562](https://github.com/apache/tvm/pull/8562) [#8599](https://github.com/apache/tvm/pull/8599)
-
-#### Codegen Backends and Runtime
-
-- LLVM backend: recover LLVM support on windows; support target feature strings in function attributes; atomic support in NVPTX, ROCm; LLVM compatibility to LLVM 12+ [#9305](https://github.com/apache/tvm/pull/9305) [#9223](https://github.com/apache/tvm/pull/9223) [#9138](https://github.com/apache/tvm/pull/9138) [#8860](https://github.com/apache/tvm/pull/8860) [#8958](https://github.com/apache/tvm/pull/8958) [#6763](https://github.com/apache/tvm/pull/6763) [#6698](https://github.com/apach [...]
-- ROCm 3.9 bitcode files search [#6865](https://github.com/apache/tvm/pull/6865)
-- Vulkan and SPIR-V refactoring and major improvement in codegen and runtime. [A critical bug fix in SPIRV codegen](https://github.com/apache/tvm/pull/8102) allows the Vulkan backend to produce correct outputs on more hardwares and drivers. Added support for querying device specific hardware parameters and capabilities, dynamic shapes, irregular ops such as sorting and NMS, UBO, fp16, and vectorization. We can now run complicated models like MaskRCNN on Vulkan end to end. [#8904](https:/ [...]
-- Metal language version upgrade (`MTLLanguageVersion2_3`), better codegen support, int64 support, various bug fixes [#7830](https://github.com/apache/tvm/pull/7830) [#7819](https://github.com/apache/tvm/pull/7819) [#7714](https://github.com/apache/tvm/pull/7714) [#7118](https://github.com/apache/tvm/pull/7118) [#7116](https://github.com/apache/tvm/pull/7116) [#7105](https://github.com/apache/tvm/pull/7105) [#7980](https://github.com/apache/tvm/pull/7980) [#8054](https://github.com/apach [...]
-- OpenCL, VTA, Verilator: refactored code generator, better error messages, various bug fixes [#7834](https://github.com/apache/tvm/pull/7834) [#7777](https://github.com/apache/tvm/pull/7777) [#7761](https://github.com/apache/tvm/pull/7761) [#7100](https://github.com/apache/tvm/pull/7100) [#6125](https://github.com/apache/tvm/pull/6125) [#6126](https://github.com/apache/tvm/pull/6126) [#6191](https://github.com/apache/tvm/pull/6191) [#7834](https://github.com/apache/tvm/pull/7834) [#8256 [...]
-- CUDA: enable `__launch_bounds__`, dynamic shared memory, TensorCore, BF16, half2, NVCC version upgrade [#9341](https://github.com/apache/tvm/pull/9341) [#8678](https://github.com/apache/tvm/pull/8678) [#7561](https://github.com/apache/tvm/pull/7561) [#7273](https://github.com/apache/tvm/pull/7273) [#7146](https://github.com/apache/tvm/pull/7146) [#7147](https://github.com/apache/tvm/pull/7147) [#7099](https://github.com/apache/tvm/pull/7099) [#7065](https://github.com/apache/tvm/pull/7 [...]
-- ARM: CMSIS-NN, Ethos-N [#8653](https://github.com/apache/tvm/pull/8653) [#7628](https://github.com/apache/tvm/pull/7628) [#8951](https://github.com/apache/tvm/pull/8951) [#7506](https://github.com/apache/tvm/pull/7506) [#7443](https://github.com/apache/tvm/pull/7443) [#7858](https://github.com/apache/tvm/pull/7858) [#6982](https://github.com/apache/tvm/pull/6982) [#8795](https://github.com/apache/tvm/pull/8795) [#8806](https://github.com/apache/tvm/pull/8806) [#8833](https://github.com [...]
-- Hexagon: build, compilation, model launcher, more target options and better runtime [#7784](https://github.com/apache/tvm/pull/7784) [#6718](https://github.com/apache/tvm/pull/6718) [#8821](https://github.com/apache/tvm/pull/8821) [#8822](https://github.com/apache/tvm/pull/8822) [#9033](https://github.com/apache/tvm/pull/9033) [#8823](https://github.com/apache/tvm/pull/8823) [#8859](https://github.com/apache/tvm/pull/8859) [#8865](https://github.com/apache/tvm/pull/8865) [#8915](https: [...]
-
-
-- WASM: Update support for latest emcc, add ffi test. [#6751](https://github.com/apache/tvm/pull/6751)
-
-#### BYOC Integration with Vendor Libraries: TensorRT, ACL, VitisAI
-
-- TensorRT initial integration, stabilization, int8 calibration, dynamism support [#6395](https://github.com/apache/tvm/pull/6395) [#7702](https://github.com/apache/tvm/pull/7702) [#7595](https://github.com/apache/tvm/pull/7595) [#7581](https://github.com/apache/tvm/pull/7581) [#7412](https://github.com/apache/tvm/pull/7412) [#7372](https://github.com/apache/tvm/pull/7372) [#9047](https://github.com/apache/tvm/pull/9047) [#8073](https://github.com/apache/tvm/pull/8073) [#8808](https://g [...]
-- Arm Compute Library (ACL) integration [#7649](https://github.com/apache/tvm/pull/7649) [#7206](https://github.com/apache/tvm/pull/7206) [#6532](https://github.com/apache/tvm/pull/6532) [#7121](https://github.com/apache/tvm/pull/7121) [#6724](https://github.com/apache/tvm/pull/6724) [#8149](https://github.com/apache/tvm/pull/8149) [#7251](https://github.com/apache/tvm/pull/7251) [#9396](https://github.com/apache/tvm/pull/9396)
-- Verilator integration [#7406](https://github.com/apache/tvm/pull/7406) [#7351](https://github.com/apache/tvm/pull/7351) [#7286](https://github.com/apache/tvm/pull/7286) [#8094](https://github.com/apache/tvm/pull/8094)
-- VitisAI integration [#6343](https://github.com/apache/tvm/pull/6343) [#7350](https://github.com/apache/tvm/pull/7350)
-- BYOC infrastructure enhancement: improving control flow, AnnotateTarget, custom codegen [#6641](https://github.com/apache/tvm/pull/6641) [#6655](https://github.com/apache/tvm/pull/6655) [#6697](https://github.com/apache/tvm/pull/6697) [#6786](https://github.com/apache/tvm/pull/6786) [#7977](https://github.com/apache/tvm/pull/7977) [#8464](https://github.com/apache/tvm/pull/8464)
-
-
-#### TVMC
-
-- MacOS support [#8396](https://github.com/apache/tvm/pull/8396)
-- AutoScheduler support [#7070](https://github.com/apache/tvm/pull/7070)
-- Support cross compiler options [#7922](https://github.com/apache/tvm/pull/7922)
-- Python scripting [#7823](https://github.com/apache/tvm/pull/7823) [#7698](https://github.com/apache/tvm/pull/7698)
-- More flexible input specification [#7366](https://github.com/apache/tvm/pull/7366) [#7788](https://github.com/apache/tvm/pull/7788)
-- More options, `--disable-pass` and `--config` [#7816](https://github.com/apache/tvm/pull/7816) [#8253](https://github.com/apache/tvm/pull/8253)
-- Allow passing optional arguments to importers [#7674](https://github.com/apache/tvm/pull/7674)
-- Model library format (MLF) support [#8086](https://github.com/apache/tvm/pull/8086) [#8331](https://github.com/apache/tvm/pull/8331)
-- More backend and library support: metal, ACL, Vulkan, OpenCL, ROCm, Vitis AI [#8282](https://github.com/apache/tvm/pull/8282) [#7508](https://github.com/apache/tvm/pull/7508) [#8359](https://github.com/apache/tvm/pull/8359) [#6831](https://github.com/apache/tvm/pull/6831) [#8896](https://github.com/apache/tvm/pull/8896) [#7577](https://github.com/apache/tvm/pull/7577)
-- Support for the new target system [#7651](https://github.com/apache/tvm/pull/7651) [#7654](https://github.com/apache/tvm/pull/7654) [#6788](https://github.com/apache/tvm/pull/6788) [#7304](https://github.com/apache/tvm/pull/7304) [#6855](https://github.com/apache/tvm/pull/6855)
-
-#### Rust Binding
-
-- Rust bindings installable via Cargo [#7503](https://github.com/apache/tvm/pull/7503) [#6678](https://github.com/apache/tvm/pull/6678) [#8631](https://github.com/apache/tvm/pull/8631) [#8665](https://github.com/apache/tvm/pull/8665)
-- Initial support for diagnostic interface [#6656](https://github.com/apache/tvm/pull/6656)
-- Fixes for using Python APIs from Rust [#7085](https://github.com/apache/tvm/pull/7085)
-- Improve NDArray, GraphRt, Relay, IRModule, Array, Attrs bindings [#6563](https://github.com/apache/tvm/pull/6563) [#6741](https://github.com/apache/tvm/pull/6741) [#7138](https://github.com/apache/tvm/pull/7138) [#8353](https://github.com/apache/tvm/pull/8353) [#7082](https://github.com/apache/tvm/pull/7082)
-- Improve error handling, error messages and fix memory leaks [#8289](https://github.com/apache/tvm/pull/8289) [#6815](https://github.com/apache/tvm/pull/6815) [#8714](https://github.com/apache/tvm/pull/8714) [#8725](https://github.com/apache/tvm/pull/8725)
-
-#### Misc
-
-- Enhanced CPP-RPC implementation: allow user supplied work dir, support of CPP-RPC server for Apple, support adb-shell style CPP-RPC [#7670](https://github.com/apache/tvm/pull/7670) [#8224](https://github.com/apache/tvm/pull/8224) [#8223](https://github.com/apache/tvm/pull/8223) [#7766](https://github.com/apache/tvm/pull/7766) [#7013](https://github.com/apache/tvm/pull/7013)
-- Use PopenWorker to handle RPC system: [#7889](https://github.com/apache/tvm/pull/7889) [#7757](https://github.com/apache/tvm/pull/7757) [#7961](https://github.com/apache/tvm/pull/7961)
-- Fold target host into target [#7462](https://github.com/apache/tvm/pull/7462) [#7791](https://github.com/apache/tvm/pull/7791) [#7534](https://github.com/apache/tvm/pull/7534) [#8835](https://github.com/apache/tvm/pull/8835)
-- Target-based intrinsic lowering and legalization [#7936](https://github.com/apache/tvm/pull/7936) [#7809](https://github.com/apache/tvm/pull/7809)
-- Add target tags for all existing CUDA GPU models [#7410](https://github.com/apache/tvm/pull/7410)
-- Linear Congruential Random Engine [#8642](https://github.com/apache/tvm/pull/8642)
-
-## 0.7
-v0.7 brings many major features. The community works together to refactor the internal code base to bring an unified IR code structure with a unified IRModule, type system and pass infrastructure. We have also bought many exciting new features, some highlights include:
-
-* Initial automatic scheduling support
-* Initial command line driver interface
-* WebGPU and webassembly support
-* Better first class rust support in the codebase
-* Intial Hexagon support
-* Bring your own codegen (BYOC) support
-
-The community also continues to bring high quality improvements to the existing modules including, but not limited to: better frontend coverage, performance, quantization, microTVM and dynamic shape support.
-
-### New Features
-#### Automatic Scheduling (Experimental)
-* Phase 0: Ansor minimum system for auto schedule generating #5962
-* Phase 1: Access Analyzer #6103
-* Phase 1: Add `follow_split` and `follow_fused_split` steps #6142
-* Phase 1: Add `pragma`/`storage_align`/`rfactor` steps #6141
-* Phase 1: Add RPC Runner #6077
-* Phase 1: Add `annotation`/`compute_at`/`compute_root`/`compute_inline` steps #6073
-* Phase 1: Add `cache_read`/`cache_write` steps #6107
-* Phase 1: Rename namspace form `auto_schedule` to `auto_scheduler` #6059
-* Phase 1: The base class for cost models #6187
-* Phase 1: feature extraction for cost models #6190
-* Phase 1: XGBoost Cost Model #6270
-* Phase 2: Basic GPU Sketch Search Policy #6269
-* Phase 2: Evolutionary Search #6310
-* Phase 2: Update heavy operations with `parallel_for` #6348
-* Parallel the InitPopulation (#6512)
-* Tutorial: Using the template-free auto-scheduler on CPU (#6488)
-
-#### BYOC
-* External codegen support in Relay (#4482), (#4544)
-* Bring Your Own Codegen Guide -- Part 1 #4602
-* Bring Your Own Codegen Guide -- Part 2 #4718
-* Relay annotation and partitioning for external compilers #4570
-* JSON Runtime with DNNL End-to-End Flow #5919
-* Handle one symbol for each runtime #5989
-* Run accelerator specific optimizations #6068
-* Arm Compute Library integration #5915
-* Retire the example json runtime #6177
-* `json_node.h` should include `data_type.h` #6224
-* Improve installation tutorial #6170
-* Add support for dense (fully connected) layer #6254
-* Introduce the Ethos-N BYOC integration #6222
-* Enable remote device via environment variables #6279
-* Improved pooling support #6248
-* Add support for quantized convolution #6335
-* CoreML codegen #5634
-
-#### Operator Coverage
-* Add `strided_set` operation (#4303)
-* Add support for conv3d (#4400), pool3d (#4478), 3d upsampling ops (#4584)
-* Add group convolution for VTA (#4421)
-* Add 1d deconvolution op (#4476)
-* Allow batch matmul to be fused into injective ops (#4537)
-* Add native depthtospace and spacetodepth operators (#4566)
-* Add CUDNN conv3d support (#4418)
-* Dilation2D operator support #5033
-* Isfinite operator #4981
-* Unravel Index operator #5082
-* Add thrust support for nms #5116
-* Resize3d, Upsample3d op support #5633
-* Add operator Correlation #5628
-* `affine_grid` and `grid_sample` #5657
-* Sparse to dense operator #5447
-* `Conv3d_transpose` op support added #5737
-* add op `crop_and_resize` #4417
-* Add bitwise ops #4815
-* Sparse to dense operator #5447
-* support dynamic NMS(Non Maximum Suppression), symbolic begin, end, and strides for strided_slice #4312
-* `Conv3d_transpose` op support added #5737
-* ReverseSequence operator #5495
-* Conv1D #4639
-* 1D Pooling #4663
-
-#### Quantization
-* Channel wise quantization - Quantize & Requantize #4629
-* Support QNN ops. #5066
-* Adding support for QNN subtract op #5153
-* TFLite QNN Tutorial #5595
-* Tutorial: Deploy Quantized Model on CUDA #4667
-* Support asymmetric per-layer quantized operators #6109
-
-#### Relay
-* Add convertlayout pass in Relay (#4335, #4600)
-* Added Merge Composite pass #4771
-* Call graph for relay #4922
-* Add inline pass #4927
-* Target annotation for external codegen #4933
-* GradientCell Relay Pass #5039
-* Add MergeCompilerRegions pass #5134
-* Non-recursive Graph Vistor and Rewriter (#4886)
-* [Blocksparse] Pipeline for lowering dense model to sparse-dense (#5377)
-* Relay op strategy #4644
-* Static Tensor Array (#5103)
-* Memory planner (part 1) #5144
-* ONNX codegen #5052
-* Add Parser 2.0 #5932, part 2 #6162
-* Basic block normal form #6152
-* Convert Layout pass. #4664
-* Pattern Language, Matcher, Rewriter, and Function Paritioner #5231
-
-#### Runtime and Backend
-* Add ADTObject POD container type (#4346)
-* TFLite RPC runtime (#4439)
-* Standardized graph runtime export (#4532)
-* MISRA-C compliant TVM runtime #3934
-* Add String container #4628
-* Introduce Virtual Memory Allocator to CRT (#5124)
-* Initial implementation of Hexagon runtime support (#5252)
-* FastRPC interface for Hexagon runtime (#5353)
-* CoreML Runtime (#5283)
-* AutoTVM + uTVM for Cortex-M7 (#5417)
-* Windows Support for cpp_rpc (#4857)
-* Implement TVMDSOOp(TensorFlow custom op) for TVM runtime (#4459)
-* WebGPU support #5545
-* TVM WebAssembly JS Runtime #5506
-* Hexagon driver for offloading kernels to simulator #5492
-* Introduce runtime::Array #5585
-* Allow non-nullable ObjectRef, introduce Optional. (#5314)
-* Introduce static slots for common objects. (#5423)
-* ntroduce RValue reference(move) support to TypedPackedFunc (#5271)
-* Introduce MetadataModule to separate code compilation/interpretation and weight initialization #5770
-* Support module based interface runtime #5753
-* Add TVM application extension with WASM runtime #5892
-* Provide guide to user who has difficulty register SEqualReduce (#5300)
-
-#### Rust Support
-* Revive the Rust + SGX refactor #4976
-* Improve Rust bindings: Map, Array, String, various IR nodes #6339
-* Rust Refactor Stage 4: Rewrite Rust graph runtime to use new APIs #5830
-* Second stage of Rust Refactor #5527
-* tvm crate stage 3 of Rust refactor #5769
-* Add first stage of updating and rewriting Rust bindings. #5526
-
-#### TIR
-* Introduce StructuralHash for the Unified IR. #5160
-* Introduce StructuralEqual Infra for the unified IR. #5154
-* Introduce ExprDeepEqual, Remove IRDeepCompare #5206
-* [TIR] Introduce BufferLoad/Store (#5205)
-* Improved massive build times caused by tir.floormod and tir.floordiv. Fixed Topi testcase. #5666
-* Buffer logger assert removed #6147
-* Enhance VerifyGPUCode #6194
-* HoistIfThenElse added #6066
-* Hybrid Script Support for TIR #6227
-* Migrate Low-level Passes to Pass Manager #5198
-* HoistIfThenElse added #6066
-* Hybrid Script Support for TIR #6227
-* Block scope hoisting added #6238
-
-#### TE
-* reverse-mode autodiff without any optimization #5121
-* Tensor Expression Debug Display (TEDD) #4651
-* Optimize and eliminate the Jacobian tensor for te.autodiff #6078
-
-#### TVMC(Experimental)
-* TVMC - A command line driver for TVM (Part 1) #6112
-* TVMC - Linting error on onnx command line driver frontend #6536
-* TVMC - Command line driver 'compile' (part 2/4) #6302
-* TVMC - Introduce 'tune' subcommand (part 3/4) #6537
-* TVMC - Introduce 'run' subcommand (part 4/4) #6578
-* TVMC - Getting started tutorial for TVMC #6597
-
-
-### Feature Improvement
-#### Accelerator and Microcontroller Support
-- Cleanup legacy verilog code (#4576)
-- uTVM support for ARM STM32F746XX boards (#4274)
-- Add --runtime=c, remove `micro_dev` target, enable LLVM backend #6145
-
-#### Arithmetic Analysis
-* Linear system and equation solver (#5171)
-* Inequalities solver #5618
-* Improve IntervalSet's floormod (#5367)
-* Remove legacy const pattern functions (#5387)
-* Handle likely in IRMutatorWithAnalyzer #5665
-* ExtendedEuclidean merge impl to int_operator #5625
-* Rewrite simplify fix for Vectorized Cooperative Fetching #5924
-
-#### AutoTVM and Graph Tuner
-* Adding ROCM schedules for TOPI (#4507)
-* NHWC conv2d schedule templates for ARM (#3859)
-* Use VM compile to extract autotvm tasks #4328
-* Download fallback schedule file if it does not exist #4671
-* Ignore error when removing tmpdir #4781
-* Fix a bug in generating the search space #4779
-* Minor bug fixes in AutoTVM for QNN graphs #4797
-* Fix autotvm customized template #5034
-* Add opt out operator for `has_multiple_inputs` for graph tuner #5000
-* Customize SI prefix in logging (#5411)
-* Update XGBoost verbosity option #5649
-* Support range in index based tuners #4870
-* Enable random fill and CPU cache flush for AutoTVM and Ansor (#6391)
-* Auto-scheduler tutorial for GPU and necessary refactor/fix (#6512)
-
-#### BYOC
-* [BYOC] Bind constant tuples in graph partitioner (#5476)
-* [BYOC] Add support for composite functions in BYOC (#5261)
-* [BYOC] Register pattern tables from external codegens (#5262)
-* [BYOC] Enhance partitioning and external codegen (#5310)
-* [BYOC] Refine AnnotateTarget and MergeCompilerRegion Passes (#5277)
-* [BYOC] Use Non-Recursive Visitor/Mutator (#5410)
-* [BYOC] Refine DNNL Codegen (#5288)
-* [BYOC] Add example of Composite + Annotate for DNNL fused op (#5272)
-* [BYOC] Prevent duplicate outputs in subgraph Tuple (#5320)
-* [BYOC] Introduce further operator support (#6355)
-* [BYOC] Support input nodes with multiple entries (#6368)
-* [BYOC] Add maximum support for float32 (#6506)
-
-#### Codegen
-* Intrinsic dispatching with OCML instead of LLVM for ROCm (#4499)
-* Make target codegen take IRModule and PrimFunc. #5107
-* Enhance CUDA codegen for SelectNode #4983
-* Vectorization for intrinsics #5101
-* [LLVM] Do not use `x86_vcvtph2ps_256` intrinsic with LLVM 11+ (#5267)
-* [LLVM] Use llvm::ElementCount with LLVM 11+ when creating vectors (#5265)
-* [LLVM] Use llvm::FunctionCallee in IRBuilder::CreateCall with LLVM 11+ (#5338)
-* [LLVM] Include Support/Host.h for declaration of getDefaultTargetTriple (#5268)
-* [LLVM] Replace calls to Type::getVectorNumElements (#5398)
-* [LLVM] Use ArrayRef in calls to CreateShuffleVector (#5399)
-* [LLVM] Use llvm::Align with LLVM 11+ to avoid warnings (#5264)
-* [CodeGen] Cleanup generated code (#5424)
-* Rename `target_id` => `target_kind` #6199
-* 64-bit RPi4b target #6211
-* Creating Target from JSON-like Configuration #6218
-* Add python binding to new JSON target construction #6315
-* Use target class in all codegens #6347
-* Initial support for Hexagon codegen #6261
-* Add --runtime=c, remove `micro_dev` target, enable LLVM backend #6145
-* Add tvm::support::hexdump() debug utility #6154
-* Adding AMD codegen unit tests (#4509)
-* Support cuda tensorcore subbyte int data type in auto tensorcore #4546
-* Handle empty LLVMModule in GetFunction #5146
-* Support int4/int8 conv2d tensor core with HWNC layout #6121
-
-#### Dynamism Support
-* Add shape function for `zero`, `zeros_like`, `ones`, `ones_like` (#4448), `tile` (#4441)
-* Support symbolic newshape for Reshape #5429
-* Support symbolic TopK, Ones, Zeros and Full #5459
-* Add `shape_of` instruction #5855
-* symbolic `max_output_size` #5844
-* Dynamic TopK Op #6008
-* Dynamic `broadcast_to`, `zeros`, `ones` #6007
-* Add dynamic reshape grad #6080
-* Keep fixed dim when unifying dynamic shape #5795
-* OneHot operation #6209
-* Add Dynamic Resize Op #6198
-* Dynamic full operator #6260
-* Dynamic upsampling relay op #6273
-* Dynamic Tile Op #5983
-
-#### Frontend and User Interface
-* TFLite parser support for `transpose_conv` (#4440), `unpack` (#4447)
-* LLDB pretty printers for relay (#4453)
-* ONNX to Relay converter op support: expand op (#4483)
-* ONNX `auto_pad` in conv and convtranspose (#4563)
-* TF to Relay converter op support (#4504) (#4551) (#4484)
-* Remove unnecessary cast of constants in ONNX converter (#4573)
-* Add support for tf.Keras networks in Relay Keras frontend #4630
-* Add conv3d #4604
-* Fix incorrect calculations in tf SLICE #4518
-* Dynamically calculate `input_stats` of any `fake_quant` range #4789
-* LSTM Support #4825
-* Add `MIRROR_PAD` operator #4822
-* use qnn helper function in softmax #4840
-* Add Resize op converter #4838
-* Add support for `TFLite_Detection_PostProcess` #4543
-* Fix tests for tflite unary elemwise operations #4913
-* GaussianDropout/Noise parsing support #4928
-* Add parser support for 'square' operator #4915
-* `make_loss` operator support #4930
-* Add parser support for `l2_normalization` #4966
-* ReadVariableOp operator support #4952
-* Check graph inputs match expected #4992
-* support multiply outputs #4980
-* TFLite: Using real image for QNN testing. #4816
-* TFLite: `FLOOR_MOD` & `FLOOR_DIV` support #4971
-* PyTorch: Upsampling op support and enable registering a user defined op conversion map #4961
-* PyTorch: fix unordered dictionary problem for python version under 3.6 #4982
-* Operator support NonZero #5073
-* Upsampling op support and enable registering a user defined op conversion map #4961
-* Check graph inputs match expected #4992
-* Add support for quantized models via QNN #4977
-* Add initial control flow support #4964
-* Remove FP32 piggy back and use QNN add/mul/concatenate #5061
-* Add missing upcast to uint8 `avg_pool` conversion #5089
-* Add initial 3D op support and test on Resnet 3D #5075
-* Fix conv2d conversion for group conv (group > 1 but != in channels) #5132
-* Add support for `max_pool1d` #5142
-* Add support for split #5174
-* `FLOOR_MOD` & `FLOOR_DIV` support #4971
-* Activation functions support #4978
-* Round op parsing support added #5022
-* DepthToSpace and SpaceToDepth support #5041
-* `TOP_K` op parser support #5051
-* ReadVariableOp operator support #4952
-* Support multiply outputs #4980
-* `reduce_any` op parsing support #4926
-* TensorFlow Parser Control Flow Enhancement #5020
-* TensorFlow Frontend support with shared params #5042
-* Support for AddV2 in Relay Tensorflow frontend converter. #5046
-* conv3d frontend operator support #5080
-* `max_pool3d` and Averagepool3d operator support #5085
-* Support for Atan/Atan2 in Relay Tensorflow frontend converter. #5104
-* Use leaky by default for LeakyReLU #5192
-* Conv3D ONNX support and `conv3D_ncdhw` x86 schedules #4949
-* Add support for FusedBatchNormV3 #5065
-* Activations for pytorch #5194
-* Dropouts And InstanceNorm support added #5203
-* [Frontend] Asymmetric padding of convolution support (#4803)
-* [ONNX]Pool3d & upsample3d op support (#5135)
-* Add TopK to ONNX Frontend (#5441)
-* Add RoiAlign to Onnx frontend (#5454)
-* [PYTORCH]AvgPool3d, MaxPool3d and Squeeze op support (#5220)
-* [PYTORCH]celu, gelu, selu activations (#5263)
-* [Pytorch]layernorm bug fix and testcase updated (#5257)
-* [PYTORCH]LayerNorm support added (#5249)
-* [PYTORCH]GroupNorm op support added (#5358)
-* [PYTORCH]Logical & Bitwise operator support (#5341)
-* [PYTORCH]Tensor creation ops support (#5347)
-* [PYTORCH]cosh,sinh,log2,log10,log1p op support (#5395)
-* [PYTORCH]Rsub, Embedded, OneHot ops support (#5434)
-* [PYTORCH]Abs, Arange, Softplus ops (#5295)
-* [PYTORCH]isNan, isinf, isfinite, ceil, clamp, round ops (#5316)
-* [PYTORCH]Activations for pytorch (#5194)
-* [PYTORCH]Repeat, Reciprocal & Reshape Op support (#5280)
-* [PYTORCH]`Reduce_ops` support added (#5308)
-* [PYTORCH]Take, Topk op support (#5332)
-* [PYTORCH]Dropouts And InstanceNorm support added (#5203)
-* [PYTORCH]Unary Ops frontend support. (#5378)
-* [Torch] Support Python list, more realistic recurrent networks (#5306)
-* [PYTORCH]where, addcdiv, addcmul op support (#5383)
-* [Torch] Add support for split (#5174)
-* [Torch] Fix up graph input handling (#5204)
-* [TFLITE]Logical not op support (#5475)
-* [TFLITE]Hard Swish & MobilnetV3 model testing (#5239)
-* [TFLITE]Gather, StridedSlice op support added (#4788)
-* [TFLITE] Match TFLite shape for SSD custom op (#5473)
-* Factor out import of common tflite.Operator in tflite frontend. (#5355)
-* [TFLite] support for FILL and `SPLIT_V` operators (#5330)
-* [TFLite] `L2_POOL_2D` operator (#5452)
-* [TFLite] Add config option to specify FlatBuffers location (#5425)
-* [TFLITE]Logical not op support (#5475)
-* [TENSORFLOW]reduce ops updated (#5180)
-* [TENSORFLOW] Fix `gather_nd` indices (#5279)
-* [TensorFlow]Improve TensorFlow Static Shape Tensor Array (#5243)
-* [KERAS]Minimum & AlphaDropout op support (#5380)
-* [KERAS]Embedding layer (#5444)
-* [KERAS]`Max_pool3d` and Averagepool3d operator support (#5085)
-* [CAFFE2]add Mul and ConvTranspose operator (#5302)
-* [MXNET]DepthToSpace & SpaceToDepth Operator (#5408)
-* [MXNET]broadcast and logical op support (#5461)
-* [MXNET] Use leaky by default for LeakyReLU (#5192)
-* [MXNET] support elemwise logic ops (#5361)
-* [Frontend|MXNet] SwapAxis operator support (#5246)
-* [RELAY] Move frontend utils (#5345)
-* [Pytorch] Fix translation of transpose when axis argument is as a list (#5451)
-* LpPool Support added #5696
-* Skip ADD inside Gemm op when vector is zero #5697
-* ReduceL1, ReduceL2, ReduceSumSquare, ReduceLogSum ops added #5721
-* MaxRoiPool, Mod & Xor op support added #5729
-* Skip multiply with 1.0f constant for GEMM import #5800
-* StatefulPartitionedCall/PartitionedCall Ops support added #5617
-* Don't add cast for batch norm when type isn't changing #5731
-* Conv3d Transpose OP added #5775
-* expand bug fix #5576
-* Support `max_pool2d_with_indices` #5549
-* Add prim::device op #5584
-* ImplicitTensorToNum support added #5603
-* Matmul fix for `batch_matmul` #5604
-* ReflectionPad2d op #5624
-* Padding op support #5638
-* Minor bug fixes #5683
-* `floor_divide` support for squeezenet #5702
-* ReplicationPad support added #5708
-* aten::norm support added #5776
-* broadcast and logical op support #5461
-* MaxPool3d and AvgPool3d Ops support added #5614
-* Softmin, trunc op support added #5715
-* conv3d and `conv3d_transpose` addedx #5814
-* Model importer to be compatible with tflite 2.1.0 #5497
-* Nit: Function names made consistent #5515
-* Select op support for tflite frontend #5486
-* `GATHER_ND` #5508
-* Quantize & Dequantize op #5394
-* Fully connected op conversion made in sync with TFLite #5510
-* `ADD_N` operator #5474
-* onnx, mxnet, pytorch mathops added #5561
-* abs, round, reciprocal, sign, softsign, `hard_sigmoid` ops support #5587
-* Gather nd bug fix for one dim support in tensorflow #5588
-* Add parser support for shape and range #5329
-* Darknet support batch size for yolo #5688
-* Improve Control Flow and TensorArray #5699
-* MXNet: Softmin, trunc op support added #5715
-* MXNet: conv3d and `conv3d_transpose` addedx #5814
-* MXNet: Add parser for `contrib.box_decode` #5967
-* Onnx: ReduceL1, ReduceL2, ReduceSumSquare, ReduceLogSum ops added #5721
-* Onnx: MaxRoiPool, Mod & Xor op support added #5729
-* Onnx: Skip multiply with 1.0f constant for GEMM import #5800
-* Onnx: Fix an issue with #5755 and add Batch norm unit tests. #5845
-* TensorFlow: StatefulPartitionedCall/PartitionedCall Ops support added #5617
-* TensorFlow: Don’t add cast for batch norm when type isn’t changing #5731
-* TensorFlow: Conv3d Transpose OP added #5775
-* Add parser support for shape and range #5329
-* Darknet support batch size for yolo #5688
-* Improve Control Flow and TensorArray #5699
-* Improve TF Parser to keep output nodes for `saved_model` #5794
-* Add parser support for `relu6`, `leaky_relu`, `relu_n1_to_1`, `log_softmax` #4805
-* Fix TF Dynamic input shape #5825
-* Support a few contrib ops in mxnet #5819
-* Improve TF Parser to keep output nodes for `saved_model` #5794
-* Add parser support for `relu6`, `leaky_relu`, `relu_n1_to_1`, `log_softmax` #4805
-* Check all unsupported ops before raising an exception #5929
-* Add Pytorch advanced indexing #6318
-* Support `index_select` #6295
-* Fix cast to long #6301
-* Fix dtype handling for modules with integer parameters #6311
-* pytorch frontend support conv1d #6203
-* Add cast to double, fix flatten conversion #6357
-* Fix aten::max and aten::min conversion #6372
-* Match pytorch 1.6 googlenet pretrained model (#6201) #6212Add unbiased variance op and corresponding support in pytorch frontend #6232
-* Implemented PADV2 Operator for TFLite and added support for constant values in PAD. #6167
-* Implemented `ONE_HOT` Operator for TFLite. #6223
-* Implemented `EXPAND_DIMS` Operator for TFLite. #6243
-* Implemented `REVERSE_V2` Operator for TFLite. #6304
-* Implemented `MATRIX_SET_DIAG` Operator for Relay/TOPI and TFLite Frontend. #6303
-* RESHAPE with dynamic shape arg in TFLite frontend #6208
-* Constant input attr added to fully connected operation in TFLite frontend #6228
-* Gather operation with indices as tensor expr in TFLite frontend #6168
-* Added support for tflite quantized maximum and minimum #6018
-* Unary ops support added in frontend #6196
-* Introduce caffe frontend for tvm #6206
-* Keras softmax and prelu fix under NHWC #6278
-* add support for MXNET numpy operators #6054
-* Refine tensorflow frontend 1.x & 2.x compatibility #6240
-* Reduceops support added to frontend #6252
-* Update precision in the ONNX `strided_slice`, update precision of ToScalar #6272
-* NHWC import support. #4899
-* Refine tensorflow frontend 1.x & 2.x compatibility #6240
-* Fix node indices attribute error for tensorflow 2.3 #6288
-* Support NMSv4 #6085
-* Support for PyTorch Non-Maximum Suppression #6314
-* ReplicationPad support added #5708
-* MXNet pre-quantized BERT #6039
-* Keep parameter names from PyTorch #5887
-* Refine LSTMBlockCell to support dynamic rnn #5963
-
-#### Relay
-* Add function attributes to IR hash (#4479)
-* Relay passes lookup overhead optimization (#4594)
-* Add `half_pixel` option to Resize op #4610
-* Skip example json runtime test when config is not set #4614
-* Test `tensor_array` in vm #4608
-* Improve `memory_allocation` pass to support multiple i/o dynamic kernels #4595
-* Add unit test for `tensor_array_split` #4619
-* Add parses support for unary elemwise ops #4634
-* Add parses support for SLICE #4502
-* Added pool autopadding and simplified converters. #4672
-* Fix meaning of `conv2d_transpose` `output_padding` parameter #4318
-* Use packed func macro for external codegen #4710
-* Fix `_parse_param` bug #4711
-* Add constant input support for elemwise ops #4666
-* Add parser support for squared difference #4652
-* Add type check to dense #4724
-* Invoke tvm::build from relay `compile_engine` and interpreter #4723
-* Broadcast condition, x, and y for Where op #4774
-* Add parser support for relational ops #4695
-* Remove duplicated BindParamByName function in VM compiler #4793
-* Use SimplifyInference for L2 Normalization. #4795
-* Expose vm OptimizeModule to Python #4800
-* Add parser support for logical operators #4642
-* Conv2D padding representation #4787
-* Add support for quantized LOGISTIC #4696
-* Fix VM compiler for while loop with free vars #4889
-* Fix bug in re-processing call node in MergeComposite pass #4879
-* Expose FunctionGetAttr to Python #4905
-* Add a PyTorch to Relay Parser #4497
-* Support data types for CSourceModuleCodegen args and output #4934
-* Clean up and refactor PyTorch frontend #4944
-* Relay pass to use fast exp/tanh #4873
-* BatchNorm support with run-time mean and variance calculation #4990
-* Reduce plevel of conv2d winograd implementation on cuda #4987
-* Add operation tan to TVM #4938
-* Outline and inline lifted functions for external codegen #4996
-* Remove primitive attribute from composite function #5014
-* Refactor Relay Python to use new FFI #5077
-* Fix relay node registration after refactor #5083
-* `Codegen_c.h` should include relay.function #5093
-* Move expr.Function to function.py #5087
-* Propagate constant to subgraphs #5094
-* Adjust strategy plevel to achieve expected performance by default #5118
-* Added a AnnotatedRegion utility class #5030
-* Support TupleGetItem in body of pattern #5106
-* Partition graph codestyle fixes #5202
-* Re-wrote the Graph Partitioner to support multiple outputs #5143
-* Fixes to MergeCompilerRegions #5195
-* Refactor build module to take IRModule #4988
-* Separate analysis and transform passes #5035
-* Relay Node::make to constructor #5128
-* relay::StructuralHash to tvm::StructuralHash #5166
-* Conditions updated to cover better user scenarios #5043
-* Replace UseDefaultCompiler with GetAttr #5088
-* Return empty CSourceModule when no `lowered_funcs` exists in Relay mod #4847
-* Clean up for memory pass to enable heterogenous execution support. (#5324)
-* Remove re-exports of tvm.transform (#5337)
-* [Refactor] Add memoized expr translator for use by backend codegen (#5325)
-* Legalize - Use Non-recursive Rewriter. (#5296)
-* Add additional check before re-using the cached match #5552
-* Remove kCompiler attr from external functions #5615
-* Pattern Language MergeComposite #5656
-* Support Tuple Output in C/DNNL Codegen #5701
-* Infer types in MergeComposite #5766
-* Convert PatternGrouper to do pre-order, non-recursive analysis #5653
-* Remove constants from partitioned functions #5663
-* Add a check for null function attributes #5674
-* Add ConstantPattern #5689
-* Conditionally Embedding Constants in Partitioned Functions #5693
-* Simplify Pattern API Implementations #5703
-* Add ShapePattern and DataTypePattern #5760
-* Remove unnecessary print #5642
-* Improve Shape Func handling for Tuple inputs #5467
-* Relay updated with String #5578
-* Fix the creation of tuple of tuples in PartitionGraph #5616
-* Preserve type information in Merge Composite #5640
-* Move `compiler_begin`/`end_op` to local static objects #5622
-* Fix `dataflow_pattern`.rewrite() hang if Match in IR #5680
-* Fix segfault in pretty print when ObjectRef is null #5681
-* Move `fallback_device` to config #5690
-* Replace `build_config` with PassContext #5698
-* Clear compile engine after task extraction #5724
-* Add `storage_order` ignore in pooling layer. #5781
-* Tweak cublas/cudnn priority level #5820
-* Skip Unknown Function Symbols #5888
-* Allow every runtime module to handle constants #5885
-* handle Tuple/TupleGetItem in first order gradient #5946
-* Add resnet-3d & Update network definitions for NHWC layout #5945
-* Use TargetNode::attrs for Target serialization #5993
-* each option of target str should only contain one ‘=’ #5988
-* Rename `target_id` => `target_kind` #6199
-* 64-bit RPi4b target #6211
-* Add resnet-3d & Update network definitions for NHWC layout #5945
-* Small bug fix for Conv1D imports. #5995
-* Move `invoke_tvm_op` and `shape_func` to vm dialect #5958
-* GRU Layer Support #6020
-* Add pass for getting calibration data from a relay module #5997
-* Merge two consecutive reshape ops #6052
-* Add operation `scatter_add` to relay, based on scatter implementation. #6030
-* i64 indices #5235
-* Port `eliminate_common_subexpr` to non-recursive form #6134
-* Fix interpreter for dyanmic shape input of `ndarray_size` #6086
-* Allow to config allocator type and refactor vm code structure #6105
-* Handle `ndarray_size` in FoldConstant #6156
-* when converting constant nodes with types of int64 or float64 #6159
-* Add ReshapeTensor instruction in the VM to replace the reshape op #6089
-* Support combine multiple dense op just into dense #6062
-* Add unbiased variance op and corresponding support in pytorch frontend #6232
-* Specify additional layouts in convert layout pass #5422
-* Safe check added for Merge Composite Call Node #5562
-* Non recursive partitioning #5493
-* Support combine multiple dense op just into dense #6062
-* Make the max number of fused ops configurable #6327
-* Implementation of the dynamic pad operator #6284
-* change device annotation from post DFS to recursive #6124
-* Make check stricter: disallow inserting function with free vars into module #6313
-* Make check stricter by using Feature. Fixed multiple bugs #6326
-* Resize support for NCHW-convertible layouts #6293
-* Make AutoDiff thread through global function #6336
-* Create Interpreter for each constant subgraph #6195
-* Add Dynamic reshape to a dynamic namespace and add DynamicToStatic Pass #5826
-* Expose relay BindParamsByName to Python #4751
-* Implement pass manager tracing API #4782
-* Move Ops in relay.op.contrib #4942
-* Conditions updated to cover better user scenarios #4951
-* [External codegen] Add test cases for fused ops with manual annotation (#4741)
-* Multiple output support, reshape, split ops added #6296
-
-#### Operator Coverage
-* Allow empty tensor for `reshape`, `tile` and `strided_slice` #4618
-* Fix meaning of `conv2d_transpose` `output_padding` parameter"; #4708
-* Remove cpp upsampling and resize op #4769
-* upsample operator 'NCHWinic' format support. #4791
-* Injective schedule improvement #4786
-* Enable vectorization on fp16 type #4867
-* Support for Int8 schedules - CUDA/x86 #5031
-* New PR to re-add tan to TVM #5025
-* Register topi schedule for Relay `fast_exp` and `fast_tanh` #5131
-* Move Dilation2d from nn to image namespace #5110
-* Use Thrust sort for argsort and topk #5097
-* Conv2d and Dense ops support on Tensor Core #5099
-* Setting workload correctly for Depthwise Spatial conv ARM. #5182
-* Adding a few missing math intrin #5011
-* Missing vectorize for depthwise conv2d. #5196
-* [TOPI] Using x86 schedules for ARM conv2d (#5334)
-* [TOPI-ARM] Do not alter layout if layout is NHWC (#5350)
-* [TOPI] Setting workload correctly for Depthwise Spatial conv ARM. (#5182)
-* [OP] Add `fast_erf` implementation (#5241)
-* [Topi] Tensorcore support for Conv3D (#5284)
-* [intrin] a few more math functions (#5468)
-* [Intrinsic] Add log1p, ldexp, atan2, hypot, nextafter, copysign (#5312)
-* [topi] Add operation relay.nn.dilate() which calls topi.nn.dilate() (#5331)
-* [Topi x86] Missing vectorize for depthwise conv2d. (#5196)
-* [TOPI x86] Adding `unroll_kw` config option for depthwise conv2d. (#5197)
-* [Topi] Breakdown topi.cc into smaller files (#5253)
-* ReduceLogSumExp Operator support #5453
-* Math ops added #5502
-* Enable blocking format in x86 conv2d and fold scale axis #5357
-* Add operation gather to relay. #5716
-* Add `storage_order` ignore in pooling layer. #5781
-* Fix bifrost spatial packing conv2d auto tune #5684
-* Fix reshape usage in ARM schedule #5732
-* Block sparse dense on cuda #5746
-* Improve CUDA softmax scheduling #5600
-* block sparse dense on cuda #5746
-* pass-by-value -> pass-by-const-reference #5783
-* Using MKL blas for quantized dense #6115
-* topi -> tvm/topi #6186
-* Use auto-tuner to improve `conv2d_gemm` performance #6117
-* Improve CUDA `conv2d_transpose_nchw` #4762
-* Add CUDA conv2d for NHWC layout #4737
-* `conv3d_ndhwc` schedule #4775
-* Fast exponent #4790
-* Add Scatter to Topi/Relay/ONNX via hybrid script #5619
-* Split MKL from BLAS. #6182
-* Change the meaning of `conv3d_transpose` `output_padding` to match `conv{1,2}d_transpose` #6065
-* Gather op support added #6013
-
-#### Runtime and Backend
-* Cythonize NDArray.copyto (#4549)
-* Unified Object System runtime refactor (#4578, #4581, #4603)
-* VM profiler: sort VM stats by time (#4601)
-* Update RPC runtime to allow remote module as arg (#4462)
-* Refactorying system lib and dso lib into library module (#4481)
-* Improve TSIM virtual memory mapping (#4545)
-* make adt tag signed #4605
-* Improve TVMBackendPackedCFunc to allow return val #4637
-* EdgeTPU runtime for Coral Boards #4698
-* Fix memory leak when using openMP #4811
-* Fix memory leakage of TVMByteArray #4856
-* Fix `TVM_DLL_EXPORT_TYPED_FUNC` to work on Windows #4955
-* Fix memory leak when using openMP #4811
-* Export GraphRuntime in `tvm_runtime.dll` #5002
-* MISRA-C compliant TVM runtime #3934
-* Update the `type_keys` to reflect the code-org #5074
-* Fix AttrEqual for Array and StrMap, double #5054
-* Export GraphRuntime in `tvm_runtime.dll` #5002
-* Fix unused-value warning #5140
-* crt error handling #5147
-* Bundle deployment with static linking #5158
-* Implemented kDLCPUPinned (cudaMallocHost) #4985
-* Explicitly cast min/max operands #5090
-* `ref_counter` -> `ref_counter_` #5184
-* Expose runtime::String to Python (#5212)
-* [FFI] Refactor runtime.String to subclass str (#5426)
-* [RUNTIME] Auto conversion from str to runtime::String in PackedFUnc (#5251)
-* [RUNTIME] Improved Packed FFI for optional. (#5478)
-* [Hexagon] Add `hexagon_posix.cc` to TVM/RT sources in the right place (#5346)
-* [FFI] Refactor runtime.String to subclass str (#5426)
-* Fix workspace #5503
-* Store nullptr PackedFunc as nullptr for better error propagation #5540
-* Improve PackedFunc robustness #5517
-* Seg fault in WorkspacePool's destructor (#5632) #5636
-* Resolve constexpr issue in debug mode. #5651
-* Add `compile_shared` option to linux compile utility fn #5751
-* Call sync in CopyFromRemote and CopyToRemote #5512
-* Fix the multihop cpu case #5522
-* Improve RPCServer AsyncIO support. #5544
-* Modularize the RPC infra #5484
-* Add `compile_shared` option to linux compile utility fn #5751
-* Overload string operators #5806
-* Only initialize required module #5926
-* if a param not in input, we should still consume it’s data #5990
-* init TVMPackedFunc’s name #6044
-* Enable auto conversion `String->DLDataType` #6214
-* Support random fill #5913
-* Use new to avoid exit-time de-allocation order #6292
-* Add `parallel_for` support to run a loop in parallel #6275
-* Solve ARM BIG.LITTLE heterogeneous multicores #4747
-* [RUNTIME] Quick fix PackedFunc String passing (#5266)
-* Introduce runtime::String::CanConvertFrom #5718
-* Restore the StrMap behavior in JSON/SHash/SEqual #5719
-* Support overriding RPCWatchdog termination behavior on Android and other platforms #6216
-* Set `NDArray::Container.shape_` in NDArray::FromDLPack (#5301)
-* Enable x86 cpu cache flush #5914
-
-#### Quantization
-* Conv2D type checking for kernel per-channel scales. #4732
-* Add missing nullptr check #4773
-* Doc fix on convolution and dequantize #4799
-* Conv2D with dilation support. #4796
-* Making `scale`/`zero_points` as expr instead of attrs. #4611
-* Make calibration faster and more memory usage friendly #4589
-* Doc fix on convolution and dequantize #4799
-* Conv2D with dilation support. #4796
-* Optimize lowering for requantize and FixedPointMultiply. #4798
-* More doc fix on quantize and convolution #4874
-* Add support for per channel weight scale in dense op #4880
-* Add support for quantized models via QNN #4977 #5013
-* Support 4D padding. #5036
-* [Requantize] Cleanup and Optimize Lowering (#5286)
-* [Topi, ARM] Disbale Winograd for quantized tensors. (#5363)
-* Adding support for TFLite QnnSubtract operator. (#5230)
-* Remove developer facing api from frontend exports. (#5375)
-* Add Quantize/Dequantize Partitioning #5940
-* Add support for quantized models via QNN #5016
-* Quanitze operation expanded to take const argument #6127
-* FP32 and Quantized Object Detection Model #5479
-* Support CallNode inputs in qnn.concatenate #5360
-* QNN support for TFLite 2.1.0 quantized models #5848
-
-#### TE
-* Tighten split's extent #4931
-* Set split node's range to minimum of ext and split factor or split np… #5044
-* Support mixing normal and cross-thread reduction (#5193)
-* Inline -> `te/schedule/operation_inline.h` (#5386)
-* Create loops according to storage scope and thread hierarchies (#5190)
-* Fix import in dump pass ir (#5327)
-* Scalar support for te.extern #6079
-
-#### TIR
-* IR readability enhancement (#4501)
-* Introduce tir::PrimFunc #5070
-* Introduce PrimFuncPass. #5139
-* [TIR] Enhance Substitute, python bindings for Substitute/PostOrderVisit (#5400)
-* [TIR] Remove ProducerConsumer and `AllocateNode::new_expr` (#5333)
-* [TRANSFORM] Enable CopyOnWrite for TIR passes. (#5309)
-* [REFACTOR] Migrate LowerTVMBuiltin, InferFragment, LowerThreadAllreduce, ThreadSync to Pass Manager (#5213)
-* [REFACTOR] Remove te::Tensor dependencies from TIR passes. (#5372)
-* [TIR] Refactor MakePackedAPI to target dependent stage. (#5326)
-* [REFACTOR] tvm.hybrid -> te.hybrid (#5223)
-* [REFACTOR] Migrate most of low-level build to use the Pass Manager. (#5225)
-* [REFACTOR] Migrate low-level passes in tvm.lower to the Pass Manager (#5364)
-* [TIR] Migrate VTA TIR passes to the new pass manager. (#5397)
-* [REFACTOR] Migrate all low-level passes to the Pass Manager. (#5233)
-* [REFACTOR] Introduce ExprDeepEqual, Remove IRDeepCompare (#5206)
-* [REFACTOR] RewriteForTensorCore -> te/schedule (#5379)
-* [REFACTOR] Remove `ir_pass` in favor of analysis/transform. (#5415)
-* text format printer considering future parsing use #5483
-* Remove buffer params from pass config. #5652
-* std::string -> String Migration in TIR nodes #5596
-* Remove `CallNode.call_type` in favor of attribute. #5937
-* Remove legacy HoistIfThenElse #5944
-* Improve Let/LetStmt support. #5949
-* Refine side effect analysis. #5954
-* `Provide->ProducerStore`, `Realize->ProducerRealize`. #5750
-* Migrate the tvm/tir/expr.h to constructor #5773
-* Migrate tir/stmt.h to use constructor. #5778
-* Cleanup unused classes #5789
-* Add tir prefix to type keys #5802
-* Enhance VerifyGPUCode #6194
-* Enforce buffer pointer var type to be consistent with dtype. #6317
-* Create a StringImm reference type #4806
-* Add init member to ReduceNode #6138
-* Add dump and print for debugging (NFC) #5207
-* Streamline Function Attr interface. #5045
-* `alpha_equal` to `structural_equal` #5161
-* Remove AttrsEqual and AttrsHash related code #5169
-* [NODE] General serialzation of leaf objects into bytes. (#5299)
-* [POC] Initial stab at `std::string->String` upgrade (#5438)
-* [TIR] Make `lower_warp_memory` support `extent(threadIdx.x) < warp_size` (#5307)
-* [PASS] dtype rewrite for indexing variables (#5092)
-* [PYTHON] Enhance `with_attr` API, cleanup MakeAPILegacy in testcases (#5335)
-* [PYTHON] Make IntImm more like an integer (#5232)
-* [IR] Move to runtime::String (#5276)
-* [IR] kExternalSymbol -> kGlobalSymbol (#5211)
-* [IR] Remove PrimExpr from String (#5311)
-* IRModule is updated with String #5523
-* IR is updated with String #5547
-* Streamline ir/op Registry #5609
-* Migrate IRModule ObjectRef to not-null #5654
-* Migrate BuildConfig to PassContext. #5668
-* relay.op.Op -> tvm.ir.Op #5705
-* Separate ArgTypeCode from DLDataTypeCode #5730
-* Remove legacy `compute_expr.h` #5738
-* Call::Halide => ProducerLoad, DSL/TIR decouple. #5743
-* `Provide->ProducerStore`, `Realize->ProducerRealize`. #5750
-* Migrate the tvm/tir/expr.h to constructor #5773
-* Migrate tir/stmt.h to use constructor. #5778
-* Migrate all Object construction to constructor. #5784
-* Cleanup unused classes #5789
-* Finish `std::string->String` updates #5793
-* Add tir prefix to type keys #5802
-* Change Call.name to Call.op(RelayExpr) #5863
-* Range/IntSet API style consistency. #5953
-* Separate ArgTypeCode from DLDataTypeCode #5730
-* Migrate all Object construction to constructor. #5784
-* Finish `std::string->String` updates #5793
-* Unify StrMapNode and MapNode #5687
-
-#### Performance Improvements
-* Int8 GEMM performance enhancement using Cublas (#4550)
-* Speedup TSIM with multi-threading (#4491)
-* Support cudnn softmax (#5214)
-* Add cuDNN grouped convolution support (#5319)
-* Winograd support for Conv3D (#5186)
-* Improve `get_valid_count` and nms performance for CUDA (#5339)
-* Optimizations of `global_ave_pool` for NHWC layout (#5450)
-* Optimization of Conv2d Winograd algorithm on Tensor #5485
-* Some performance improvement to VM #5901
-* Optimize x86 `conv3d_ndhwc` using data packing approach. #4866
-* Improve NHWC depthwise convolution for AArch64 #6095
-* Improve quantized convolution performance for armv8 architectures #5754
-
-#### Documentation
-* Adding benchmark log format doc (#4366)
-* Add Ninja build system to installation docs (#4554)
-* Doc/comment fixes (#4452, #4463, #4469, #4493, #4397, #4580, #4585, #4591)
-* Fix doc after moving to unified IR #4835
-* Introduction to module serialization #4564
-* ConvertLayout - Call RemoveUnunsedFunctions. #4834
-* Fix bugs that override `n_trials` #4842
-* Update the vm doc #4868
-* Refine the example description of `max/min/sum/tag_scope` #4974
-* Fix vta tutorial #4809
-* Introduce how to add hardware backend to FAQ #4898
-* Update API docs to reflect the status after the refactor. #4907
-* Fix sphinx warnings #4917
-* Fix Sphinx Warnings (RST indent, cross-ref, and image scale) #4920
-* Fix Sphinx Warning: the target found for cross-reference #4925
-* Sphinx -- Introduce alias detection. #4954
-* Fix Warnings from #4942 #4959
-* Fix sphinx precheck #4967
-* Move `git_howto` to rst, add Stage documents to te #5055
-* Add doc for Relay op strategy #5078
-* Update relay docs #5112
-* Include a tarball of docs, add a security faq #5119
-* Cleanup docs before rebuild #5127
-* Minimize necessary doc change #5129
-* Various sphinx related fix. #5168
-* Point docs to the ASF site. #5178
-* Use https link #5183
-* Reduce artifcats generated by sphinx gallery #5208
-* Refine the example description of `max/min/sum/tag_scope` #4974
-* Description updated for pooling attributes #5091
-* [DOCS] Migrate some markdowns to rst, fix sphinx3 warnings (#5416)
-* [DOCS] Misc docs improvements (#5222)
-* [DOCS] Bring relay docs to the top-level flat view (#5343)
-* [DOCS] Reduce artifcats generated by sphinx gallery (#5208)
-* [DOCS] Use https link (#5183)
-* [DOCSTRING]missing function parameters updated (#5228)
-* [DOCS] Migrate HLS documents from md to rst (#5419)
-* [Tutorial, QNN] Add tutorial for loading quantized PyTorch model (#5321)
-* [Docs] VTA install doc migration from md to rst (#5442)
-* [Docs] compiler version in docs (#5281)
-* Remove legacy `compute_expr.h` #5738
-* `TVM_REGISTER_API` -> `TVM_REGISTER_GLOBAL` #4768
-
-#### Bug Fixes
-* Add bfloat16 typeflag support (#4525)
-* MSVC / Windows fixes (#4455, #4569)
-* Fix Makefile for `howto_deploy` (#4457)
-* Fix GCC 4.8 compact (#4461)
-* Fix search path to build `libtvm_topi.so` (#4467)
-* Fix for `conv2d_transpose` CUDA compilation (#4472)
-* Fix for LLVM 10.0 codegen (#4480, #4515)
-* Fix alter op layout when calling global var (#4454)
-* Fix `float2half_rn` support for cuda compute capabilities < 53 (#4489)
-* Fix compile errors for OpenCL backends (#4492)
-* Fix serialization precision loss (#4503)
-* Fix hybrid script to support array of tensors (#4494)
-* Fix annotation for multiply op (#4458)
-* Fix Dockerfile for linter CI (#4506)
-* Fix TF resize for dynamic size models (#4510)
-* Fix `bias_add` gradient (#4516)
-* Fix tanH unit test function call (#4517)
-* Fix extra reshape parameter for ONNX (#4524)
-* Fix crash caused by empty TOPI config (#4520)
-* Fix ONNX shape op type to use int64 (#4528)
-* Fix crash in TSIM virtual memory driver (#4527)
-* Replace deprecated python library in setup script (#4533)
-* Fix NMS `max_output_size` loop (#4541)
-* Fix style in IR mutator and IR visitor (#4561)
-* Fix compiler warning (#4559)
-* Fix to get end to end inference on Chisel VTA (#4574)
-* Fix LLVM build by adding missing intrinsics headers (#4575)
-* Fix context creation in quantization (#4582)
-* Fix NDArray SaveDLTensor signature (#4586)
-* Fix dense pack schedule for x86 (#4539)
-* Fix for broadcast tensor of scalar type (#4577)
-* Datatype refactor (#4513, #4560)
-* Add const qualifiers for NDArray container (#4590)
-* Fix TF <= 1.12 compatibility (#4593)
-* Fix for graph debug runtime (#4598)
-* Disable copy constructor for external codegen (#4597)
-* Make ADT tag signed (#4605)
-* Added declare of aluBits for TensorAlu #4624
-* Get around limitation of g++-4.8 #4626
-* Bugfix StmtMutator IfThenElse #4609
-* Remove unecessary rdynamic #4613
-* Resolve constexpr related link error in debug mode #4641
-* Asymmetric padding #4511
-* Reduce data size of asymmetric padding testcase #4658
-* Fix Base64OutStream portability issue #4668
-* Fix `topi.nn.global_pool` layout="NHWC" #4656
-* Also package core.rly #4679
-* fskip of EliminateCommonSubexpr cannot always return false #4620
-* Fix Python syntax error in `start_rpc_server_to_tracker.py` #4682
-* os.path --> osp to match the import #4681
-* GitHub actions/checkout@v1 --> v2 #4680
-* Fix Python syntax error AGAIN in `start_rpc_server_to_tracker.py` #4685
-* Use ==/!= to compare str, bytes, and int literals #4686
-* Rename `start_rpc_server_to_tracker.py` to `start_rpc_server_to_tracker.sh` #4689
-* GitHub Action lint Python code for syntax errors #4688
-* Generate blob use LLVM directly #4657
-* Reduce input size to fix oom #4653
-* Fix RemoveUnusedFunctions pass #4700
-* Link the math library by default #4713
-* Update mainline version to 0.7.dev0 #4720
-* Add SizeVar representing non-neg valued variable in a tensor shape #4684
-* Fix the compile problem of `cpp_rpc` #4725
-* JSON upgrader to upgrade serialized json. #4730
-* Fallback schedule for Int8 depthwise. #4733
-* Fix dense x86 schedule #4728
-* Fix demo dockerfile build failed #4744
-* Improve CUDA vectorizer #4736
-* Add .asf.yaml for github info #4761
-* Fix padding in pooling op #4738
-* Remove `run_infer_type` duplicates #4766
-* pooling.cc improvements #4767
-* Export `builtin_fp16` on Windows #4731
-* Fix Tensorflow conv3d pad bug, add non-cubic data and kernel tests #4772
-* Bump prebuilt-image version in demo dockerfile #4770
-* Update `tune_simple_template.py` #4778
-* Explicitly link to cublasLt if it exists #4776
-* Fix hasattr by extracting Python error type from Windows error message #4780
-* Replace os.path.exists with try...except...else #4784
-* Make sure to visit the arguments of inlined functions #4783
-* Parse additional exception strings #4785
-* Fix #4670: add bias for fc layer #4801
-* Change color channel from BGR to RGB for darknet preprocessing #4794
-* Fix -Wextra #4804
-* Fix vta tutorial #4809
-* Minor bug fixes in AutoTVM for QNN graphs #4797
-* Fixed subprocess creation under windows #4820
-* Improve tol to resolve flaky case #4836
-* Fixed process termination routine in windows #4844
-* `test_cuddn` flaky #4846
-* Mxnet parser for Qnn dialect #4714
-* Enhance `cc.cross_compiler` #4817
-* Fixed crash caused by reversing bitwise operations #4852
-* Reverse some changes made for `intel_graphics/conv2d.py` in PR #4849 #4853
-* const auto p -> const auto& p #4861
-* Fix onnx import bugs #4750
-* Explicit llvm::StringRef to std::string conversion #4859
-* Update the runtime PackedFunc for module #4871
-* Improve antlr import error message #4888
-* Fix `alpha_equal` bug for attribute check #4897
-* Fix issues in cuda codegen #4876
-* Fixed: Bitwise ops on floats causing wrong code generation and crashes. #4892
-* Fix `tvm.target.generic_func` runtime detection #4910
-* `topi/tests/python/test_topi_sort.py::test_argsort` #4891
-* Use opencv reisze method for preprocessing of image in darknet #4883
-* Fix build breaks with StringRef changes #4923
-* Remove unnecessary spliting in the cached chunk #4935
-* Fixing an Infinite Loop case in UnmatchedChecker. #4881
-* Remove SGX toolchain installation from CI Dockerfile #4948
-* Fix tedd tutorial after strategy change #4947
-* Allow customize MKLDNN library location #4814
-* Added CopyFromBytes and CopyToBytes convenience methods to NDArray. Fixed typos. #4970
-* Fix gcn tutorial failure #4994
-* Fix stride default value None in torch.nn.functional.avg_pool #4984
-* Fix ROCm strategy for winograd conv selection #5001
-* Fix `get_valid_count` flaky test for cuda #4901
-* Change Scala Linter scalafmt => scalastyle #4998
-* Kill from tvm import te #5007
-* Chisel fixes and de10nano support #4986
-* Fix gpu not found when running TVM docker #4975
-* Fixes for pylint==2.4.4 #4849
-* Fix unordered dictionary problem for python version under 3.6 #4982
-* Fix gcn tutorial failure #4994
-* Fix stride default value None in `torch.nn.functional.avg_pool` #4984
-* Fix ROCm strategy for winograd conv selection #5001
-* Early checking added and new test cases added for schedule fuse #5010
-* Fixed div by zero core dump. Fixed rounding intrinsics on int crash #5026
-* Test case modified for int type #5012
-* Bug Fix for ARM CPUs. Lower strict assumption. #5063
-* Triage the testcases to fit the new namespaces #5071
-* Add colors to `compute_at` edges and thread/block indices. #5111
-* Temporary fix to the stack overflow issue in autotvm task extraction #5019
-* Fix compilation of If-Elses #5040
-* Fix CompilerAttrs #5109
-* Fix the existing test cases before refactoring. #5122
-* Fixed bug where shifting by out-of-bounds value results in no compute code being emitted. #5115
-* Fix for issue #4831. The `data_min_idx` and `data_max_idx` were flipped. #5136
-* Duplicate likely nodes added when loop axis split unevenly #5084
-* Fix incorrect name of calibration mode #5150
-* Remove contrib spatial pack schedule of depthwise convolution #5148
-* Fix annotate pass static variable #5023
-* Fixed ConvTranspose2D parsing #5157
-* Nullptr check #5176
-* rocm: fix miopen convolutions #5179
-* rocm: fix `dense_rocblas` in strategy, topi #5191
-* Fix CRT static test bug (#5293)
-* Fix perf regression of tir refactor (#5258)
-* Bugfix in tensorflow `space_to_batch_nd` (#5175)
-* Compilation warnings fixed for 32bit and 64bit compilation (#5349)
-* Fix hang in MergeCompilerRegions (#5227)
-* Fixes to MergeCompilerRegions (#5195)
-* Fix generation of LLVM intrinsics (#5282)
-* Fix setting up hints for getaddrinfo (#2872)
-* Add ConstantNode to IsAtomic (#5457)
-* Fix String SEqual (#5275)
-* Fix fuse over functions that are handled by external codegen (#5365)
-* Fix memory leak when accessing NDArray (#5413)
-* Remove the duplicate PrintIR pass in Relay (#5403)
-* Fix `lower_warp_memory` (#5247)
-* Fix `lower_warp_memory` when there are >1 warp buffers (#5368)
-* Fix intel conv2d auto tune (#5200)
-* Fix FuseBatchNorm output cast error if `need_cast` is True #4894
-* Fix an assertion exposed by loop vectorizer #4916
-* Fix error message #4945
-* Fix for recursive let #5757
-* Fix Calibration Pass to Support Modules with Multiple Functions #5768
-* Fix what looks like bizzare copy-paste issue #6010
-* Fix bug in `transpose_shape_func` #6180
-* Fix bugs in CUDA codegen (#5209)
-* Don’t remove() TemporaryFile in del. (#5414)
-* Fix `test_ir_type`. (#5390)
-* Fix multiple identical inputs bug (#5389)
-* Add cuda target check to dense tensorcore schedule. (#5376)
-* T2 test fixups (#5391)
-* Fix miopen padding (#5433)
-* Misc fixes for ROCm (#5431)
-* Fix copy constructor (#5237)
-* Corrected TVM autotuning on GPU (#5432)
-* Fix vector load (#5226)
-* Minor bugfix in `message_passing.cc` (#5254)
-* Fix a bug when vectorized load&store was involved for… (#5428)
-* Fix to skip node not in graph. (#5238)
-* Fix #5388 [VULKAN] vkBuffer released before memory copy command se… (#5418)
-* Fix a minor error in `device_annotation` (#5291)
-* Fix scalar’s ndim is 0 (#5344)
-* Fix the runtime raise error #5586
-* Fixed bug in attribute parsing for pool layers. #5582
-* AutoTVM incorrect measurement #5511
-* fix a min/max simplify bug #5761
-* Rename `tvm_dso_op` to `libtvm_dso_op` #5714
-* Fix generating types like float44 and float88 #5722
-* Avoid downloading when `TOPHUB_LOCATION` is NONE #5720
-* codegen llvm: move nvptx-specific intrinsic handling into `codegen_nvptx` #5726
-* ROCm warp shuffles and reductions #5727
-* fix small bug about `dense_grad` #5695
-* Clarify downstream consistency of TVMArgTypeCode #5742
-* Fix gelu in PyTorch frontend, tighten numerical checks #5763
-* Make batch matrix multiplication on GPU tunable #5752
-* update vulkan build rule #5777
-* aten::norm support added #5776
-* Edit onnx parser to infer values in post order #5755
-* Support symbolic inputs of Fill #5762
-* support `aten::type_as` in the pytorch frontend #5787
-* Temporary disable fp16 `type_as` test for PyTorch Frontend #5799
-* Add config switch for nn.dense layer type. #5801
-* Move cpu-only frontend tests to a CPU stage #5807
-* Pin hand landmark network to version 0.7.4. #5813
-* Limit number of threads in all jobs #5815
-* Error msg update #5818
-* fix relay.build to not change the module argument in place #5822
-* Fix InferType when module contains Prelude #5797
-* Add a combine `batch_matmul` pass #5791
-* RepeatVector, Conv3DTranspose op support added #5833
-* Fix converting serialized quantized models #5839
-* ffi (Object): make class dict visible in instances #5843
-* Additional canonicalization added for AddNode #5846
-* Suppress the warning messages when compile engine selects impls #5821
-* fix #5849 #5851
-* Introduce POD-C Compliant tvm::Map #5740
-* Add bfloat16 #5601
-* Add Python Classes for all Attrs #5853
-* Fix map assign issue in CI test #5854
-* Introduce Target Id Registry #5838
-* Update `has_dtype/has_shape` to pattern lang doc #5847
-* Add `nn.batch_flatten` as quantizable. #5805
-* Fail early before running invalid dynamic graphs #5856
-* Improve type handling in PyTorch frontend #5834
-* HotFix the python intrin rule #5895
-* add a few gradients #5899
-* Add Binary Intrinsic ops to TIR Ops in C++ #5900
-* Allow implicit conversion in TVM FFI to tvm::Bool #5907
-* PyTorch frontend: fix handling of duplicate use of a model weight #5897
-* Don’t multiply by constant 1 uselessly in dense #5911
-* Support any index matching for TupleGetItem #5909
-* Add MicroTVM tutorial using the STM32F746 discovery board #5655
-* Fix serialization of inf float value #5912
-* Fix CPU Thread Binding for Multiple Sockets #5918
-* CUDA device API & VerifyGPUCode pass update #5898
-* Update install.rst #5858
-* Two small fixes to AMDCPU codegen for LLVM 10+ and ROCm 3.5+ #5920
-* Add LegalizeInvalidAttach to legalize the `compute_at` location after split or fuse #591
-* Don’t rewrite expressions used outside of the pattern #5930
-* Add TupleGetItem to CSE #5931
-* Various update for CoreML codegen #5934
-* Update date in the NOTICE #5943
-* Raise right error in tensorflow split op #5951
-* Add rm xla attributes in tf docs #5950
-* Fix OpenCL `get_valid_counts` errors due to intrinsic `atomic_add` #5857
-* Amendments for gradients #5941
-* Fix the meaning of `conv{1,2}d_transpose` `output_padding` parameter. #5758
-* Make first order gradient graphs more efficient #5959
-* Raise an exception when extern function does not return Stmt #5964
-* Improve docker/bash.sh to handle git worktrees #5970
-* Install DNNL (OneDNN) to CI Environment #5936
-* Add Dynamic reshape to a dynamic namespace and add DynamicToStatic Pass #5826
-* Add meshgrid op in Relay, TOPI, Pytorch frontend #5961
-* Print right number of parentheses for LoadNode #5965
-* Migrate data structure of TargetNode #5960
-* Remove redundant function CreateBufferVecPtr #5982
-* Fix string argument mismatch in GraphRuntimeCodegen #5933
-* VectorType::get with two parameters is deprecated in LLVM 11+ #5984
-* Fix Compilation Error in CRT #5713
-* Fix runtime::String backward compatibility in JSON #5725
-* Allow RPCWrappedFunc to rewrite runtime::String as std::string #5796
-* Fix reshape #5739
-* Fix building with LLVM-10 on macOS #5859
-* Add cuda 11 to `contrib.nvcc.find_libdevice_path()` #5902
-* Fix sequential cpp test #5745
-* Infer types in MergeComposite #5766
-* Fix recursive let for well formed check #5780
-* Recover global state after `test_util.py` #5824
-* Fix bug in rpc ring buffer shrink #5516
-* Fix remote device sync #5538
-* Fix bug in rpc ring buffer shrink (#5516) #5537
-* RPC Server error fix on Pynq FPGA #5607
-* Fix FloorMod Simplifier #5509
-* Fix Python debugger segfaults with TVM built with LLVM #5685
-* Fix Compilation Error in CRT #5713
-* Fix runtime::String backward compatibility in JSON #5725
-* Allow RPCWrappedFunc to rewrite runtime::String as std::string #5796
-* Fix reshape #5739
-* Make "none" DataType explicit #5491
-* Change "scalar" and "stack" in IDL from "inrout" to "in" #5487
-* Link necessary libraries when building runtime for Android #5496
-* Fixes for wasm32 target #5489
-* Reset target and wait for runtime initialization on connect. #5499
-* Bump tophub rocm version #5504
-* Improve commentary for RingBuffer #5518
-* Add unit tests for ONNX PRelu and fix importer to pass them. #5521
-* LRN only supports 4D tensors, remove it from `alter_op_layout` #5520
-* Fix an issue with ONNX Upsample #5530
-* Cache PrimExpr instead of raw pointers in bound analyzer #5533
-* fix a few bugs with shape inference and types in the ONNX importer #5534
-* Add Onnx Pad v11 #5539
-* Changes to `cpp_rpc` to make it work on Android (+ Hexagon offloading) #5535
-* Fix to reduce RAM size during loading model #5507
-* Fix MakeLoopNest for warp memory #5382
-* Load platform specific lib for tvmdsoop instead of the hard-coded tvm_dso_op.so #5542
-* Add tests for running micro on native arm hardware #5546
-* Apparently, ONNX Conv with no 'pads' defaults to zero padding #5548
-* clang-format the h,cc,m files. #5557
-* Fix conv2d alter op for arm cpu #5532
-* Fix topi test for non tensorcore CI. #5563
-* Add clang-format and nodejs to ci-lint #5567
-* Enable clang-format. #5572
-* Allow `ubuntu_install_darknet.sh` to work in both 18.04 and 16.04 #5574
-* Add a quantized conv2 unit test for the tflite front-end #5558
-* Fix JSON graph dumping. #5591
-* Warp level reduction support for CUDA #5498
-* One more fix for concurrency count #5589
-* Improve robustness of the docs build #5583
-* Phase out WebGL #5570
-* Fix vulkansdk in the ci-gpu and upgrade to 1.2.135 #5566
-* Update ci-cpu to bionic #5554
-* Overestimate binary size for microTVM compiled binaries. #5590
-* Fix bug and re-enable RPC execution test #5436
-* Add ostream formatters for TargetPtr/TargetVal. #5592
-* Fix cross thread reduction #5551
-* Fix TVMArray layout on device #5599
-* Add debug mode to tempdir() #5581
-* Represent alignment information in LLVM IR #5598
-* Fix codegen for warp shuffle intrinsics #5606
-* Fix Topological Order calculation for DFPattern Language #5612
-* Global MaxPool3d and AvgPool3d support #5098
-* Fix build error of iOS RPC #5621
-* isn't a CallNode sometimes #5623
-* Introduce config to PassContext. #5631
-* CMAKE fix #5630
-* Label Pattern Partitions #5627
-* Extend AttrPattern to support CallNode and FunctionNode attributes #5637
-* Increase bss section size. #5660
-* Add buffer name when creating tensor bindings #5670
-* µtvm debug improvements #5648
-* enable `amd_apu` device on vulkan target #5659
-* Support TupleWrapper as direct ancestor of control flow ops #5639
-* add tvm.micro pydoc to sphinx #5661
-* Add a regression testcase for #5674 #5677
-* Fix C++ RPC build problem on Linux #5671
-* Add a check Callback to the Pattern Paritioner #5646
-* Call previous excepthook in `tvm_excepthook`. #5675
-* Fix the shift column for `scale_shift_nchw` and `scale_shift_nhwc` in C topi #5679
-* Support more dtypes for TVMDSOOp #5694
-* In `memory_plan`, check if value is not None, instead of just checking value as boolean. #5700
-* Fix flaky `test_topi_pooling.py:test_adaptive_pool` #5736
-* Fix the values for `test_fmod` since it fails way too often otherwise #5723
-* fix small bug about `dense_grad` #5695
-* Fix sequential cpp test #5745
-* Add Scatter to Topi/Relay/ONNX via hybrid script #5619
-* Clean WASM environment before build #5759
-* Fix gelu in PyTorch frontend, tighten numerical checks #5763
-* fix #5686: remove a overstrict assert in MakeAllreduce (#5686) #5785
-* Improve Pattern Language Docs #5676
-* Add missing expr visitor for any #6082
-* Remove the tvm web from version update #6122
-* Clear relay cache after every build & Clear warning message cache after autotvm task extraction #6131
-* avoid unexpected throw in AttrInitEntry #6128
-* Verify that tensor reshape is valid. #6215
-* Use LocalRunner by default in the tutorial tune_relay_cuda.py #6001
-* Undefined names: import os for line 324 & import re for line 308 #6003
-* GitHub Actions upgrade to actions/setup-python@v2 #6002
-* Only pass pythonpath for ci images #6005
-* Auto-convert shuffle with single index to “extract element” #6006
-* Cache object refs in loop partitioner instead of object pointers #6004
-* Fix `test_arith_solve_linear_inequality.py::test_multi_equal` #6014
-* MXNet frontend support for AMP cast op #5976
-* Demo showing how to run a pruned model. #5975
-* Move compiler related registry items to `vta/build_module.py` #6012
-* Pin keras version #6032
-* Fix in `arm_cpu/conv2d_alter_op` for NHWC quantized #6027
-* Add creation of Hexagon device in RPC client #6035
-* Terminate basic block after “ret” instruction #6036
-* µTVM CRT modifications for on-device RPC server #5921
-* Create TBAA information based on the unrelying buffer type #6046
-* Add support for tflite `arg_min` and `arg_max` #5992
-* Fix `fully_connected` converter when batch size is not 1 #6038
-* Fix a primitive check error #5991
-* Refactor to expose MakeOp functions to C++ #6047
-* Fix `conv2_gemm` after target structure update #6037
-* Remove use of designated initializers from `hexagon_module.cc` #6055
-* Build crttest and cpptest separately. #6057
-* Fix pytorch frontend prim::Constant issue #6051
-* update frontend tutorials to new model based runtime interface #6063
-* Remove unnecessary std::cout #6072
-* Fix error message in Buffer::vstore, NFC #6056
-* Fix FSIM Compile Error. #6070
-* Improve vector simplification for float operands #6043
-* Fix LocalBuilder on macOS with python 3.8. #6083
-* Add missing test for fast erf #6058
-* Fixed point multiplication improvements for AArch64 #5980
-* Fix code generation bugs for C/CUDA & Improve VerifyGPUCode pass #6041
-* Delete declaration of unused `op_node` #6102
-* Load configs even it has no entity #6100
-* Update SGX example Cargo.toml #6067
-* Add default value for option `USE_DNNL_CODEGEN` in the cmake #6099
-* Update installation doc with minor improvements #6104
-* lint: add opencl .cl file type #6092
-* Clean up conversions between TVM and Rust functions #6114
-* Improve reduction schedule on arm CPUs #6110
-* Register Shape Func for Some Operators to Handle Dynamic Shapes #5955
-* Fix variable name conflict with OpenCL keyword #6048
-* Some rust cleanups #6116
-* Option to specify alternate directory to output build to #6016
-* Add `get_num_inputs` to GraphRuntime #6118
-* TFLite quantized conv test #6084
-* Fix autotvm on the `conv2d_nchw_winograd.mali` operator #6130
-* add attr option mfloat-abi for arm32 #6123
-* Fix CUDA Library Tuning #6132
-* Add missing RPC sources after refactor #6113
-* Correct `runtime.load_module` #6161
-* Improve error messages in graph tuner, graph runtime, and module loader. #6148
-* Fix some shape mismatches between TF and Relay #6166
-* Improve doc string #6176
-* Fix incorrect function signature in header #6172
-* Fix alignment of note #6181
-* Implemented PADV2 Operator for TFLite and added support for constant values in PAD. #6167
-* Unary ops support added in frontend #6196
-* Change the meaning of `conv3d_transpose` `output_padding` to match `conv{1,2}d_transpose` #6065
-* Fix compile warnings. #6204
-* Fix -mfloat-abi=soft compilation for ARM with OpenCL target #6150
-* Match pytorch 1.6 googlenet pretrained model (#6201) #6212
-* Mod operator, bug fix #6160
-* RESHAPE with dynamic shape arg in TFLite frontend #6208
-* Fix compilation error with cuda 11 #6213
-* Fix `port_end` wrong default value 9199 to 9099 for keeping same with source code #6220
-* Std op without specified dimensions support #6226
-* fix crt building and running error #6231
-* Implemented `ONE_HOT` Operator for TFLite. #6223)
-* Avoid unexpected throw in AttrInitEntry #6128
-* Added casting to hybrid script doc and fixed pass infra doc #6174
-* Fix compile warnings. #6204
-* Fix -mfloat-abi=soft compilation for ARM with OpenCL target #6150
-* Mod operator, bug fix #6160
-* Fix compilation error with cuda 11 #6213
-* Fix `port_end` wrong default value 9199 to 9099 for keeping same with source code #6220
-* Std op without specified dimensions support #6226
-* Verify that tensor reshape is valid. #6215
-* Fix crt building and running error #6231
-* Fix `conv2d_transpose` output padding #6236
-* Fix cuda half math function is undefined: hpow, htanh #6225
-* Fix division range estimation error in simplifier #6244
-* Fix newer GCC compiler warnings. #6257
-* Support `_contrib_SyncBatchNorm` #6245
-* Fix reduction #6250
-* Add apt repository for clang-11 and llvm-11 #6256
-* Update tutorial to new TARGET as `micro_dev` is no more #6262
-* Fix clang-format #6264
-* Trivial fix, up the rodata section for the discovery board to 512 bytes. #6259
-* Fix cuda half math function is undefined: hpow, htanh #6253
-* Add dilation in x86 NCHWc depthwise conv support #6267
-* Decrease test times by introducing testing model #6235
-* Add support for parsing the any dimension. #6277
-* Improve error messages for memory verifier and gpu memory verifier #6281
-* Reflect Compile-Time CMake Options into libtvm.so #6280
-* Add cmake options into libinfo #6286
-* Update slice to infer attributes when not graph inputs #6276
-* Use rpc.LocalSession for simple tests #6294
-* Fix random fail #6312
-* Fix resize test #6298
-* Fix cython FFI compact with np.int64 #6321
-* Fix relay vm optimize #6322
-* Changed TVMCTVMContext to TVMContext #6306
-* Make able to compile with MSVC #6341
-* ROCm changed name of library and removed the old one in ROCm 3.7 release. #6345
-* Compatible for ROCm before 3.7 #6359
-* Use clear name that is separate from ASF brand for cache #6360
-* Fix `Dockerfile.demo_android` #6361
-* Fx sparse dense schedule on cuda #5803
-* Fix strategy for sparse dense cuda #5782
-* Fix x86 conv2d template when tuning with unpacked layout #5938
-* Fix the filter width parameter in `depthwise_conv2d` #6081
-* Fix reshape usage in ARM schedule #5732
-* Missing header #4865
-* Fix `conv2d_transpose` output padding #6236
-* Simplify reduce expression in te.gradient #6611
-
-### API Changes
-* `tvm.module` -> `tvm.runtime.module`
-* `tvm.module.load` -> `tvm.runtime.load_module`
-* `tvm.module.enabled` -> `tvm.runtime.enabled`
-* `tvm.module.system_lib` -> `tvm.runtime.system_lib`
-* `tvm.relay.Module` -> `tvm.IRModule`
-* `tvm.create_schedule` -> `tvm.te.create_schedule`
-* `tvm.placeholder` -> `tvm.te.placeholder`
-* `tvm.compute` -> `tvm.te.compute`
-
-### Deprecation
-* Deprecate NNVM (#4535, #4562, #4565, #4571)
-* Deprecate FreeStmt #5890
-* Remove legacy `compute_expr.h` #5738
-* Deprecate OpenGL #5711, #5712
-
-## 0.6
-
-### Relay in Production
-Relay is a functional, differentiable programming language designed to be an expressive intermediate representation for machine learning systems. Relay supports algebraic data types, closures, control flow, and recursion, allowing it to directly represent more complex models than computation graph-based IRs (e.g., NNVM) can. In TVM v0.6, Relay is in stable phase and is ready for production.
-
-* Algebraic Data Types (ADT) support (#2442, #2575). ADT provides an expressive, efficient, and safe way to realize recursive computation (e.g., RNN). Refer to https://tvm.apache.org/docs/langref/relay_adt.html for more information.
-* Pass manager for Relay (#2546, #3226, #3234, #3191)
-* Most frameworks have been supported in Relay, including ONNX, Keras, Tensorflow, Caffe2, CoreML, NNVMv1, MXNet (#2246).
-* Explicitly manifest memory and tensor allocations in Relay. (#3560)
-
-### Relay Virtual Machine
-The Relay Virtual Machine (Relay VM) is the new generation of runtime to strike a balance between performance and flexibility when deploying and executing Relay programs. Previously, the graph runtime is able to utilize the fully static nature of the input graphs to perform aggressive optimization such as fully static allocation, and optimal memory reuse. When we introduce models which make use of control-flow, recursion, dynamic shapes, dynamic allocation we must change how execution works.
-
-Relay VM is now usable and is able to achieve decent performance for a various of models and targets.
-
-* Design (#2810 #2915) and a first version of implementation (#2889),
-* Add VM runtime for Relay and compiler support (#3120, #3121, #2889, #3139)
-* Relay VM (pattern matching #3470, port to python #3391, serialization #3647)
-* Relay VM Profiler (#3727)
-* Support execution on devices for Relay VM (#3678)
-* [Relay][VM] Add more passes to VMCompiler (#4058)
-* [relay][vm] Separate VM runtime with executable (#4100)
-* Port VM, VM compiler, and Object into Python (#3391)
-* VM: Add AllocTensor instruction and better instruction printer (#3306)
-* [Relay][VM][Interpreter] Enable first-class constructors in VM and interpreter via eta expansion. (#4218)
-* [Relay][VM] Clean up the VM and VM profiler code (#4391)
-
-### Training
-Relay is designed to natively support first-order and higher-order differentiation. The automatic differentiation infrastructure is now usable and a count of operators with gradient support are available in v0.6 release.
-
-* Higher order reverse mode automatic differentiation that work with control flow (#2496)
-* Higher order continuation passing style (#3456, #3485 )
-* Relay gradient registration (clip #3509, `max_pool2d` and `avg_pool2d` #3601)
-* Relay AD algorithm (#3585)
-* Relay Training - allow gradient to return a tuple (#3600), numerical gradient check (#3630)
-* Improve AD for concatenate (#3729)
-* [Relay][Training] Add missing gradient check to gradient pass (#4169)
-* As a part of Relay's automatic differentiation system, we are adding primal gradients for Relay operators. Please refer to #2562 for tracking the progress.
-* Gradient for Conv2d (#3636)
-* Add gradient operators (#3857, #3894, #3901, #3915)
-* Add gradient for log-softmax (#4069)
-* [Relay][Training] Add gradient for Crossentropy (#3925)
-* [Relay][Training] Add and fix gradients (#4126)
-
-### Quantization
-
-Low-bit inference is getting more and more popular as it benefits both the performance and storage usage. TVM now supports two types of quantization. 1. Automatic quantizaion takes floating-point precision model, does per-layer calibration and generates low-bit model. 2. TVM also imports pre-quantized model from Tensorflow and MXNet, a new dialect QNN is introduced to handle further lowering to normal operators.
-
-* Automatic Quantization
- - Low-bit automatic quantization supported. (#2116). The workflow includes annotation, calibration and transformation.
- - Refactor quantization codebase and fix model accuracy. (#3543)
- - KL-divergence-based per-layer calibration. (#3538)
- - Add option to select which convolution layers are quantized. (#3173)
- - [Relay][Quantize] Integrate data-aware calibration into quantization. (#4295)
-* Pre-quantized model support (QNN operators and legalize pass).
- - Add a legalize pass to Relay (#3672)
- - Qnn Concatenate, quantize, dequantize and requantize operators (#3819, #3730, #3745, #3531)
- - QNNtoRelay & QNNLegalize Pass utility (#3838, #3782)
- - Requantize: Optimize lowering for some corner cases. (#3864)
- - New quantized operator support: conv2d, add, dense (#3580, #3736, #3896, #3910)
- - Do type checking for the input and kernel in the qnn conv2d (#3904)
- - Legalize and AlterOpLayout for Intel int8. (#3961)
- - Renaming tests to follow the Relay nomenclature. (#3975)
- - Fix padding changes due to #3739 (#3989)
- - Memorizing quantize node mapping to avoid duplicated simulated quantization (#3233)
- - Infrastructure to support pre-quantized models (QNN) (#3971).
- - [Relay][AlterOp] NHWC to NCHWc support for Pool, concatenate, sum. (#4059)
- - [TOPI][x86] Cascade lake support. (#4123)
- - [TOPI][x86] Legalize - Support int8xint8 convolution to use VNNI inst (#4196)
- - Qnn dequantize with min max using Mxnet flavor to support Mxnet prequantized models. (#3945)
- - Improve the lowering of Qnn Dense (#4213)
- - Adding support for dequantizing from int32 to float32. (#4130)
- - [QNN] Refactor fixed point multiplication in requantize (#4073)
- - [Relay][Quantize] Use fixed point mulplications (#4160)
- - Add support for quantized multiply to Relay (#4141)
- - Use legalize to handle NHWC layout for `arm_cpu` (#3754)
- - [QNN][Legalize] Specialize for Platforms w/o fast Int8 support (#4307)
- - [QNN] Use Int16 upcast in Fallback Conv2D. (#4329)
- - Retain input kernel scales in QNN dialect (#4292)
- - [QNN] Lowering for Depthwise Convolution. (#4351)
- - [QNN][TFLite] Parsing QNN Add op. Adding MobilenetV2. (#4142)
- - [QNN][TFLite] Parsing TFLite quantized models. (#3900)
- - Added tflite frontend support for quantized mean. (#4339)
- - [Relay][Legalize] Legalize `conv2d_transpose` for NHWC (#4399)
-
-### Accelerator and Microcontroller Support
-
-TSIM is introduced to improve software and hardware integration and simulation accuracy. It integrates the hardware development process into the software stack. TSIM enables VTA to provide a more accurate performance feedback, i.e. clock cycles, compared to the traditional functional model of a hardware accelerator. Moreover, Chisel implementation for VTA is availale and it runs on top of TSIM.
-
-There has been a proliferation of resource-constrained and embedded devices that do not have operating systems or a mature software stack. MicroTVM is intended to support TVM on such bare-metal devices.
-
-* [TSIM] Enabling Cycle-Accurate Hardware Simulation for VTA (#3010, #3206, #3242)
-* Chisel implementation for VTA and runs on top of TSIM (#3258, #3347)
-* MicroTVM (#3227)
-* Relay Compilation + AutoTVM compatible operator libraries for VTA (#3135)
-* ChangeBatch pass for batched VTA compilation (#3656, #3660)
-* VTA fast simulator statistics (#3481)
-* TSIM improvements and fixes (#3505)
-* Chisel VTA enhancements and fixes (32bit support #3558, alu instruction generation #3592, coherence support #3593, separate types #3605, tensor issue/commit #3637, uop load request #3643, uop dma requests #3654)
-* VTA Runtime refactor for non-shared memory FPGAs (#3590)
-* VTA HLS codebase refactor for Ultra96 (#3496)
-* VTA support for batched inference (#3661)
-* VTA bitstream compilation for Intel FPGA (#3494)
-* TSIM: Introduce Virtual Memory for TSIM Driver (#3686)
-* Parallel TSIM hardware compilation with macOS and debug support (#3797)
-* Chisel: scale dram base address in hardware instead of runtime (#3772)
-* Chisel: run all unittests by default (#3766)
-* Chisel: improved Data Gen, Added ALU Test (#3743)
-* Chisel dependencies for TSIM CI (#3721)
-* Chisel: Added Module Unit Test Infrastructure (#3698)
-* Add ISA BitPat generation (#3891)
-* de10-nano driver (#3394)
-* Extending Vision model coverage compilation for VTA (#3740)
-* Conv2d transpose (deconvolution) operator support (#3777)
-* Support TLPP in function simulator. (#3555)
-* [VTA][Chisel] TSIM VTA Source Refactor (#4163)
-* [VTA][TSIM] Serial GEMM Application Added (#4082)
-
-### Rust Support
-Rust language support in TVM includes two parts. 1. The frontend wraps the current C API and exposes a Rust programming model. 2. The backend serves as an alternative to C++ runtime. It privdes a standalone WASM module and security support, e.g., SGX.
-
-* Rust frontend (#2292).
-* Unify types between bindings and pure Rust impl (#2616)
-* Rust: load syslib modules at compile time (#3274)
-* Rustify PackedFunc & Friends (#2969)
-* Rust DSO module (#2976)
-
-### Operator Support
-* A special operator `annotation.stop_fusion` to prevent it being fused with previous expressions (#2624).
-* `batch_matmul` supported (#2561).
-* `reverse_reshape` supported (#2503).
-* Faster-RCNN proposal operator for CUDA (#2420).
-* Vision operator for YOLO `yolo_reorg` (#1941).
-* `slice` operator for MXNet (#2662).
-* `arange` supported (#2621).
-* Vision operator `roi_align` (#2618).
-* `where` operator for MXNet (#2647).
-* Deformable conv2d (#2908)
-* Faster-RCNN Proposal OP (#2725)
-* ROI Pool operator (#2811)
-* Gluoncv SSD support on CPU (#2353)
-* shape, reverse, and sign op (#2749, #2800, #2775)
-* tile and repeat op (#2720)
-* logical operators (#2743, #2453)
-* stack op (#2729)
-* NCHWc upsampling (#2806)
-* clip and wrap mode support in take (#2858)
-* AlterLayout support for `intel_graphics` conv2d , depthwise conv2d (#2729, #2806)
-* Add foldr1 operator (#2928)
-* Add rsqrt operator (#2949)
-* Add clip and wrap mode support in take (#2858)
-* `Gather_nd` exposed to relay (#2945)
-* `bitserial_conv2d` move to autotvm template and updates (#2819)
-* Port x86 NCHWc to AutoTVM for Task Extraction (#2664)
-* Implement relay `nn.bias_add` compute in C++ (#3027)
-* Rename output tensors for better readability (#3006)
-* int8 dense on CUDA & Dense op quantization (#2877)
-* Bitserial dense operators for CPU (#3051)
-* Enhance upsample operator to adapt onnx opset v9 (#2968)
-* Add adaptive pooling operator (#3085)
-* Add all operator (#3124)
-* Add cblas `batch_matmul` (#3210)
-* Add packing for int8 1x1 convolution and support the int8 group convolution on X86 (#2991)
-* Add op size (#3094)
-* x86 TOPI (`roi_align` #3475, `conv2d_transpose` #3491)
-* Intel INT8 (dilation in conv2d #3510, type checking #3516)
-* Reinterpretation of tensor elements (#3599)
-* Spase-Dense for block-sparse multiplication (#3566)
-* Winograd matrix computation (#3553)
-* CUDA schedule for `pool_grad` (#3622), `group_conv2d` (#3663)
-* Bitserial operations conv2d, dense and bitpack (#3844)
-* Improve numeric gradient check (#3856)
-* Resize rework ([3788](#3788))
-* Improve `conv2d_transpose` CUDA schedule template (#3796)
-* SpaceToDepth and MirrorPad Operators (#3718)
-* Add variance and layer norm op (#3700)
-* Add `sparse_transpose` for Square CSR matrices (#3707)
-* TOPI: Memoize winograd matrix (#3687)
-* New TOPI operators: `erf`, `logical_and`, `logical_or`, `logical_not`, `isnan` (#3702, #3929, #3979)
-* Improve `ceil_divide` in tile/split (#3842)
-* [Relay][Frontend][TF] Add tensor array ops (#3798, #4309)
-* [TF][Op] Op where (#4045)
-* [TOPI]Add op argwhere (#3994)
-* [Relay] `crossentropy_with_logits` and its gradient (#4075)
-* [Relay][Op] Enhance Upsample Operator to support float scales (#4206)
-* [Relay][Op] Add instance norm op (#4004)
-
-### Frontend and User Interface
-* Frontend darknet (#2773)
-* Support tf.gather (#2935)
-* Support tf.where (#2936)
-* Adding ADD operator to tflite frontend for compiling the MobileNetV2 (#2919)
-* Support SpaceToBatchND/BatchToSpaceND in Tensorflow frontend (#2943)
-* Simplify TF `get_output_names` (#3025)
-* TF Tile Round Sign Pow Exp Reverse (#2960)
-* Gluncv SSD support on the GPU (#2784)
-* Allow an op as loop var in Tensorflow (#3056)
-* Add `FULLY_CONNECTED` op into tflite frontend (#3019)
-* Add MXNet converter for RNN layer ops (#3125)
-* Add log op in tf frontend (#3111)
-* Add SoftPlus Sqrt in Tensorflow frontend (#3187)
-* Add onnx elemwise greater/less (#3186)
-* Add PlaceholderWithDefault (limited) implementation in TensorFlow (#3184)
-* Support `tf.math.reduce_prod` (#3166)
-* Better shape inference in TensorFlow Frontend (#3176)
-* Get list of unsupported ONNX operators (#2995)
-* Implement ONNX MaxPool-v8 and MaxPool-v10 (#3114)
-* Convert TFLite NCHW to NHWC (#3141)
-* Add Crop op converter (#3241)
-* TFLite frontend operator support: PAD, RESIZE, MUL, Reduce (min, max, mean, prod), LOGISTIC, elemwise operators (Sub, Divide, Power, Max, Min) (#3310, #3370, #3304, #3421, #3313, #3357)
-* Tensorflow frontend operator support: Abs, FloorDiv, GatherND, LeftShift, LogSoftmax, Max, Min, Mod, RightShift, ZerosLike, TruncateMod, Neg, ClipByValue, ResizeNearestNeighbor (#3270, #3211, #3393)
-* TFLite: Add `fused_activation_function` for ADD, SUB, MUL, DIV (#3372)
-* Support bidirectional RNN layer for MXNet (#3397)
-* TFLite operator support (pack #3521, split #3520 )
-* Keras operator support (permute, softmax #3618)
-* TF operator support (BatchMatMul #3634)
-* TFLite frontend operator support: tile, transpose (#3814, #3705)
-* ONNX frontend operator support: PReLU for NNVM, Not, Sign, Equal (#3813, #3836, #3760)
-* Keras frontend operator support: Dot (#3668)
-* Add more cases to Keras `_convert_reshape` (#3846)
-* TensorFlow frontend operator support: OneHot, log1p, cos, sin (#3781, #3614)
-* Support BatchMatMul with input dimensions larger than 3 for TensorFlow (#3732)
-* ONNX new operator support: And, Tile, Erf (#3878, #3941, #3988)
-* MXNet new operator support: pad, conv1d, deconv1d (#3739)
-* TFLite new operator support: `batch_to_space_nd`, `space_to_batch_nd`, tanh, greater, relu (#3850, #3996, #3963, #4022)
-* TFLite: Support depthwise convolution multiplier greater than 1 (#3922)
-* Keras: Fix ReLU in Keras Converter missed the case (#3917)
-* Keras: frontend upsample and 1 channel conv2d fixes (#3937)
-* Tensorflow: Convert scalar Const into tvm.relay.const (#3885)
-* TensorFlow: Add support for SquaredDifference (#3930)
-* [relay][frontend] clean up tf frontend (#3710)
-* [Relay][Topi][TensorFlow][ONNX][Lang] Add support for Any op (#4205)
-* [Relay][Frontend][ONNX] Add support for op Where (#4184)
-* [Relay][TopHub] Add switch to disable TopHub download (#4015)
-* Add parser support for CAST tflite operator (#4096)
-* Add parses support for `zeros_like` tflite operator (#4042)
-* Add parser support for SUM tflite operator (#4182)
-* Add support for tf.assert (as no-op) and `tf.no_op` to TF Relay frontend. (#4172)
-* [Relay][Frontend][ONNX] New Operators and Opsets to Support BERT (#4197)
-* [Relay][Params] Add APIs for storing and retrieving parameters from individual functions. (#4194)
-* Add `build_create_shared_func` to tvm/contrib/cc.py (#3840)
-* Tensorflow saved model for NNVM ([#2493](#2493/) and Relay ([#2586](#2586/)).
-* Introduced `HybridModule` (#2477) so that normal TVM schedule can be compiled to hybrid target, run and dumped to Hybrid Script.
-* Relay ][Frontend][Tensorflow] add operator `add_n` (#4181)
-* [Relay][Frontend][Tensorflow] StopGradient (#4238)
-* [Relay][Frontend][ONNX] Add support for broadcasting to Where and MatMul (#4267)
-* [TFLite] Support PRelu (#4298)
-* [Frontend][MxNet] support mxnet cond op (#4311)
-* Add support for `quant.mul` operator in tflite frontend (#4283)
-* [Relay][Frontend][ONNX] operator support: DepthToSpace, SpaceToDepth (#4271)
-* [Relay][Frontend][Tensorflow]Add `conv2d_transpose`. (#4300)
-* [Frontend]Add TensorFlow FloorMod (#4308)
-
-### Runtime and Backend Support
-* Make external library extend TVM's NDArray more easily (#2613).
-* Improvements for NNPACK integratation, includes ci test, winograd (#2846, #2868, #2856, #2721)
-* Improvements for OpenCL runtime (#2741, #2737)
-* GraphRuntime: Enable sharing parameters of a model among multiple threads (#3384)
-* Android runtime argsort support (#3472)
-* GraphRuntime enhancements (`set_input_zero_copy` #3416)
-* A new minimal runtime implementation (~12kb .text on ARMv7/x86) for TVM.
-* Add AVX512VNNI support for TVM (#3388)
-* Enable miopen Group Convolution (#3987)
-* Minimal runtime (~12kb .text on ARMv7/x86) for subset of TVM models (#3567)
-* [RUNTIME] Separate runtime related contrib into runtime/contrib (#4207)
-* [topi] add ARM v8.2 udot (uint8) support (#3978)
-* [codegen] Add multiple operands and function support when using fp16 compilation (#4056)
-* [TOPI] Added support for Mali Bifrost target (#4047)
-* [topi] enable fp16 sort for arm (#4084)
-* Add OpenOCD Low-Level Device (RISC-V Support) (#3756)
-* Add wave 32 bc for AMD ROCm backend (#3984)
-* [RUNTIME] Support C++ RPC (#4281)
-* [TOPI][OP] Support Faster-RCNN Proposal OP on CPU (#4297)
-* [TVM][RUNTIME] A minimum example to generate external library wrappers for DSOModule (#4280)
-
-### Language and Architecture
-* Support custom datatypes (#2900)
-* Add the acc16 intrinsic support (#3081)
-* Handle float16 constants & fix BatchNorm (#3260)
-* Structural hash - incorporate the var type into its hash (#3267)
-* Relay C++ Build Module (#3082, #3144, #3174)
-* Enable decorating python class to be a Relay Pass (#3364)
-* Make Partial Eval support interprocedural optimization and termination check. (#3033)
-* Introduce feature manager to Relay. (#3236)
-* Use Relay parser to define the Relay prelude (#3043)
-* Mechanism to detect incomplete expression match in Relay (#3203)
-* EQ/NE operators support for StringImm expressions (#3283)
-* Mechanism to detect incomplete expression match in Relay (#3203)
-* Introduce CanonicalizeCast pass to formally reduce memory overhead introduced by fused cast operations (#3280)
-* Support overloading comparison operations in Relay (#3168)
-* Mac count: provide a pass to calculate the number of multiply-accumulate operations in a network (#2609).
- - support for `conv_2d_transpose` (#3469)
- - [Relay][Pass] Count MAC for BatchMatMul (#4157)
- - Detect depthwise conv2d in `mac_count` pass (#3083)
-* Add Tuple pattern (#3596)
-* Text format support for ADTs and prelude (#3863, #3939)
-* Add new IR pass CombineParallelDense (#3862)
-* Add support for `EQ` op in the deduce bound and the loop partition (#3775)
-* Introduce base-class IRMutatorWithAnalyzer (#3969)
-* Define more standard global functions in the prelude of relay program, includes foldr1, hd, tl, nth, list update (#2928, #2917, #2771, #2866)
-* Add SkipVectorize pass (#3222, #3228)
-* [Relay][Pass] Add pass to remove unused functions in relay module (#4334)
-
-### Symbolic shape enhancement
-* Add shape function for symbolic shape. It enables certain cases for broadcast with symbolic shapes. (#3606)
-* [tvm][any] broadcast with values other than one (#3967)
-* Symbolic shape support (broadcast op #3389)
-* Support reshape for dynamic shape in tf converter (#4185)
-* Runtime Shape Functions (#4179)
-
-### Language and Architecture
-* An optimization pass to eliminate expressions which have the same functionality and same inputs (#2639).
-* Refactor text printer to add stream-like API and FunctionType support (#2605, #2882)
-* Build a scaffold for structured error handling (#2838). The new mechanism detects and rewrites error messages so that c++ and python stack trace are unified and not redundant. Guideslines and conventions for error handling is also discussed.
-* Higher order reverse mode automatic differentiation that work with control flow (#2496)
-* Integer arithmetic analyzers, includes modular set analysis, const integer bound analysis and rewrite simplifier (#2904, #2851, #2768, #2722, #2668, #2860)
-* Improve operator fusion for TupleGetItem in relay (#2914, #2929
-* Compute FLOP of autotvm template for int8 models (#2776)
-* Common subexpression elimination pass in Relay (#2639)
-* Improve quantization in Relay (#2723)
-* Refactor `build_func` in measure module of autotvm to better support cross compiler (#2927)
-* Quantize all fields of concatenate (#2913)
-* Remove stale verilog generator (#2964)
-* Improve Relay printing (#2984, #2881, #3030, #3041)
-* Add `min_num_branches` option in CombineParallelConv2D (#2961)
-* Add `expr_visitor`, fix `expr_functor` exponential blowup problem (#2988)
-* Support Deriving channels when it is not provided in AlterLayout. (#2972)
-* Enhance BoundDeduce algorithm (#2795)
-* Enhance loop partition algorithm (#2956)
-* Better tuple fusion implementation (#3092)
-* Enhance fusion rule that starts from elemwise and broadcast (#2932)
-* Remove `on_device` op after annotation in heterogeneous pass (#3204)
-* Improve canonical and rewrite simplifier (#3132, #3149)
-* Capture constant external python variables in hybrid script (#3157)
-* Remove Peano nats from the prelude (#3045)
-* Macro to define NodeRef methods, constructor style example (#3224)
-* Consistent RAII scoping API (#3231)
-* Register all operators' attributes in Python (#3175)
-* Add module supoort in relay.build (#3424)
-* Relay pass infrastructure improvement (#3319, #3336, #3430, #3353)
-* Migrate Relay passes to pass manager (#3323, #3289, #3251, #3406)
-* Improve heterogeneous annotation by using visitor (#3261)
-* Support export ADT value in Python (#3299)
-* Extend TensorComputeOp to allow scalar inputs (#3300)
-* Transitioning low-level IR away from HalideIR (#3533, #3535)
-* Tags for ADT constructors (#3369)
-* IR dumping for debugging (#3493)
-* Pretty printer and parser roundtrip (#3460, #3536)
-* Relay type checking (conv2d weight dimension #3511, any shape #3221)
-* Relay Module enhancements (remove free variables #3476)
-* LLVM DWARF debug information (#3420)
-* Printer for Layout/BijectiveLayout (#3582)
-* Type inference escape hatch (#3571)
-* Making iterators compatible with constructors of STL containers (#3624)
-* Moving Conv, Dense, Concatenate InferTypes to header (#3783)
-* Simplify casts of constants 0 and 1 (#3758)
-* Conditionally replace reduction init axis. (#3408)
-* Improve Partial Evaluator (#3749, #3703)
-* Strict mode in Relay pattern matching (#3620)
-* Quit and clean when TVM is interrupted (#3640)
-* Make Type Relation catch more errors (#3899, #3699)
-* Refactor the way we interface between different modules of Relay (#3906)
-* Introduce `schedule_injective_from_existing` and unify external schedules for all targets (#3983)
-* [NODE][REFACTOR] Refactor reflection system in node. (#4189)
-* Unify node system and object (#4161, #4115, #4128)
-* [Relay][Refactor] Rename Datatype to ADT (#4156)
-* [Relay] fix exponential blowup in interpreter (#3559)
-* [Relay] Fix memory leak in the interpreter (#4155)
-* [rpc] use callback func to do send & recv (#4147)
-* Add `lift_if_then_else` pass to improve loop partitioning (#3865)
-* Decrease the complexity of CalcDep from exponential to linear (#4053)
-* [IR] Make iterators compatible with constructors of STL containers (#3624)
-* [Relay][Pass] Avoid FoldConstant folding some ops (#4245)
-* [Relay][Prelude] More dtypes support in `tensor_t` (#4233)
-* [NODE][REFACTOR] Rename IRFunctor->NodeFunctor, use func pointer (#4247)
-* [RUNTIME][REFACTOR] Use object protocol to support runtime::Module (#4289)
-* [CodeGen] Add build config option `disable_assert` to control whether to generate assert. (#4340)
-
-### Arithmetic Analysis
-* Formalize Integer Arithmetic Analysis (RFC: #2588). It is aiming to perform better context-dependent analysis, bound analysis, centralized arithmetic logic and arithmetic simplification. (#3272, #3463, #3464, #3368, #3503, #3504 , #3502, #3479 , #3568)
-* Introduce FloorDiv/Mod, TruncDiv/Mod, and IndexDiv/Mod for better arithmetic simplification (#3976, #3986, #4000, #4014, #4008, #4028)
-* [ARITH] Use floordiv for the deduce bound (#4025)
-* [Simplifier] Rewrite simplification rule to eliminate unnecessary conditionals. (#4076)
-
-### Runtime and Backend Support
-* Provide error msg for failure function call in tvm4j (#2967)
-* Expose backtrace symbols in Debug mode (#3001)
-* C++ GraphRuntimeCodegen, Deprecate Python2 (#2986)
-* Ensure interpreted functions can take values that are not TensorValues (#3015)
-* Make OpenCL runtime Compatible with OpenCL2.0 (#2897)
-* Handle INF and NAN in CUDA and OpenCL (#3194)
-* Update debug graph runtime for more precise layerwise timing (#3232)
-* ROCM support (llvm printing #3662, ld.lld finding #3664, save to file #3665)
-* Threadpool: make `spin_count` configurable (#3577)
-* RPC worker children termination (#3669)
-* Vulkan runtime reimplementation (stream approach) (#3849)
-* Vulkan backend supports Call::reinterpret and vectorized comparison (#3795)
-* Support MKL on Windows (#3837)
-* Vulkan IR builder (bool to float #3513)
-* Force `code_object_v2` for amd gpu backend (#4099)
-* [Codegen][cuda-fp16] fallback to fp32 simulation when cuda arch < sm53 (#4268)
-* Fix and refactoring for AMD gpu backend (#4305, #4321, #4341, #4342)
-* [Debugger] Sorting op-time breakdown for quicker analysis. (#4352)
-* [nvcc] enable multiple arch in one fatbin (#4377)
-* [RUNTIME] Move module export to the function level. (#4405)
-
-
-### Frontend and User Interface
-* Relay now supports saving and loading parameter dictionaries. (#2620)
-* Add `max_num_threads` to Hybrid Script, which allows users to get max number of threads for GPU targets ([#2672](#2672/)).
-* Improvements for tensorflow frontend (#2830, #2757, #2586), includes decompiling tf control flow (#2830)
-* Improvements for mxnet frontend (#2844, #2777, #2772, #2706, #2704, #2709,, #2739)
-* Improvements for keras frontend (#2842, #2854)
-* Improvements for DarkNet frontend (#2673)
-* Improvements for ONNX frontend (#2843, #2840)
-* Better profile result dump in Chrome Tracing format (#2922, #2863)
-* Unified error handling in NNVM and Relay frontends (#2828)
-* Improve NNVM to Relay conversion (#2734)
-* Remove `input_0d_mismatch` special handling for TF Frontend(#3087)
-* Bumped ONNX version from 1.1.0 to 1.4.1 (#3286)
-* Simplify parameter handling in Tensorflow frontend (#2993)
-* CoreML improvement for image scaler and padding (#3800)
-* Clean up TensorFlow frontend (#3710)
-* Darknet: Solve tvm parsing darknet resnext failure bug (#3778)
-* Frontend changes `get_workload` - (#3483)
-* [TF][Relay][Op] Pass module when infer shape (#4287)
-
-### AutoTVM
-* Support override in `register_topi_compute` and `register_topi_schedule`. (#3292)
-* Improve graph tuner dealing with Tuple. (#3649)
-* Add AutoTVM template for conv2d Intel int8. (#3955)
-* Add AutoTVM template for dense on CUDA. (#3923)
-* Add AutoTVM template for conv2d on Intel graphics. (#3839)
-* Optimizing autotvm task extraction speed. (#4138)
-* [AutoTVM] Add `batch_matmul` to tunable operations. (#4242)
-* Selecting tuning templates when extracting task. (#4338)
-
-### Performance Improvements
-* Enable AlterOpLayout pass for x86 on Relay (#2585). It is essential to get decent performance for CNN-based model on Intel CPUs.
-* Better intrinsic matching for x86 CPU and ARM CPU, includes variants of vcvtph2ps and vmlal.s16 (#2925, #2748).
-* Improve injective schedule for ARM CPU(#2801)
-* Core functionality for Graph tuner (#2184)
-* Fast tanh implementation (#3255)
-* Improve multi-batch conv2d on x86 (#3308)
-* Improve `non_max_suppression` and `get_valid_counts` for CPU (#3305)
-* Improve `roi_align` performance for CPU (#3296)
-* Improve `nms` and `get_valid_count` performance (#3282)
-* Graph tuner for multiple subgraph (#3490)
-* For sparsity, fast transpose for square CSR matrices has been now merged, which is a good start point for more general sparse type support.
-* Reduce `set_input` and `set_input_zero_copy` overhead (#3805)
-* Parallelize batch axis for ARM (#3931)
-* Support cuBLAS BatchMatMul (#3936)
-* Add AVX512VNNI support for TVM (#3388)
-* Enhance tuning space of split (#3949)
-* Enable miopen transpose convolution and fp16 support (#3952)
-* Improve `conv2d_transpose` schedule on X86 and CUDA (#3948)
-* Expose llvm.nearbyint intrinsic (#4001)
-* [TOPI][X86] Pool operator parallel support. (#4090)
-* Improve layout for several operators (#4103, #4040, #4080)
-* [Relay][VM] Fix constant folding issue in VM compiler (#4077)
-* [relay][vm] Reuse allocated device memory (#4170)
-* [Runtime] Enable option to use OpenMP thread pool (#4089)
-* [PERF] Parallelize reduction for CPU (#4158)
-* [TOPI] Tunable Template for Conv2D HWCN on CUDA (#4168)
-* [TOPI] Add valid auto tvm for Intel Graphics (#4078)
-* [TOPI] FIFO buffer op, to accelerate sequence modeling with dilated convolutions (#4039)
-* TensorCore Support using Intrinsic (#4136)
-* Auto TensorCore CodeGen (#4234)
-* Use cblas for dense and `batch_matmul` (#3787)
-* Update TOPI softmax compute and CPU schedule (#3680)
-* [VTA] Performance optimize, remove unnecessary contigious memory use. (#4246)
-* [TOPI][AlterOpLayout][ARM] Enabling NHWC to NCHW layout transformation. (#4249)
-* [PERF] Parallelize reduction for CPU (#4158)
-* [ThreadPool] Solve thread transitions issue (#4344)
-
-### Documentation
-* Tutorials for deep learning frameworks support in Relay.
-* Tutorial for running AutoTVM with Relay (#2594).
-* Document for Algebraic Data Types (#2575).
-* Move NNVM tutorials to Relay (#2783, #2785, #2766, #2693)
-* Documentation on operators (#2761)
-* Add gradient operator tutorial docs (#2751)
-* Add compiler pass tutorial docs (#2746)
-* Add Android Tutorial (#2977)
-* Developer documentation for InferBound pass (#3126)
-* Add missing targets to `target_name` documentation (#3128)
-* Various documentation improvements (#3133)
-* Add VM doc (#3188)
-* Update documents for TSim (#3409, #3318, #3302, #3343, #3206)
-* Improve tvm4j document describing LLVM support (#3404)
-* Tutorial migration to Python3 (#3498)
-* Android RPC README (#3500)
-* Documentation for Relay opcode (#3522)
-* Tutorial for pass manager (#3515)
-* Minimum version of Python in docs (#3588)
-* Relay pass infra (#3583)
-* X86 Autotune tutorial improvements (#3609)
-* YOLOv3 tiny Darknet tutorial (#3674)
-* SSD doc to avoid confusion (#3677)
-* Tutorial: Build a Graph Convolutional Network on TVM (#3681)
-* Add docs for analysis namespace (#3985)
-* [tutorial] Relay pass infra tutorial (#4083)
-* [DOCS] Add TensorFlow frontend docs (#4154)
-* Tutorial: update Building a Graph Convolutional Network tutorial (#4060)
-* [Docs] Add dependency of compilation with LLVM (#4117)
-* [Documentation]Fix example code in comment of `tvm.build_module.build()` (#4195)
-* TSIM: add virtual memory support to examples (#3868)
-* Relay pass infra tutorial (#4083)
-* Fix the TF tutorial to run against TF2.0 and TF1.x (#4104)
-* Add `topi.nn.fifo_buffer` to TVM doc (#4343)
-* License statement (#4345, #4359, #4401, #4402, #4408, #4409, #4410, #4414, #4431)
-
-### Build and Test
-* Increate the robuteness of CI test (#2841, #2798, #2793, #2788, #2781, #2727, #2710, #2711, #2923)
-* Improve conda build (#2742)
-* Add caffe2 nnvm frontend to CI (#3018)
-* Use bridge network and expose port on macOS when launch docker image (#3086)
-* Run DarkNet tests (#2673)
-* Add file type check (#3116)
-* Always run cpptest during build to ensure library correctness (#3147)
-* Handle more file types in ASF header (#3235)
-* Add `test_forward_ssd_mobilenet_v1` to `tflite/test_forward` (#3350)
-* Add Azure build pipeline (#3458, #3459)
-* Update ci-gpu to v0.52 (#3374)
-* Enable more visible symbols by default (#3365)
-* Separate out legacy as a stage in CI (#3337)
-* Simplify build script, remove python 2 support (#3419)
-* Ignore rust cargo lock files in rat (#3314)
-* Improve CUDA Conda package build (#3281)
-* Update CMakeLists.txt to be more flexible to find the third parties libraries (#3354)
-* Docker update conda package (#3344), requests and pillow (#3495), Android demo (#3499), rat install (#3527), ARM support (#3546), LLVM (#3590)
-* Relay-to-Python testing (#3156)
-* Code refactoring/remove (#3523, #3667)
-* Zero-rank testing (#3612)
-* CMake compilation (#3611, #3650, google test #3628)
-* Standalone wheel build for TOPI (#3657)
-* Fixing performance issues in PassUpDomain when fusing and splitting axes (#3073)
-* conda recipe (#3791)
-* Allow users to specify download directory (#3803)
-* Update docs for installation for CUDA (#3832)
-* Update `hybrid_script.rst` (#3799)
-* Acknowledge Halide attributions (#3824)
-* Add psutil dependency (#3780)
-* Temporary disable rust test (#3809)
-* Solve occasional CI issue when pad value is all 0 (#3801)
-* Towards TSIM CI testing (#3704)
-* Use pip3 for python3 (#3742)
-* Update docker image `ci_cpu,i386` to include verilator (#3738)
-* Remove sccache from Rust install (#3728)
-* Update dmlc-core to the latest commit (#3716)
-* Update GPU docker (#3709)
-* Add an option to build with -pthread (#3671)
-* Add DGL to `{ci_gpu, demo_cpu, demo_gpu}` docker images (#3692)
-* Use pytest instead of nosetest (#3524)
-* Enable NHWC of `relay.testing.mobilenet` (#3886)
-* Add .hsaco save/load for `tesnor_expr` Tutorial (#3852)
-* Support LLVM trunk (#3907)
-* Remove GTest cmake flag from install docs (#3953)
-* Allow `USE_LLVM` to take extra arguments (#3954)
-* [CI] Pin NNPack pthreadtools version (#4152)
-* [TOPI] Fix flaky testcase for check round (#4211)
-* [CI] Move gpu docker binary to cuda10 (#4229)
-* [CI] use llvm9 for the gpu tests (#4224)
-* [CI] Update GPU docker to cuda10 (#4228)
-* [Relay] Install Relay Prelude program in package install (#4227)
-* [relay] use `time_evaluator` for measurement (#4191)
-* [Relay] Improve build error when no lowered funcs are produced (#4132)
-* [llvm] switch to use Align for llvm trunk (#4051)
-* [CUDA] Update `have_int8` condition to run on compute capability 7.x devices (#4214)
-* [DOCKER] Pin torchvision==0.4.1 (#4140)
-* [DOCKER] torch install depends on future package (#4098)
-* [CodeGen] Disable -mfloat-abi hard option for LLVM < 6.0 (#4071)
-* Add a python how to example of deploying tvm module with tvm runtime only (#4094)
-* Hide symbols from dependent libraries if `HIDE_PRIVATE_SYMBOLS` is ON. (#4041)
-* [BUILD] Disable utvm standalone runtime by default (#4240)
-* Fix TSIM compile error in Linux (add missing -fPIC flag) (#3876)
-* Add scalafmt and format existing scala codebase (#3880)
-* Update TFLite wheel version to 1.13.1 (#3435)
-* Remove PEP498 f-string new feature for support python3.5 (#4250)
-* Require LLVM >= 9 for AMDGPU backend (#4253)
-* Rename ml.dmlc.tvm to org.apache.tvm (#4290)
-* [Test][TF][Relay] Fix argument preparation for vm test mode (#4296)
-* Add test for the `qnn_add` operator (#4282)
-* [CI][DOCKER] Add ONNX runtime dep (#4314)
-* [CI][DOCKER] Upgrade image to include onnx runtime (#4313)
-* [CI] Set workspace to be per executor (#4336)
-* [Build][Windows] Fix Windows build by including cctype (#4319)
-* [Contrib] Add MKL DNN option (#4323)
-* [Test][Relay][Pass] Add test case for lambda lift (#4317)
-* Remove Python imp module as it is deprecated (#4275)
-* Bump up CUDA log version in tophub.py (#4347)
-* Add rule for clean in APPs (#4364)
-* [Relay tests] Temporary Attr Update for Order-Independent Testing (#4357)
-* [CI] Avoid content-length request in test data download (#4375)
-* Compare all outputs in TFLite `test_forward_ssd_mobilenet_v1` (#4373)
-
-### Bug Fixes
-* [RELAY] Fix `get_int_tuple`. (#2691)
-* [ARITH] Select support for integer set analysis. (#2687)
-* [Relay] Fix error in ANF (too aggressively inline atomic expression and create free variable). (#2665)
-* [Hybrid Script] Fix name conflict and attached scope problem. (#2649)
-* [Relay] Fix ANF for reference and pattern matching. (#2637)
-* [Relay] Fix fusion bug when call symbol that is not an operator. (#2630)
-* Fix missing <sstream> header file. (#2629)
-* [Relay]Fix the bug in heterogeneous annotation which mistakenly steps into the fused op. (#2622)
-* [AutoTVM] Fix incorrect localhost usage in RPC mode. (#2619)
-* [NNVM] Fix incorrectly getting layout attribute as a tuple. (#2610)
-* [Relay] Fix mutating IF expression. (#2601)
-* [Tutorial] Fix downloaded file path. (#2590)
-* [Storage] Fix int32 overflow bug when input is big. (#2580)
-* [NNVM] Fix non-identity problem for FInplaceIdentity. (#2572)
-* [Golang] Fix compilation error. (#2558)
-* [Tensor Expression] Fix missing reduction init predicates. (#2495)
-* [Relay] Fix missing argument for NCHWc in Relay. (#2627)
-* [TOPI] Fix `Nms_ir` data race. (#2600)
-* Fix `compute_inline` with multiple outputs (#2934)
-* [TEXPR][PASS] Fix thread all reduce to avoid write after read hazzard (#2937)
-* [FRONTEND][TENSORFLOW] bug fix for tensorflow official slim models. (#2864)
-* [FRONTEND][ONNX] Some bug fixes and Shape operator fixed for relay. (#2850)
-* Turn on `USE_SORT` by default (#2916)
-* [DOCKER] Upgrade ci-cpu to latest v0.50 (#2901)
-* [TESTS] Import script robustness (set -u) (#2896)
-* [Relay] Fix name of bias in testing.mlp (#2892)
-* [TESTS] Improve script robustness (#2893)
-* Add dense schedules to `__init__` for cpu (#2855)
-* [Apps] [howto_deploy] fix cxx-flags order and build directory (#2888)
-* [Relay] Add TVM_DLL for ANF/GNF conversion #2883
-* [Relay] Fix Relay ARM CPU depthwise spatial pack schedule alter op layout issue. (#2861)
-* Fix setting up hints for getaddrinfo (#2872)
-* Add missing sgx includes (#2878)
-* Fix error reporting for missing axis (#2835)
-* Fix an OrderDict initilization bug. (#2862)
-* Fix Xcode 10 metal compile error (#2836)
-* tvmrpc: Fix includes (#2825)
-* Fix `init_proj.py`: Team ID expected (#2824)
-* [DOCKER] Fix git clone failure. (#2816)
-* upgrade java style-check due to CVE-2019-9658 (#2817)
-* [Relay][Quantization] Fix duplicated simulated quantization (#2803)
-* [Bugfix] Repeat and tile bug fixed, relay tests added (#2804)
-* Fix caffe2 relay frontend (#2733)
-* Fix a bug in nnvm to relay converter. (#2756)
-* Ensure loop count is a constant before trying to unroll. (#2797)
-* xcode.py: Decode bytes before output #2833
-* [WIN] Fix a bug in `find_llvm` when specify llvm-config (#2758)
-* [DLPACK] fix flaky ctypes support (#2759)
-* [Bugfix][Relay][Frontend] Fix bug in mxnet converter for `slick_like` (#2744)
-* [DOCS] Fix tutorial (#2724)
-* [TOPI][Relay] Fix default `out_dtype` for `conv2d_NCHWc` and Relay (#2702)
-* [Relay] fix checkwellform (#2705)
-* fix prelu, now can use on 2d input and add one test (#2875)
-* [CODEGEN][OPENCL] Fix compile error about ternary expression. (#2821)
-* Fix Placeholder issue (#2834)
-* Fix makedirs() condition in contrib (#2942)
-* Add missing #!/bin/bash directive (#2951)
-* Bilinear resize bug fix from PR #2777 (#2857)
-* Fix `bias_add` default axis (#2829)
-* Remove empty ty.rs (#2958)
-* fix undefined reference to dlopen, etc (#2957)
-* Removed deprecated `std::unary_function` (#2962)
-* Add output format to ndk build func (#2999)
-* Fix java checkstyle version (#2998)
-* Fix relay invariant error message (#3011)
-* Fix for caffe2 nnvm frontend (#2996)
-* Fix rust resnet example (#3000)
-* Fix x||!x for comparisons in rewrite simplifier (#3029)
-* Fix BatchMatMulRel typerelation (#3032)
-* Update dmlc-core, fix default ctors of NodeEntry (#3017)
-* Fix Fuse (#3035)
-* Fix PostOrderVisit signature (#3048)
-* Fix winograd nnpack fp16 (#3046)
-* Fix some typos (#3063, #3112)
-* Fix `group_conv2d` unit test (#3113)
-* Fix bug in ONNX importer (#3084)
-* Fixing a doc nit (#3123)
-* Fix type code error for StringImm (#3050)
-* Fix bug of wrongly generated `device_map` (#2990)
-* use `unordered_map` instead of map in ANF (#3024)
-* Fix PRelu layout in Relay (#3013)
-* Minor addition to graph runtime debug (#3129)
-* Fix mali conv2d performance regression (#3131)
-* Fix dense autotvm template registration in ROCm (#3136)
-* Fix `conv2d_transpose` (#3138)
-* Fix python lint warnings (#3145)
-* Some fixes for golang latest version compiler #3119 (#3182)
-* Add more syncs to fix flaky test caused by `get_valid_counts` (#3151)
-* Fix AlterLayout Pass (#3155)
-* Fix a multithreaded bug in llvm LazyInitJIT (#3158)
-* Fix a tensorflow test bug. (#3165)
-* Fix concat for ARM (#3061)
-* Handle vectorize for LE statement (#3137)
-* Raise exception `group_conv2d_nchw` not supported (#3195)
-* Quick fix of VTA FPGA Toolchain Installation documentation (#3196)
-* Check file exists before removing it (#3178)
-* Fix a bug of flatten in ONNX to Relay converter (#3180)
-* Fix converter where initializers were not registered as nodes (#3143)
-* Fix bug in cast to bool (#3207)
-* Hotfix `build_module` creation (#3198)
-* Fix sort changing original input data issue (#3212)
-* Fix bug in vta runtime DepPop function (#3208)
-* Fix resize nearest with fractional scaling (#3244)
-* Fix `vta_conv2d` crash issue after change `vta_config.json` (#3213)
-* Fix a memory leak in OpManager (#3263)
-* PkgConfig cause crash in PYNQ board due to link library (#3257)
-* Fix Error messages in tflite.py (#3320)
-* Fix typos in docs and comments (#3309, #3376)
-* Bugfix min/max const canonicalize rule (#3386)
-* Return module from frontend for autotvm (#3401)
-* Fix constant and reshape in ONNX (#3387)
-* Default verilator location fix (#3324)
-* Fix autodiff for conditional expression (#3453)
-* Gramatical improvements to `tensor_expr_get_started` (#3330)
-* Fix AutoTVM data structure bug (#3462)
-* Fix MXNet RNN without providing state initialization as input (#3326)
-* Fix flaky test on topk and quantize pass (#3362)
-* Add VTA PYNQ `metal_test` bitstream program logic and fix compilation issue. (#3400)
-* Fix VTA function Vivado Compile Error. (#3375)
-* Fix VTA DRAM functionality issue. (#3278)
-* Fix reshape precompute and type error in ONNX frontend (#3230)
-* Fix interpreter argument conversion for tuples. (#3349)
-* Fix code generation for packed functions + tuples in VM (#3287)
-* Fix memory leak in Relay interpreter (#3448)
-* Fix x86 depthwise conv2d `alter_op_layout` (#3264)
-* Create closure object for GlobalVar (#3411)
-* Fix getting global var in prelude (#3405)
-* Fix rfactor bugs which related to predicate and loop partition (#3382, #3444)
-* Fix the bug in AutoTVM where SimulatedAnnealingOptimizer sometimes finds useless candidate (#3413)
-* Fix name conflict in PartialEval (#3402)
-* Fix int bound analysis bug for modular (#3288)
-* Check arg positiveness for modular rules (#3279)
-* Fixes failure of `sum` and `all` on `axis=0` (#3422)
-* Fix package path in tflite test (#3427)
-* Fix Windows build (#3429)
-* Fix `LSTMBlockCell` in Tensorflow frontend (#3410)
-* TF fix where output index is ignored (#3622)
-* Runtime fix for custom datatypes (#3471)
-* Relay build module warnings (#3452)
-* Relay partial evaluator (#3482)
-* Pynq AutoTVM tracker (#3497, #3578)
-* A normal form test (#3525)
-* Lint issue (#3519, #3615 )
-* Any shape testing (#3528)
-* Android `posix_memalign` (#3532)
-* Quantization `add_rewrite` and UnifyDTypeScale (#3534)
-* Bound inference fix (#3526)
-* Tensorflow NCHW data format (#3514)
-* First order gradient (#3550)
-* JS load module example (#3556)
-* Build error (#3552)
-* Relay VM debug statements (#3565)
-* C++ lambda expr (#3570)
-* Handling of tempdir if subprocess is killed (#3574)
-* Remove tabs in Chisel source (#3603)
-* Relay VM DataTypeObject (#3604)
-* Removing prints (#3616)
-* Average Pool2D Bug (#3607)
-* Missing header in `cuda_device_api.cc` (#3621)
-* Tensorflow frontend fix where `output_shape` is None (#3632)
-* Winograd accuracy fix (#3644)
-* Fix comment (#3646)
-* Zero-input op fix for recursive traversals (#3623)
-* Python 3.5 compatibility (#3675)
-* Fix infinite recursive `device_api.ext_dev` call in VTA. (#3843)
-* Fix `depth_mult` for TensorFlow frontend (#3676)
-* Fix database APIs for AutoTVM (#3821)
-* Fix axis of softmax in Keras (#3834)
-* Fix VTA TensorLoad module (#3841)
-* Fix inconsistent python/cpp API behavior for `if_then_else`, power (#3829)
-* Fix code comment of operators in ONNX frontend (#3830)
-* Added repo for llvm-9 to fix missing dependency issue (#3826)
-* Fix typo in Relay text parser (#3785)
-* Fix tvm const warnings (#3817)
-* Add gfx906 bc (#3808)
-* Fixed onnx test failures when run on a cpu backend (#3764)
-* Fix ArgBinder assert order (#3794)
-* Fix for NoneType Target for quantization (#3792)
-* Fix out-of-date quantization realize (#3790)
-* Fix Qnn concatenate InferType (#3779)
-* Fix dense tuning (#3768)
-* Fix `visit_pattern` in ExprMutator (#3769)
-* Fix Chisel Scala style (#3765)
-* Fix some pass docs (#3767)
-* Fix mistype in rpc tutorial (#3763)
-* Fix tvm.scan follow by tvm.compute segfault (#3723)
-* Fix the potential index overflow in where operator (#3751)
-* Revert `compile_cmd` kwarg name change (#3746)
-* Update tophub (#3752)
-* Fix typo in `ir_pass.h` (#3741)
-* Bug fix for VME Shell (#3737)
-* Fix missing apt https transport support (#3735)
-* Take zero extent loops as NoOp and remove it (#3724)
-* Fix mxnet converter for hybridblock and add `div_sqrt_dim` (#3701)
-* Fix partial eval unit test name (#3719)
-* Fix conv2d schedule code (#3648, #3717)
-* Remove thread related headers (#3713)
-* Fix FunctionPass (#3712)
-* Export tvm::relay::OpRegistry::OpRegistry (#3711)
-* Fix Metal reinterpret (#3706)
-* Fix `gather_nd` in Relay (#3442)
-* Fix error in partial evaluator (#3693)
-* Align the naming rule for OpAttributeUnImplemented (#3695)
-* Enable the sparse schedule (#3651)
-* Fix typo names in Caffe2 frontend (#3685)
-* Make tests multi-process friendly. (#3683)
-* Fix typo in README.md (#3684)
-* Fix doc rendering (#3897)
-* Add test script starter command to document (#3993)
-* Add type solver unit tests for unifying quantified funcs (#3947)
-* Change Vivado install instructions to version 2018.3 (#4003)
-* Add a link to the defining network description of auto-tuning tutorial (#4023)
-* Additional MXNet Convolution and Deconvolution tests (#4026)
-* Adding support to check if an attribute is present or not without having to get the value (#3957)
-* Fix parser for cast. (#3873)
-* Fix operator fusion for multiple output (#3871)
-* Remove extern C warpper for cuBLAS (#3877)
-* Fix int32 range overflow by using int64 (#3870)
-* Remove duplicate resize (#3902)
-* Fix blas cmake for mac os (#3898)
-* Add another MKL name alias for MKL installed through pypi (#3853)
-* Numpy compatible dtype inference for `tvm.convert` and `tvm.const` (#3861)
-* Remove incorrect check for LLVM in C codegen test (#3921)
-* Fix exponential blowup in interpreter (#3559)
-* Fix CUDA int8x4 vectorize (#3928)
-* Make buffer auto broadcast independent to the order of input args (#3956)
-* Fix benchmark layout in graph tuner (#3926)
-* Fix Android Demo LLVM version (#3962)
-* Cast filepath arguments to string (#3968)
-* Fixes "common" sub crate using nightly and main (#3965)
-* Changes to make tensorize work. These changes also fix the previously broken test. (#3981)
-* Remove FLOP computation when calling 3rd party library (#4005)
-* Use a more intuitive way to limit the #ops in a group (#4018)
-* Add more `pad_mode` support for onnx converter (#4029)
-* Impose a max op limit to the op fusion pass (#4002)
-* Fixes issue with CPP enums (#4019)
-* Int64 shape handling for outputs. (#4031)
-* [PYTHON] Fix installation for generated grammar (#4223)
-* [Bugfix] Fix target host for vm compiler (#4057)
-* [Fix][VM] Fix VM invoke with `set_params` (#4079)
-* [Fix] Fix a few bugs when dtype is fp16 (#4088)
-* [Relay][Frontend][TF] Fix Size operator (#4175)
-* [cmake][ANTLR] Support setting path to ANTLR jar (#4176)
-* Fix infer type of kernel in dense. (#4125)
-* [Relay] Fix match case in Python-side expr functor (#4037)
-* Split `adaptive_pool2d_avg` into sum and div (#4186)
-* [AutoTVM] Fix Split Factors when `no_tail` is off (#4044)
-* Fix extent one for the `post_stmt` in loop partition (#3734)
-* [TOPI] Fix bug in intel graphics auto tune (#4093)
-* [ARITH] Fix lowering of `floormod(x, y) != 0` (#4127)
-* [ARITH] Fix the rule `y < x && x <= y` (#4220)
-* [Bugfix][TF] reset graph after getting tag of savedmodel (#4055)
-* [Fix] Fix the logic of the number of nodes checking in op fusion (#4074)
-* [VTA] hotfix for de10-nano driver (#4081)
-* Fixing tensor not found issue in bitserial operator (#4095)
-* Fix wrong `n_trial` number in autotvm tutorials' progress bar if `n_trial` is larger then config space. (#4070)
-* [PATCH] Fix undefined `__floatdihf` in libtvmruntime.so on aarch64. (#4119)
-* [ARITH] Fix lowering of FloorMod (#4236)
-* [Relay][Frontend][Tensorflow] Fix GatherV2 (#4238)
-* Fix typing.Deque import error for Python 3.5 (#4254)
-* [VTA] Hotfix for padded load test in Chisel VTA (#4264)
-* [Contrib] Fix error message at `callback_get_section_size()` (#4221)
-* [TOPI] Fix bug in Winograd on CUDA (#4260)
-* AutoTVM: Fix hang/crash issues on feature extraction (#3689)
-* [TOPI][CUDA] Fix Winograd Kernel Size Support (#4276)
-* [Relay][Frontend][Tensorflow] Fix type assignment for 'tf.range' operator (#4294)
-* Fix incorrect call to Unicode Win32 InetPton (#4306)
-* [Relay][Frontend][Keras] handle `batch_norm` op params well (#4310)
-* [VTA] fix error when `memory_id` is `VTA_MEM_ID_OUT` (#4330)
-* [Doc][fix] fix sphinx parsing for pass infra tutorial (#4337)
-* [Codegen] remove fp16 function override for cuda (#4331)
-* [TFLite] Fix Prelu unified shape error (#4326)
-* [Relay][Frontend][TF] Fix transpose when axes is not a param (#4327)
-* [VTA] Bug fix for padded load with large inputs (#4293)
-* Fix inconsistent operator tag name (#4134)
-* Fix for a specific case when loop partitioning with indivisble. (#4243)
-* Send list as argument to `schedule_conv2d` (#4358)
-* [Docker] Fix TVM folder name for installing on Android and OpenCL. (#4363)
-* Fix TFLite Reshape assert (#4320)
-* [Relay][Frontend][TF] Fix slice when begin or size is not Const (#4372)
-* Fix compilaton of bfloat16 on Windows (#4415)
-
-### Known Issues
-
-* The performance of Relay VM is not good enough on GPU, due to memeory allocation overhead which will be resolved later.
-* TFlite rounding vs tvm rounding causing differences in accuracy and potentially off by 1 errors. For reference #3900
-* TFlite pre-quantized network support is still a work in progress and the project would welcome further contributions.
-* TSIM build requires `python` command exist on the host. See [forum discussion](https://discuss.tvm.ai/t/vta-build-failure/4790) for details.
-* Tensorflow control flow has not been fully supported in the frontend converter.
-* `topi.floor_div` is inconsistent with floor division semantic when result number is close to an integer.
-
-
-### Depreciations
-* Deprecating python2 support and following release (v0.6). (#2994, #2986)
-* NNVM is deprecated and will be removed in a future version. (#4333, #4368)
-
-
-## 0.5
-This release features several major improvements. Some of the highlights are: Arbitrary bits quantization algorithm; High-level auto-differentiable programming IR -- Relay.
-
-- Fully featured 8-bit network support
- - 8bit quantizer
- - Arbitrary bits quantization algorithm
- - Intel cpu support
- - ARM cpu support
-- NVidia GPU 8-bit kernel
- - int8 gemm recipe
- - int8 conv2d
- - Autotvm integration
-- Automated tuning and scheduling
- - AutoTVM optimizations for mobile GPUs
- - AutoTVM optimizations for CUDA
- - AutoTVM optimizations for x86
-- Initial release of the differentiable programming IR, Relay
- - Generic & informative Relay error reporting #2408
- - Relay IR text format support #1781
- - Support control flows
- - A Normal Form Canonicalization #2251
- - Type system support
- - End to end compilation
- * Frontend support: Caffe2 #2507 , CoreML #2476 , Keras #2376 , MXNet #2163 , ONNX, TFLite #2365
- * Operator coverage #1799 #2051
- - FoldScaleAxis #2020
- - SimplifyInference #2033
- - CombineParallelConv2D #2089
- - InstrumentBoundCheckers pass #2079
- - Bind & FoldConstant #2100
- - Alter Op Layout #2150
- - General OpFusion #2090
-- CodeGen
- - Gcc / g++ compatible C code generator for TVM #2161
- - Device type annotation for heterogeneous compilation #2361
- - Cache packed func ptr, lift alloca #2070
- - Generalize compute to tensor region #1476
-- Runtime
- - Relay interpreter and compiler #1954
- - Heterogeneous runtime #1695
- - Language bindings: Golang runtime #1470 , Rust runtime #1597
- - Add min_repeat_ms to time_evaluator #2200
- - Bundled interpreter demonstration #2297
- - Enable PlanMemory in the graph runtime #2120
-- Language Binding
- - Rust frontend #2292
-- VTA
- - Improved RPC for VTA #2043
-- Hybrid python programming model
- - Support for scheduling #2416
- - Support for Inter-function call #2287
- - Backend support #2477
-- TOPI
- - Initial support for sparse tensor computation
- - Improve ARM CPU depthwise convolution performance #2345
- - Port winograd ops to relay #2356
- - Add faster-rcnn proposal op #2420
-- Tutorials and docs
- - Relay language docs #2232
- - Tutorials on how to use SGX backend
- - How to write a pass in python
- - General lowering flow of TVM
- - How to do tensorize
- - TFLite frontend tutorial #2508
- - Keras seq2seq model for translation tutorial #1815
- - Committer guide and tips #2468
- - Code review guideline on API designs #2459
-
-
-
-## 0.4
-
-This release features several major improvements. The high-level graph optimizer is now part of TVM repo. Some of the highlights are: Initial support of AutoTVM for automated optimization; customized accelerator backend VTA.
-
-- Tensor operator primitives
- - Introduce attrs field to operator primitives(e.g. compute) to store additional metadata, the attrs can be used as hint for scheduling
-- Enable embedding of asm micro-kernels
-- Hybrid python programming model
- - python AST based IR builder interface
- - support GPU programs
-- AutoTVM, Automated tuning, and scheduling
- - basic autotvm infra
- - GPU IR verifier
- - basic autotuning tutorial
- - topi integration
-- ARM support
- - winograd support
- - initial support of ARM autotuning records
-- TOPI Vision
- - Generic GPU sort support(useful for vision)
- - SSD operator support
-- TOPI numpy consistency
- - Rename all binary operators for numpy consistecy: broadcast_add-> add, broadcast_sub -> substract, broadcast_mul -> multiply, broadcast_div->divide
- - New operators: slice, LRN, equal, not_equal, less, greater
- - tutorials on topi
-- Initial low-bit operator support support
- - Optimized popcount generation on ARM
- - general bit-serial convolution and GEMM
- - optimized low bit kernels
- - parallel optimization
-- New topi backend optimization for intel graphics
-- Adapt AVX schedules for SSE target
-- VTA: customized accelerator backend
- - custom hardware backend example
- - tutorials on how to use customized accelerator
-- Initial experimental support for HLS backend
-- Bugfix in SPIRV code generator for vulkan
-- libdevice support, enable NVPTX backend
-- Introduce NDArrayContainer for managed NDarray
-- RPC and Device API
- - Support communication between big/small endian machines.
- - RPC and device API protocol upgrade (this is a non-backward compatible change) to support big-small endian communication. This is a non-backward compatible change, need to use the latest version of TVM runtime with the RPC
- - graduate rpc from contrib, tvm.contrib.rpc->tvm.rpc
- -Support tracker in Android RPC, add fault tolerance for AutoTVM
-- BIG.LITTLE aware threadpool
-- tvm4j graph runtime that runs end to end workload in java
-- DLPack support
- - Support from_dlpack and to_dlpack
- - Enables bridges to pytorch
-- Enable link of stackvm in runtime
-- Tensorflow graphdef frontend
-- Keras frontend
- - improved to support reuse layers, add activations
-- ONNX
- - gather, LRN
-- CoreML frontend
- - Support C-RNN and activation functions
-- Fix grads for sum and expand_like
-- Enhanced operator fusion for multiple elemwise branches
-- Separate nnvm fusion and compilation pass
-- Unified build system to cmake, customizable cmake path for vulkan, rocm, cuda
-
-
-## 0.3
-
-This release features numerous improvements in TOPI and backends. We make the first step toward object detection support in TOPI, featuring operators necessary for YOLO and SSDs. The topi now supports numpy-style API and operator overloading. RPC is significantly improved to support resource allocation and using a pool of devices. We are adding two new backends: WebGL for running GPUs on the browser, and Vulkan for running on next-generation graphics API.
-
-- TOPI Vision operators
- - SSD support
- - YOLO support
- - NMS operator support in vision
-- TOPI general numpy-style operators
- - numpy style operator overload in topi
- - more operators: flip, take
- - dilation support on conv2d and depthwise
-- 8bit support
- - ARM 8bit gemm
- - ARM 8bit conv
-- Low bit operator support
- - popcount intrinsics
- - 1-bit fully connected
-- Contrib: MPSDNN fully-connected and conv2d support
-- Better RPC support
- - RPC Tracker support to allow centralized resource management
- - RPC protocol upgrade (this is a non-backward compatible change) to support timeout in the proxy
- - This is a breaking change, need to use the latest version of TVM runtime with the RPC
- - Fault-tolerant to early server termination with correct exception propagated
- - RPC support enabled for ROCm AMDGPUs
-- Tutorials and docs
- - How to deploy to android devices.
-- Optimizations for hardware backends
- - intel CPU (AVX and AVX512)
-- Schedule Primitives
- - rfactor now support factor_axis to specify the factored dimension in the result
- - cache_write now support multiple output operators
- - enable warp memory which generates shuffle instructions
-- Framework bridge
- - MXNet bridge supported
-- C++ compiler API support
- - build migration
- - topi migration to c++
- - Target system in c++
-- WebGL backend
- - runtime and codegen
- - topi integration
- - end to end pipeline on the browser
-- Vulkan backend
- - vulkan runtime
- - spirv code generator
-- Security
- - intel SGX runtime support
- - multi-threaded SGX runtime
-- LLVM 7.0 support
-- Robustness
- - VerifyMemory to verify incorrect GPU schedules that writes into GPU memory from cpu
- - Verify compute formulas
-- Better CPU parallel runtime
-
-## 0.2
-
-This release comes with a complete set of TOPI support for NNVM compiler, which allows compilation of end to end workloads.
-We also make major improvements in supporting new backends: ROCm for AMDGPUs and ARM GPU.
-
-- Backend support
- - Support LLVM mainline(4.0, 5.0, 6.0)
- - Support ROCM stack for AMD GPUs
- - More robust OpenCL support for ARM GPUs
-- Android RPC runtime
-- Multi-threading optimization for ARM
- - multi-threaded depthwise
- - multi-threaded conv2d
-- New schedule primitives
- - storage_align for shared memory alignment
- - double_buffer
-- UnrollLoop : more robust version of unroll loop, count maximum steps that can be unrolled.
-- Full set of TOPI operators
- - Introduce tvm.target to specify target options for compilation better.
- - broadcast/ reduction operators
- - pooling and global pooling
- - Generic target support for topi
- - schedule with external libraries
-- End to end deep learning pipelines for CPU, GPU, ARM GPU
-- Tutorials
- - How to load compiled module in any language runtime
- - How to use java runtime
-- Contrib library: MIOpen, CuDNN
-- Ongoing items that contains functioning pieces
- - WebGL backend
- - C++ compiler support
- - MPS DNN
- - low bit support, introduced popcount
-
-
-## 0.1
-
-- Language runtime
- - python
- - javascript
- - java
- - c++
-- Backend
- - arm, x86
- - javascript, wasm
- - CUDA
- - opencl
- - Metal
-- DNN Library integration
-- RPC runtime
-- TOPI operator pipeline python
-- TOPI operator pipeline in C++
-- Rough perf of the TOPI GPU pipeline
-- Rough pref of TOPI CPU pipeline
-- End to end graph executors
-
-
-## Initial version
-
-- Pack libary into shared library.
-- External function and contrib libraries
-- DLPack integration support
-- AOT and module system
-- Basic code structure ready.
diff --git a/cmake/config.cmake b/cmake/config.cmake
index 17f587efd4..b552cd1187 100644
--- a/cmake/config.cmake
+++ b/cmake/config.cmake
@@ -329,9 +329,6 @@ set(USE_MRVL OFF)
# Whether to use QHL library
set(USE_HEXAGON_QHL OFF)
-# Whether to use ONNX codegen
-set(USE_TARGET_ONNX OFF)
-
# Whether enable BNNS runtime
set(USE_BNNS OFF)
diff --git a/cmake/modules/LibInfo.cmake b/cmake/modules/LibInfo.cmake
index 3881247ae1..b94bb42b25 100644
--- a/cmake/modules/LibInfo.cmake
+++ b/cmake/modules/LibInfo.cmake
@@ -114,7 +114,6 @@ function(add_lib_info src_file)
TVM_INFO_USE_SORT="${USE_SORT}"
TVM_INFO_USE_SPIRV_KHR_INTEGER_DOT_PRODUCT="${USE_SPIRV_KHR_INTEGER_DOT_PRODUCT}"
TVM_INFO_USE_STACKVM_RUNTIME="${USE_STACKVM_RUNTIME}"
- TVM_INFO_USE_TARGET_ONNX="${USE_TARGET_ONNX}"
TVM_INFO_USE_TENSORFLOW_PATH="${USE_TENSORFLOW_PATH}"
TVM_INFO_USE_TENSORRT_CODEGEN="${USE_TENSORRT_CODEGEN}"
TVM_INFO_USE_TENSORRT_RUNTIME="${USE_TENSORRT_RUNTIME}"
diff --git a/cmake/modules/contrib/ArmComputeLib.cmake b/cmake/modules/contrib/ArmComputeLib.cmake
index 83d5e2da1d..681a425ea4 100644
--- a/cmake/modules/contrib/ArmComputeLib.cmake
+++ b/cmake/modules/contrib/ArmComputeLib.cmake
@@ -20,9 +20,7 @@
# which is common with arm devices, we need to be able to cross-compile
# a relay graph on x86 for AArch and then run the graph on AArch.
if(USE_ARM_COMPUTE_LIB)
- tvm_file_glob(GLOB ACL_RELAY_CONTRIB_SRC src/relay/backend/contrib/arm_compute_lib/*.cc)
tvm_file_glob(GLOB ACL_RUNTIME_MODULE src/runtime/contrib/arm_compute_lib/acl_runtime.cc)
- list(APPEND COMPILER_SRCS ${ACL_RELAY_CONTRIB_SRC})
if(NOT USE_ARM_COMPUTE_LIB_GRAPH_EXECUTOR)
list(APPEND COMPILER_SRCS ${ACL_RUNTIME_MODULE})
diff --git a/cmake/modules/contrib/ONNX.cmake b/cmake/modules/contrib/ONNX.cmake
deleted file mode 100644
index 55405f33ca..0000000000
--- a/cmake/modules/contrib/ONNX.cmake
+++ /dev/null
@@ -1,22 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements. See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership. The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License. You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied. See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-if(USE_TARGET_ONNX)
- message(STATUS "Build with contrib.codegen_onnx")
- tvm_file_glob(GLOB ONNX_CONTRIB_SRC src/runtime/contrib/onnx/onnx_module.cc)
- list(APPEND RUNTIME_SRCS ${ONNX_CONTRIB_SRC})
-endif(USE_TARGET_ONNX)
diff --git a/conftest.py b/conftest.py
index 88e21f4941..81c1f90be4 100644
--- a/conftest.py
+++ b/conftest.py
@@ -30,22 +30,7 @@ REPO_ROOT = Path(__file__).resolve().parent
# that should be allocated to test shards in a round-robin fashion. These are
# taken from the 20 (arbitrary number) of tests as from
# https://ci.tlcpack.ai/job/tvm/job/main/2907/testReport
-_slowest_tests = [
- "tests/python/topi/python/test_topi_conv2d_int8.py::test_conv2d_nchw[int8]",
- "tests/python/topi/python/test_topi_conv2d_int8.py::test_conv2d_nchw[uint8]",
- "tests/python/topi/python/test_topi_upsampling.py::test_upsampling3d",
- "tests/python/topi/python/test_topi_upsampling.py::test_upsampling3d",
- "tests/python/topi/python/test_topi_conv2d_int8.py::test_conv2d_nchw[int8]",
- "tests/python/topi/python/test_topi_conv2d_int8.py::test_conv2d_nchw[uint8]",
- "tests/python/topi/python/test_topi_conv2d_NCHWc.py::test_conv2d_NCHWc",
- "tests/python/topi/python/test_topi_conv2d_hwnc_tensorcore.py::test_conv2d_hwnc_tensorcore",
- "tests/python/contrib/test_tensorrt.py::test_binary[compile]",
- "tests/python/topi/python/test_topi_conv2d_NCHWc.py::test_conv2d_NCHWc",
- "tests/python/relay/test_py_converter.py::test_global_recursion",
- "tests/python/relay/test_op_level6.py::test_topk",
- "tests/python/topi/python/test_topi_conv2d_winograd.py::test_conv2d_nchw",
- "tests/python/relay/test_py_converter.py::test_global_recursion",
-]
+_slowest_tests = []
HARDCODED_ALLOCATIONS = {}
for idx, test in enumerate(_slowest_tests):
HARDCODED_ALLOCATIONS[test] = idx
diff --git a/docs/arch/index.rst b/docs/arch/index.rst
index 717876d2db..4df78f6c3c 100644
--- a/docs/arch/index.rst
+++ b/docs/arch/index.rst
@@ -170,7 +170,7 @@ Summary and Discussions
In summary, the key data structures in the compilation flows are:
-- IRModule: contains relay.Function and tir.PrimFunc
+- IRModule: contains relax.Function and tir.PrimFunc
- runtime.Module: contains runtime.PackedFunc
Most parts of the compilation are transformations among the key data structures.
@@ -254,7 +254,7 @@ The components in `tvm/ir` are shared by `tvm/relay` and `tvm/tir`, notable ones
- PassContext and Pass
- Op
-Different variants of functions(e.g. relay.Function and tir.PrimFunc) can co-exist in an IRModule.
+Different variants of functions(e.g. relax.Function and tir.PrimFunc) can co-exist in an IRModule.
While these variants may not have the same content representation, they use the same data structure to represent types.
As a consequence, we use the same data structure to represent function (type) signatures of these variants.
The unified type system allows one function variant to call another function
diff --git a/docs/arch/pass_infra.rst b/docs/arch/pass_infra.rst
index 1e320dceba..ae92874fc7 100644
--- a/docs/arch/pass_infra.rst
+++ b/docs/arch/pass_infra.rst
@@ -193,7 +193,7 @@ optimization passes, e.g., function-level passes, module-level passes, and
sequential passes. Each subclass itself could act as a pass manager. For
instance, they could collect the required passes and execute them or build
a dependency graph based on the given metadata. The full definition of them
-can be found in `src/relay/ir/transform.cc`_ and `src/ir/transform.cc`_.
+can be found in `src/ir/transform.cc`_.
Module-Level Passes
^^^^^^^^^^^^^^^^^^^
@@ -267,12 +267,6 @@ of passes for execution.
Module operator()(const Module& mod, const PassContext& pass_ctx) const final;
};
-Only a few passes currently in Relay are put in this group. For example,
-``FoldScaleAxis`` requires to dispatch ``ForwardFoldScaleAxis`` and
-``BackwardFoldScaleAxis`` internally. In addition, ``BackwardFoldScaleAxis`` is
-recommended to be fulfilled first. This pass, hence, is an ideal candidate for
-``SequentialPass``.
-
The following code shows how individual passes in a sequential pass are invoked.
Essentially, we sequentially execute each pass in a sequential pass using the
order that they were appended to the pass list.
@@ -311,7 +305,7 @@ pass is registered with an API endpoint as we will show later.
Pass GetPass(const std::string& pass_name) {
using tvm::runtime::Registry;
- std::string fpass_name = "relay._transform." + pass_name;
+ std::string fpass_name = "relax.transform." + pass_name;
const auto* f = Registry::Get(fpass_name);
ICHECK(f != nullptr) << "Cannot find " << fpass_name
<< "to create the pass " << pass_name;
@@ -350,8 +344,8 @@ Pass Registration
We've covered the concept of different level of passes and the context used for
compilation. It would be interesting to see how easily users can register
a pass. Let's take const folding as an example. This pass has already been
-implemented to fold constants in a Relay function (found in
-`src/relay/transforms/fold_constant.cc`_).
+implemented to fold constants in a Relax function (found in
+`src/relax/transforms/fold_constant.cc`_).
An API was provided to perform the ``Expr`` to ``Expr`` transformation.
@@ -368,7 +362,7 @@ indicates that no prerequisite is required for this pass. Otherwise, the pass
developer has to identify and list them.
Meanwhile, a pass API endpoint is registered with the name
-``relay._transform.FoldConstant``. This pass, therefore, becomes an entry in the
+``"relax.transform.FoldConstant``. This pass, therefore, becomes an entry in the
registry that can be accessed by both C++ (e.g. the ``GetPass`` above) and
Python when needed.
@@ -378,19 +372,17 @@ Python when needed.
Pass FoldConstant() {
runtime::TypedPackedFunc<Function(Function, IRModule, PassContext)> pass_func =
- [=](Function f, IRModule m, PassContext pc) {
- return Downcast<Function>(FoldConstant(f));
- };
- return CreateFunctionPass(pass_func, 2, "FoldConstant", {});
+ [=](Function f, IRModule m, PassContext pc) { return ConstantFolder::Fold(f, m); };
+ return CreateFunctionPass(pass_func, 0, "FoldConstant", {});
}
- TVM_REGISTER_GLOBAL("relay._transform.FoldConstant")
+ TVM_REGISTER_GLOBAL("relax.transform.FoldConstant")
.set_body_typed(FoldConstant);
} // namespace transform
To allow other C++ modules to apply this pass, we declare a free function in
-`include/tvm/relay/transform.h`_ as the following:
+`include/tvm/relax/transform.h`_ as the following:
.. code:: c++
@@ -543,7 +535,7 @@ Python Frontend
Only some simple APIs are needed for the frontend side. For example, we can
provide users the following APIs to create and execute a pass (full
-implementation is provided in `python/tvm/relay/transform/transform.py`_ and
+implementation is provided in `python/tvm/relax/transform/transform.py`_ and
`python/tvm/ir/transform.py`_). The backend
receives the information and decides which function it should use to create
a Pass object.
@@ -586,96 +578,6 @@ loop unrolling pass
Please refer to `src/tir/transforms/unroll_loop.cc`_ for more details.
-Pass Objects
-^^^^^^^^^^^^
-
-``Pass`` is the base class of all pass objects. All methods here are just simple
-wrappers that were implemented in the backend. They are defined for users to
-conveniently interact with the base class in Python. Only a ``__call__`` is
-defined in the pass base class to make the subclasses as callable objects so
-that they can be invoked easily (e.g., ``pass_xx(arg)``) for execution.
-
-.. code:: python
-
- @register_relay_node
- class Pass(RelayNode):
- def __call__(self, mod):
- return _transform.RunPass(self, mod)
-
-Some auxiliary APIs are provided to enable easy creation of passes from
-the Python frontend and to let the pass infra control the execution. For
-example, ``module_pass``, ``function_pass``, and ``sequential`` are provided to
-users so that they can customize their own pass or pass pipeline.
-
-For all the passes that are implemented in the C++ backend, we provide
-corresponding Python APIs in `python/tvm/ir/transform.py`_ and
-`python/tvm/relay/transform/transform.py`_, respectively. For instance,
-const folding has a Python API like the following:
-
-.. code:: python
-
- def FoldConstant():
- return _transform.FoldConstant()
-
-Users can build a pass through decoration like the following:
-
-.. code:: python
-
- @relay.transform.module_pass(opt_level=2)
- def transform(mod, ctx):
- tp = relay.TensorType((10,), "float32")
- x = relay.var("x", tp)
- gv = relay.GlobalVar("abs")
- func = relay.Function([x], relay.abs(x))
- new_mod = tvm.IRModule({gv: func})
- new_mod.update(mod)
- return new_mod
-
- module_pass = transform
- assert isinstance(module_pass, transform.ModulePass)
- assert module_pass.info.opt_level == 2
-
-The ``transform`` function here adds an ``abs`` function to the input module,
-but it could be any customized optimizations at the module level. After
-creating this ``module_pass``, users can apply it on any Relay module. For
-example, we can build an empty module and apply this pass to add an ``abs``
-function.
-
-.. code:: python
-
- mod = tvm.IRModule()
- mod = module_pass(mod)
-
-Correspondingly, we also offer such functionality for ``function_pass``. For
-instance, an example function-level pass could be written as the following:
-
-.. code:: python
-
- @relay.transform.function_pass(opt_level=1)
- class TestReplaceFunc:
- def __init__(self, new_func):
- self.new_func = new_func
- def transform_function(self, func, mod, ctx):
- # Just for demo purposes
- # Transform func to new_func
- return self.new_func
-
- x = relay.var("x", shape=(10, 20))
- f1 = relay.Function([x], x)
- f2 = relay.Function([x], relay.log(x))
- # fpass is now a special pass that replaces every
- # function to f1
- fpass = TestReplaceFunc(f1)
- # Now every function in input_mod is replaced by f1
- res_mod = fpass(input_mod)
-
-
-Alternatively, users can also directly register a pass without using the
-decorators and then invoke it. For more examples about how to customize your own
-optimization pipeline and debug Relay and tir passes, please refer to the
-`use pass infra`_ tutorial.
-
-
.. _pass_instrument_py_frontend:
Pass Instrument
@@ -741,17 +643,17 @@ new ``PassInstrument`` are called.
.. _include/tvm/support/with.h: https://github.com/apache/tvm/blob/main/include/tvm/support/with.h
-.. _src/relay/ir/transform.cc: https://github.com/apache/tvm/blob/main/src/relay/ir/transform.cc
+.. _src/relax/ir/transform.cc: https://github.com/apache/tvm/blob/main/src/relax/ir/transform.cc
.. _src/ir/transform.cc: https://github.com/apache/tvm/blob/main/src/ir/transform.cc
.. _src/ir/instrument.cc: https://github.com/apache/tvm/blob/main/src/ir/instrument.cc
-.. _src/relay/transforms/fold_constant.cc: https://github.com/apache/tvm/blob/main/src/relay/transforms/fold_constant.cc
+.. _src/relax/transforms/fold_constant.cc: https://github.com/apache/tvm/blob/main/src/relax/transforms/fold_constant.cc
-.. _python/tvm/relay/transform/transform.py: https://github.com/apache/tvm/blob/main/python/tvm/relay/transform/transform.py
+.. _python/tvm/relax/transform/transform.py: https://github.com/apache/tvm/blob/main/python/tvm/relax/transform/transform.py
-.. _include/tvm/relay/transform.h: https://github.com/apache/tvm/blob/main/include/tvm/relay/transform.h
+.. _include/tvm/relax/transform.h: https://github.com/apache/tvm/blob/main/include/tvm/relax/transform.h
.. _python/tvm/ir/transform.py: https://github.com/apache/tvm/blob/main/python/tvm/ir/transform.py
diff --git a/include/tvm/ir/affine_type.h b/include/tvm/ir/affine_type.h
deleted file mode 100644
index 5726e9eec1..0000000000
--- a/include/tvm/ir/affine_type.h
+++ /dev/null
@@ -1,150 +0,0 @@
-/*
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements. See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership. The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * "License"); you may not use this file except in compliance
- * with the License. You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations
- * under the License.
- */
-
-/*!
- * \file tvm/ir/affine_type.h
- * \brief Quantized Tensor Types.
- */
-#ifndef TVM_IR_AFFINE_TYPE_H_
-#define TVM_IR_AFFINE_TYPE_H_
-
-#include <tvm/ir/expr.h>
-#include <tvm/ir/type.h>
-
-namespace tvm {
-
-/*!
- * \brief AffineType representation
- * \sa AffineType
- */
-class AffineTypeNode : public Object {
- public:
- /*!
- * \brief Span that points to the original source code.
- * Reserved debug information.
- */
- mutable Span span;
-
- static constexpr const char* _type_key = "AffineType";
- static constexpr const bool _type_has_method_sequal_reduce = true;
- static constexpr const bool _type_has_method_shash_reduce = true;
- TVM_DECLARE_BASE_OBJECT_INFO(AffineTypeNode, Object);
-};
-
-/*!
- * \brief Managed reference to AffineTypeNode.
- * \sa AffineTypeNode
- */
-class AffineType : public ObjectRef {
- public:
- TVM_DEFINE_OBJECT_REF_METHODS(AffineType, ObjectRef, AffineTypeNode);
-};
-
-/*!
- * \brief TensorAffineType representation
- * \sa TensorAffineType
- *
- * This Type represents a quantized integer tensor that can be converted
- * back to real space via the x_real = scale * (x_quant - zero_point)
- */
-class TensorAffineTypeNode : public AffineTypeNode {
- public:
- /*! \brief The scale of this type */
- RelayExpr scale;
- /*! \brief The zero point of this type */
- RelayExpr zero_point;
- /*! \brief The data type of this type */
- DataType dtype;
- /*! \brief The axis for per-channel quantization */
- int axis;
-
- void VisitAttrs(tvm::AttrVisitor* v) {
- v->Visit("scale", &scale);
- v->Visit("zero_point", &zero_point);
- v->Visit("dtype", &dtype);
- v->Visit("axis", &axis);
- }
-
- bool SEqualReduce(const TensorAffineTypeNode* other, SEqualReducer equal) const {
- equal->MarkGraphNode();
- return equal(scale, other->scale) && equal(zero_point, other->zero_point) &&
- equal(dtype, other->dtype) && equal(axis, other->axis);
- }
-
- void SHashReduce(SHashReducer hash_reduce) const {
- hash_reduce->MarkGraphNode();
- hash_reduce(scale);
- hash_reduce(zero_point);
- hash_reduce(dtype);
- hash_reduce(axis);
- }
-
- static constexpr const char* _type_key = "TensorAffineType";
- TVM_DECLARE_BASE_OBJECT_INFO(TensorAffineTypeNode, AffineTypeNode);
-};
-
-/*!
- * \brief Managed reference to AffineTypes.
- * \sa AffineTypeNode
- */
-class TensorAffineType : public AffineType {
- public:
- TVM_DLL TensorAffineType(RelayExpr scale, RelayExpr zero_point, DataType dtype, int axis);
-
- TVM_DEFINE_OBJECT_REF_METHODS(TensorAffineType, AffineType, TensorAffineTypeNode);
-};
-
-/*!
- * \brief TupleAffineType representation
- * \sa TupleAffineType
- */
-class TupleAffineTypeNode : public AffineTypeNode {
- public:
- /*! \brief The types of this tuple*/
- Array<TensorAffineType> types;
-
- void VisitAttrs(tvm::AttrVisitor* v) { v->Visit("types", &types); }
-
- bool SEqualReduce(const TupleAffineTypeNode* other, SEqualReducer equal) const {
- equal->MarkGraphNode();
- return equal(types, other->types);
- }
-
- void SHashReduce(SHashReducer hash_reduce) const {
- hash_reduce->MarkGraphNode();
- hash_reduce(types);
- }
-
- static constexpr const char* _type_key = "TupleAffineType";
- TVM_DECLARE_BASE_OBJECT_INFO(TupleAffineTypeNode, AffineTypeNode);
-};
-
-/*!
- * \brief Managed reference to TupleAffineTypes.
- * \sa TupleAffineType
- */
-class TupleAffineType : public AffineType {
- public:
- TVM_DLL TupleAffineType(Array<TensorAffineType> types);
-
- TVM_DEFINE_OBJECT_REF_METHODS(TupleAffineType, AffineType, TupleAffineTypeNode);
-};
-
-} // namespace tvm
-#endif // TVM_IR_AFFINE_TYPE_H_
diff --git a/include/tvm/ir/expr.h b/include/tvm/ir/expr.h
index efde523851..b3b4e8ab32 100644
--- a/include/tvm/ir/expr.h
+++ b/include/tvm/ir/expr.h
@@ -353,13 +353,13 @@ TVM_DLL PrimExpr operator~(PrimExpr a);
/*!
* \brief Base node of all non-primitive expressions.
*
- * RelayExpr supports tensor types, functions and ADT as
- * first class citizens. The life-cycle of the corresponding
+ * RelaxExpr supports tensor and functions as first class citizen.
+ * The life-cycle of the corresponding
* objects are implicitly managed by the language.
*
- * \sa RelayExpr
+ * \sa RelaxExpr
*/
-class RelayExprNode : public BaseExprNode {
+class RelaxExprNode : public BaseExprNode {
public:
/*!
* \brief Stores the result of type inference(type checking).
@@ -393,55 +393,18 @@ class RelayExprNode : public BaseExprNode {
template <typename TTypeNode>
inline const TTypeNode* type_as() const;
- /*!
- * \brief The virtual device (VirtualDevice) for this node (the result of device planning).
- * For first-order expressions (non functions), this describes where the result of evaluating the
- * expression should be stored. Note that currently, all composite first-order values (tuples,
- * references, ADTs) must be stored on the same virtual device. This means that it is not possible
- * to store two tuple fields on different devices, so we only need one virtual device for these
- * types.
- *
- * For expressions that have the function type, the virtual device describes where the result of
- * the call to the function or closure is stored (instead of where the function itself is stored).
- * For example, the virtual device of f = fn(x) { body } is the virtual device of f(y), not where
- * the function itself is stored. Note that f(y)'s virtual device will be the same as the virtual
- * device of body. For more details, see the documentation in
- * src/relay/transforms/device_planner.cc.
- *
- * The VirtualDevice's Target field describes how the body of the function should be compiled.
- *
- * Set to VirtualDevice::FullyUnconstrained by default.
- *
- * \note Unfortunately, the type of virtual_device_ needs to be ObjectRef to avoid a circular
- * import.
- */
- mutable ObjectRef virtual_device_;
-
- /*!
- * \return The virtual device (VirtualDevice).
- * If the virtual device is not defined, returns VirtualDevice::FullyUnconstrained().
- * Note that for function types, the virtual device is the device where the result of a
- * call to the function is stored, not where the function itself lives.
- * For example, the virtual device of f = fn(x) { body } is the virtual device of f(y), not where
- * the function itself is stored. Note that f(y)'s virtual device will be the same as the virtual
- * device of body.
- *
- * See the documentation of the virtual_device_ field (above) for more details.
- */
- VirtualDevice virtual_device() const;
-
- static constexpr const char* _type_key = "RelayExpr";
+ static constexpr const char* _type_key = "RelaxExpr";
static constexpr const uint32_t _type_child_slots = 22;
- TVM_DECLARE_BASE_OBJECT_INFO(RelayExprNode, BaseExprNode);
+ TVM_DECLARE_BASE_OBJECT_INFO(RelaxExprNode, BaseExprNode);
};
/*!
- * \brief Managed reference to RelayExprNode.
- * \sa RelayExprNode
+ * \brief Managed reference to RelaxExprNode.
+ * \sa RelaxExprNode
*/
-class RelayExpr : public BaseExpr {
+class RelaxExpr : public BaseExpr {
public:
- TVM_DEFINE_OBJECT_REF_METHODS(RelayExpr, BaseExpr, RelayExprNode);
+ TVM_DEFINE_OBJECT_REF_METHODS(RelaxExpr, BaseExpr, RelaxExprNode);
};
class GlobalVar;
@@ -453,14 +416,13 @@ class GlobalVar;
*
* \sa GlobalVarNode
*/
-class GlobalVarNode : public RelayExprNode {
+class GlobalVarNode : public RelaxExprNode {
public:
/*! \brief The name of the variable, this only acts as a hint. */
String name_hint;
void VisitAttrs(AttrVisitor* v) {
v->Visit("name_hint", &name_hint);
- v->Visit("virtual_device_", &virtual_device_);
v->Visit("span", &span);
v->Visit("_checked_type_", &checked_type_);
v->Visit("struct_info_", &struct_info_);
@@ -477,18 +439,18 @@ class GlobalVarNode : public RelayExprNode {
}
static constexpr const char* _type_key = "GlobalVar";
- TVM_DECLARE_FINAL_OBJECT_INFO(GlobalVarNode, RelayExprNode);
+ TVM_DECLARE_FINAL_OBJECT_INFO(GlobalVarNode, RelaxExprNode);
};
/*!
* \brief Managed reference to GlobalVarNode.
* \sa GlobalVarNode
*/
-class GlobalVar : public RelayExpr {
+class GlobalVar : public RelaxExpr {
public:
TVM_DLL explicit GlobalVar(String name_hint, Type type = {}, Span span = {});
- TVM_DEFINE_OBJECT_REF_METHODS(GlobalVar, RelayExpr, GlobalVarNode);
+ TVM_DEFINE_OBJECT_REF_METHODS(GlobalVar, RelaxExpr, GlobalVarNode);
TVM_DEFINE_OBJECT_REF_COW_METHOD(GlobalVarNode);
};
@@ -747,15 +709,15 @@ class Range : public ObjectRef {
};
// implementations
-inline const Type& RelayExprNode::checked_type() const {
+inline const Type& RelaxExprNode::checked_type() const {
ICHECK(checked_type_.defined()) << "internal error: the type checker has "
<< "not populated the checked_type "
- << "field for " << GetRef<RelayExpr>(this);
+ << "field for " << GetRef<RelaxExpr>(this);
return this->checked_type_;
}
template <typename TTypeNode>
-inline const TTypeNode* RelayExprNode::type_as() const {
+inline const TTypeNode* RelaxExprNode::type_as() const {
static_assert(std::is_base_of<TypeNode, TTypeNode>::value,
"TType must be a special case of type");
ICHECK(checked_type_.defined())
@@ -872,7 +834,7 @@ struct PackedFuncValueConverter<tvm::FloatImm> {
* integer arguments, instead of runtime::Int. For backwards
* compatibility where the callee has been updated to expected a
* runtime::Int, the caller has not been updated to provide a
- * runtime::Int (e.g. relay script parsing), and the auto-unboxing of
+ * runtime::Int, and the auto-unboxing of
* runtime::Int does not apply (e.g. making an `Array<runtime::Int>`),
* allow the IntImm to be generated.
*/
diff --git a/include/tvm/ir/function.h b/include/tvm/ir/function.h
index 3845409968..2282cb979b 100644
--- a/include/tvm/ir/function.h
+++ b/include/tvm/ir/function.h
@@ -136,7 +136,7 @@ constexpr const char* kGlobalSymbol = "global_symbol";
*
* \sa BaseFunc
*/
-class BaseFuncNode : public RelayExprNode {
+class BaseFuncNode : public RelaxExprNode {
public:
/*! \brief Additional attributes storing the meta-data */
DictAttrs attrs;
@@ -220,16 +220,16 @@ class BaseFuncNode : public RelayExprNode {
static constexpr const char* _type_key = "BaseFunc";
static constexpr const uint32_t _type_child_slots = 2;
- TVM_DECLARE_BASE_OBJECT_INFO(BaseFuncNode, RelayExprNode);
+ TVM_DECLARE_BASE_OBJECT_INFO(BaseFuncNode, RelaxExprNode);
};
/*!
* \brief Managed reference to BaseFuncNode.
* \sa BaseFuncNode
*/
-class BaseFunc : public RelayExpr {
+class BaseFunc : public RelaxExpr {
public:
- TVM_DEFINE_OBJECT_REF_METHODS(BaseFunc, RelayExpr, BaseFuncNode);
+ TVM_DEFINE_OBJECT_REF_METHODS(BaseFunc, RelaxExpr, BaseFuncNode);
};
} // namespace tvm
diff --git a/include/tvm/ir/memory_pools.h b/include/tvm/ir/memory_pools.h
deleted file mode 100644
index ebab13cf3a..0000000000
--- a/include/tvm/ir/memory_pools.h
+++ /dev/null
@@ -1,361 +0,0 @@
-/*
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements. See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership. The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * "License"); you may not use this file except in compliance
- * with the License. You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations
- * under the License.
- */
-
-/*!
- * \file tvm/ir/memory_pools.h
- * \brief The object definition for relay.build argument type of memory pools
- */
-#ifndef TVM_IR_MEMORY_POOLS_H_
-#define TVM_IR_MEMORY_POOLS_H_
-
-#include <tvm/runtime/registry.h>
-#include <tvm/target/target.h>
-
-struct TVMConstantInfo;
-namespace tvm {
-
-/*!
- * \brief Describes a pool of memory accessible by one or more targets.
- */
-struct PoolInfoNode : public Object {
- public:
- /*! \brief The name of the memory pool */
- String pool_name;
- /*! \brief The expected size hint to be used by the allocator.
- * The size_hint_bytes is set to kUnrestrictedPoolSizeHint
- * to indicate the pool is not size restricted.
- */
- Integer size_hint_bytes;
- /*! \brief The clock frequency of the memory in Hz */
- Integer clock_frequency_hz;
- /*! \brief The read bandwidth in bytes/cycle */
- Integer read_bandwidth_bytes_per_cycle;
- /*! \brief The write bandwidth in bytes/cycle */
- Integer write_bandwidth_bytes_per_cycle;
- /*! \brief The read latency in cycles */
- Integer read_latency_cycles;
- /*! \brief The write latency in cycles */
- Integer write_latency_cycles;
- /*! \brief The burst length in bytes for each Target */
- Map<Target, Integer> target_burst_bytes;
- /*! \brief Whether pool is internally generated.
- * The internal pools will be generated as part of
- * the entry point code generation of the executor
- */
- bool is_internal = false;
-
- /*! \brief The targets linked to the pool */
- Array<Target> targets;
-
- void VisitAttrs(tvm::AttrVisitor* v) {
- v->Visit("pool_name", &pool_name);
- v->Visit("targets", &targets);
- v->Visit("size_hint_bytes", &size_hint_bytes);
- v->Visit("clock_frequency_hz", &clock_frequency_hz);
- v->Visit("read_bandwidth_bytes_per_cycle", &read_bandwidth_bytes_per_cycle);
- v->Visit("write_bandwidth_bytes_per_cycle", &write_bandwidth_bytes_per_cycle);
- v->Visit("read_latency_cycles", &read_latency_cycles);
- v->Visit("write_latency_cycles", &write_latency_cycles);
- v->Visit("target_burst_bytes", &target_burst_bytes);
- v->Visit("is_internal", &is_internal);
- }
-
- bool SEqualReduce(const PoolInfoNode* other, SEqualReducer equal) const {
- return equal(pool_name, other->pool_name) && equal(size_hint_bytes, other->size_hint_bytes) &&
- equal(clock_frequency_hz, other->clock_frequency_hz) &&
- equal(read_bandwidth_bytes_per_cycle, other->read_bandwidth_bytes_per_cycle) &&
- equal(write_bandwidth_bytes_per_cycle, other->write_bandwidth_bytes_per_cycle) &&
- equal(read_latency_cycles, other->read_latency_cycles) &&
- equal(write_latency_cycles, other->write_latency_cycles) &&
- equal(target_burst_bytes, other->target_burst_bytes) &&
- equal(is_internal, other->is_internal);
- }
-
- void SHashReduce(SHashReducer hash_reduce) const {
- hash_reduce(pool_name);
- hash_reduce(size_hint_bytes);
- hash_reduce(clock_frequency_hz);
- hash_reduce(read_bandwidth_bytes_per_cycle);
- hash_reduce(write_bandwidth_bytes_per_cycle);
- hash_reduce(read_latency_cycles);
- hash_reduce(write_latency_cycles);
- hash_reduce(target_burst_bytes);
- hash_reduce(is_internal);
- }
-
- static constexpr const char* _type_key = "ir.PoolInfo";
- TVM_DECLARE_BASE_OBJECT_INFO(PoolInfoNode, Object);
-};
-
-/*!
- * \brief The string parameter to indicate read and write access to a pool
- * This needs to be kept in sync with PoolInfo.READ_WRITE_ACCESS in
- * python/tvm/ir/memory_pools.py
- */
-static constexpr const char* kTargetPoolReadWriteAccess = "rw";
-
-/*!
- * \brief The string parameter to indicate read only access to a pool
- * This needs to be kept in sync with PoolInfo.READ_ONLY_ACCESS in
- * python/tvm/ir/memory_pools.py
- */
-static constexpr const char* kTargetPoolReadOnlyAccess = "ro";
-
-/*! \brief The PoolSize is unrestricted for the memory planner */
-static const int kUnrestrictedPoolSizeHint = -1;
-
-/*! \brief The clock frequency is not known */
-static const int kUnknownClockFrequency = -1;
-
-/*! \brief The read bandwidth is not known */
-static const int kUnknownReadBandwidth = -1;
-
-/*! \brief The write bandwidth is not known */
-static const int kUnknownWriteBandwidth = -1;
-
-/*! \brief Base class for WorkspacePoolInfo and ConstantPoolInfo */
-class PoolInfo : public ObjectRef {
- protected:
- TVM_DLL PoolInfo(String pool_name, Integer size_hint_bytes = kUnrestrictedPoolSizeHint,
- Integer clock_frequency_hz = kUnknownClockFrequency,
- Integer read_bandwidth_bytes_per_cycle = kUnknownReadBandwidth,
- Integer write_bandwidth_bytes_per_cycle = kUnknownWriteBandwidth,
- Integer read_latency_cycles = 0, Integer write_latency_cycles = 0,
- Map<Target, Integer> target_burst_bytes = {}, Bool is_internal = Bool(false));
-
- public:
- TVM_DEFINE_OBJECT_REF_METHODS(PoolInfo, ObjectRef, PoolInfoNode);
-};
-
-/*!
- * \brief Describes a pool of memory properties
- */
-struct PoolInfoPropertiesNode : public Object {
- /*! \brief The expected size hint to be used by the allocator.
- * The size_hint_bytes is set to kUnrestrictedPoolSizeHint
- * to indicate the pool is not size restricted.
- */
- Integer size_hint_bytes = kUnrestrictedPoolSizeHint;
- /*! \brief The clock frequency of the memory in Hz */
- Integer clock_frequency_hz = kUnknownClockFrequency;
- /*! \brief The read bandwidth in bytes/cycle */
- Integer read_bandwidth_bytes_per_cycle = kUnknownReadBandwidth;
- /*! \brief The write bandwidth in bytes/cycle */
- Integer write_bandwidth_bytes_per_cycle = kUnknownWriteBandwidth;
- /*! \brief The read latency in cycles */
- Integer read_latency_cycles = 0;
- /*! \brief The write latency in cycles */
- Integer write_latency_cycles = 0;
- /*! \brief The burst length in bytes for each Target */
- Map<Target, Integer> target_burst_bytes{};
- /*! \brief Whether pool is internally generated.
- * The internal pools will be generated as part of
- * the entry point code generation of the executor
- */
- bool is_internal = false;
-
- void VisitAttrs(tvm::AttrVisitor* v) {
- v->Visit("size_hint_bytes", &size_hint_bytes);
- v->Visit("clock_frequency_hz", &clock_frequency_hz);
- v->Visit("read_bandwidth_bytes_per_cycle", &read_bandwidth_bytes_per_cycle);
- v->Visit("write_bandwidth_bytes_per_cycle", &write_bandwidth_bytes_per_cycle);
- v->Visit("read_latency_cycles", &read_latency_cycles);
- v->Visit("write_latency_cycles", &write_latency_cycles);
- v->Visit("target_burst_bytes", &target_burst_bytes);
- v->Visit("is_internal", &is_internal);
- }
-
- bool SEqualReduce(const PoolInfoPropertiesNode* other, SEqualReducer equal) const {
- return equal(size_hint_bytes, other->size_hint_bytes) &&
- equal(clock_frequency_hz, other->clock_frequency_hz) &&
- equal(read_bandwidth_bytes_per_cycle, other->read_bandwidth_bytes_per_cycle) &&
- equal(write_bandwidth_bytes_per_cycle, other->write_bandwidth_bytes_per_cycle) &&
- equal(read_latency_cycles, other->read_latency_cycles) &&
- equal(write_latency_cycles, other->write_latency_cycles) &&
- equal(target_burst_bytes, other->target_burst_bytes) &&
- equal(is_internal, other->is_internal);
- }
-
- void SHashReduce(SHashReducer hash_reduce) const {
- hash_reduce(size_hint_bytes);
- hash_reduce(clock_frequency_hz);
- hash_reduce(read_bandwidth_bytes_per_cycle);
- hash_reduce(write_bandwidth_bytes_per_cycle);
- hash_reduce(read_latency_cycles);
- hash_reduce(write_latency_cycles);
- hash_reduce(target_burst_bytes);
- hash_reduce(is_internal);
- }
-
- static constexpr const char* _type_key = "ir.PoolInfoProperties";
- TVM_DECLARE_FINAL_OBJECT_INFO(PoolInfoPropertiesNode, Object);
-};
-
-class PoolInfoProperties : public ObjectRef {
- public:
- TVM_DLL PoolInfoProperties(Integer size_hint_bytes,
- Integer clock_frequency_hz = kUnknownClockFrequency,
- Integer read_bandwidth_bytes_per_cycle = kUnknownReadBandwidth,
- Integer write_bandwidth_bytes_per_cycle = kUnknownWriteBandwidth,
- Integer read_latency_cycles = 0, Integer write_latency_cycles = 0,
- Map<Target, Integer> target_burst_bytes = {},
- Bool is_internal = Bool(false));
- TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS(PoolInfoProperties, ObjectRef, PoolInfoPropertiesNode);
-};
-
-/* \brief Represents RW memory area */
-struct WorkspacePoolInfoNode : public PoolInfoNode {
- void VisitAttrs(tvm::AttrVisitor* v) { PoolInfoNode::VisitAttrs(v); }
-
- bool SEqualReduce(const WorkspacePoolInfoNode* other, SEqualReducer equal) const {
- return PoolInfoNode::SEqualReduce(other, equal);
- }
-
- void SHashReduce(SHashReducer hash_reduce) const { PoolInfoNode::SHashReduce(hash_reduce); }
-
- static constexpr const char* _type_key = "ir.WorkspacePoolInfo";
- TVM_DECLARE_FINAL_OBJECT_INFO(WorkspacePoolInfoNode, PoolInfoNode);
-};
-
-class WorkspacePoolInfo : public PoolInfo {
- public:
- TVM_DLL WorkspacePoolInfo(
- String pool_name, Array<Target> targets,
- PoolInfoProperties properties = PoolInfoProperties(kUnrestrictedPoolSizeHint));
- TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS(WorkspacePoolInfo, PoolInfo, WorkspacePoolInfoNode);
-};
-
-/*
- * \brief The ConstantInfoNode contains numeric literal in RO pool
- * Used to initialise RO memory in ConstantPoolInfo
- */
-struct ConstantInfoNode : public Object {
- String name_hint;
- Integer byte_offset;
- runtime::NDArray data;
-
- void VisitAttrs(tvm::AttrVisitor* v) {
- v->Visit("name_hint", &name_hint);
- v->Visit("byte_offset", &byte_offset);
- v->Visit("data", &data);
- }
-
- bool SEqualReduce(const ConstantInfoNode* other, SEqualReducer equal) const {
- return equal(name_hint, other->name_hint) && equal(byte_offset, other->byte_offset) &&
- equal(data, other->data);
- }
-
- void SHashReduce(SHashReducer hash_reduce) const {
- hash_reduce(name_hint);
- hash_reduce(byte_offset);
- hash_reduce(data);
- }
-
- static constexpr const char* _type_key = "ir.ConstantInfo";
- static constexpr bool _type_has_method_sequal_reduce = true;
- static constexpr bool _type_has_method_shash_reduce = true;
- TVM_DECLARE_FINAL_OBJECT_INFO(ConstantInfoNode, Object);
-};
-
-class ConstantInfo : public ObjectRef {
- public:
- TVM_DLL ConstantInfo(const struct ::TVMConstantInfo* data);
- ConstantInfo(String name, Integer byte_offset, runtime::NDArray data);
- TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS(ConstantInfo, ObjectRef, ConstantInfoNode);
-};
-
-/* \brief ConstantPoolInfoNode represents an RO memory area initialized with
- * data from constant_info_array */
-struct ConstantPoolInfoNode : public PoolInfoNode {
- Array<ConstantInfo> constant_info_array;
-
- void VisitAttrs(tvm::AttrVisitor* v) {
- PoolInfoNode::VisitAttrs(v);
- v->Visit("constant_info_array", &constant_info_array);
- }
-
- bool SEqualReduce(const ConstantPoolInfoNode* other, SEqualReducer equal) const {
- return PoolInfoNode::SEqualReduce(other, equal) &&
- equal(constant_info_array, other->constant_info_array);
- }
-
- void SHashReduce(SHashReducer hash_reduce) const {
- PoolInfoNode::SHashReduce(hash_reduce);
- hash_reduce(constant_info_array);
- }
-
- static constexpr const char* _type_key = "ir.ConstantPoolInfo";
- TVM_DECLARE_FINAL_OBJECT_INFO(ConstantPoolInfoNode, PoolInfoNode);
-};
-
-class ConstantPoolInfo : public PoolInfo {
- public:
- TVM_DLL ConstantPoolInfo(
- String pool_name, Array<Target> targets, Array<ConstantInfo> constant_info_array,
- PoolInfoProperties properties = PoolInfoProperties(kUnrestrictedPoolSizeHint));
- TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS(ConstantPoolInfo, PoolInfo, ConstantPoolInfoNode);
-};
-
-/* \brief A container for WorkspacePoolInfo objects */
-struct WorkspaceMemoryPoolsNode : public Object {
- Array<PoolInfo> pools;
-
- void VisitAttrs(tvm::AttrVisitor* v) { v->Visit("pools", &pools); }
-
- bool SEqualReduce(const WorkspaceMemoryPoolsNode* other, SEqualReducer equal) const {
- return equal(pools, other->pools);
- }
-
- void SHashReduce(SHashReducer hash_reduce) const { hash_reduce(pools); }
-
- static constexpr const char* _type_key = "ir.WorkspaceMemoryPools";
- TVM_DECLARE_FINAL_OBJECT_INFO(WorkspaceMemoryPoolsNode, Object);
-};
-
-class WorkspaceMemoryPools : public ObjectRef {
- public:
- TVM_DLL WorkspaceMemoryPools(Array<PoolInfo> pools);
- TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS(WorkspaceMemoryPools, ObjectRef, WorkspaceMemoryPoolsNode);
-};
-
-/* \brief A container for ConstantPoolInfo objects */
-struct ConstantMemoryPoolsNode : public Object {
- Array<ConstantPoolInfo> pools;
-
- void VisitAttrs(tvm::AttrVisitor* v) { v->Visit("pools", &pools); }
-
- bool SEqualReduce(const ConstantMemoryPoolsNode* other, SEqualReducer equal) const {
- return equal(pools, other->pools);
- }
-
- void SHashReduce(SHashReducer hash_reduce) const { hash_reduce(pools); }
-
- static constexpr const char* _type_key = "ir.ConstantMemoryPools";
- TVM_DECLARE_FINAL_OBJECT_INFO(ConstantMemoryPoolsNode, Object);
-};
-
-class ConstantMemoryPools : public ObjectRef {
- public:
- TVM_DLL ConstantMemoryPools(Array<ConstantPoolInfo> pools);
- TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS(ConstantMemoryPools, ObjectRef, ConstantMemoryPoolsNode);
-};
-
-} // namespace tvm
-
-#endif // TVM_IR_MEMORY_POOLS_H_
diff --git a/include/tvm/ir/module.h b/include/tvm/ir/module.h
index b3895ee725..0338096d70 100644
--- a/include/tvm/ir/module.h
+++ b/include/tvm/ir/module.h
@@ -277,7 +277,7 @@ class IRModule : public ObjectRef {
* \brief As for \p FromExprInContext, but assuming \p expr is bound to 'main' and no
* imports.
*/
- TVM_DLL static IRModule FromExpr(const RelayExpr& expr,
+ TVM_DLL static IRModule FromExpr(const RelaxExpr& expr,
const Map<GlobalVar, BaseFunc>& global_funcs = {});
/*!
@@ -310,42 +310,6 @@ namespace attr {
*/
constexpr const char* kModuleName = "mod_name";
-/*!
- * \brief Executor targeted by the module
- *
- * Type: Executor
- *
- * \sa tvm::relay::Executor
- */
-constexpr const char* kExecutor = "executor";
-
-/*!
- * \brief Runtime target of the module
- *
- * Type: Runtime
- *
- * \sa tvm::relay::Runtime
- */
-constexpr const char* kRuntime = "runtime";
-
-/*!
- * \brief workspace memory pools of the module
- *
- * Type: WorkspaceMemoryPools
- *
- * \sa tvm::WorkspaceMemoryPools
- */
-constexpr const char* kWorkspaceMemoryPools = "workspace_memory_pools";
-
-/*!
- * \brief constant memory pools of the module
- *
- * Type: ConstantMemoryPools
- *
- * \sa tvm::ConstantMemoryPools
- */
-constexpr const char* kConstantMemoryPools = "constant_memory_pools";
-
/*
* \brief All the runtime::NDArrays extracted from PrimFunc tir::AllocateConst nodes. The
* node will record the index into this array. See also kConstNameToConstant below, which is
diff --git a/include/tvm/ir/op.h b/include/tvm/ir/op.h
index a703f16f5a..7fbd1cbb84 100644
--- a/include/tvm/ir/op.h
+++ b/include/tvm/ir/op.h
@@ -55,7 +55,7 @@ class OpAttrMap;
*
* \sa Op
*/
-class OpNode : public RelayExprNode {
+class OpNode : public RelaxExprNode {
public:
/*! \brief name of the operator */
String name;
@@ -111,7 +111,7 @@ class OpNode : public RelayExprNode {
}
static constexpr const char* _type_key = "Op";
- TVM_DECLARE_FINAL_OBJECT_INFO(OpNode, RelayExprNode);
+ TVM_DECLARE_FINAL_OBJECT_INFO(OpNode, RelaxExprNode);
private:
/*! \return the internal attr registry index. */
@@ -135,7 +135,7 @@ class OpNode : public RelayExprNode {
* \brief Managed reference class to OpNode.
* \sa OpNode
*/
-class Op : public RelayExpr {
+class Op : public RelaxExpr {
public:
/*!
* \brief Get additional registered attribute about operators.
@@ -160,7 +160,7 @@ class Op : public RelayExpr {
*/
TVM_DLL static const Op& Get(const String& op_name);
- TVM_DEFINE_OBJECT_REF_METHODS(Op, RelayExpr, OpNode)
+ TVM_DEFINE_OBJECT_REF_METHODS(Op, RelaxExpr, OpNode)
private:
/*!
@@ -286,7 +286,7 @@ class OpAttrMap : public AttrRegistryMap<Op, ValueType> {
* or if expr is not an Op.
* \return the const reference to the content value.
*/
- inline ValueType get(const RelayExpr& expr, ValueType def_value) const;
+ inline ValueType get(const RelaxExpr& expr, ValueType def_value) const;
using TParent = AttrRegistryMap<Op, ValueType>;
using TParent::count;
@@ -381,7 +381,7 @@ inline OpRegEntry& OpRegEntry::set_attr( // NOLINT(*)
// member functions of OpAttrMap
template <typename ValueType>
-inline ValueType OpAttrMap<ValueType>::get(const RelayExpr& expr, ValueType def_value) const {
+inline ValueType OpAttrMap<ValueType>::get(const RelaxExpr& expr, ValueType def_value) const {
ICHECK(expr.defined());
if (const OpNode* op = expr.as<OpNode>()) {
return this->map_.get(GetRef<Op>(op), def_value);
diff --git a/include/tvm/ir/tensor_type.h b/include/tvm/ir/tensor_type.h
deleted file mode 100644
index 7a700258f2..0000000000
--- a/include/tvm/ir/tensor_type.h
+++ /dev/null
@@ -1,127 +0,0 @@
-/*
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements. See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership. The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * "License"); you may not use this file except in compliance
- * with the License. You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations
- * under the License.
- */
-
-/*!
- * \file tvm/ir/tensor_type.h
- * \brief Polymorphic tensor types.
- */
-#ifndef TVM_IR_TENSOR_TYPE_H_
-#define TVM_IR_TENSOR_TYPE_H_
-
-#include <tvm/ir/expr.h>
-#include <tvm/ir/type.h>
-
-namespace tvm {
-/*!
- * \brief Base of all Tensor types
- * This container can hold TensorType or GenericTensorType.
- * \sa BaseTensorType, TensorTypeNode
- */
-class BaseTensorTypeNode : public TypeNode {
- public:
- static constexpr const char* _type_key = "relay.BaseTensorType";
- static constexpr const uint32_t _type_child_slots = 1;
- TVM_DECLARE_BASE_OBJECT_INFO(BaseTensorTypeNode, TypeNode);
-};
-
-/*!
- * \brief Managed reference to BaseTensorTypeNode.
- * \sa BaseTensorTypeNode.
- */
-class BaseTensorType : public Type {
- public:
- TVM_DEFINE_OBJECT_REF_METHODS(BaseTensorType, Type, BaseTensorTypeNode);
-};
-
-/*!
- * \brief This is the most commonly used type in relay.
- * TensorType have a fixed dimension, data type.
- *
- * The elements of shape can be either IntImm(constant integer),
- * or any symbolic integer expression.
- * The symbolic integer allows generic shape inference in certain cases.
- * \sa TensorType
- */
-class TensorTypeNode : public BaseTensorTypeNode {
- public:
- /*!
- * \brief The shape of the tensor,
- * represented by PrimExpr(tvm::Expr).
- */
- Array<PrimExpr> shape;
- /*! \brief The content data type */
- DataType dtype;
-
- void VisitAttrs(tvm::AttrVisitor* v) {
- v->Visit("shape", &shape);
- v->Visit("dtype", &dtype);
- v->Visit("span", &span);
- }
-
- bool SEqualReduce(const TensorTypeNode* other, SEqualReducer equal) const {
- return equal(shape, other->shape) && equal(dtype, other->dtype);
- }
-
- void SHashReduce(SHashReducer hash_reduce) const {
- hash_reduce(shape);
- hash_reduce(dtype);
- }
-
- /*! \brief Return product of elements in the shape.
- * \return (d1 * d_2 ... * d_n) if shape is (d_1, d_2, ..., d_n) and 1 if shape size is zero.
- */
- TVM_DLL PrimExpr Size() const;
-
- static constexpr const char* _type_key = "relay.TensorType";
- TVM_DECLARE_FINAL_OBJECT_INFO(TensorTypeNode, BaseTensorTypeNode);
-};
-
-/*!
- * \brief Managed reference to TensorTypeNode.
- * \sa TensorTypeNode.
- */
-class TensorType : public Type {
- public:
- /*!
- * \brief Constructor.
- * \param shape The shape of the tensor.
- * \param dtype The runtime dtype of the tensor's elements.
- */
- TVM_DLL TensorType(Array<PrimExpr> shape, DataType dtype);
-
- /*!
- * \brief Construct an scalar containing elements of dtype.
- * \param dtype The runtime dtype of the tensor's elements.
- * \return THe constructed type.
- */
- TVM_DLL static TensorType Scalar(DataType dtype);
-
- TVM_DEFINE_OBJECT_REF_METHODS(TensorType, Type, TensorTypeNode);
-};
-
-// The following fields contains advanced typing
-// Only keep the class name and reserved for future usage.
-class GenericTensorType;
-// stores a DataType.
-class GenericDataType;
-// stores a DataType.
-class GenericShape;
-
-} // namespace tvm
-#endif // TVM_IR_TENSOR_TYPE_H_
diff --git a/include/tvm/ir/type_functor.h b/include/tvm/ir/type_functor.h
index eb213b17df..2c145e480b 100644
--- a/include/tvm/ir/type_functor.h
+++ b/include/tvm/ir/type_functor.h
@@ -24,7 +24,7 @@
#ifndef TVM_IR_TYPE_FUNCTOR_H_
#define TVM_IR_TYPE_FUNCTOR_H_
-#include <tvm/ir/tensor_type.h>
+#include <tvm/ir/type.h>
#include <tvm/node/functor.h>
#include <string>
@@ -75,7 +75,6 @@ class TypeFunctor<R(const Type& n, Args...)> {
return vtable(n, this, std::forward<Args>(args)...);
}
// Functions that can be overriden by subclass
- virtual R VisitType_(const TensorTypeNode* op, Args... args) TYPE_FUNCTOR_DEFAULT;
virtual R VisitType_(const FuncTypeNode* op, Args... args) TYPE_FUNCTOR_DEFAULT;
virtual R VisitType_(const TupleTypeNode* op, Args... args) TYPE_FUNCTOR_DEFAULT;
virtual R VisitType_(const PrimTypeNode* op, Args... args) TYPE_FUNCTOR_DEFAULT;
@@ -90,7 +89,6 @@ class TypeFunctor<R(const Type& n, Args...)> {
static FType InitVTable() {
FType vtable;
// Set dispatch
- TVM_TYPE_FUNCTOR_DISPATCH(TensorTypeNode);
TVM_TYPE_FUNCTOR_DISPATCH(FuncTypeNode);
TVM_TYPE_FUNCTOR_DISPATCH(TupleTypeNode);
TVM_TYPE_FUNCTOR_DISPATCH(PrimTypeNode);
@@ -106,7 +104,6 @@ class TypeFunctor<R(const Type& n, Args...)> {
*/
class TVM_DLL TypeVisitor : public TypeFunctor<void(const Type& n)> {
public:
- void VisitType_(const TensorTypeNode* op) override;
void VisitType_(const FuncTypeNode* op) override;
void VisitType_(const TupleTypeNode* op) override;
void VisitType_(const PrimTypeNode* op) override;
@@ -119,7 +116,6 @@ class TVM_DLL TypeVisitor : public TypeFunctor<void(const Type& n)> {
class TVM_DLL TypeMutator : public TypeFunctor<Type(const Type& n)> {
public:
Type VisitType(const Type& t) override;
- Type VisitType_(const TensorTypeNode* op) override;
Type VisitType_(const FuncTypeNode* op) override;
Type VisitType_(const TupleTypeNode* op) override;
Type VisitType_(const PrimTypeNode* op) override;
diff --git a/include/tvm/node/structural_equal.h b/include/tvm/node/structural_equal.h
index f88b7e63b3..2dd732c72a 100644
--- a/include/tvm/node/structural_equal.h
+++ b/include/tvm/node/structural_equal.h
@@ -100,11 +100,9 @@ class ObjectPathPair : public ObjectRef {
* - Normal node: equality is recursively defined without the restriction
* of graph nodes.
*
- * Vars(tir::Var, TypeVar) and non-constant relay expression nodes are graph nodes.
- * For example, it means that `%1 = %x + %y; %1 + %1` is not structurally equal
- * to `%1 = %x + %y; %2 = %x + %y; %1 + %2` in relay.
+ * Vars(tir::Var, relax::Var) nodes are graph nodes.
*
- * A var-type node(e.g. tir::Var, TypeVar) can be mapped as equal to another var
+ * A var-type node(e.g. tir::Var) can be mapped as equal to another var
* with the same type if one of the following condition holds:
*
* - They appear in a same definition point(e.g. function argument).
diff --git a/include/tvm/relax/dataflow_pattern.h b/include/tvm/relax/dataflow_pattern.h
index f7094b2212..df9fdcad97 100644
--- a/include/tvm/relax/dataflow_pattern.h
+++ b/include/tvm/relax/dataflow_pattern.h
@@ -461,7 +461,7 @@ class CallPatternNode : public DFPatternNode {
public:
/*!
* \note The op field can be:
- * - relay::Op which corresponds to the primitive operators.
+ * - relax::Op which corresponds to the primitive operators.
* - user defined functions (Function, GlobalVar, Var).
*/
DFPattern op; /*!< The operator (function) being invoked */
diff --git a/include/tvm/relax/expr.h b/include/tvm/relax/expr.h
index 9afddeb807..fb6f0e40b1 100644
--- a/include/tvm/relax/expr.h
+++ b/include/tvm/relax/expr.h
@@ -35,8 +35,8 @@
namespace tvm {
namespace relax {
-using Expr = RelayExpr;
-using ExprNode = RelayExprNode;
+using Expr = RelaxExpr;
+using ExprNode = RelaxExprNode;
/*!
* \brief The unique identifier of variables.
*
diff --git a/include/tvm/relax/type.h b/include/tvm/relax/type.h
index 210730bec6..ed832f6a02 100644
--- a/include/tvm/relax/type.h
+++ b/include/tvm/relax/type.h
@@ -26,7 +26,6 @@
#include <tvm/ir/attrs.h>
#include <tvm/ir/env_func.h>
-#include <tvm/ir/tensor_type.h>
#include <tvm/ir/type.h>
#include <tvm/runtime/registry.h>
#include <tvm/tir/expr.h>
@@ -67,26 +66,12 @@ class ShapeType : public Type {
TVM_DEFINE_NOTNULLABLE_OBJECT_REF_METHODS(ShapeType, Type, ShapeTypeNode);
};
-class ObjectTypeNode : public TypeNode {
- public:
- void VisitAttrs(tvm::AttrVisitor* v) { v->Visit("span", &span); }
-
- bool SEqualReduce(const ObjectTypeNode* other, SEqualReducer equal) const { return true; }
-
- void SHashReduce(SHashReducer hash_reduce) const { hash_reduce(0); }
-
- static constexpr const char* _type_key = "relax.ObjectType";
- TVM_DECLARE_FINAL_OBJECT_INFO(ObjectTypeNode, TypeNode);
-};
-
-class ObjectType : public Type {
- public:
- TVM_DLL ObjectType(Span span = Span());
-
- TVM_DEFINE_NOTNULLABLE_OBJECT_REF_METHODS(ObjectType, Type, ObjectTypeNode);
-};
-
-class DynTensorTypeNode : public BaseTensorTypeNode {
+/*!
+ * \brief Dynamic version of TensorType
+ *
+ * Use relax::TensorStructInfo for more detailed (possibly dynamic) shape constrains
+ */
+class TensorTypeNode : public TypeNode {
public:
/*!
* \brief The number of dimensions of the tensor, use -1 to denote tensor with unknwon number of
@@ -102,7 +87,7 @@ class DynTensorTypeNode : public BaseTensorTypeNode {
v->Visit("span", &span);
}
- bool SEqualReduce(const DynTensorTypeNode* other, SEqualReducer equal) const {
+ bool SEqualReduce(const TensorTypeNode* other, SEqualReducer equal) const {
return equal(ndim, other->ndim) && equal(dtype, other->dtype);
}
@@ -116,14 +101,14 @@ class DynTensorTypeNode : public BaseTensorTypeNode {
inline bool IsUnknownDtype() const { return dtype.is_void(); }
static constexpr const char* _type_key = "relax.DynTensorType";
- TVM_DECLARE_FINAL_OBJECT_INFO(DynTensorTypeNode, BaseTensorTypeNode);
+ TVM_DECLARE_FINAL_OBJECT_INFO(TensorTypeNode, TypeNode);
};
/*!
- * \brief Managed reference to DynTensorTypeNode.
- * \sa DynTensorTypeNode.
+ * \brief Managed reference to TensorTypeNode.
+ * \sa TensorTypeNode.
*/
-class DynTensorType : public Type {
+class TensorType : public Type {
public:
/*!
* \brief Constructor.
@@ -131,14 +116,36 @@ class DynTensorType : public Type {
* \param dtype The runtime dtype of the tensor's elements.
* \param span The span.
*/
- TVM_DLL DynTensorType(int ndim, DataType dtype, Span span = Span());
+ TVM_DLL TensorType(int ndim, DataType dtype, Span span = Span());
/*!
- * \brief Create a DynTensorType with unknown ndim.
+ * \brief Create a TensorType with unknown ndim.
*/
- TVM_DLL static DynTensorType CreateUnknownNDim(DataType dtype, Span span = Span());
+ TVM_DLL static TensorType CreateUnknownNDim(DataType dtype, Span span = Span());
+
+ TVM_DEFINE_OBJECT_REF_METHODS(TensorType, Type, TensorTypeNode);
+};
- TVM_DEFINE_OBJECT_REF_METHODS(DynTensorType, Type, DynTensorTypeNode);
+using TensorTypeNode = TensorTypeNode;
+using TensorType = TensorType;
+
+class ObjectTypeNode : public TypeNode {
+ public:
+ void VisitAttrs(tvm::AttrVisitor* v) { v->Visit("span", &span); }
+
+ bool SEqualReduce(const ObjectTypeNode* other, SEqualReducer equal) const { return true; }
+
+ void SHashReduce(SHashReducer hash_reduce) const { hash_reduce(0); }
+
+ static constexpr const char* _type_key = "relax.ObjectType";
+ TVM_DECLARE_FINAL_OBJECT_INFO(ObjectTypeNode, TypeNode);
+};
+
+class ObjectType : public Type {
+ public:
+ TVM_DLL ObjectType(Span span = Span());
+
+ TVM_DEFINE_NOTNULLABLE_OBJECT_REF_METHODS(ObjectType, Type, ObjectTypeNode);
};
class PackedFuncTypeNode : public TypeNode {
diff --git a/include/tvm/runtime/logging.h b/include/tvm/runtime/logging.h
index 708f4bcaa9..440c7a4383 100644
--- a/include/tvm/runtime/logging.h
+++ b/include/tvm/runtime/logging.h
@@ -488,9 +488,9 @@ inline bool DebugLoggingEnabled() {
* Filenames are canonicalized to be w.r.t. the src/ dir of the TVM tree. (VLOG's should not
* appear under include/).
*
- * To enable file \p relay/foo.cc up to level 2 and \p ir/bar.cc for level 0 only set:
+ * To enable file \p ir/bar.cc for level 0 only set:
* \code
- * TVM_LOG_DEBUG="relay/foo.cc=2,ir/bar.cc=0"
+ * TVM_LOG_DEBUG="ir/bar.cc=0"
* \endcode
*
* To enable all files up to level 3 but disable \p ir/bar.cc set:
diff --git a/include/tvm/target/generic_func.h b/include/tvm/target/generic_func.h
deleted file mode 100644
index bd498616f9..0000000000
--- a/include/tvm/target/generic_func.h
+++ /dev/null
@@ -1,171 +0,0 @@
-/*
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements. See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership. The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * "License"); you may not use this file except in compliance
- * with the License. You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations
- * under the License.
- */
-
-/*!
- * \file tvm/target/generic_func.h
- * \brief Generic function that can be specialzied on a per target basis.
- */
-#ifndef TVM_TARGET_GENERIC_FUNC_H_
-#define TVM_TARGET_GENERIC_FUNC_H_
-
-#include <tvm/runtime/packed_func.h>
-#include <tvm/support/with.h>
-#include <tvm/target/target.h>
-
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-namespace tvm {
-
-class GenericFuncNode;
-
-/*!
- * \brief Generic function that can be specialized on a per-target basis.
- */
-class GenericFunc : public ObjectRef {
- public:
- GenericFunc() {}
- explicit GenericFunc(ObjectPtr<Object> n) : ObjectRef(n) {}
-
- /*!
- * \brief Set the default function implementaiton.
- * \param value The default function
- * \param allow_override If true, this call may override a previously registered function. If
- * false, an error will be logged if the call would override a previously registered function.
- * \return reference to self.
- */
- TVM_DLL GenericFunc& set_default(const runtime::PackedFunc value, bool allow_override = false);
- /*!
- * \brief Register a specialized function
- * \param tags The tags for this specialization
- * \param value The specialized function
- * \param allow_override If true, this call may override previously registered tags. If false,
- * an error will be logged if the call would override previously registered tags.
- * \return reference to self.
- */
- TVM_DLL GenericFunc& register_func(const std::vector<std::string>& tags,
- const runtime::PackedFunc value, bool allow_override = false);
- /*!
- * \brief Call generic function by directly passing in unpacked format.
- * \param args Arguments to be passed.
- * \tparam Args arguments to be passed.
- *
- * \code
- * // Example code on how to call generic function
- * void CallGeneric(GenericFunc f) {
- * // call like normal functions by pass in arguments
- * // return value is automatically converted back
- * int rvalue = f(1, 2.0);
- * }
- * \endcode
- */
- template <typename... Args>
- inline runtime::TVMRetValue operator()(Args&&... args) const;
- /*!
- * \brief Invoke the relevant function for the current target context, set by set_target_context.
- * Arguments are passed in packed format.
- * \param args The arguments to pass to the function.
- * \param ret The return value
- */
- TVM_DLL void CallPacked(runtime::TVMArgs args, runtime::TVMRetValue* ret) const;
- /*!
- * \brief Get the packed function specified for the current target context.
- */
- TVM_DLL PackedFunc GetPacked() const;
- /*!
- * \brief Find or register the GenericFunc instance corresponding to the give name
- * \param name The name of the registered GenericFunc
- * \return The GenericFunc instance
- */
- TVM_DLL static GenericFunc Get(const std::string& name);
-
- /*!
- * \brief Add a GenericFunc instance to the registry
- * \param func The GenericFunc instance
- * \param name The name of the registered GenericFunc
- */
- TVM_DLL static void RegisterGenericFunc(GenericFunc func, const std::string& name);
-
- /*!
- * \brief access the internal node container
- * \return the pointer to the internal node container
- */
- inline GenericFuncNode* operator->();
-
- // declare container type
- using ContainerType = GenericFuncNode;
-
- // Internal class.
- struct Manager;
-
- private:
- friend struct Manager;
-};
-
-template <typename... Args>
-inline runtime::TVMRetValue GenericFunc::operator()(Args&&... args) const {
- const int kNumArgs = sizeof...(Args);
- const int kArraySize = kNumArgs > 0 ? kNumArgs : 1;
- TVMValue values[kArraySize];
- int type_codes[kArraySize];
- runtime::detail::for_each(runtime::TVMArgsSetter(values, type_codes),
- std::forward<Args>(args)...);
- runtime::TVMRetValue rv;
- CallPacked(runtime::TVMArgs(values, type_codes, kNumArgs), &rv);
- return rv;
-}
-
-/*!
- * \brief Represents a generic function that can be specialized on a per-target basis.
- */
-class GenericFuncNode : public Object {
- public:
- /*! \brief name of the function */
- std::string name_;
- /* \brief the generic builder */
- runtime::PackedFunc generic_func_;
- /* \brief map from keys to registered functions */
- std::unordered_map<std::string, runtime::PackedFunc> dispatch_dict_;
-
- void VisitAttrs(AttrVisitor* v) {}
-
- static constexpr const char* _type_key = "GenericFunc";
- TVM_DECLARE_FINAL_OBJECT_INFO(GenericFuncNode, Object);
-};
-
-inline GenericFuncNode* GenericFunc::operator->() {
- return static_cast<GenericFuncNode*>(get_mutable());
-}
-
-#define TVM_GENERIC_FUNC_REG_VAR_DEF static TVM_ATTRIBUTE_UNUSED ::tvm::GenericFunc& __mk_##TVM
-
-/*!
- * \def TVM_REGISTER_GENERIC_FUNC
- * \brief Register a new generic function, or set a device-specific variant
- * of the corresponding function.
- *
- * \param name The name of the function
- */
-#define TVM_REGISTER_GENERIC_FUNC(name) \
- TVM_STR_CONCAT(TVM_GENERIC_FUNC_REG_VAR_DEF, __COUNTER__) = ::tvm::GenericFunc::Get(#name)
-
-} // namespace tvm
-#endif // TVM_TARGET_GENERIC_FUNC_H_
diff --git a/include/tvm/target/virtual_device.h b/include/tvm/target/virtual_device.h
index 8ebcbf69bb..5c6db4bc24 100644
--- a/include/tvm/target/virtual_device.h
+++ b/include/tvm/target/virtual_device.h
@@ -76,12 +76,8 @@ constexpr int kInvalidDeviceType = -1;
* device_type must equal \p target->GetTargetDeviceType().
*
* Note that currently we assume if a function returns its result on a particular (virtual) device
- * then the function body is also executed on that device. See the overview comment in
- * src/relay/transforms/device_planner.cc for more details.
+ * then the function body is also executed on that device.
*
- * By 'data' we include both tensors and additional supporting datastructures such as shapes,
- * Relay ADT items (including tuples), Relay references, and Relay closures. Typically non-tensor
- * data must reside on a 'CPU'-like host device with good support for scalars.
*
* By 'execution' we include both (fused) primitive operators, and all the Relay expressions
* surrounding them which coordinates data and control flow. Again, typically non-primitive
diff --git a/include/tvm/tir/expr.h b/include/tvm/tir/expr.h
index 28cb022151..a157516f53 100644
--- a/include/tvm/tir/expr.h
+++ b/include/tvm/tir/expr.h
@@ -886,7 +886,7 @@ class CallNode : public PrimExprNode {
* - It can be tvm::Op which corresponds to the primitive operators(intrinsics).
* - It can also be another function in the IRModule (GlobalVar).
*/
- RelayExpr op;
+ RelaxExpr op;
/*! \brief The arguments. */
Array<PrimExpr> args;
@@ -917,7 +917,7 @@ class CallNode : public PrimExprNode {
*/
class Call : public PrimExpr {
public:
- TVM_DLL Call(DataType dtype, RelayExpr op, Array<PrimExpr> args, Span span = Span());
+ TVM_DLL Call(DataType dtype, RelaxExpr op, Array<PrimExpr> args, Span span = Span());
TVM_DEFINE_OBJECT_REF_METHODS(Call, PrimExpr, CallNode);
TVM_DEFINE_OBJECT_REF_COW_METHOD(CallNode);
};
diff --git a/python/setup.py b/python/setup.py
index 6cca9060bd..30b1e2174d 100644
--- a/python/setup.py
+++ b/python/setup.py
@@ -208,7 +208,7 @@ if not CONDA_BUILD and not INPLACE_BUILD:
def get_package_data_files():
# Relay standard libraries
- return ["relay/std/prelude.rly", "relay/std/core.rly"]
+ return []
def long_description_contents():
diff --git a/python/tvm/contrib/cutlass/build.py b/python/tvm/contrib/cutlass/build.py
index ba598e9b22..9dba930726 100644
--- a/python/tvm/contrib/cutlass/build.py
+++ b/python/tvm/contrib/cutlass/build.py
@@ -39,7 +39,7 @@ logger = logging.getLogger("cutlass")
def has_cutlass():
"""Returns true if the CUTLASS custom codegen is available"""
- return tvm.get_global_func("relay.ext.cutlass.create_c_source_module", True) is not None
+ return tvm.get_global_func("relax.ext.cutlass", True) is not None
def _get_cutlass_path():
diff --git a/python/tvm/contrib/hexagon/session.py b/python/tvm/contrib/hexagon/session.py
index 50064e42ba..2456aa244e 100644
--- a/python/tvm/contrib/hexagon/session.py
+++ b/python/tvm/contrib/hexagon/session.py
@@ -28,11 +28,6 @@ from tvm import relax
from tvm import rpc as _rpc
from tvm.contrib import utils
import tvm.contrib.hexagon as hexagon
-from tvm.relay.backend.executor_factory import (
- ExecutorFactoryModule,
- AOTExecutorFactoryModule,
- GraphExecutorFactoryModule,
-)
from .tools import export_module, HEXAGON_SIMULATOR_NAME
@@ -206,86 +201,6 @@ class Session:
)
return self._rpc.get_function("tvm.hexagon.load_module")(str(remote_file_path))
- def get_graph_executor(
- self,
- graph_json: str,
- module_name: Union[str, pathlib.Path, tvm.runtime.Module],
- ):
- """Create a local GraphModule which consumes a remote libmod.
-
- The session must be established (via __enter__) prior to
- calling this function.
-
- Parameters
- ----------
- module_name : Union[str, pathlib.Path, tvm.runtime.Module]
- The remote module filename, following the same restrictions
- as `load_module`.
- graph_json : str
- The string with the graph JSON.
-
- Returns
- -------
- GraphModule :
- Runtime graph module that can be used to execute the graph.
-
- """
- graph_mod = self.load_module(module_name)
- self._set_device_type(graph_mod)
- return tvm.contrib.graph_executor.create(graph_json, graph_mod, self.device)
-
- def get_aot_executor(
- self,
- module_file: Union[str, pathlib.Path],
- ):
- """Create a local GraphModule which consumes a remote libmod.
- The session must be established (via __enter__) prior to
- calling this function.
- Parameters
- ----------
- module_file : Union[str, pathlib.Path]
- The remote module filename, following the same restrictions
- as `load_module`. The filename should be an absolute path.
- Returns
- -------
- GraphModule :
- Runtime graph module that can be used to execute the graph.
- """
- # Temporary workaround for https://github.com/apache/tvm/issues/13741
- self.aot_mod = self.load_module(module_file)
- return tvm.runtime.executor.AotModule(self.aot_mod["default"](self.device))
-
- def get_graph_debug_executor(
- self,
- graph_json: str,
- module_name: Union[str, pathlib.Path, tvm.runtime.Module],
- dump_root: Union[str, pathlib.Path] = None,
- ):
- """Create a local GraphModuleDebug which consumes a remote libmod.
-
- Parameters
- ----------
- graph_json : str
- The string with the graph JSON.
- module_name : Union[str, pathlib.Path, tvm.runtime.Module]
- The remote module filename, following the same restrictions
- as `load_module`.
- session : Session
- Remote session. The session must be established (via __enter__)
- prior to calling this function.
-
- Returns
- -------
- GraphModuleDebug :
- Runtime debug graph module that can be used to debug the graph.
- """
-
- graph_debug_mod = self.load_module(module_name)
- self._set_device_type(graph_debug_mod)
- return tvm.contrib.debugger.debug_executor.create(
- graph_json, graph_debug_mod, self.device, dump_root=str(dump_root)
- )
-
def get_executor_from_factory(
self, module: Union[ExecutorFactoryModule, relax.Executable, str], hexagon_arch: str = "v68"
):
@@ -294,17 +209,13 @@ class Session:
Parameters
----------
- module : Union[ExecutorFactoryModule, relax.Executable, str]
+ module : Union[relax.Executable]
The module to upload to the remote
session and load.
hexagon_arch : str
The hexagon arch to be used
"""
- if isinstance(module, AOTExecutorFactoryModule):
- return self._aot_executor_from_factory(module)
- if isinstance(module, GraphExecutorFactoryModule):
- return self._graph_executor_from_factory(module)
if isinstance(module, (relax.Executable, str)):
return self._relax_vm_executable_executor(module, hexagon_arch=hexagon_arch)
@@ -332,32 +243,6 @@ class Session:
else:
self._requires_cpu_device = False
- def _graph_executor_from_factory(
- self,
- module: Union[str, pathlib.Path, GraphExecutorFactoryModule],
- ):
- """Create a local GraphModule which consumes a remote libmod.
-
- The session must be established (via __enter__) prior to
- calling this function.
-
- Parameters
- ----------
-
- module : GraphExecutorFactoryModule
-
- The graph executor module to upload to the remote and load.
- This will typically be the output of `tvm.relay.build`,
- when passing `executor=Executor("graph")`.
-
- Returns
- -------
- GraphModule :
- Runtime graph module that can be used to execute the graph.
-
- """
- return self.get_graph_executor(module.get_graph_json(), module.get_lib())
-
def _relax_vm_executable_executor(
self, vm_exec: Union[relax.Executable, str], hexagon_arch: str
):
@@ -397,75 +282,6 @@ class Session:
path = self.upload(path_exec, "exec.so")
return self._rpc.get_function("tvm.hexagon.load_module")(str(path))
- def _aot_executor_from_factory(
- self,
- module: Union[str, pathlib.Path, AOTExecutorFactoryModule],
- ):
- """Create a local GraphModule which consumes a remote libmod.
-
- The session must be established (via __enter__) prior to
- calling this function.
-
- Parameters
- ----------
-
- module : AOTExecutorFactoryModule
-
- The graph executor module to upload to the remote and load.
- This will typically be the output of `tvm.relay.build`,
- when passing `executor=Executor("aot")`.
-
- Returns
- -------
- GraphModule :
- Runtime graph module that can be used to execute the graph.
-
- """
-
- hexagon_arch = set(
- target.mcpu.replace("hexagon", "")
- for target in module.target
- if "hexagon" in target.keys
- )
-
- self._set_device_type(module)
-
- for target in module.target:
- target_type = str(target).split()[0]
-
- assert hexagon_arch, "No hexagon target architecture found"
- assert len(hexagon_arch) == 1, f"Inconsistent hexagon architecture found, {hexagon_arch}"
- hexagon_arch = hexagon_arch.pop()
-
- with tempfile.TemporaryDirectory() as temp_dir:
- temp_dir = pathlib.Path(temp_dir)
- binary_name = "test_binary.so"
- binary_path = temp_dir / binary_name
-
- if target_type == "hexagon":
- module.export_library(
- str(binary_path),
- fcompile=hexagon.create_aot_shared,
- fpack_imports=hexagon.pack_imports,
- hexagon_arch=hexagon_arch,
- )
- elif target_type == "llvm":
- module.export_library(
- str(binary_path),
- fcompile=hexagon.create_shared,
- fpack_imports=hexagon.pack_imports,
- cc=hexagon.hexagon_clang_plus(),
- )
- else:
- raise ValueError(
- "Incorrect Target kind.\n"
- "Target kind should be from these options: [hexagon, llvm]."
- )
-
- remote_file_path = self.upload(binary_path, binary_name)
-
- return self.get_aot_executor(remote_file_path)
-
def get_profile_output(self, mode: str, path: str):
assert isinstance(mode, str), f"Invalid mode type, {type(mode)} != str"
assert isinstance(path, str), f"Invalid path type, {type(path)} != str"
diff --git a/python/tvm/contrib/msc/core/transform/transform.py b/python/tvm/contrib/msc/core/transform/transform.py
index c6d7113f44..e78b5cb714 100644
--- a/python/tvm/contrib/msc/core/transform/transform.py
+++ b/python/tvm/contrib/msc/core/transform/transform.py
@@ -53,7 +53,7 @@ def SetExprName(
var_names = var_names or {}
var_names = {k: msc_utils.legalize_expr_name(v) for k, v in var_names.items()}
return relax_api.SetRelaxExprName(entry_name, target, var_names) # type: ignore
- return relay_api.SetRelayExprName(entry_name) # type: ignore
+ return relay_api.SetRelaxExprName(entry_name) # type: ignore
def BindExprName(
@@ -75,7 +75,7 @@ def BindExprName(
ret: tvm.ir.transform.Pass
"""
- return relay_api.BindRelayExprName(name_key, seperator, entry_name) # type: ignore
+ return relay_api.BindRelaxExprName(name_key, seperator, entry_name) # type: ignore
def SetExprLayout(allow_missing: bool = True, entry_name: str = "main") -> tvm.ir.transform.Pass:
diff --git a/python/tvm/contrib/target/__init__.py b/python/tvm/contrib/target/__init__.py
deleted file mode 100644
index 13a83393a9..0000000000
--- a/python/tvm/contrib/target/__init__.py
+++ /dev/null
@@ -1,16 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements. See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership. The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License. You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied. See the License for the
-# specific language governing permissions and limitations
-# under the License.
diff --git a/python/tvm/contrib/target/coreml.py b/python/tvm/contrib/target/coreml.py
deleted file mode 100644
index d8846ce5f1..0000000000
--- a/python/tvm/contrib/target/coreml.py
+++ /dev/null
@@ -1,233 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements. See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership. The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License. You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied. See the License for the
-# specific language governing permissions and limitations
-# under the License.
-# pylint: disable=invalid-name, unused-argument, import-outside-toplevel, missing-function-docstring
-"""Utility to compile CoreML models"""
-
-import os
-import shutil
-
-import tvm._ffi
-from ...relay.expr_functor import ExprVisitor
-from .. import xcode, coreml_runtime
-
-
-def _convert_add(builder, name, inputs, outputs, args, attrs):
- builder.add_elementwise(name=name, input_names=inputs, output_name=outputs[0], mode="ADD")
-
-
-def _convert_multiply(builder, name, inputs, outputs, args, attrs):
- builder.add_elementwise(name=name, input_names=inputs, output_name=outputs[0], mode="MULTIPLY")
-
-
-def _convert_clip(builder, name, inputs, outputs, args, attrs):
- builder.add_clip(
- name=name,
- input_name=inputs[0],
- output_name=outputs[0],
- min_value=attrs.a_min,
- max_value=attrs.a_max,
- )
-
-
-def _convert_batch_flatten(builder, name, inputs, outputs, args, attrs):
- builder.add_flatten_to_2d(name=name, input_name=inputs[0], output_name=outputs[0])
-
-
-def _convert_expand_dims(builder, name, inputs, outputs, args, attrs):
- if attrs.axis >= 0:
- axes = list(range(attrs.axis, attrs.axis + attrs.num_newaxis))
- else:
- axes = list(range(attrs.axis - attrs.num_newaxis + 1, attrs.axis + 1))
-
- builder.add_expand_dims(name=name, input_name=inputs[0], output_name=outputs[0], axes=axes)
-
-
-def _convert_relu(builder, name, inputs, outputs, args, attrs):
- builder.add_activation(
- name=name, non_linearity="RELU", input_name=inputs[0], output_name=outputs[0]
- )
-
-
-def _convert_softmax(builder, name, inputs, outputs, args, attrs):
- builder.add_softmax_nd(
- name=name, input_name=inputs[0], output_name=outputs[0], axis=int(attrs["axis"])
- )
-
-
-def _convert_conv2d(builder, name, inputs, outputs, args, attrs):
- weight = args[1].data.numpy()
- if attrs["kernel_layout"] == "OIHW":
- # convert to 'HWIO'
- weight = weight.transpose([2, 3, 1, 0])
- kh, kw, kc, oc = weight.shape
-
- builder.add_convolution(
- name=name,
- kernel_channels=kc,
- output_channels=oc,
- height=kh,
- width=kw,
- stride_height=int(attrs["strides"][0]),
- stride_width=int(attrs["strides"][0]),
- border_mode="valid",
- groups=int(attrs["groups"]),
- W=weight,
- b=None,
- has_bias=False,
- input_name=inputs[0],
- output_name=outputs[0],
- dilation_factors=[int(v) for v in attrs["dilation"]],
- padding_top=int(attrs["padding"][0]),
- padding_bottom=int(attrs["padding"][2]),
- padding_left=int(attrs["padding"][1]),
- padding_right=int(attrs["padding"][3]),
- )
-
-
-def _convert_global_avg_pool2d(builder, name, inputs, outputs, args, attrs):
- builder.add_pooling(
- name=name,
- height=1,
- width=1,
- stride_height=1,
- stride_width=1,
- layer_type="AVERAGE",
- padding_type="VALID",
- input_name=inputs[0],
- output_name=outputs[0],
- is_global=True,
- )
-
-
-_convert_map = {
- "add": _convert_add,
- "multiply": _convert_multiply,
- "clip": _convert_clip,
- "expand_dims": _convert_expand_dims,
- "nn.relu": _convert_relu,
- "nn.batch_flatten": _convert_batch_flatten,
- "nn.softmax": _convert_softmax,
- "nn.conv2d": _convert_conv2d,
- "nn.global_avg_pool2d": _convert_global_avg_pool2d,
-}
-
-
-class CodegenCoreML(ExprVisitor):
- """
- A visitor to traverse subgraphs and build Core ML models.
- """
-
- def __init__(self, model_name, function):
- import coremltools
- from coremltools.models.neural_network import NeuralNetworkBuilder
-
- ExprVisitor.__init__(self)
- self.model_name = model_name
- self.function = function
- self.out_map = {}
- self.model_inputs_ = []
- self.buf_idx_ = 0
-
- # Update inputs and outputs after we visit all the nodes.
- # Set dummy values for now.
- # TODO: support multiple outputs
- inputs = [("", coremltools.models.datatypes.Array(1)) for _ in self.function.params]
- outputs = [("", coremltools.models.datatypes.Array(1))]
- self.builder = NeuralNetworkBuilder(inputs, outputs, disable_rank5_shape_mapping=True)
-
- def visit_constant(self, const):
- output = "buf_" + str(self.buf_idx_)
- self.builder.add_load_constant_nd(
- name=output,
- output_name=output,
- constant_value=const.data.numpy(),
- shape=const.data.shape,
- )
- self.buf_idx_ = self.buf_idx_ + 1
- self.out_map[const] = [output]
-
- def visit_var(self, var):
- name = var.name_hint
- shape = [int(n) for n in var.type_annotation.shape]
- dtype = var.type_annotation.dtype
- self.model_inputs_.append((name, shape, dtype))
- self.out_map[var] = [name]
-
- def visit_call(self, call):
- inputs = []
- for arg in call.args:
- super().visit(arg)
- for out in self.out_map[arg]:
- inputs.append(out)
- outputs = ["buf_" + str(self.buf_idx_)]
- op_name = call.op.name
- layer_name = op_name + "_" + str(self.buf_idx_)
-
- assert op_name in _convert_map, f"{op_name} is not supported"
- _convert_map[op_name](self.builder, layer_name, inputs, outputs, call.args, call.attrs)
-
- self.buf_idx_ = self.buf_idx_ + 1
- self.out_map[call] = outputs
-
- def compile(self, out_dir):
- """
- Build a Core ML model and compile it with Xcode toolchain.
- """
- import coremltools
- from coremltools.proto.Model_pb2 import ArrayFeatureType
-
- FEATURE_TYPE_MAP = {
- "float32": ArrayFeatureType.FLOAT32,
- "float64": ArrayFeatureType.DOUBLE,
- "int32": ArrayFeatureType.INT32,
- }
-
- input_names, input_dims, input_dtypes = zip(*self.model_inputs_)
- self.builder.set_input(input_names, input_dims)
- for i, dtype in enumerate(input_dtypes):
- assert dtype in FEATURE_TYPE_MAP
- input_desc = self.builder.spec.description.input
- input_desc[i].type.multiArrayType.dataType = FEATURE_TYPE_MAP[dtype]
-
- output_dim = [int(n) for n in self.function.ret_type.shape]
- self.builder.set_output(self.out_map[self.function.body], [output_dim])
- for i, dtype in enumerate([self.function.ret_type.dtype]):
- assert dtype in FEATURE_TYPE_MAP
- output_desc = self.builder.spec.description.output
- output_desc[i].type.multiArrayType.dataType = FEATURE_TYPE_MAP[dtype]
-
- model = coremltools.models.MLModel(self.builder.spec)
- xcode.compile_coreml(model, self.model_name, out_dir)
-
-
-@tvm._ffi.register_func("relay.ext.coremlcompiler")
-def coreml_compiler(func):
- """
- Create a CoreML runtime from a Relay module.
- """
- assert isinstance(func, tvm.relay.function.Function)
- model_dir = os.getcwd()
- name = str(func.attrs.global_symbol)
- builder = CodegenCoreML(name, func)
- builder.visit(func.body)
- mlmodelc_path = f"{model_dir}/{name}.mlmodelc"
- if os.path.exists(mlmodelc_path):
- shutil.rmtree(mlmodelc_path)
- builder.compile(model_dir)
-
- dev = tvm.cpu(0)
- return coreml_runtime.create(name, mlmodelc_path, dev).module
diff --git a/python/tvm/ir/__init__.py b/python/tvm/ir/__init__.py
index 3e893099f4..b74e9954d9 100644
--- a/python/tvm/ir/__init__.py
+++ b/python/tvm/ir/__init__.py
@@ -18,7 +18,6 @@
"""Common data structures across all IR variants."""
from . import diagnostics, instrument, transform
-from .affine_type import TensorAffineType, TupleAffineType
from .attrs import Attrs, DictAttrs, make_node
from .base import (
EnvFunc,
@@ -33,12 +32,11 @@ from .base import (
structural_hash,
)
from .container import Array, Map
-from .expr import BaseExpr, GlobalVar, PrimExpr, Range, RelayExpr
+from .expr import BaseExpr, GlobalVar, PrimExpr, Range, RelaxExpr
from .function import BaseFunc, CallingConv
from .global_info import GlobalInfo, DummyGlobalInfo, VDevice
from .module import IRModule
from .op import Op, register_intrin_lowering, register_op_attr
-from .tensor_type import TensorType
from .type import (
FuncType,
PointerType,
diff --git a/python/tvm/ir/affine_type.py b/python/tvm/ir/affine_type.py
deleted file mode 100644
index 24126f94b9..0000000000
--- a/python/tvm/ir/affine_type.py
+++ /dev/null
@@ -1,79 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements. See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership. The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License. You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied. See the License for the
-# specific language governing permissions and limitations
-# under the License.
-"""Types for quantized Tensors."""
-import tvm._ffi
-
-from . import _ffi_api
-from .base import Node
-
-
-class AffineType(Node):
- """The base class of Affine Types."""
-
- def __eq__(self, other):
- """Compare two types for structural equivalence."""
- return bool(tvm.ir.structural_equal(self, other))
-
- def __ne__(self, other):
- return not self.__eq__(other)
-
- def __str__(self):
- from tvm.relay import pretty_print # pylint: disable=import-outside-toplevel
-
- return pretty_print(self)
-
-
-@tvm._ffi.register_object("TensorAffineType")
-class TensorAffineType(AffineType):
- """The quantized type of a tensor, with scale, zero point, and datatype
-
- The real space value is calculated as x = x_q * scale + zero_point
-
- Parameters
- ----------
- scale: Expr
- The scale
-
- zero_point: Expr
- The zero_point
-
- dtype : str
- The content data type.
-
- axis : int
- The axis for per-channel quantization.
- """
-
- def __init__(self, scale, zero_point, dtype, axis=-1):
- self.__init_handle_by_constructor__(
- _ffi_api.TensorAffineType, scale, zero_point, dtype, axis
- )
-
-
-@tvm._ffi.register_object("TupleAffineType")
-class TupleAffineType(AffineType):
- """Affine types of a node with multiple outputs
-
- Parameters
- ----------
- types : List[TensorAffineType]
- The shape of the Tensor
-
- """
-
- def __init__(self, types):
- self.__init_handle_by_constructor__(_ffi_api.TupleAffineType, types)
diff --git a/python/tvm/ir/attrs.py b/python/tvm/ir/attrs.py
index 6afb383c9f..1dd43de91e 100644
--- a/python/tvm/ir/attrs.py
+++ b/python/tvm/ir/attrs.py
@@ -24,7 +24,7 @@ from . import _ffi_api
@tvm._ffi.register_object
class Attrs(Object):
- """Attribute node, which is mainly use for defining attributes of relay operators.
+ """Attribute node, which is mainly use for defining attributes of operators.
Used by function registered in python side, such as compute, schedule and alter_layout.
Attrs is passed as the first argument to these functions.
diff --git a/python/tvm/ir/base.py b/python/tvm/ir/base.py
index 535b97e62d..2097423cbe 100644
--- a/python/tvm/ir/base.py
+++ b/python/tvm/ir/base.py
@@ -157,11 +157,9 @@ def structural_equal(lhs, rhs, map_free_vars=False):
- Normal node: equality is recursively defined without the restriction
of graph nodes.
- Vars(tir::Var, TypeVar) and non-constant relay expression nodes are graph nodes.
- For example, it means that `%1 = %x + %y; %1 + %1` is not structurally equal
- to `%1 = %x + %y; %2 = %x + %y; %1 + %2` in relay.
+ Vars(tir::Var, relax::Var) are graph nodes.
- A var-type node(e.g. tir::Var, TypeVar) can be mapped as equal to another var
+ A var-type node(e.g. tir::Var) can be mapped as equal to another var
with the same type if one of the following condition holds:
- They appear in a same definition point(e.g. function argument).
diff --git a/python/tvm/ir/expr.py b/python/tvm/ir/expr.py
index 1dcb9f6cf6..d140b5867c 100644
--- a/python/tvm/ir/expr.py
+++ b/python/tvm/ir/expr.py
@@ -42,16 +42,16 @@ class PrimExpr(BaseExpr):
dtype: str
-class RelayExpr(BaseExpr):
+class RelaxExpr(BaseExpr):
"""Base class of all non-primitive expressions."""
@property
def checked_type(self):
- """Get the checked type of tvm.relay.Expr.
+ """Get the checked type of tvm.relax.Expr.
Returns
-------
- checked_type : tvm.relay.Type
+ checked_type : tvm.ir.Type
The checked type.
"""
ret = self._checked_type_
@@ -72,7 +72,7 @@ class RelayExpr(BaseExpr):
@tvm._ffi.register_object("GlobalVar")
-class GlobalVar(RelayExpr):
+class GlobalVar(RelaxExpr):
"""A global variable in the IR.
GlobalVar is used to refer to the global functions
@@ -89,12 +89,12 @@ class GlobalVar(RelayExpr):
def __init__(self, name_hint: str, type_annot: Optional[Type] = None):
self.__init_handle_by_constructor__(_ffi_api.GlobalVar, name_hint, type_annot)
- def __call__(self, *args: RelayExpr) -> BaseExpr:
+ def __call__(self, *args: RelaxExpr) -> BaseExpr:
"""Call the global variable.
Parameters
----------
- args: List[RelayExpr]
+ args: List[RelaxExpr]
The arguments to the call.
Returns
@@ -105,7 +105,7 @@ class GlobalVar(RelayExpr):
# pylint: disable=import-outside-toplevel
# TODO(@relax-team): replace with Relax base class after it's introduced
- if all(isinstance(x, RelayExpr) for x in args):
+ if all(isinstance(x, RelaxExpr) for x in args):
from tvm import relax
return relax.Call(self, args)
@@ -185,12 +185,12 @@ class Range(Node, Scriptable):
# TODO(@relax-team): remove when we have a RelaxExpr base class
-def is_relax_expr(expr: RelayExpr) -> bool:
- """check if a RelayExpr is a Relax expresssion.
+def is_relax_expr(expr: RelaxExpr) -> bool:
+ """check if a RelaxExpr is a Relax expresssion.
Parameters
----------
- expr : RelayExpr
+ expr : RelaxExpr
The expression to check.
Returns
diff --git a/python/tvm/ir/function.py b/python/tvm/ir/function.py
index a4ea094c74..8527ce66f0 100644
--- a/python/tvm/ir/function.py
+++ b/python/tvm/ir/function.py
@@ -21,7 +21,7 @@ from typing import Union, Dict
from enum import IntEnum
import tvm.runtime
from tvm.runtime.object import Object
-from .expr import RelayExpr
+from .expr import RelaxExpr
from .attrs import DictAttrs
from . import _ffi_api
@@ -34,7 +34,7 @@ class CallingConv(IntEnum):
DEVICE_KERNEL_LAUNCH = 2
-class BaseFunc(RelayExpr):
+class BaseFunc(RelaxExpr):
"""Base class of all functions."""
@property
diff --git a/python/tvm/ir/instrument.py b/python/tvm/ir/instrument.py
index 4402d0fc22..72d4777194 100644
--- a/python/tvm/ir/instrument.py
+++ b/python/tvm/ir/instrument.py
@@ -213,7 +213,7 @@ def pass_instrument(pi_cls=None):
skip_annotate = SkipPass("AnnotateSpans")
with tvm.transform.PassContext(instruments=[skip_annotate]):
- tvm.relay.build(mod, "llvm")
+ tvm.build(mod, "llvm")
"""
def create_pass_instrument(pi_cls):
@@ -249,8 +249,7 @@ class PassTimingInstrument(tvm.runtime.Object):
timing_inst = PassTimingInstrument()
with tvm.transform.PassContext(instruments=[timing_inst]):
- relay_mod = relay.transform.InferType()(relay_mod)
- relay_mod = relay.transform.FoldScaleAxis()(relay_mod)
+ relax_mod = relax.transform.FuseOps()(relax_mod)
# before exiting the context, get profile results.
profiles = timing_inst.render()
"""
diff --git a/python/tvm/ir/json_compact.py b/python/tvm/ir/json_compact.py
index 756dbc4992..17c55db81a 100644
--- a/python/tvm/ir/json_compact.py
+++ b/python/tvm/ir/json_compact.py
@@ -141,18 +141,6 @@ def create_updater_08_to_09():
node_map = {
# Base IR
"GlobalVar": _initialize_virtual_device,
- "relay.Var": _initialize_virtual_device,
- "relay.Function": _initialize_virtual_device,
- "relay.Tuple": _initialize_virtual_device,
- "relay.Call": _initialize_virtual_device,
- "relay.Let": _initialize_virtual_device,
- "relay.If": _initialize_virtual_device,
- "relay.TupleGetItem": _initialize_virtual_device,
- "relay.RefCreate": _initialize_virtual_device,
- "relay.RefRead": _initialize_virtual_device,
- "relay.RefWrite": _initialize_virtual_device,
- "relay.Match": _initialize_virtual_device,
- "relay.Constant": _initialize_virtual_device,
}
return create_updater(node_map, "0.8", "0.9")
@@ -171,138 +159,6 @@ def create_updater_07_to_08():
return create_updater(node_map, "0.7", "0.8")
-def create_updater_06_to_07():
- """Create an update to upgrade json from v0.6 to v0.7
-
- Returns
- -------
- fupdater : function
- The updater function
- """
-
- def _ftype_var(item, nodes):
- vindex = int(item["attrs"]["var"])
- item["attrs"]["name_hint"] = nodes[vindex]["attrs"]["name"]
- # set vindex to null
- nodes[vindex]["type_key"] = ""
- del item["attrs"]["var"]
- assert item["type_key"].startswith("relay.")
- item["type_key"] = item["type_key"][len("relay.") :]
- return item
-
- def _rename(new_name):
- def _convert(item, _):
- item["type_key"] = new_name
- return item
-
- return _convert
-
- def _update_tir_var(new_name):
- def _convert(item, _):
- item["type_key"] = new_name
- item["attrs"]["type_annotation"] = "0"
- return item
-
- return _convert
-
- def _update_global_key(item, _):
- if "global_key" in item:
- item["repr_str"] = item["global_key"]
- del item["global_key"]
- return item
-
- def _update_from_std_str(key):
- def _convert(item, nodes):
- str_val = item["attrs"][key]
- jdata = json.loads(tvm.ir.save_json(tvm.runtime.String(str_val)))
- root_idx = jdata["root"]
- val = jdata["nodes"][root_idx]
- sidx = len(nodes)
- nodes.append(val)
- item["attrs"][key] = f"{sidx}"
- return item
-
- return _convert
-
- node_map = {
- # Base IR
- "SourceName": _update_global_key,
- "EnvFunc": _update_global_key,
- "relay.Op": [_update_global_key, _rename("Op")],
- "relay.TypeVar": [_ftype_var, _update_from_std_str("name_hint")],
- "TypeVar": _update_from_std_str("name_hint"),
- "relay.Id": [_update_from_std_str("name_hint")],
- "relay.GlobalTypeVar": [_ftype_var, _update_from_std_str("name_hint")],
- "GlobalTypeVar": _update_from_std_str("name_hint"),
- "relay.Type": _rename("Type"),
- "relay.TupleType": _rename("TupleType"),
- "relay.TypeConstraint": _rename("TypeConstraint"),
- "relay.FuncType": _rename("FuncType"),
- "relay.IncompleteType": _rename("IncompleteType"),
- "relay.TypeRelation": _rename("TypeRelation"),
- "relay.TypeCall": _rename("TypeCall"),
- "relay.Constructor": _update_from_std_str("name_hint"),
- "relay.Module": _rename("IRModule"),
- "relay.SourceName": _rename("SourceName"),
- "relay.Span": _rename("Span"),
- "relay.GlobalVar": [_rename("GlobalVar"), _update_from_std_str("name_hint")],
- "GlobalVar": _update_from_std_str("name_hint"),
- "relay.Pass": _rename("transform.Pass"),
- "relay.PassInfo": _rename("transform.PassInfo"),
- "relay.PassContext": _rename("transform.PassContext"),
- "relay.ModulePass": _rename("transform.ModulePass"),
- "relay.Sequential": _rename("transform.Sequential"),
- "StrMap": _rename("Map"),
- # TIR
- "Variable": [_update_tir_var("tir.Var"), _update_from_std_str("name")],
- "SizeVar": [_update_tir_var("tir.SizeVar"), _update_from_std_str("name")],
- "StringImm": [_rename("tir.StringImm"), _update_from_std_str("value")],
- "Cast": _rename("tir.Cast"),
- "Add": _rename("tir.Add"),
- "Sub": _rename("tir.Sub"),
- "Mul": _rename("tir.Mul"),
- "Div": _rename("tir.Div"),
- "Mod": _rename("tir.Mod"),
- "FloorDiv": _rename("tir.FloorDiv"),
- "FloorMod": _rename("tir.FloorMod"),
- "Min": _rename("tir.Min"),
- "Max": _rename("tir.Max"),
- "EQ": _rename("tir.EQ"),
- "NE": _rename("tir.NE"),
- "LT": _rename("tir.LT"),
- "LE": _rename("tir.LE"),
- "GT": _rename("tir.GT"),
- "GE": _rename("tir.GE"),
- "And": _rename("tir.And"),
- "Or": _rename("tir.Or"),
- "Not": _rename("tir.Not"),
- "Select": _rename("tir.Select"),
- "BufferLoad": _rename("tir.BufferLoad"),
- "Ramp": _rename("tir.Ramp"),
- "Broadcast": _rename("tir.Broadcast"),
- "Shuffle": _rename("tir.Shuffle"),
- "Call": [_rename("tir.Call"), _update_from_std_str("name")],
- "Let": _rename("tir.Let"),
- "Any": _rename("tir.Any"),
- "LetStmt": _rename("tir.LetStmt"),
- "AssertStmt": _rename("tir.AssertStmt"),
- "BufferStore": _rename("tir.BufferStore"),
- "BufferRealize": _rename("tir.BufferRealize"),
- "Allocate": _rename("tir.Allocate"),
- "IfThenElse": _rename("tir.IfThenElse"),
- "Evaluate": _rename("tir.Evaluate"),
- "Prefetch": _rename("tir.Prefetch"),
- "AttrStmt": [_rename("tir.AttrStmt"), _update_from_std_str("attr_key")],
- "Layout": [_rename("tir.Layout"), _update_from_std_str("name")],
- "Buffer": [
- _rename("tir.Buffer"),
- _update_from_std_str("name"),
- _update_from_std_str("scope"),
- ],
- }
- return create_updater(node_map, "0.6", "0.7")
-
-
def upgrade_json(json_str):
"""Update json from a historical version.
diff --git a/python/tvm/ir/module.py b/python/tvm/ir/module.py
index ed7ea44399..8347e218be 100644
--- a/python/tvm/ir/module.py
+++ b/python/tvm/ir/module.py
@@ -97,7 +97,7 @@ class IRModule(Node, Scriptable):
return self._add(var, val, True)
def _add(self, var, val, update=True):
- if isinstance(val, _expr.RelayExpr):
+ if isinstance(val, _expr.RelaxExpr):
if isinstance(var, string_types):
if _ffi_api.Module_ContainGlobalVar(self, var):
var = _ffi_api.Module_GetGlobalVar(self, var)
@@ -151,7 +151,7 @@ class IRModule(Node, Scriptable):
var: GlobalVar
The global variable.
- func: tvm.relay.Function
+ func: tvm.ir.BaseFunc
The function to be inserted.
"""
return _ffi_api.Module_UpdateFunction(self, var, func)
@@ -231,7 +231,7 @@ class IRModule(Node, Scriptable):
Parameters
----------
- expr: RelayExpr
+ expr: RelaxExpr
The starting expression
global_funcs: Optional[dict]
diff --git a/python/tvm/ir/op.py b/python/tvm/ir/op.py
index dae97f114b..932aef24c6 100644
--- a/python/tvm/ir/op.py
+++ b/python/tvm/ir/op.py
@@ -19,11 +19,11 @@
import tvm._ffi
from . import _ffi_api
-from .expr import RelayExpr
+from .expr import RelaxExpr
@tvm._ffi.register_object("Op")
-class Op(RelayExpr):
+class Op(RelaxExpr):
"""Primitive operator in the IR."""
def __init__(self):
diff --git a/python/tvm/ir/tensor_type.py b/python/tvm/ir/tensor_type.py
deleted file mode 100644
index 495e0fe868..0000000000
--- a/python/tvm/ir/tensor_type.py
+++ /dev/null
@@ -1,61 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements. See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership. The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License. You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied. See the License for the
-# specific language governing permissions and limitations
-# under the License.
-"""Type relation and function for type checking."""
-import tvm._ffi
-
-from . import _ffi_api
-from .type import Type
-
-
-@tvm._ffi.register_object("relay.TensorType")
-class TensorType(Type):
- """A concrete TensorType in Relay.
-
- This is the type assigned to tensors with a known dtype and shape.
- For example, a tensor of `float32` and `(5, 5)`.
-
- Parameters
- ----------
- shape : List[tvm.ir.PrimExpr]
- The shape of the Tensor
-
- dtype : Optional[str]
- The content data type.
- """
-
- def __init__(self, shape, dtype="float32"):
- self.__init_handle_by_constructor__(_ffi_api.TensorType, shape, dtype)
-
- @property
- def concrete_shape(self):
- """Get shape of the type as concrete tuple of int.
-
- Returns
- -------
- shape : List[int]
- The concrete shape of the Type.
-
- Raises
- ------
- TypeError : If the shape is symbolic
- """
- return tuple(int(x) for x in self.shape)
-
- def __str__(self):
- from tvm.relay import pretty_print # pylint: disable=import-outside-toplevel
-
- return pretty_print(self)
diff --git a/python/tvm/ir/transform.py b/python/tvm/ir/transform.py
index b2937acaa7..c00a4de6da 100644
--- a/python/tvm/ir/transform.py
+++ b/python/tvm/ir/transform.py
@@ -365,15 +365,13 @@ def module_pass(pass_func=None, opt_level=None, name=None, required=None, tracea
.. code-block:: python
- @relay.transform.module_pass
+ @tvm.ir.transform.module_pass
class CustomPipeline:
def __init__(self, enable_fold):
self.enable_fold = enable_fold
- self.cse = relay.transform.EliminateCommonSubexpr()
- self.const_fold = relay.transform.FoldConstant()
+ self.const_fold = relax.transform.FoldConstant()
def transform_module(self, mod, ctx):
- mod = self.cse(mod, ctx)
if self.enable_fold:
mod = self.const_fold(mod, ctx)
return mod
@@ -389,15 +387,9 @@ def module_pass(pass_func=None, opt_level=None, name=None, required=None, tracea
.. code-block:: python
- @relay.transform.module_pass(opt_level=2)
+ @tvm.ir.transform.module_pass(opt_level=2)
def transform(mod, ctx):
- tp = relay.TensorType((10,), "float32")
- x = relay.var("x", tp)
- gv = relay.GlobalVar("var")
- func = relay.Function([x], relay.abs(x))
- new_mod = tvm.IRModule({gv: func})
- new_mod.update(mod)
- return new_mod
+ return relax.transform.FoldConstant(mod)
module_pass = transform
assert isinstance(module_pass, transform.ModulePass)
diff --git a/python/tvm/ir/type.py b/python/tvm/ir/type.py
index 8ecf1ffb4a..914275415c 100644
--- a/python/tvm/ir/type.py
+++ b/python/tvm/ir/type.py
@@ -94,10 +94,10 @@ class FuncType(Type):
Parameters
----------
- arg_types : List[tvm.relay.Type]
+ arg_types : List[tvm.ir.Type]
The argument types
- ret_type : tvm.relay.Type
+ ret_type : tvm.ir.Type
The return type.
"""
diff --git a/python/tvm/meta_schedule/testing/custom_builder_runner.py b/python/tvm/meta_schedule/testing/custom_builder_runner.py
deleted file mode 100644
index 0e8ee435f7..0000000000
--- a/python/tvm/meta_schedule/testing/custom_builder_runner.py
+++ /dev/null
@@ -1,173 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements. See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership. The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License. You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied. See the License for the
-# specific language governing permissions and limitations
-# under the License.
-"""Customized builder and runner methods"""
-# pylint: disable=import-outside-toplevel
-
-from typing import TYPE_CHECKING, Dict, List, Optional, Union, Callable
-
-if TYPE_CHECKING:
- import numpy as np # type: ignore
- from tvm.ir import IRModule
- from tvm.meta_schedule.runner import EvaluatorConfig, RPCConfig
- from tvm.runtime import Device, Module, NDArray
- from tvm.target import Target
- from tvm.runtime.vm import Executable
-
-
-def build_relay(
- mod: "IRModule",
- target: "Target",
- params: Dict[str, "NDArray"],
-) -> "Module":
- """Build a Relay IRModule
-
- Parameters
- ----------
- mod : IRModule
- The Relay IRModule to build.
- target : Target
- The target to build the module for.
- params : Dict[str, NDArray]
- The parameter dict to build the module with.
-
- Returns
- -------
- mod : runtime.Module
- The built module.
- """
- from tvm.relay.build_module import _build_module_no_factory as relay_build
- from tvm.runtime import Module
-
- result = relay_build(mod, target=target, target_host=None, params=params)
- assert isinstance(result, Module)
- return result
-
-
-def build_relay_with_tensorrt(
- mod: "IRModule",
- target: "Target",
- params: Dict[str, "NDArray"],
-) -> "Module":
- """Build a Relay IRModule with TensorRT BYOC
-
- Parameters
- ----------
- mod : IRModule
- The Relay IRModule to build.
-
- target : Target
- The target to build the module for.
-
- params : Dict[str, NDArray]
- The parameter dict to build the module with.
-
- Returns
- -------
- mod : runtime.Module
- The built module.
- """
- from tvm.ir.transform import PassContext
- from tvm.relay.build_module import _build_module_no_factory as relay_build
- from tvm.relay.op.contrib import tensorrt
- from tvm.runtime import Module
-
- mod = tensorrt.partition_for_tensorrt(mod, params)
- with PassContext(opt_level=3):
- result = relay_build(mod, target=target, target_host=None, params=params)
- assert isinstance(result, Module)
- return result
-
-
-def run_with_graph_executor(
- rt_mod: "Module",
- device: "Device",
- evaluator_config: "EvaluatorConfig",
- repeated_args: List["NDArray"],
-) -> List[float]:
- """Run a Relay module with GraphExecutor
-
- Parameters
- ----------
- rt_mod : Module
- The Relay module to run.
- device : Device
- The device to run the module on.
- evaluator_config : EvaluatorConfig
- The evaluator configuration to run the module with.
- repeated_args : List[NDArray]
- The list of repeated arguments to run the module with.
-
- Returns
- -------
- results : List[float]
- The list of results.
- """
- import itertools
-
- from tvm.contrib.graph_executor import GraphModule
-
- graph_mod = GraphModule(rt_mod["default"](device))
- evaluator = graph_mod.module.time_evaluator(
- func_name="run",
- dev=device,
- number=evaluator_config.number,
- repeat=evaluator_config.repeat,
- min_repeat_ms=evaluator_config.min_repeat_ms,
- f_preproc="cache_flush_cpu_non_first_arg"
- if evaluator_config.enable_cpu_cache_flush
- else "",
- )
- repeated_costs = []
- for args in repeated_args:
- profile_result = evaluator(*args)
- repeated_costs.append(profile_result.results)
- costs = [float(cost) for cost in itertools.chain.from_iterable(repeated_costs)]
- return costs
-
-
-def run_module_via_rpc(
- rpc_config: "RPCConfig",
- lib: Union["Module", "Executable"],
- dev_type: str,
- args: Union[Dict[int, "np.ndarray"], Dict[str, "np.ndarray"]],
- continuation: Callable,
- backend: Optional[str] = "graph",
-):
- """Execute a tvm.runtime.Module on RPC remote"""
- # pylint: disable=import-outside-toplevel
- import os
- import tempfile
-
- from tvm.contrib.tar import tar
- from tvm.runtime import ndarray
-
- # pylint: enable=import-outside-toplevel
-
- with tempfile.TemporaryDirectory() as tmp_dir:
- filename = os.path.join(tmp_dir, "tvm_tmp_mod." + tar.output_format)
- if backend == "vm":
- code, lib = lib.save()
- lib.export_library(filename, fcompile=tar)
- session = rpc_config.connect_server()
- session.upload(filename)
- _, filename = os.path.split(filename)
- rt_mod = session.load_module(filename)
- if backend == "vm":
- rt_mod = session.get_function("runtime.Load_Executable")(code, rt_mod)
- dev = session.device(dev_type=dev_type, dev_id=0)
- nd_args = {k: ndarray.array(v, dev) for k, v in args.items()}
- return continuation(rt_mod, dev, nd_args)
diff --git a/python/tvm/meta_schedule/testing/dataset_collect_models.py b/python/tvm/meta_schedule/testing/dataset_collect_models.py
deleted file mode 100644
index 8992f73d28..0000000000
--- a/python/tvm/meta_schedule/testing/dataset_collect_models.py
+++ /dev/null
@@ -1,85 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements. See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership. The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License. You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied. See the License for the
-# specific language governing permissions and limitations
-# under the License.
-# pylint: disable=missing-docstring
-
-import argparse
-import os
-from typing import List, Tuple
-
-from tqdm import tqdm # type: ignore
-from tvm.meta_schedule.testing.relay_workload import get_network
-
-
-# pylint: disable=too-many-branches
-def _build_dataset() -> List[Tuple[str, List[int]]]:
- network_keys = []
- for name in [
- "resnet_18",
- "resnet_50",
- "mobilenet_v2",
- "mobilenet_v3",
- "wide_resnet_50",
- "resnext_50",
- "densenet_121",
- "vgg_16",
- ]:
- for batch_size in [1, 4, 8]:
- for image_size in [224, 240, 256]:
- network_keys.append((name, [batch_size, 3, image_size, image_size]))
- # inception-v3
- for name in ["inception_v3"]:
- for batch_size in [1, 2, 4]:
- for image_size in [299]:
- network_keys.append((name, [batch_size, 3, image_size, image_size]))
- # resnet3d
- for name in ["resnet3d_18"]:
- for batch_size in [1, 2, 4]:
- for image_size in [112, 128, 144]:
- network_keys.append((name, [batch_size, 3, image_size, image_size, 16]))
- # bert
- for name in ["bert_tiny", "bert_base", "bert_medium", "bert_large"]:
- for batch_size in [1, 2, 4]:
- for seq_length in [64, 128, 256]:
- network_keys.append((name, [batch_size, seq_length]))
- # dcgan
- for name in ["dcgan"]:
- for batch_size in [1, 4, 8]:
- for image_size in [64]:
- network_keys.append((name, [batch_size, 3, image_size, image_size]))
- return network_keys
-
-
-def main():
- model_cache_dir = args.model_cache_dir
- try:
- os.makedirs(model_cache_dir, exist_ok=True)
- except OSError:
- print(f"Directory {model_cache_dir} cannot be created successfully.")
- keys = _build_dataset()
- for name, input_shape in tqdm(keys):
- get_network(name=name, input_shape=input_shape, cache_dir=model_cache_dir)
-
-
-if __name__ == "__main__":
- parser = argparse.ArgumentParser() # pylint: disable=invalid-name
- parser.add_argument(
- "--model_cache_dir",
- type=str,
- help="Please provide the full path to the model cache dir.",
- )
- args = parser.parse_args() # pylint: disable=invalid-name
- main()
diff --git a/python/tvm/meta_schedule/testing/dataset_extract_tasks.py b/python/tvm/meta_schedule/testing/dataset_extract_tasks.py
deleted file mode 100644
index f299227aa3..0000000000
--- a/python/tvm/meta_schedule/testing/dataset_extract_tasks.py
+++ /dev/null
@@ -1,104 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements. See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership. The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License. You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied. See the License for the
-# specific language governing permissions and limitations
-# under the License.
-# pylint: disable=missing-docstring
-
-import argparse
-import glob
-import json
-import os
-from tqdm import tqdm # type: ignore
-
-import tvm
-from tvm import meta_schedule as ms
-from tvm.ir import save_json
-from tvm.meta_schedule.testing.relay_workload import _load_cache
-from tvm.runtime import load_param_dict
-
-
-def _parse_args():
- parser = argparse.ArgumentParser()
- parser.add_argument(
- "--model_cache_dir", type=str, help="Please provide the full path to the model cache dir."
- )
- parser.add_argument(
- "--task_cache_dir", type=str, help="Please provide the full path to save extracted tasks."
- )
- parser.add_argument(
- "--target", type=str, default="cuda", help="Please specify the target hardware for tuning."
- )
- return parser.parse_args()
-
-
-# pylint: disable=too-many-locals
-def extract_and_save_tasks(cache_file):
- """Extract tuning tasks and cache the nonspatial ones in the given directory.
-
- Parameters
- ----------
- cache_file : str
- The filename of the cached model.
-
- Returns
- -------
- None
- """
-
- mod, params_bytearray, _ = _load_cache(args.model_cache_dir, cache_file)
- params = load_param_dict(params_bytearray)
- try:
- extracted_tasks = ms.relay_integration.extract_tasks(mod, target=args.target, params=params)
- except tvm.error.TVMError as error:
- print(str(error))
- return
- task_cache_path = os.path.join(
- args.task_cache_dir, cache_file.split(".")[0] + "_extracted_tasks.json"
- )
- is_spatial = tvm.get_global_func("tir.schedule.IsSpatialPrimFunc")
- with open(task_cache_path, "w", encoding="utf8") as file:
- for i, task in enumerate(extracted_tasks):
- subgraph = task.dispatched[0]
- prim_func = subgraph[subgraph.get_global_vars()[0]]
- if not is_spatial(prim_func):
- subgraph_str = save_json(subgraph)
- json_obj = [task.task_name, json.loads(subgraph_str)]
- json_str = json.dumps(json_obj)
- assert "\n" not in json_str, "Failed to generate single line string."
- if i == len(extracted_tasks) - 1:
- file.write(json_str)
- else:
- file.write(json_str + "\n")
-
-
-args = _parse_args() # pylint: disable=invalid-name
-
-
-def main():
- if not os.path.isdir(args.model_cache_dir):
- raise Exception("Please provide a correct model cache dir.")
- try:
- os.makedirs(args.task_cache_dir, exist_ok=True)
- except OSError:
- print(f"Directory {args.task_cache_dir} cannot be created successfully.")
-
- paths = glob.glob(os.path.join(args.model_cache_dir, "*.json")) # pylint: disable=invalid-name
- for path in tqdm(paths):
- filename = path.split("/")[-1]
- extract_and_save_tasks(filename)
-
-
-if __name__ == "__main__":
- main()
diff --git a/python/tvm/meta_schedule/testing/dataset_sample_candidates.py b/python/tvm/meta_schedule/testing/dataset_sample_candidates.py
deleted file mode 100644
index 6da11bf5e9..0000000000
--- a/python/tvm/meta_schedule/testing/dataset_sample_candidates.py
+++ /dev/null
@@ -1,187 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements. See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership. The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License. You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied. See the License for the
-# specific language governing permissions and limitations
-# under the License.
-# pylint: disable=missing-docstring
-
-import argparse
-import glob
-import json
-import os
-from typing import List
-
-from tqdm import tqdm # type: ignore
-import tvm
-from tvm import meta_schedule as ms
-from tvm.ir import load_json
-from tvm.target import Target
-
-
-def _parse_args():
- parser = argparse.ArgumentParser()
- parser.add_argument(
- "--task_cache_dir", type=str, help="Please provide the full path to the extracted tasks."
- )
- parser.add_argument(
- "--candidate_cache_dir",
- type=str,
- help="Please provide the full path to save the sampled candidates.",
- )
- parser.add_argument(
- "--target",
- type=str,
- default="nvidia/geforce-rtx-3070",
- help="Please specify the target hardware for tuning.\
- Note: for generating dataset, the hardware does not need to be present.",
- )
- parser.add_argument(
- "--init_population_size",
- type=int,
- default=256,
- help="The initial population size used in evolutionary search.",
- )
- parser.add_argument(
- "--num_samples_per_task",
- type=int,
- default=400,
- help="The number of samples to gather per tuning task.",
- )
- parser.add_argument(
- "--num_trials_per_iter",
- type=int,
- default=64,
- help="The number of trials per iteration in evolutionary search.",
- )
- parser.add_argument(
- "--max_trials_per_task",
- type=int,
- default=400,
- help="The maximum number of trials per task in evolutionary search.",
- )
- parser.add_argument(
- "--max_retry_per_task",
- type=int,
- default=10,
- help="The maximum number of retry attempts allowed.",
- )
- parser.add_argument(
- "--file_group",
- type=int,
- default=0,
- help="To enable running multiple scripts in parallel, files [idx * 10 : (idx + 1) * 10]\
- in the sorted file list from the given directory will be run.",
- )
- return parser.parse_args()
-
-
-# pylint: disable=too-many-locals
-def sample_candidates(task, task_name, model_name):
- """Randomly sample candidates for a task and save the candidates in the given directory.
-
- Parameters
- ----------
- task : IRModule
- The initial ir module used for generating the search space.
- task_name : str
- The name of the task.
- model_name : str
- The name of the model.
-
- Returns
- -------
- None
- """
- candidate_path = os.path.join(
- args.candidate_cache_dir, model_name, task_name + "_candidates.json"
- )
- workload_path = os.path.join(args.candidate_cache_dir, model_name, task_name + "_workload.json")
- database = ms.database.JSONDatabase(
- path_workload=workload_path,
- path_tuning_record=candidate_path,
- )
- sample_init_population = tvm.get_global_func(
- "meta_schedule.SearchStrategyEvolutionarySearchSampleInitPopulation"
- )
- evolve_with_cost_model = tvm.get_global_func(
- "meta_schedule.SearchStrategyEvolutionarySearchEvolveWithCostModel"
- )
- strategy = ms.search_strategy.EvolutionarySearch(init_measured_ratio=0.0)
- target = Target(args.target)
- context = ms.TuneContext(
- mod=task,
- target=target,
- space_generator="post-order-apply",
- search_strategy=strategy,
- task_name=task_name,
- )
- context.initialize()
- context.pre_tuning(
- max_trials=args.max_trials_per_task,
- num_trials_per_iter=args.num_trials_per_iter,
- design_spaces=context.generate_design_space(),
- database=database,
- cost_model=ms.cost_model.RandomModel(), # type: ignore
- )
-
- all_states: List[tvm.tir.Schedule] = []
- num_retry, itr = 0, 0
- states = sample_init_population(strategy, args.init_population_size)
- while len(all_states) < args.num_samples_per_task and num_retry < args.max_retry_per_task:
- states = evolve_with_cost_model(strategy, states, len(states))
- all_states += states
- if len(states) == 0:
- states = sample_init_population(strategy, args.init_population_size)
- num_retry += 1
- else:
- num_retry = 0
- print(f"iter: {itr}, number of states sampled: {len(all_states)}")
- itr += 1
- all_states = all_states[: args.num_samples_per_task]
-
- workload = ms.database.Workload(context.mod)
- database.commit_workload(context.mod)
- for state in all_states:
- database.commit_tuning_record(ms.database.TuningRecord(state.trace, workload))
-
-
-args = _parse_args() # pylint: disable=invalid-name
-
-
-def main():
- if not os.path.isdir(args.task_cache_dir):
- raise Exception("Please provide a correct task cache dir.")
- try:
- os.makedirs(args.candidate_cache_dir, exist_ok=True)
- except OSError:
- print(f"Directory {args.candidate_cache_dir} cannot be created successfully.")
-
- task_paths = sorted(glob.glob(os.path.join(args.task_cache_dir, "*.json")))[
- args.file_group * 10 : (args.file_group + 1) * 10
- ]
- print(f"Selected models: {task_paths}")
- for num, task_path in enumerate(task_paths):
- print(f"Processing model {num} ...")
- with open(task_path, "rb") as file:
- tasks = file.readlines()
- model_name = task_path.split("/")[-1][len("relay-") :][: -len("_extracted_tasks.json")]
- os.makedirs(os.path.join(args.candidate_cache_dir, model_name), exist_ok=True)
- for task_str in tqdm(tasks):
- task_name, task_mod = json.loads(task_str)
- task_mod = load_json(json.dumps(task_mod))
- sample_candidates(task_mod, task_name, model_name)
-
-
-if __name__ == "__main__":
- main()
diff --git a/python/tvm/meta_schedule/testing/tune_onnx.py b/python/tvm/meta_schedule/testing/tune_onnx.py
deleted file mode 100644
index 2100f0e7c9..0000000000
--- a/python/tvm/meta_schedule/testing/tune_onnx.py
+++ /dev/null
@@ -1,196 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements. See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership. The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License. You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied. See the License for the
-# specific language governing permissions and limitations
-# under the License.
-# pylint: disable=missing-docstring
-import argparse
-import json
-import logging
-
-import onnx # type: ignore
-import tvm
-from tvm import meta_schedule as ms
-from tvm.meta_schedule.testing.custom_builder_runner import run_module_via_rpc
-from tvm.relay.frontend import from_onnx
-from tvm.support import describe
-from tvm.testing.utils import strtobool
-
-from .tune_utils import create_timer, generate_input_data
-
-
-def _parse_args():
- args = argparse.ArgumentParser()
- args.add_argument(
- "--model-name",
- type=str,
- required=True,
- )
- args.add_argument(
- "--onnx-path",
- type=str,
- required=True,
- )
- args.add_argument(
- "--input-shape",
- type=str,
- required=True,
- help='example: `[{"name": "input1", "dtype": "int64", "shape": [1, 1, 8]}]',
- )
- args.add_argument(
- "--target",
- type=str,
- required=True,
- )
- args.add_argument(
- "--num-trials",
- type=int,
- required=True,
- )
- args.add_argument(
- "--rpc-host",
- type=str,
- required=True,
- )
- args.add_argument(
- "--rpc-port",
- type=int,
- required=True,
- )
- args.add_argument(
- "--rpc-key",
- type=str,
- required=True,
- )
- args.add_argument(
- "--work-dir",
- type=str,
- required=True,
- )
- args.add_argument(
- "--number",
- type=int,
- default=3,
- )
- args.add_argument(
- "--repeat",
- type=int,
- default=1,
- )
- args.add_argument(
- "--min-repeat-ms",
- type=int,
- default=100,
- )
- args.add_argument(
- "--adaptive-training",
- type=lambda x: bool(strtobool(x)),
- help="example: True / False",
- default=True,
- )
- args.add_argument(
- "--cpu-flush",
- type=lambda x: bool(strtobool(x)),
- help="example: True / False",
- required=True,
- )
- args.add_argument(
- "--backend",
- type=str,
- choices=["graph", "vm"],
- help="example: graph / vm",
- required=True,
- )
- parsed = args.parse_args()
- parsed.target = tvm.target.Target(parsed.target)
- parsed.input_shape = json.loads(parsed.input_shape)
- parsed.rpc_config = ms.runner.RPCConfig(
- tracker_host=parsed.rpc_host,
- tracker_port=parsed.rpc_port,
- tracker_key=parsed.rpc_key,
- session_timeout_sec=600,
- )
- return parsed
-
-
-logging.basicConfig(
- format="%(asctime)s.%(msecs)03d %(levelname)s %(message)s", datefmt="%Y-%m-%d %H:%M:%S"
-)
-logging.getLogger("tvm.meta_schedule").setLevel(logging.DEBUG)
-ARGS = _parse_args()
-
-
-def main():
- describe()
- print(f"Workload: {ARGS.model_name}")
-
- onnx_model = onnx.load(ARGS.onnx_path)
- shape_dict = {}
- for item in ARGS.input_shape:
- print(f" input_name : {item['name']}")
- print(f" input_shape: {item['shape']}")
- print(f" input_dtype: {item['dtype']}")
- shape_dict[item["name"]] = item["shape"]
- mod, params = from_onnx(onnx_model, shape_dict, freeze_params=True)
- input_data = {
- item["name"]: generate_input_data(item["shape"], item["dtype"]) for item in ARGS.input_shape
- }
-
- with ms.Profiler() as profiler:
- database = ms.relay_integration.tune_relay(
- mod=mod,
- target=ARGS.target,
- params=params,
- work_dir=ARGS.work_dir,
- max_trials_global=ARGS.num_trials,
- num_trials_per_iter=64,
- runner=ms.runner.RPCRunner( # type: ignore
- rpc_config=ARGS.rpc_config,
- evaluator_config=ms.runner.EvaluatorConfig(
- number=ARGS.number,
- repeat=ARGS.repeat,
- min_repeat_ms=ARGS.min_repeat_ms,
- enable_cpu_cache_flush=ARGS.cpu_flush,
- ),
- alloc_repeat=1,
- ),
- cost_model=ms.cost_model.XGBModel( # type: ignore
- extractor=ms.feature_extractor.PerStoreFeature(),
- adaptive_training=ARGS.adaptive_training,
- ),
- strategy=ms.search_strategy.EvolutionarySearch(),
- )
- lib = ms.relay_integration.compile_relay(
- database=database,
- mod=mod,
- target=ARGS.target,
- params=params,
- backend=ARGS.backend,
- )
-
- print("Tuning Time:")
- print(profiler.table())
-
- run_module_via_rpc(
- rpc_config=ARGS.rpc_config,
- lib=lib,
- dev_type=ARGS.target.kind.name,
- args=input_data,
- continuation=create_timer(ARGS.backend),
- backend=ARGS.backend,
- )
-
-
-if __name__ == "__main__":
- main()
diff --git a/python/tvm/meta_schedule/testing/tune_relay.py b/python/tvm/meta_schedule/testing/tune_relay.py
deleted file mode 100644
index 98eddf793f..0000000000
--- a/python/tvm/meta_schedule/testing/tune_relay.py
+++ /dev/null
@@ -1,213 +0,0 @@
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements. See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership. The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License. You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied. See the License for the
-# specific language governing permissions and limitations
-# under the License.
-# pylint: disable=missing-docstring
-import argparse
-import json
-import logging
-from typing import Dict
-
-import numpy as np # type: ignore
-import tvm
-from tvm import meta_schedule as ms
-from tvm.meta_schedule.testing.custom_builder_runner import run_module_via_rpc
-from tvm.meta_schedule.testing.relay_workload import get_network
-from tvm.meta_schedule.testing.tune_utils import create_timer, generate_input_data
-from tvm.support import describe
-from tvm.testing.utils import strtobool
-
-
-def _parse_args():
- args = argparse.ArgumentParser()
- args.add_argument(
- "--workload",
- type=str,
- required=True,
- )
- args.add_argument(
- "--input-shape",
- type=str,
- required=True,
- )
- args.add_argument(
- "--target",
- type=str,
- required=True,
- )
- args.add_argument(
- "--num-trials",
- type=int,
- required=True,
- )
- args.add_argument(
- "--rpc-host",
- type=str,
- required=True,
- )
- args.add_argument(
- "--rpc-port",
- type=int,
- required=True,
- )
- args.add_argument(
- "--rpc-key",
- type=str,
- required=True,
- )
- args.add_argument(
- "--work-dir",
- type=str,
- required=True,
- )
- args.add_argument(
- "--layout",
- type=str,
- default=None,
- )
- args.add_argument(
- "--cache-dir",
- type=str,
- default=None,
- )
- args.add_argument(
- "--number",
- type=int,
- default=3,
- )
- args.add_argument(
- "--repeat",
- type=int,
- default=1,
- )
- args.add_argument(
- "--min-repeat-ms",
- type=int,
- default=100,
- )
- args.add_argument(
- "--adaptive-training",
- type=lambda x: bool(strtobool(x)),
- help="example: True / False",
- default=True,
- )
- args.add_argument(
- "--cpu-flush",
- type=lambda x: bool(strtobool(x)),
- help="example: True / False",
- required=True,
- )
- args.add_argument(
- "--backend",
- type=str,
- choices=["graph", "vm"],
- help="example: graph / vm",
- required=True,
- )
- parsed = args.parse_args()
- parsed.target = tvm.target.Target(parsed.target)
- parsed.input_shape = json.loads(parsed.input_shape)
- parsed.rpc_config = ms.runner.RPCConfig(
- tracker_host=parsed.rpc_host,
- tracker_port=parsed.rpc_port,
- tracker_key=parsed.rpc_key,
- session_timeout_sec=600,
- )
- return parsed
-
-
-logging.basicConfig(
- format="%(asctime)s.%(msecs)03d %(levelname)s %(message)s", datefmt="%Y-%m-%d %H:%M:%S"
-)
-logging.getLogger("tvm.meta_schedule").setLevel(logging.DEBUG)
-ARGS = _parse_args()
-
-
-def main():
- describe()
- print(f"Workload: {ARGS.workload}")
-
- mod, params, (input_name, input_shape, input_dtype) = get_network(
- ARGS.workload,
- ARGS.input_shape,
- layout=ARGS.layout,
- cache_dir=ARGS.cache_dir,
- )
- input_info = [
- {
- "name": input_name,
- "shape": input_shape,
- "dtype": input_dtype,
- },
- ]
- input_data: Dict[str, np.ndarray] = {
- item["name"]: generate_input_data( # type: ignore
- item["shape"], # type: ignore
- item["dtype"], # type: ignore
- )
- for item in input_info
- }
- for item in input_info:
- print(f" input_name : {item['name']}")
- print(f" input_shape: {item['shape']}")
- print(f" input_dtype: {item['dtype']}")
-
- with ms.Profiler() as profiler:
- database = ms.relay_integration.tune_relay(
- mod=mod,
- target=ARGS.target,
- work_dir=ARGS.work_dir,
- max_trials_global=ARGS.num_trials,
- num_trials_per_iter=64,
- params=params,
- runner=ms.runner.RPCRunner( # type: ignore
- rpc_config=ARGS.rpc_config,
- evaluator_config=ms.runner.EvaluatorConfig(
- number=ARGS.number,
- repeat=ARGS.repeat,
- min_repeat_ms=ARGS.min_repeat_ms,
- enable_cpu_cache_flush=ARGS.cpu_flush,
- ),
- alloc_repeat=1,
- ),
- cost_model=ms.cost_model.XGBModel( # type: ignore
- extractor=ms.feature_extractor.PerStoreFeature(),
- adaptive_training=ARGS.adaptive_training,
- ),
- strategy=ms.search_strategy.EvolutionarySearch(),
- )
- lib = ms.relay_integration.compile_relay(
- database=database,
- mod=mod,
- target=ARGS.target,
- params=params,
- backend=ARGS.backend,
- )
-
- print("Tuning Time:")
- print(profiler.table())
-
- run_module_via_rpc(
- rpc_config=ARGS.rpc_config,
- lib=lib,
- dev_type=ARGS.target.kind.name,
- args=input_data,
- continuation=create_timer(ARGS.backend),
- backend=ARGS.backend,
- )
-
-
-if __name__ == "__main__":
- main()
diff --git a/python/tvm/meta_schedule/testing/tune_utils.py b/python/tvm/meta_schedule/testing/tune_utils.py
index 96de3a9711..46d61eaf78 100644
--- a/python/tvm/meta_schedule/testing/tune_utils.py
+++ b/python/tvm/meta_schedule/testing/tune_utils.py
@@ -66,131 +66,6 @@ def generate_input_data(
raise ValueError("Unsupported input datatype!")
-def create_timer(backend: str) -> Callable:
- """Create a function to run and benchmark the performance of whole given runtime module,
- or Executable in relay vm.
-
- Parameters
- ----------
- backend : str
- The backend to use, graph / vm.
-
- Returns
- -------
- func : Callable
- The function to benchmark the workload.
- """
-
- def f_timer(
- rt_mod: Union[tvm.runtime.Module, tvm.runtime.vm.Executable],
- dev: tvm.runtime.Device,
- input_data: Dict[str, NDArray],
- ) -> None:
- """Run and benchmark the given runtime module, print out the result.
-
- Parameters
- ----------
- rt_mod : Union[tvm.runtime.Module, tvm.runtime.vm.Executable]
- The runtime module or vm executable.
- dev : tvm.runtime.Device
- The device type to run workload.
- input_data : Dict[str, np.ndarray]
- The input data as a dictionary.
- """
- from tvm.contrib.graph_executor import GraphModule # pylint:disable=import-outside-toplevel
- from tvm.runtime.vm import VirtualMachine # pylint:disable=import-outside-toplevel
-
- try:
- if backend == "vm":
- vm = VirtualMachine(rt_mod, dev) # pylint: disable=invalid-name
- ftimer = vm.benchmark(
- dev, min_repeat_ms=500, repeat=5, number=1, end_to_end=False, **input_data
- )
- elif backend == "graph":
- mod = GraphModule(rt_mod["default"](dev))
- for input_name, input_value in input_data.items():
- mod.set_input(input_name, input_value)
- ftimer = mod.module.time_evaluator(
- "run", dev, min_repeat_ms=500, repeat=5, number=1
- )()
- else:
- raise ValueError(f"Backend {backend} not supported in f_timer!")
-
- results = list(np.array(ftimer.results) * 1000.0) # type: ignore
-
- print("Running time in time_evaluator: ", results)
- print("-------------------------------")
- print(f" Min (ms) : {min(results)}")
- print(f" Max (ms) : {max(results)}")
- print(f" Median (ms) : {median(results)}")
- print(f"Average (ms) : {sum(results) / len(results)}")
- except Exception as exc: # pylint: disable=broad-except
- print(
- f"Run module f_timer via RPC failed, exception: {exc}",
- )
-
- return f_timer
-
-
-def create_time_per_layer(graph: str) -> Callable:
- """Create a function to run and benchmark the per-layer performance of given runtime module,
- given the graph output of the module from graph compiler.
-
- Parameters
- ----------
- graph : str
- The json format graph output of the module from graph compiler.
-
- Returns
- -------
- func : Callable
- The function using the json format graph.
- """
-
- def f_time_per_layer(
- rt_mod: tvm.runtime.Module,
- dev: tvm.runtime.Device,
- input_data: Dict[str, NDArray],
- ) -> None:
- """Run and benchmark the per-layer performance of given runtime module,
- print out the result.
-
- Parameters
- ----------
- rt_mod : tvm.runtime.Module
- The runtime module.
- dev : tvm.runtime.Device
- The device type to run workload.
- input_data : Dict[str, np.ndarray]
- The input data as a dictionary.
- """
- # pylint:disable=import-outside-toplevel
- from tvm.contrib.debugger.debug_executor import create
-
- # pylint:enable=import-outside-toplevel
-
- try:
- mod = create(graph, rt_mod, dev)
- for input_name, input_value in input_data.items():
- mod.set_input(input_name, input_value)
- graph_nodes = [n["name"] for n in json.loads(graph)["nodes"]]
- graph_time = mod.run_individual(number=10, repeat=1, min_repeat_ms=5000)
-
- print("Running time of each layer:")
- print("---------------------------")
- print("|graph_nodes| = ", len(graph_nodes))
- print("|graph_time| = ", len(graph_time))
-
- for k, v in zip(graph_nodes, graph_time):
- print(k, float(v) * 1e6, "us")
- except Exception as exc: # pylint: disable=broad-except
- print(
- f"Run module f_time_per_layer via RPC failed, exception: {exc}",
- )
-
- return f_time_per_layer
-
-
def create_calculator(backend: str) -> Callable:
"""Create a function to fetch the computing result of running the given runtime module.
diff --git a/python/tvm/relax/__init__.py b/python/tvm/relax/__init__.py
index dd3245441b..471a4ba9d3 100644
--- a/python/tvm/relax/__init__.py
+++ b/python/tvm/relax/__init__.py
@@ -55,7 +55,7 @@ from .ty import (
Type,
ObjectType,
ShapeType,
- DynTensorType,
+ TensorType,
TupleType,
FuncType,
PackedFuncType,
diff --git a/python/tvm/relax/analysis/analysis.py b/python/tvm/relax/analysis/analysis.py
index edcf02bf6a..af0772ea6c 100644
--- a/python/tvm/relax/analysis/analysis.py
+++ b/python/tvm/relax/analysis/analysis.py
@@ -334,7 +334,7 @@ def post_order_visit(expr, fvisit):
Parameters
----------
- expr : tvm.relay.Expr
+ expr : tvm.relax.Expr
The input expression.
fvisit : function
diff --git a/python/tvm/relax/expr.py b/python/tvm/relax/expr.py
index 190df42860..1b2797214f 100644
--- a/python/tvm/relax/expr.py
+++ b/python/tvm/relax/expr.py
@@ -38,7 +38,7 @@ from . import _ffi_api
# It is a workaround for mypy: https://github.com/python/mypy/issues/7866#issuecomment-549454370
# This feature is not supported until python 3.10:
# https://docs.python.org/3.10/whatsnew/3.10.html#pep-613-typealias
-Expr = Union[tvm.ir.RelayExpr]
+Expr = Union[tvm.ir.RelaxExpr]
Type = Union[tvm.ir.Type] # pylint: disable=invalid-name
GlobalVar = Union[tvm.ir.GlobalVar]
diff --git a/python/tvm/relax/frontend/onnx/onnx_frontend.py b/python/tvm/relax/frontend/onnx/onnx_frontend.py
index e1f09233bc..dd4b8a4254 100644
--- a/python/tvm/relax/frontend/onnx/onnx_frontend.py
+++ b/python/tvm/relax/frontend/onnx/onnx_frontend.py
@@ -2426,7 +2426,7 @@ class MaxUnpool(OnnxOpConverter):
total_output_shape = output_shape
elif pads is not None:
- # Get pads in the proper format for relay.
+ # Get pads in the proper format
pads = _np.concatenate([[0, 0, 0, 0], list(pads)], axis=0)
pads = _np.reshape(pads, [-1, 2])
# Compute the total padding per axis.
diff --git a/python/tvm/relax/frontend/torch/fx_translator.py b/python/tvm/relax/frontend/torch/fx_translator.py
index 4c06a3e76f..724bb3fc20 100644
--- a/python/tvm/relax/frontend/torch/fx_translator.py
+++ b/python/tvm/relax/frontend/torch/fx_translator.py
@@ -290,7 +290,6 @@ class TorchFXImporter(BaseFXGraphImporter):
# input, size=None, scale_factor=None, mode='nearest', align_corners=None,
# recompute_scale_factor=None, antialias=False)
# (TODO) this is a temporary implementation for interpolate that only considers NCHW layout
- # it basically replicates the implementation in tvm.relay.frontend.pytorch
data = self.env[node.args[0]]
size = (
node.args[1]
diff --git a/python/tvm/relax/op/qdq.py b/python/tvm/relax/op/qdq.py
index 6842e9e378..6231b5f9b9 100644
--- a/python/tvm/relax/op/qdq.py
+++ b/python/tvm/relax/op/qdq.py
@@ -35,7 +35,7 @@ def quantize(data: Expr, scale: Expr, zero_point: Expr, axis: int = -1, out_dtyp
scale : tvm.relax.Expr
The output scale.
- zero_point : tvm.relay.Expr
+ zero_point : tvm.relax.Expr
The output zero_point.
axis : int
@@ -70,7 +70,7 @@ def dequantize(
scale : tvm.relax.Expr
The input scale.
- zero_point : tvm.relay.Expr
+ zero_point : tvm.relax.Expr
The input zero_point.
axis : int
diff --git a/python/tvm/relax/testing/ast_printer.py b/python/tvm/relax/testing/ast_printer.py
index 4c670bbe74..40e617b38f 100644
--- a/python/tvm/relax/testing/ast_printer.py
+++ b/python/tvm/relax/testing/ast_printer.py
@@ -248,13 +248,13 @@ class ASTPrinter(ExprFunctor):
return self.build_ast_node("PackedFuncType")
if isinstance(type_node, tvm.ir.PrimType):
return self.build_ast_node("PrimType", dtype=type_node.dtype)
- if isinstance(type_node, relax.DynTensorType):
+ if isinstance(type_node, relax.TensorType):
fields = {}
if type_node.ndim is not None:
fields["ndim"] = str(type_node.ndim)
if type_node.dtype != "":
fields["dtype"] = type_node.dtype
- return self.build_ast_node("DynTensorType", **fields)
+ return self.build_ast_node("TensorType", **fields)
if isinstance(type_node, relax.TupleType):
return self.build_ast_node(
"TupleType", fields=self.build_list(map(self.visit_type_, type_node.fields))
diff --git a/python/tvm/relax/ty.py b/python/tvm/relax/ty.py
index 05492d6a9c..b0afb06943 100644
--- a/python/tvm/relax/ty.py
+++ b/python/tvm/relax/ty.py
@@ -17,7 +17,7 @@
# pylint: disable=invalid-name, unused-import
"""The type nodes of the Relax language."""
import tvm._ffi
-from tvm.ir import Type, TensorType, TupleType, FuncType, Span
+from tvm.ir import Type, TupleType, FuncType, Span
from . import _ffi_api
@@ -47,7 +47,7 @@ class ObjectType(Type):
@tvm._ffi.register_object("relax.DynTensorType")
-class DynTensorType(Type):
+class TensorType(Type):
"""A dynamic tensor type in Relax.
This is the type assigned to tensors with a known dtype and unknown shape.
@@ -62,9 +62,7 @@ class DynTensorType(Type):
"""
def __init__(self, ndim=-1, dtype="float32", span: Span = None) -> None:
- self.__init_handle_by_constructor__(
- _ffi_api.DynTensorType, ndim, dtype, span # type: ignore
- )
+ self.__init_handle_by_constructor__(_ffi_api.TensorType, ndim, dtype, span) # type: ignore
@tvm._ffi.register_object("relax.PackedFuncType")
diff --git a/python/tvm/testing/utils.py b/python/tvm/testing/utils.py
index 9eb0c23db4..cc1b8fb1d7 100644
--- a/python/tvm/testing/utils.py
+++ b/python/tvm/testing/utils.py
@@ -1022,6 +1022,7 @@ requires_aprofile_aem_fvp = Feature(
compile_time_check=_aprofile_aem_fvp_compile_time_check,
)
+
# check cpu features
def _has_cpu_feat(features):
cpu = codegen.llvm_get_system_cpu()
@@ -2182,25 +2183,3 @@ class CompareBeforeAfter:
f"or an instance of `tvm.tir.PrimFunc`. "
f"Instead, received {type(expected)}."
)
-
-
-class _control_span_filling:
- def __init__(self, on=True):
- self._on = on
- self._pass_ctx = tvm.transform.PassContext(config={"relay.frontend.fill_span": self._on})
-
- def __enter__(self):
- self._pass_ctx.__enter__()
-
- def __exit__(self, exc_type, exc_val, exc_tb):
- self._pass_ctx.__exit__(exc_type, exc_val, exc_tb)
-
-
-class enable_span_filling(_control_span_filling):
- def __init__(self):
- super().__init__()
-
-
-class disable_span_filling(_control_span_filling):
- def __init__(self):
- super().__init__(on=False)
diff --git a/python/tvm/tir/analysis/analysis.py b/python/tvm/tir/analysis/analysis.py
index e98c176dd0..ce869d32d2 100644
--- a/python/tvm/tir/analysis/analysis.py
+++ b/python/tvm/tir/analysis/analysis.py
@@ -248,34 +248,6 @@ def undefined_vars(node: Union[Stmt, PrimExpr], defs: Optional[List[Var]] = None
return _ffi_api.UndefinedVars(node, defs) # type: ignore # pylint: disable=no-member
-def get_prim_func_arg_and_result_memory_constraints(
- func: PrimFunc, relay_func_type: Object
-) -> List[str]:
- """Returns the memory (aka storage) scope constraints for all the arguments and result
- of func. However the result will be w.r.t. the func's representation as a Relay Function
- of relay_func_type before lowering and conversion to DPS.
-
- Visible for testing.
-
- Parameters
- ----------
- func: tvm.tir.PrimFunc
- The function to retrieve constraints from.
-
- relay_func_type: tvm.relay.FuncType
- The type of the Relay Function from which the func was derived.
-
- Returns
- -------
- result: List[AnyStr]
- Memory scope constraints for funcs args and result in Relay form. The empty string
- denotes 'no constraint'.
- """
- return _ffi_api.GetPrimFuncArgAndResultMemoryConstraints( # type: ignore # pylint: disable=no-member
- func, relay_func_type
- )
-
-
def verify_well_formed(obj: Union[PrimFunc, IRModule], assert_mode: bool = True) -> bool:
"""Verify if the given TIR is well-formed. The verification includes:
- Check if expressions not contain vars that is defined outside the block.
diff --git a/python/tvm/topi/signal.py b/python/tvm/topi/signal.py
index 64a804a851..162b9a735e 100644
--- a/python/tvm/topi/signal.py
+++ b/python/tvm/topi/signal.py
@@ -35,7 +35,7 @@ def stft(
This gives frequency components of the signal as they change over time.
Parameters
----------
- data : relay.Expr
+ data : te.Tensor
Either a 1-D tensor or a 2-D batch tensor.
n_fft : int
The size of Fourier transform
@@ -43,7 +43,7 @@ def stft(
The distance between neighboring sliding window frames
win_length : int
The size of window frame and STFT filter
- window : relay.Expr
+ window : te.Tensor
A 1-D tensor window frame
normalized : bool
Whether to return the normalized STFT results
@@ -51,7 +51,7 @@ def stft(
Whether to return onesided result or fill with conjugate symmetry
Returns
-------
- output : relay.Expr
+ output : te.Tensor
Tensor containing the STFT result
Examples
--------
@@ -60,7 +60,7 @@ def stft(
data = [1, 2, 3, 4, 5, 6]
window = [4, 3, 2]
[n_fft, hop_length, win_length, normalized, onesided] = [3, 3, 3, False, True]
- relay.stft(data, n_fft, hop_length, win_length, window, normalized, onesided)
+ topi.stft(data, n_fft, hop_length, win_length, window, normalized, onesided)
-> [[[15.0000, 0.0000], [34.0000, 0.0000]], [[ 4.5000, 0.8660], [ 1.0000, -1.7321]]]
"""
@@ -136,10 +136,10 @@ def dft(
Parameters
----------
- re_data : relay.Expr
+ re_data : te.Tensor
N-D tensor, real part of the input signal.
- im_data : relay.Expr
+ im_data : te.Tensor
N-D tensor, imaginary part of the input signal.
If the signal is real, then the values of this tensor are zeros.
@@ -148,9 +148,9 @@ def dft(
Returns
-------
- re_output : relay.Expr
+ re_output : te.Tensor
The Fourier Transform of the input (Real part).
- im_output : relay.Expr
+ im_output : te.Tensor
The Fourier Transform of the input (Imaginary part).
"""
diff --git a/python/tvm/topi/sparse_reshape.py b/python/tvm/topi/sparse_reshape.py
index b25bd854a7..6bf3857e2e 100644
--- a/python/tvm/topi/sparse_reshape.py
+++ b/python/tvm/topi/sparse_reshape.py
@@ -32,19 +32,19 @@ def sparse_reshape(
Parameters
----------
- sparse_indices : relay.Expr
+ sparse_indices : te.Expr
A 2-D tensor[N, n_dim] of integers containing location of sparse values, where N is the
number of sparse values and n_dim is the number of dimensions of the dense_shape
- prev_shape : relay.Expr
+ prev_shape : te.Expr
A 1-D tensor containing the previous shape of the dense tensor
- new_shape : relay.Expr
+ new_shape : te.Expr
A 1-D tensor containing the new shape of the dense tensor
Returns
-------
- result: relay.Expr
+ result: te.Expr
Output tensor.
Examples
@@ -58,7 +58,7 @@ def sparse_reshape(
[1, 2, 3]]
prev_shape = [2, 3, 4]
new_shape = [9, -1]
- new_sparse_indices, new_shape = relay.sparse_reshape(
+ new_sparse_indices, new_shape = topi.sparse_reshape(
sparse_indices, prev_shape, new_shape)
new_sparse_indices = [[0, 0],
[0, 1],
diff --git a/python/tvm/topi/testing/one_hot.py b/python/tvm/topi/testing/one_hot.py
index 0c4b0600fa..4b9c3432da 100644
--- a/python/tvm/topi/testing/one_hot.py
+++ b/python/tvm/topi/testing/one_hot.py
@@ -48,7 +48,7 @@ def one_hot(indices, on_value, off_value, depth, axis, dtype):
Returns
-------
- ret : relay.Expr
+ ret : tvm.te.Tensor
The one-hot tensor.
"""
oshape = []
diff --git a/python/tvm/topi/transform.py b/python/tvm/topi/transform.py
index 6101c2e57d..b8605aa58a 100644
--- a/python/tvm/topi/transform.py
+++ b/python/tvm/topi/transform.py
@@ -793,12 +793,12 @@ def one_hot(indices, on_value, off_value, depth, axis, dtype):
axis : int
Axis to fill.
- dtype : relay.DataType
+ dtype : str
Data type of the output tensor.
Returns
-------
- ret : relay.Expr
+ ret : tvm.te.Tensor
The one-hot tensor.
Examples
@@ -807,7 +807,7 @@ def one_hot(indices, on_value, off_value, depth, axis, dtype):
indices = [0, 1, 2]
- relay.one_hot(indices, 3) =
+ topi.one_hot(indices, 3) =
[[1, 0, 0],
[0, 1, 0],
[0, 0, 1]]
@@ -823,15 +823,15 @@ def unravel_index(indices, shape):
Parameters
----------
- indices : relay.Expr
+ indices : tvm.te.Tensor
An integer array containing indices.
- shape : relay.Expr
+ shape : tvm.te.Tensor
The shape of the array.
Returns
-------
- result : relay.Expr
+ result : tvm.te.Tensor
The tuple of coordinate arrays.
"""
@@ -874,10 +874,10 @@ def matrix_set_diag(data, diagonal, k=0, align="RIGHT_LEFT"):
Parameters
----------
- data : relay.Expr
+ data : tvm.te.Tensor
Input Tensor.
- diagonal : relay.Expr
+ diagonal : tvm.te.Tensor
Values to be filled in the diagonal.
k : int or tuple of int, optional
@@ -897,7 +897,7 @@ def matrix_set_diag(data, diagonal, k=0, align="RIGHT_LEFT"):
Returns
-------
- result : relay.Expr
+ result : tvm.te.Tensor
New tensor with given diagonal values.
Examples
@@ -964,7 +964,7 @@ def sliding_window(data, axis, window_shape, strides):
Parameters
----------
- data : relay.Expr
+ data : tvm.te.Tensor
The input data to the operator.
axis : int
@@ -983,7 +983,7 @@ def sliding_window(data, axis, window_shape, strides):
Returns
-------
- result : relay.Expr
+ result : tvm.te.Tensor
The resulting tensor.
"""
return cpp.sliding_window(data, axis, window_shape, strides)
@@ -1011,7 +1011,7 @@ def trilu(data, k, upper):
Returns
-------
- ret : relay.Expr
+ ret : tvm.te.Tensor
The new tensor with appropriate diagonals set to zero.
Examples
@@ -1022,7 +1022,7 @@ def trilu(data, k, upper):
[3, 4, 5],
[6, 7, 8]]
- relay.trilu(x, True, 0) =
+ topi.trilu(x, True, 0) =
[[0, 1, 2],
[0, 4, 5],
[0, 0, 8]]
diff --git a/src/contrib/msc/core/ir/graph_builder.cc b/src/contrib/msc/core/ir/graph_builder.cc
index 27115cb130..1cc0c4af6a 100644
--- a/src/contrib/msc/core/ir/graph_builder.cc
+++ b/src/contrib/msc/core/ir/graph_builder.cc
@@ -844,7 +844,7 @@ void RelaxWeightsExtractor::VisitExpr_(const relax::CallNode* op) {
}
void RelayFuncAttrGetter::VisitExpr_(const relay::CallNode* op) {
- RelayExprVisitor::VisitExpr_(op);
+ RelaxExprVisitor::VisitExpr_(op);
if (op->attrs.defined()) {
Map<String, String> attrs;
AttrGetter getter(&attrs);
@@ -1107,7 +1107,7 @@ void RelayGraphBuilder::VisitExpr_(const relay::FunctionNode* op) {
if (name_opt.defined()) {
StartFuncScope(SpanUtils::GetAttr(op->span, msc_attr::kName));
}
- RelayExprVisitor::VisitExpr_(op);
+ RelaxExprVisitor::VisitExpr_(op);
if (HasFuncScope()) {
AddNode(GetRef<relay::Function>(op));
EndFuncScope();
@@ -1120,7 +1120,7 @@ void RelayGraphBuilder::VisitExpr_(const relay::CallNode* op) {
if (name_opt.defined()) {
for (size_t i = 0; i < op->args.size(); i++) {
if (!expr_tensor_map_.count(op->args[i])) {
- RelayExprVisitor::VisitExpr(op->args[i]);
+ RelaxExprVisitor::VisitExpr(op->args[i]);
}
ICHECK(expr_tensor_map_.count(op->args[i]))
<< "Can not find argument " << relay::PrettyPrint(op->args[i]);
@@ -1128,7 +1128,7 @@ void RelayGraphBuilder::VisitExpr_(const relay::CallNode* op) {
}
}
}
- RelayExprVisitor::VisitExpr_(op);
+ RelaxExprVisitor::VisitExpr_(op);
if (!HasFuncScope() && op->op->IsInstance<OpNode>()) {
try {
AddNode(GetRef<relay::Call>(op));
@@ -1144,12 +1144,12 @@ void RelayGraphBuilder::VisitExpr_(const relay::CallNode* op) {
}
void RelayGraphBuilder::VisitExpr_(const relay::TupleNode* val) {
- RelayExprVisitor::VisitExpr_(val);
+ RelaxExprVisitor::VisitExpr_(val);
AddNode(GetRef<relay::Tuple>(val));
}
void RelayGraphBuilder::VisitExpr_(const relay::TupleGetItemNode* val) {
- RelayExprVisitor::VisitExpr_(val);
+ RelaxExprVisitor::VisitExpr_(val);
AddNode(GetRef<relay::TupleGetItem>(val));
}
diff --git a/src/contrib/msc/core/ir/graph_builder.h b/src/contrib/msc/core/ir/graph_builder.h
index 269a8a213c..9fd855455c 100644
--- a/src/contrib/msc/core/ir/graph_builder.h
+++ b/src/contrib/msc/core/ir/graph_builder.h
@@ -48,9 +48,9 @@ namespace tvm {
namespace contrib {
namespace msc {
-using Expr = tvm::RelayExpr;
+using Expr = tvm::RelaxExpr;
using RelaxExprVisitor = tvm::relax::ExprVisitor;
-using RelayExprVisitor = tvm::relay::ExprVisitor;
+using RelaxExprVisitor = tvm::relay::ExprVisitor;
using tvm::runtime::NDArray;
@@ -348,7 +348,7 @@ class RelaxWeightsExtractor : public RelaxExprVisitor {
IRModule ref_module_;
};
-class RelayFuncAttrGetter : public RelayExprVisitor {
+class RelayFuncAttrGetter : public RelaxExprVisitor {
public:
/*! \brief Get the attributes as Map<String, String>*/
Map<String, String> GetAttrs(const Expr& expr) {
@@ -381,7 +381,7 @@ class RelayFuncScope {
Array<String> func_weights_;
};
-class RelayGraphBuilder : public RelayExprVisitor {
+class RelayGraphBuilder : public RelaxExprVisitor {
public:
/*!
* \brief The constructor of RelayGraphBuilder
@@ -391,7 +391,7 @@ class RelayGraphBuilder : public RelayExprVisitor {
*/
explicit RelayGraphBuilder(const IRModule& ref_module, const String& name,
const std::string& options = "")
- : RelayExprVisitor() {
+ : RelaxExprVisitor() {
ref_module_ = ref_module;
if (options.size() > 0) {
std::istringstream is(options);
@@ -444,7 +444,7 @@ class RelayGraphBuilder : public RelayExprVisitor {
std::stack<RelayFuncScope> func_scopes_;
};
-class RelayWeightsExtractor : public RelayExprVisitor {
+class RelayWeightsExtractor : public RelaxExprVisitor {
public:
/*! \brief Visit the constant and save weights*/
Map<MSCTensor, NDArray> GetWeights(const relay::Function& func);
diff --git a/src/contrib/msc/core/transform/layout_utils.h b/src/contrib/msc/core/transform/layout_utils.h
index e7781a95a8..787c73cc84 100644
--- a/src/contrib/msc/core/transform/layout_utils.h
+++ b/src/contrib/msc/core/transform/layout_utils.h
@@ -37,7 +37,7 @@ namespace tvm {
namespace contrib {
namespace msc {
-using Expr = tvm::RelayExpr;
+using Expr = tvm::RelaxExpr;
using namespace tvm::relax;
/*!
diff --git a/src/contrib/msc/core/transform/rewrite_utils.h b/src/contrib/msc/core/transform/rewrite_utils.h
index 2693a6ccd2..307581b274 100644
--- a/src/contrib/msc/core/transform/rewrite_utils.h
+++ b/src/contrib/msc/core/transform/rewrite_utils.h
@@ -37,7 +37,7 @@ namespace tvm {
namespace contrib {
namespace msc {
-using Expr = tvm::RelayExpr;
+using Expr = tvm::RelaxExpr;
using namespace tvm::relax;
/*!
diff --git a/src/contrib/msc/core/transform/set_expr_name.cc b/src/contrib/msc/core/transform/set_expr_name.cc
index 163d868335..0418513c14 100644
--- a/src/contrib/msc/core/transform/set_expr_name.cc
+++ b/src/contrib/msc/core/transform/set_expr_name.cc
@@ -339,9 +339,9 @@ namespace relay {
/*!
* \brief Name setter for Relay
*/
-class RelayExprNameSetter : public ExprVisitor {
+class RelaxExprNameSetter : public ExprVisitor {
public:
- explicit RelayExprNameSetter(const IRModule& ref_module) : ref_module_(ref_module) {}
+ explicit RelaxExprNameSetter(const IRModule& ref_module) : ref_module_(ref_module) {}
void VisitExpr_(const ConstantNode* op) final {
ExprVisitor::VisitExpr_(op);
@@ -460,16 +460,16 @@ class RelayExprNameSetter : public ExprVisitor {
IRModule ref_module_;
}; // class ExprNameSetter
-void SetRelayExprName(const IRModule& ref_module, const Expr& e) {
- RelayExprNameSetter(ref_module).VisitExpr(e);
+void SetRelaxExprName(const IRModule& ref_module, const Expr& e) {
+ RelaxExprNameSetter(ref_module).VisitExpr(e);
}
/*!
* \brief Name binder for Relay
*/
-class RelayExprNameBinder : public ExprVisitor {
+class RelaxExprNameBinder : public ExprVisitor {
public:
- explicit RelayExprNameBinder(const String& name_key, const String& seperator)
+ explicit RelaxExprNameBinder(const String& name_key, const String& seperator)
: name_key_(name_key), seperator_(seperator) {}
void VisitExpr_(const ConstantNode* op) final {
@@ -523,33 +523,33 @@ class RelayExprNameBinder : public ExprVisitor {
String seperator_;
}; // class ExprNameBinder
-void BindRelayExprName(const Expr& e, const String& name_key, const String& seperator) {
- RelayExprNameBinder(name_key, seperator).VisitExpr(e);
+void BindRelaxExprName(const Expr& e, const String& name_key, const String& seperator) {
+ RelaxExprNameBinder(name_key, seperator).VisitExpr(e);
}
namespace transform {
-Pass SetRelayExprName(const String& entry_name) {
+Pass SetRelaxExprName(const String& entry_name) {
runtime::TypedPackedFunc<IRModule(IRModule, PassContext)> pass_func = [=](IRModule m,
PassContext pc) {
- relay::SetRelayExprName(m, m->Lookup(entry_name));
+ relay::SetRelaxExprName(m, m->Lookup(entry_name));
return m;
};
- return CreateModulePass(pass_func, 0, "SetRelayExprName", {});
+ return CreateModulePass(pass_func, 0, "SetRelaxExprName", {});
}
-TVM_REGISTER_GLOBAL("relay._transform.SetRelayExprName").set_body_typed(SetRelayExprName);
+TVM_REGISTER_GLOBAL("relay._transform.SetRelaxExprName").set_body_typed(SetRelaxExprName);
-Pass BindRelayExprName(const String& name_key, const String& seperator, const String& entry_name) {
+Pass BindRelaxExprName(const String& name_key, const String& seperator, const String& entry_name) {
runtime::TypedPackedFunc<IRModule(IRModule, PassContext)> pass_func = [=](IRModule m,
PassContext pc) {
- relay::BindRelayExprName(m->Lookup(entry_name), name_key, seperator);
+ relay::BindRelaxExprName(m->Lookup(entry_name), name_key, seperator);
return m;
};
- return CreateModulePass(pass_func, 0, "BindRelayExprName", {});
+ return CreateModulePass(pass_func, 0, "BindRelaxExprName", {});
}
-TVM_REGISTER_GLOBAL("relay._transform.BindRelayExprName").set_body_typed(BindRelayExprName);
+TVM_REGISTER_GLOBAL("relay._transform.BindRelaxExprName").set_body_typed(BindRelaxExprName);
} // namespace transform
} // namespace relay
diff --git a/src/contrib/msc/core/utils.h b/src/contrib/msc/core/utils.h
index 7fb9c87a99..9bcaba2a27 100644
--- a/src/contrib/msc/core/utils.h
+++ b/src/contrib/msc/core/utils.h
@@ -37,7 +37,7 @@ namespace tvm {
namespace contrib {
namespace msc {
-using Expr = tvm::RelayExpr;
+using Expr = tvm::RelaxExpr;
using RelaxCall = tvm::relax::Call;
using RelayCall = tvm::relay::Call;
diff --git a/src/ir/affine_type.cc b/src/ir/affine_type.cc
deleted file mode 100644
index 87235fe20a..0000000000
--- a/src/ir/affine_type.cc
+++ /dev/null
@@ -1,82 +0,0 @@
-/*
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements. See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership. The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * "License"); you may not use this file except in compliance
- * with the License. You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations
- * under the License.
- */
-
-/*!
- * \file src/ir/affine_type.cc
- * \brief The Type information for quantized nodes.
- */
-#include <tvm/ir/affine_type.h>
-#include <tvm/runtime/registry.h>
-#include <tvm/tir/op.h>
-
-namespace tvm {
-
-using tvm::ReprPrinter;
-using namespace tvm::runtime;
-
-TensorAffineType::TensorAffineType(RelayExpr scale, RelayExpr zero_point, DataType dtype,
- int axis) {
- ObjectPtr<TensorAffineTypeNode> n = make_object<TensorAffineTypeNode>();
- n->scale = std::move(scale);
- n->zero_point = std::move(zero_point);
- n->dtype = std::move(dtype);
- n->axis = std::move(axis);
- data_ = std::move(n);
-}
-
-TVM_REGISTER_NODE_TYPE(TensorAffineTypeNode);
-
-TVM_REGISTER_GLOBAL("ir.TensorAffineType")
- .set_body_typed([](RelayExpr scale, RelayExpr zero_point, DataType dtype, int axis) {
- return TensorAffineType(scale, zero_point, dtype, axis);
- });
-
-TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable)
- .set_dispatch<TensorAffineTypeNode>([](const ObjectRef& ref, ReprPrinter* p) {
- auto* node = static_cast<const TensorAffineTypeNode*>(ref.get());
- p->stream << "TensorAffineType(" << node->scale << ", " << node->zero_point << ", "
- << node->dtype << ", " << node->axis << ")";
- });
-
-TupleAffineType::TupleAffineType(Array<TensorAffineType> types) {
- ObjectPtr<TupleAffineTypeNode> n = make_object<TupleAffineTypeNode>();
- n->types = std::move(types);
- data_ = std::move(n);
-}
-
-TVM_REGISTER_NODE_TYPE(TupleAffineTypeNode);
-
-TVM_REGISTER_GLOBAL("ir.TupleAffineType").set_body_typed([](Array<TensorAffineType> types) {
- return TupleAffineType(types);
-});
-
-TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable)
- .set_dispatch<TupleAffineTypeNode>([](const ObjectRef& ref, ReprPrinter* p) {
- auto* node = static_cast<const TupleAffineTypeNode*>(ref.get());
- p->stream << "TupleAffineType([";
- for (size_t i = 0; i < node->types.size(); ++i) {
- p->stream << node->types[i];
- if (i < node->types.size() - 1) {
- p->stream << ", ";
- }
- }
- p->stream << "])";
- });
-
-} // namespace tvm
diff --git a/src/ir/diagnostic.cc b/src/ir/diagnostic.cc
index 8eeb4b3e6f..2c9d7e8e93 100644
--- a/src/ir/diagnostic.cc
+++ b/src/ir/diagnostic.cc
@@ -263,8 +263,6 @@ void ReportAt(const DiagnosticContext& context, std::ostream& out, const Span& s
// If the source name is not in the current source map, sources were not annotated.
if (it == context->module->source_map->source_map.end()) {
LOG(FATAL) << "The source maps are not populated for this module. "
- << "Please use `tvm.relay.transform.AnnotateSpans` to attach source maps for error "
- "reporting.\n"
<< "Error: " << diagnostic->message;
}
diff --git a/src/ir/function.cc b/src/ir/function.cc
index 3c92787530..1afe195a95 100644
--- a/src/ir/function.cc
+++ b/src/ir/function.cc
@@ -48,11 +48,6 @@ TVM_REGISTER_GLOBAL("ir.BaseFuncWithAttrs")
if (func->IsInstance<tir::PrimFuncNode>()) {
return WithAttrs(Downcast<tir::PrimFunc>(std::move(func)), attr_map);
}
- if (const auto* f = runtime::Registry::Get("relay.ir.FuncWithAttrs")) {
- if (Optional<BaseFunc> ret = (*f)(func, attr_map)) {
- return ret.value();
- }
- }
if (const auto* f = runtime::Registry::Get("relax.FuncWithAttrs")) {
if (Optional<BaseFunc> ret = (*f)(func, attr_map)) {
return ret.value();
diff --git a/src/ir/memory_pools.cc b/src/ir/memory_pools.cc
deleted file mode 100644
index 912ad80b3c..0000000000
--- a/src/ir/memory_pools.cc
+++ /dev/null
@@ -1,242 +0,0 @@
-/*
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements. See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership. The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * "License"); you may not use this file except in compliance
- * with the License. You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations
- * under the License.
- */
-
-/*!
- * \file src/ir/memory_pools.cc
- * \brief The object definition for relay.build argument type of memory pools
- */
-
-#include <tvm/ir/memory_pools.h>
-
-namespace tvm {
-
-PoolInfo::PoolInfo(String pool_name, Integer size_hint_bytes, Integer clock_frequency_hz,
- Integer read_bandwidth_bytes_per_cycle, Integer write_bandwidth_bytes_per_cycle,
- Integer read_latency_cycles, Integer write_latency_cycles,
- Map<Target, Integer> target_burst_bytes, Bool is_internal) {
- auto poolinfo_node = make_object<PoolInfoNode>();
- poolinfo_node->pool_name = pool_name;
- poolinfo_node->size_hint_bytes = size_hint_bytes;
- poolinfo_node->clock_frequency_hz = clock_frequency_hz;
- poolinfo_node->read_bandwidth_bytes_per_cycle = read_bandwidth_bytes_per_cycle;
- poolinfo_node->write_bandwidth_bytes_per_cycle = write_bandwidth_bytes_per_cycle;
- poolinfo_node->read_latency_cycles = read_latency_cycles;
- poolinfo_node->write_latency_cycles = write_latency_cycles;
- poolinfo_node->target_burst_bytes = target_burst_bytes;
- poolinfo_node->is_internal = is_internal;
- data_ = std::move(poolinfo_node);
-}
-
-TVM_REGISTER_NODE_TYPE(PoolInfoNode);
-
-TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable)
- .set_dispatch<PoolInfoNode>([](const ObjectRef& ref, ReprPrinter* p) {
- auto* node = static_cast<const PoolInfoNode*>(ref.get());
- p->stream << "PoolInfoNode(\n"
- << " pool_name=" << node->pool_name
- << ",\n size_hint_bytes=" << node->size_hint_bytes
- << ",\n clock_frequency_hz=" << node->clock_frequency_hz
- << ",\n read_bandwidth_bytes_per_cycle=" << node->read_bandwidth_bytes_per_cycle
- << ",\n write_bandwidth_bytes_per_cycle=" << node->write_bandwidth_bytes_per_cycle
- << ",\n read_latency_cycles=" << node->read_latency_cycles
- << ",\n write_latency_cycles=" << node->write_latency_cycles
- << ",\n target_burst_bytes=" << node->target_burst_bytes << ")";
- });
-
-PoolInfoProperties::PoolInfoProperties(Integer size_hint_bytes, Integer clock_frequency_hz,
- Integer read_bandwidth_bytes_per_cycle,
- Integer write_bandwidth_bytes_per_cycle,
- Integer read_latency_cycles, Integer write_latency_cycles,
- Map<Target, Integer> target_burst_bytes, Bool is_internal) {
- auto poolinfo_properties_node = make_object<PoolInfoPropertiesNode>();
- poolinfo_properties_node->size_hint_bytes = size_hint_bytes;
- poolinfo_properties_node->clock_frequency_hz = clock_frequency_hz;
- poolinfo_properties_node->read_bandwidth_bytes_per_cycle = read_bandwidth_bytes_per_cycle;
- poolinfo_properties_node->write_bandwidth_bytes_per_cycle = write_bandwidth_bytes_per_cycle;
- poolinfo_properties_node->read_latency_cycles = read_latency_cycles;
- poolinfo_properties_node->write_latency_cycles = write_latency_cycles;
- poolinfo_properties_node->target_burst_bytes = target_burst_bytes;
- poolinfo_properties_node->is_internal = is_internal;
- data_ = std::move(poolinfo_properties_node);
-}
-
-TVM_REGISTER_NODE_TYPE(PoolInfoPropertiesNode);
-TVM_REGISTER_GLOBAL("ir.PoolInfoProperties")
- .set_body_typed([](Integer size_hint_bytes, Integer clock_frequency_hz,
- Integer read_bandwidth_bytes_per_cycle,
- Integer write_bandwidth_bytes_per_cycle, Integer read_latency_cycles,
- Integer write_latency_cycles, Map<Target, Integer> target_burst_bytes) {
- return PoolInfoProperties(size_hint_bytes, clock_frequency_hz, read_bandwidth_bytes_per_cycle,
- write_bandwidth_bytes_per_cycle, read_latency_cycles,
- write_latency_cycles, target_burst_bytes, Bool(false));
- });
-
-TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable)
- .set_dispatch<PoolInfoPropertiesNode>([](const ObjectRef& ref, ReprPrinter* p) {
- auto* node = static_cast<const PoolInfoPropertiesNode*>(ref.get());
- p->stream << "PoolInfoPropertiesNode(\n"
- << " size_hint_bytes=" << node->size_hint_bytes
- << ",\n clock_frequency_hz=" << node->clock_frequency_hz
- << ",\n read_bandwidth_bytes_per_cycle=" << node->read_bandwidth_bytes_per_cycle
- << ",\n write_bandwidth_bytes_per_cycle=" << node->write_bandwidth_bytes_per_cycle
- << ",\n read_latency_cycles=" << node->read_latency_cycles
- << ",\n write_latency_cycles=" << node->write_latency_cycles
- << ",\n target_burst_bytes=" << node->target_burst_bytes << ")";
- });
-
-WorkspacePoolInfo::WorkspacePoolInfo(String pool_name, Array<Target> targets,
- PoolInfoProperties properties) {
- auto poolinfo_node = make_object<WorkspacePoolInfoNode>();
- poolinfo_node->pool_name = pool_name;
- poolinfo_node->size_hint_bytes = properties->size_hint_bytes;
- poolinfo_node->targets = targets;
- poolinfo_node->clock_frequency_hz = properties->clock_frequency_hz;
- poolinfo_node->read_bandwidth_bytes_per_cycle = properties->read_bandwidth_bytes_per_cycle;
- poolinfo_node->write_bandwidth_bytes_per_cycle = properties->write_bandwidth_bytes_per_cycle;
- poolinfo_node->read_latency_cycles = properties->read_latency_cycles;
- poolinfo_node->write_latency_cycles = properties->write_latency_cycles;
- poolinfo_node->target_burst_bytes = properties->target_burst_bytes;
- poolinfo_node->is_internal = properties->is_internal;
- data_ = std::move(poolinfo_node);
-}
-
-TVM_REGISTER_NODE_TYPE(WorkspacePoolInfoNode);
-TVM_REGISTER_GLOBAL("ir.WorkspacePoolInfo")
- .set_body_typed([](String pool_name, Array<Target> targets, PoolInfoProperties properties) {
- return WorkspacePoolInfo(pool_name, targets, properties);
- });
-
-TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable)
- .set_dispatch<WorkspacePoolInfoNode>([](const ObjectRef& ref, ReprPrinter* p) {
- auto* node = static_cast<const WorkspacePoolInfoNode*>(ref.get());
- p->stream << "WorkspacePoolInfoNode(\n"
- << " pool_name=" << node->pool_name << ",\n targets=" << node->targets
- << ",\n size_hint_bytes=" << node->size_hint_bytes
- << ",\n clock_frequency_hz=" << node->clock_frequency_hz
- << ",\n read_bandwidth_bytes_per_cycle=" << node->read_bandwidth_bytes_per_cycle
- << ",\n write_bandwidth_bytes_per_cycle=" << node->write_bandwidth_bytes_per_cycle
- << ",\n read_latency_cycles=" << node->read_latency_cycles
- << ",\n write_latency_cycles=" << node->write_latency_cycles
- << ",\n target_burst_bytes=" << node->target_burst_bytes
- << ",\n is_internal=" << node->is_internal << ")"
- << "\n";
- });
-
-ConstantInfo::ConstantInfo(String name_hint, Integer byte_offset, runtime::NDArray data) {
- auto constant_info_node = make_object<ConstantInfoNode>();
- constant_info_node->name_hint = name_hint;
- constant_info_node->byte_offset = byte_offset;
- constant_info_node->data = data;
- data_ = std::move(constant_info_node);
-}
-
-TVM_REGISTER_NODE_TYPE(ConstantInfoNode);
-TVM_REGISTER_GLOBAL("ir.ConstantInfo")
- .set_body_typed([](String name_hint, Integer byte_offset, runtime::NDArray data) {
- return ConstantInfo(name_hint, byte_offset, data);
- });
-TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable)
- .set_dispatch<ConstantInfoNode>([](const ObjectRef& ref, ReprPrinter* p) {
- auto* node = static_cast<const ConstantInfoNode*>(ref.get());
- p->stream << "ConstantInfoNode(\n"
- << "name_hint=" << node->name_hint << ",\n byte_offset=" << node->byte_offset
- << ",\n data=" << node->data << ")";
- });
-
-ConstantPoolInfo::ConstantPoolInfo(String pool_name, Array<Target> targets,
- Array<ConstantInfo> constant_info_array,
- PoolInfoProperties properties) {
- auto constant_poolinfo_node = make_object<ConstantPoolInfoNode>();
- constant_poolinfo_node->pool_name = pool_name;
- constant_poolinfo_node->constant_info_array = constant_info_array;
- constant_poolinfo_node->targets = targets;
-
- constant_poolinfo_node->size_hint_bytes = properties->size_hint_bytes;
- constant_poolinfo_node->clock_frequency_hz = properties->clock_frequency_hz;
- constant_poolinfo_node->read_bandwidth_bytes_per_cycle =
- properties->read_bandwidth_bytes_per_cycle;
- constant_poolinfo_node->write_bandwidth_bytes_per_cycle =
- properties->write_bandwidth_bytes_per_cycle;
- constant_poolinfo_node->read_latency_cycles = properties->read_latency_cycles;
- constant_poolinfo_node->write_latency_cycles = properties->write_latency_cycles;
- constant_poolinfo_node->target_burst_bytes = properties->target_burst_bytes;
- constant_poolinfo_node->is_internal = properties->is_internal;
- data_ = std::move(constant_poolinfo_node);
-}
-
-TVM_REGISTER_NODE_TYPE(ConstantPoolInfoNode);
-TVM_REGISTER_GLOBAL("ir.ConstantPoolInfo")
- .set_body_typed([](String pool_name, Array<Target> targets,
- Array<ConstantInfo> constant_info_array, PoolInfoProperties properties) {
- return ConstantPoolInfo(pool_name, targets, constant_info_array, properties);
- });
-
-TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable)
- .set_dispatch<ConstantPoolInfoNode>([](const ObjectRef& ref, ReprPrinter* p) {
- auto* node = static_cast<const ConstantPoolInfoNode*>(ref.get());
- p->stream << "ConstantPoolInfoNode(\n"
- << " pool_name=" << node->pool_name << ",\n targets=" << node->targets
- << ",\n constant_info_array=" << node->constant_info_array
- << ",\n size_hint_bytes=" << node->size_hint_bytes
- << ",\n clock_frequency_hz=" << node->clock_frequency_hz
- << ",\n read_bandwidth_bytes_per_cycle=" << node->read_bandwidth_bytes_per_cycle
- << ",\n write_bandwidth_bytes_per_cycle=" << node->write_bandwidth_bytes_per_cycle
- << ",\n read_latency_cycles=" << node->read_latency_cycles
- << ",\n write_latency_cycles=" << node->write_latency_cycles
- << ",\n target_burst_bytes=" << node->target_burst_bytes
- << ",\n is_internal=" << node->is_internal << ")"
- << "\n";
- });
-
-WorkspaceMemoryPools::WorkspaceMemoryPools(Array<PoolInfo> pools) {
- auto workspace_memory_pools_node = make_object<WorkspaceMemoryPoolsNode>();
- workspace_memory_pools_node->pools = pools;
- data_ = std::move(workspace_memory_pools_node);
-}
-
-TVM_REGISTER_NODE_TYPE(WorkspaceMemoryPoolsNode);
-TVM_REGISTER_GLOBAL("ir.WorkspaceMemoryPools").set_body_typed([](Array<PoolInfo> pools) {
- return WorkspaceMemoryPools(pools);
-});
-
-TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable)
- .set_dispatch<WorkspaceMemoryPoolsNode>([](const ObjectRef& ref, ReprPrinter* p) {
- auto* node = static_cast<const WorkspaceMemoryPoolsNode*>(ref.get());
- p->stream << "WorkspaceMemoryPoolsNode(\n"
- << "pools=" << node->pools << ")";
- });
-
-ConstantMemoryPools::ConstantMemoryPools(Array<ConstantPoolInfo> pools) {
- auto constant_memory_pools_node = make_object<ConstantMemoryPoolsNode>();
- constant_memory_pools_node->pools = pools;
- data_ = std::move(constant_memory_pools_node);
-}
-
-TVM_REGISTER_NODE_TYPE(ConstantMemoryPoolsNode);
-TVM_REGISTER_GLOBAL("ir.ConstantMemoryPools").set_body_typed([](Array<ConstantPoolInfo> pools) {
- return ConstantMemoryPools(pools);
-});
-
-TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable)
- .set_dispatch<ConstantMemoryPoolsNode>([](const ObjectRef& ref, ReprPrinter* p) {
- auto* node = static_cast<const ConstantMemoryPoolsNode*>(ref.get());
- p->stream << "ConstantMemoryPoolsNode(\n"
- << "pools=" << node->pools << ")";
- });
-} // namespace tvm
diff --git a/src/ir/module.cc b/src/ir/module.cc
index b27dec5719..54a10b84cf 100644
--- a/src/ir/module.cc
+++ b/src/ir/module.cc
@@ -211,7 +211,7 @@ IRModule IRModuleNode::ShallowCopy() {
return IRModule(this->functions, this->source_map, this->attrs, this->global_infos);
}
-IRModule IRModule::FromExpr(const RelayExpr& expr,
+IRModule IRModule::FromExpr(const RelaxExpr& expr,
const tvm::Map<GlobalVar, BaseFunc>& global_funcs) {
auto mod = IRModule(global_funcs);
String gv_name;
@@ -266,10 +266,7 @@ TVM_REGISTER_GLOBAL("ir.Module_Clone").set_body_typed([](IRModule mod) -> IRModu
TVM_REGISTER_GLOBAL("ir.Module_Add")
.set_body_typed([](IRModule mod, GlobalVar var, ObjectRef val, bool update) -> IRModule {
- ICHECK(val->IsInstance<RelayExprNode>());
- if (const auto* f = runtime::Registry::Get("relay.ir.IRModuleAdd")) {
- return (*f)(mod, var, val, update);
- }
+ ICHECK(val->IsInstance<RelaxExprNode>());
mod->Add(var, Downcast<BaseFunc>(val), update);
return mod;
});
diff --git a/src/ir/source_map.cc b/src/ir/source_map.cc
index 721a30affa..339b08d6ad 100644
--- a/src/ir/source_map.cc
+++ b/src/ir/source_map.cc
@@ -28,8 +28,6 @@
namespace tvm {
-TVM_REGISTER_PASS_CONFIG_OPTION("relay.frontend.fill_span", Bool);
-
ObjectPtr<Object> GetSourceNameNode(const String& name) {
// always return pointer as the reference can change as map re-allocate.
// or use another level of indirection by creating a unique_ptr
diff --git a/src/ir/tensor_type.cc b/src/ir/tensor_type.cc
deleted file mode 100644
index 0fab0acb89..0000000000
--- a/src/ir/tensor_type.cc
+++ /dev/null
@@ -1,66 +0,0 @@
-/*
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements. See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership. The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * "License"); you may not use this file except in compliance
- * with the License. You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations
- * under the License.
- */
-
-/*!
- * \file src/ir/tensor_type.cc
- * \brief The type system AST nodes of Relay.
- */
-#include <tvm/ir/tensor_type.h>
-#include <tvm/runtime/registry.h>
-#include <tvm/tir/op.h>
-
-namespace tvm {
-
-using tvm::ReprPrinter;
-using namespace tvm::runtime;
-
-TensorType::TensorType(Array<PrimExpr> shape, DataType dtype) {
- ObjectPtr<TensorTypeNode> n = make_object<TensorTypeNode>();
- n->shape = std::move(shape);
- n->dtype = std::move(dtype);
- data_ = std::move(n);
-}
-
-TensorType TensorType::Scalar(DataType dtype) { return TensorType({}, dtype); }
-
-PrimExpr TensorTypeNode::Size() const {
- if (shape.size() == 0) {
- return tir::make_const(DataType::Int(64), 1);
- }
-
- PrimExpr size = shape[0];
- for (size_t i = 1; i < shape.size(); ++i) {
- size *= shape[i];
- }
- return size;
-}
-
-TVM_REGISTER_NODE_TYPE(TensorTypeNode);
-
-TVM_REGISTER_GLOBAL("ir.TensorType").set_body_typed([](Array<PrimExpr> shape, DataType dtype) {
- return TensorType(shape, dtype);
-});
-
-TVM_STATIC_IR_FUNCTOR(ReprPrinter, vtable)
- .set_dispatch<TensorTypeNode>([](const ObjectRef& ref, ReprPrinter* p) {
- auto* node = static_cast<const TensorTypeNode*>(ref.get());
- p->stream << "TensorType(" << node->shape << ", " << node->dtype << ")";
- });
-
-} // namespace tvm
diff --git a/src/ir/transform.cc b/src/ir/transform.cc
index f0b879acbc..0ed80310eb 100644
--- a/src/ir/transform.cc
+++ b/src/ir/transform.cc
@@ -465,8 +465,6 @@ Pass GetPass(const String& pass_name) {
if (pass_name.operator std::string().find("transform.") != std::string::npos) {
f = Registry::Get(pass_name);
} else if ((f = Registry::Get("transform." + pass_name))) {
- // pass
- } else if ((f = Registry::Get("relay._transform." + pass_name))) {
}
ICHECK(f != nullptr) << "Cannot use " << pass_name << " to create the pass";
return (*f)();
@@ -686,11 +684,6 @@ TVM_REGISTER_GLOBAL("transform.OverrideInstruments")
Pass PrintIR(String header, bool show_meta_data) {
auto pass_func = [header, show_meta_data](IRModule mod, const PassContext& ctx) {
- if (const auto* f = runtime::Registry::Get("relay.ir.PrintIR")) {
- if ((*f)(mod, header, show_meta_data)) {
- return mod;
- }
- }
LOG(INFO) << "PrintIR(" << header << "):\n" << mod;
return mod;
};
diff --git a/src/ir/type_functor.cc b/src/ir/type_functor.cc
index 1bfc435cff..774c9d8f24 100644
--- a/src/ir/type_functor.cc
+++ b/src/ir/type_functor.cc
@@ -27,8 +27,6 @@
namespace tvm {
-void TypeVisitor::VisitType_(const TensorTypeNode* op) {}
-
void TypeVisitor::VisitType_(const FuncTypeNode* op) {
for (auto arg_type : op->arg_types) {
this->VisitType(arg_type);
@@ -57,11 +55,6 @@ Array<Type> TypeMutator::MutateArray(Array<Type> arr) {
return arr.Map([this](const Type& ty) { return VisitType(ty); });
}
-Type TypeMutator::VisitType_(const TensorTypeNode* op) {
- // TODO(tvm-team) recursively visit to replace Var
- return GetRef<Type>(op);
-}
-
Type TypeMutator::VisitType_(const FuncTypeNode* op) {
bool changed = false;
diff --git a/src/relax/analysis/graph_partitioner.cc b/src/relax/analysis/graph_partitioner.cc
index 53f66f4216..69408afbf4 100644
--- a/src/relax/analysis/graph_partitioner.cc
+++ b/src/relax/analysis/graph_partitioner.cc
@@ -225,11 +225,6 @@ size_t GraphPartitioner::CountFusedNodesWithNewChild(IndexedForwardGraph::Node*
return target->FindRoot()->num_nodes + CountNodesUptoSink_(child, dom_parent);
}
-size_t GraphPartitioner::CountAdditionalArgs_(const TensorTypeNode* ttype, bool with_strides) {
- // TODO(@syfeng): need to clean this up
- return 0;
-}
-
size_t GraphPartitioner::CountArgs_(IndexedForwardGraph::Node* src,
const IndexedForwardGraph& graph, bool update_postpone) {
std::unordered_set<Group*> visited_groups;
@@ -274,11 +269,6 @@ size_t GraphPartitioner::CountArgsLimit_(const IndexedForwardGraph::Node* child)
size_t output_args = 0;
while (outputs_list != nullptr) {
output_args++;
- if (auto call_node = GetRef<ObjectRef>(outputs_list->value.node->ref).as<CallNode>()) {
- if (const auto* ttype = call_node->checked_type().as<TensorTypeNode>()) {
- output_args += CountAdditionalArgs_(ttype, false);
- }
- }
outputs_list = outputs_list->next;
}
return (max_function_args_ > output_args) ? max_function_args_ - output_args : 0;
@@ -302,26 +292,16 @@ void GraphPartitioner::InitGroups(const IndexedForwardGraph& graph) {
for (auto& it : call_node->args) {
if (it.as<VarNode>() || it.as<TupleGetItemNode>()) {
args_num++;
- if (const auto* ttype = it.as<ExprNode>()->checked_type().as<TensorTypeNode>()) {
- args_num += CountAdditionalArgs_(ttype);
- }
}
}
} else if (auto tuple_node = GetRef<ObjectRef>(obj).as<TupleNode>()) {
for (auto& it : tuple_node->fields) {
if (it.as<VarNode>() || it.as<TupleGetItemNode>()) {
args_num++;
- if (const auto* ttype = it.as<ExprNode>()->checked_type().as<TensorTypeNode>()) {
- args_num += CountAdditionalArgs_(ttype);
- }
}
}
} else if (GetRef<ObjectRef>(obj).as<VarNode>()) {
args_num++;
- if (const auto* ttype =
- GetRef<ObjectRef>(obj).as<ExprNode>()->checked_type().as<TensorTypeNode>()) {
- args_num += CountAdditionalArgs_(ttype);
- }
}
return args_num;
};
diff --git a/src/relax/analysis/graph_partitioner.h b/src/relax/analysis/graph_partitioner.h
index 6abf10c5a6..cad9923488 100644
--- a/src/relax/analysis/graph_partitioner.h
+++ b/src/relax/analysis/graph_partitioner.h
@@ -26,6 +26,7 @@
#define TVM_RELAX_ANALYSIS_GRAPH_PARTITIONER_H_
#include <tvm/relax/op_attr_types.h>
+#include <tvm/relax/type.h>
#include <unordered_map>
#include <unordered_set>
@@ -266,11 +267,6 @@ class GraphPartitioner {
void CommitFuse(IndexedForwardGraph::Node* src, IndexedForwardGraph::Node* sink);
size_t CountNodesUptoSink_(IndexedForwardGraph::Node* src, IndexedForwardGraph::Node* sink);
- // Count the number of additional arguments. In the case of dynamic shape,
- // generated function takes several additional arguments, such as the sizes of
- // the dynamic dimensions and strides.
- // This function calculates the number of such additional arguments.
- size_t CountAdditionalArgs_(const TensorTypeNode* ttype, bool with_strides = true);
// Calculate the number of arguments for the node.
size_t CountArgs_(IndexedForwardGraph::Node* src, const IndexedForwardGraph& graph,
bool update_postpone = true);
diff --git a/src/relax/analysis/struct_info_analysis.cc b/src/relax/analysis/struct_info_analysis.cc
index 50931cbf38..d44252e86f 100644
--- a/src/relax/analysis/struct_info_analysis.cc
+++ b/src/relax/analysis/struct_info_analysis.cc
@@ -48,7 +48,7 @@ class StaticTypeDeriver : public StructInfoFunctor<Type(const StructInfo&)> {
}
Type VisitStructInfo_(const TensorStructInfoNode* op) final {
- return DynTensorType(op->ndim, op->dtype);
+ return TensorType(op->ndim, op->dtype);
}
// module: distributed
@@ -87,7 +87,7 @@ StructInfo StructInfoFromType(const Type& type) {
return PrimStructInfo(prim_type->dtype, prim_type->span);
} else if (const ShapeTypeNode* shape_type = type.as<ShapeTypeNode>()) {
return ShapeStructInfo(shape_type->ndim, type->span);
- } else if (const DynTensorTypeNode* tensor_type = type.as<DynTensorTypeNode>()) {
+ } else if (const TensorTypeNode* tensor_type = type.as<TensorTypeNode>()) {
return TensorStructInfo(tensor_type->dtype, tensor_type->ndim);
} else if (const TupleTypeNode* tuple_type = type.as<TupleTypeNode>()) {
Array<StructInfo> fields;
diff --git a/src/relax/ir/dataflow_matcher.cc b/src/relax/ir/dataflow_matcher.cc
index 417a78f0d0..baac1e4c23 100644
--- a/src/relax/ir/dataflow_matcher.cc
+++ b/src/relax/ir/dataflow_matcher.cc
@@ -594,7 +594,7 @@ bool DFPatternMatcher::VisitDFPattern_(const PrimArrPatternNode* op, const Expr&
bool DFPatternMatcher::VisitDFPattern_(const DataTypePatternNode* op, const Expr& expr) {
// no need to jump, as var.dtype == value.dtype
auto expr_type = expr.as<ExprNode>()->checked_type();
- if (const DynTensorTypeNode* tensor_type = expr_type.as<DynTensorTypeNode>()) {
+ if (const TensorTypeNode* tensor_type = expr_type.as<TensorTypeNode>()) {
return (StructuralEqual()(op->dtype, tensor_type->dtype)) && VisitDFPattern(op->pattern, expr);
}
return false;
diff --git a/src/relax/ir/expr.cc b/src/relax/ir/expr.cc
index ca97744c51..4c75666657 100644
--- a/src/relax/ir/expr.cc
+++ b/src/relax/ir/expr.cc
@@ -337,7 +337,7 @@ Constant::Constant(runtime::NDArray data, Optional<StructInfo> struct_info_annot
} else {
TensorStructInfo tinfo(ShapeExpr(values), n->data.DataType(), VDevice(), span);
n->struct_info_ = tinfo;
- n->checked_type_ = DynTensorType(tinfo->ndim, tinfo->dtype);
+ n->checked_type_ = TensorType(tinfo->ndim, tinfo->dtype);
}
data_ = std::move(n);
diff --git a/src/relax/ir/type.cc b/src/relax/ir/type.cc
index 49ef1d7163..82b95b556b 100644
--- a/src/relax/ir/type.cc
+++ b/src/relax/ir/type.cc
@@ -50,26 +50,26 @@ TVM_REGISTER_NODE_TYPE(ObjectTypeNode);
TVM_REGISTER_GLOBAL("relax.ObjectType").set_body_typed([](Span span) { return ObjectType(span); });
-DynTensorType::DynTensorType(int ndim, DataType dtype, Span span) {
- ObjectPtr<DynTensorTypeNode> n = make_object<DynTensorTypeNode>();
+TensorType::TensorType(int ndim, DataType dtype, Span span) {
+ ObjectPtr<TensorTypeNode> n = make_object<TensorTypeNode>();
n->ndim = std::move(ndim);
n->dtype = std::move(dtype);
n->span = span;
data_ = std::move(n);
}
-DynTensorType DynTensorType::CreateUnknownNDim(DataType dtype, Span span) {
- ObjectPtr<DynTensorTypeNode> n = make_object<DynTensorTypeNode>();
+TensorType TensorType::CreateUnknownNDim(DataType dtype, Span span) {
+ ObjectPtr<TensorTypeNode> n = make_object<TensorTypeNode>();
n->ndim = -1;
n->dtype = std::move(dtype);
n->span = std::move(span);
- return DynTensorType(std::move(n));
+ return TensorType(std::move(n));
}
-TVM_REGISTER_NODE_TYPE(DynTensorTypeNode);
+TVM_REGISTER_NODE_TYPE(TensorTypeNode);
-TVM_REGISTER_GLOBAL("relax.DynTensorType").set_body_typed([](int ndim, DataType dtype, Span span) {
- return DynTensorType(ndim, dtype, span);
+TVM_REGISTER_GLOBAL("relax.TensorType").set_body_typed([](int ndim, DataType dtype, Span span) {
+ return TensorType(ndim, dtype, span);
});
PackedFuncType::PackedFuncType(Span span) {
diff --git a/src/relax/transform/fold_constant.cc b/src/relax/transform/fold_constant.cc
index d6da79c484..ff193acf14 100644
--- a/src/relax/transform/fold_constant.cc
+++ b/src/relax/transform/fold_constant.cc
@@ -185,7 +185,7 @@ class ConstantFolder : public ExprMutator {
bool output_not_tuple = call->sinfo_args.size() == 1;
// Pattern 0: call constant function, const argument with const shape.
if (func && arr_args && shape && output_not_tuple) {
- DynTensorType ret_type = Downcast<DynTensorType>(call->checked_type());
+ TensorType ret_type = Downcast<TensorType>(call->checked_type());
// value_or will return value if it is not null, otherwise return or
return ConstEvaluateCallTIR(func.value(), arr_args.value(), shape.value(), ret_type->dtype)
.value_or({});
diff --git a/src/relax/transform/fuse_tir.cc b/src/relax/transform/fuse_tir.cc
index 8fba546281..37f3520d3d 100644
--- a/src/relax/transform/fuse_tir.cc
+++ b/src/relax/transform/fuse_tir.cc
@@ -973,7 +973,7 @@ class FusedTIRConstructor : public ExprVisitor {
/*! \brief Get DynTensor numbers from recursive Tuples. */
static size_t GetTotalTensorSize(const Type& type) {
- if (type.as<DynTensorTypeNode>()) {
+ if (type.as<TensorTypeNode>()) {
return 1;
} else if (const auto* tuple_type = type.as<TupleTypeNode>()) {
size_t num = 0;
@@ -982,7 +982,7 @@ class FusedTIRConstructor : public ExprVisitor {
}
return num;
} else {
- LOG(FATAL) << "DynTensorType and TupleType are expect, but got: " << type;
+ LOG(FATAL) << "TensorType and TupleType are expect, but got: " << type;
return 0;
}
}
diff --git a/src/runtime/contrib/bnns/bnns_wrp.h b/src/runtime/contrib/bnns/bnns_wrp.h
index b31e97e554..f395561a7f 100644
--- a/src/runtime/contrib/bnns/bnns_wrp.h
+++ b/src/runtime/contrib/bnns/bnns_wrp.h
@@ -287,7 +287,7 @@ class TView {
return res;
}
- /** Check if view is empty and doesn't relay to any tensor */
+ /** Check if view is empty and doesn't map to any tensor */
operator bool() const { return origin_ != nullptr; }
/** Get BNNS descriptor for particular View. Batch and Party attributed are ignored. */
diff --git a/src/runtime/contrib/cblas/gemm_common.h b/src/runtime/contrib/cblas/gemm_common.h
index 91341976bd..2c7738e458 100644
--- a/src/runtime/contrib/cblas/gemm_common.h
+++ b/src/runtime/contrib/cblas/gemm_common.h
@@ -120,7 +120,7 @@ inline void CallU8S8S32Gemm(TVMArgs args, TVMRetValue* ret, TGemmOp op) {
bool transb = args[4];
// Set the sgemm attributes. Currently, support is limited to CblasFixOffset with all offsets
- // equal to 0. This is sufficient for relay dense.
+ // equal to 0. This is sufficient for dense.
std::string offset_ctype = "CblasFixOffset";
int16_t offset_a = 0;
int16_t offset_b = 0;
diff --git a/src/runtime/contrib/onnx/onnx_module.cc b/src/runtime/contrib/onnx/onnx_module.cc
deleted file mode 100644
index 813211ca7c..0000000000
--- a/src/runtime/contrib/onnx/onnx_module.cc
+++ /dev/null
@@ -1,80 +0,0 @@
-/*
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements. See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership. The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * "License"); you may not use this file except in compliance
- * with the License. You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations
- * under the License.
- */
-
-/*!
- * \file onnx_module.cc
- * \brief ONNX Module without runtime support
- */
-#include <tvm/runtime/packed_func.h>
-#include <tvm/runtime/registry.h>
-
-namespace tvm {
-namespace codegen {
-using namespace tvm::runtime;
-
-class ONNXSourceModuleNode : public runtime::ModuleNode {
- public:
- explicit ONNXSourceModuleNode(const std::string& code, const std::string& symbol,
- const Array<String>& const_vars)
- : code_(code), symbol_(symbol), const_vars_(const_vars) {}
- const char* type_key() const { return "onnx"; }
-
- /*! \brief Get the property of the runtime module .*/
- int GetPropertyMask() const final { return ModulePropertyMask::kRunnable; };
-
- PackedFunc GetFunction(const String& name, const ObjectPtr<Object>& sptr_to_self) final {
- if (name == "get_symbol") {
- return PackedFunc(
- [sptr_to_self, this](TVMArgs args, TVMRetValue* rv) { *rv = this->symbol_; });
- } else if (name == "get_const_vars") {
- return PackedFunc(
- [sptr_to_self, this](TVMArgs args, TVMRetValue* rv) { *rv = this->const_vars_; });
- } else {
- LOG(FATAL) << "ONNX Source module cannot execute, to get executable module"
- << " build TVM with 'onnx' runtime support";
- return PackedFunc(nullptr);
- }
- }
-
- String GetSource(const String& format) final { return code_; }
-
- void SaveToFile(const String& path, const String& format) final {
- ICHECK_EQ(format, "onnx") << "Can only save to onnx format";
- ICHECK_NE(code_.length(), 0);
- const PackedFunc* to_onnx_ = runtime::Registry::Get("relay.ext.onnx.save_to_file");
- (*to_onnx_)(code_, path, format);
- }
-
- protected:
- String code_;
- std::string symbol_;
- Array<String> const_vars_;
-};
-
-Module ONNXSourceModuleNodeCreate(const String& code, const String& symbol,
- const Array<String>& const_vars) {
- auto n = make_object<ONNXSourceModuleNode>(code.operator std::string(),
- symbol.operator std::string(), const_vars);
- return Module(n);
-}
-
-TVM_REGISTER_GLOBAL("runtime.ONNXModuleCreate").set_body_typed(ONNXSourceModuleNodeCreate);
-
-} // namespace codegen
-} // namespace tvm
diff --git a/src/runtime/contrib/tensorrt/tensorrt_calibrator.h b/src/runtime/contrib/tensorrt/tensorrt_calibrator.h
index 523676b947..491714acd9 100755
--- a/src/runtime/contrib/tensorrt/tensorrt_calibrator.h
+++ b/src/runtime/contrib/tensorrt/tensorrt_calibrator.h
@@ -16,7 +16,7 @@
* under the License.
* file runtime/contrib/tensorrt/tensorrt_builder.h
- * brief Contains TensorRTBuilder class which can be used to convert a relay
+ * brief Contains TensorRTBuilder class which can be used to convert a graph
* program into a TRT engine which can be used for inference.
*/
diff --git a/src/runtime/contrib/tensorrt/tensorrt_ops.h b/src/runtime/contrib/tensorrt/tensorrt_ops.h
index e2ef341b4a..950d9d183a 100644
--- a/src/runtime/contrib/tensorrt/tensorrt_ops.h
+++ b/src/runtime/contrib/tensorrt/tensorrt_ops.h
@@ -198,7 +198,7 @@ class TensorRTOpConverter {
};
/*!
- * \brief Get the map of available TensorRTOpConverters, where the key is the name of the relay op.
+ * \brief Get the map of available TensorRTOpConverters, where the key is the name of the op.
* \return Map of TensorRTOpConverters.
*/
const std::unordered_map<std::string, std::unique_ptr<TensorRTOpConverter>>& GetOpConverters();
diff --git a/src/script/printer/ir/ir.cc b/src/script/printer/ir/ir.cc
index ac3bc5584c..f98d64dfa3 100644
--- a/src/script/printer/ir/ir.cc
+++ b/src/script/printer/ir/ir.cc
@@ -16,7 +16,7 @@
* specific language governing permissions and limitations
* under the License.
*/
-#include <tvm/ir/tensor_type.h>
+#include <tvm/ir/type.h>
#include "./utils.h"
@@ -142,13 +142,6 @@ TVM_STATIC_IR_FUNCTOR(IRDocsifier, vtable)
return IR(d, "Op")->Call({LiteralDoc::Str(op->name, p->Attr("name"))});
});
-TVM_STATIC_IR_FUNCTOR(IRDocsifier, vtable)
- .set_dispatch<TensorType>("", [](TensorType type, ObjectPath p, IRDocsifier d) -> Doc {
- return IR(d, "TensorType")
- ->Call({d->AsDoc<ExprDoc>(type->shape, p->Attr("shape")),
- LiteralDoc::DataType(type->dtype, p->Attr("dtype"))});
- });
-
TVM_STATIC_IR_FUNCTOR(IRDocsifier, vtable)
.set_dispatch<FuncType>("", [](FuncType func_type, ObjectPath p, IRDocsifier d) -> Doc {
return IR(d, "FuncType")
@@ -168,11 +161,6 @@ TVM_STATIC_IR_FUNCTOR(IRDocsifier, vtable)
});
std::string ReprPrintIRModule(const ObjectRef& mod, const PrinterConfig& cfg) {
- if (const auto* f = runtime::Registry::Get("relay.ir.PrintRelayModule")) {
- if (Optional<String> s = (*f)(mod)) {
- return s.value();
- }
- }
return ReprPrintIR(mod, cfg);
}
diff --git a/src/script/printer/relax/type.cc b/src/script/printer/relax/type.cc
index d13d90b1d5..4388678435 100644
--- a/src/script/printer/relax/type.cc
+++ b/src/script/printer/relax/type.cc
@@ -36,8 +36,8 @@ TVM_STATIC_IR_FUNCTOR(IRDocsifier, vtable)
});
TVM_STATIC_IR_FUNCTOR(IRDocsifier, vtable)
- .set_dispatch<relax::DynTensorType>(
- "", [](relax::DynTensorType n, ObjectPath n_p, IRDocsifier d) -> Doc {
+ .set_dispatch<relax::TensorType>(
+ "", [](relax::TensorType n, ObjectPath n_p, IRDocsifier d) -> Doc {
return Relax(d, "Tensor")
->Call({}, {"ndim", "dtype"},
{LiteralDoc::Int(n->ndim, n_p->Attr("ndim")),
@@ -80,7 +80,7 @@ TVM_STATIC_IR_FUNCTOR(IRDocsifier, vtable)
TVM_SCRIPT_REPR(relax::ShapeTypeNode, ReprPrintRelax);
TVM_SCRIPT_REPR(relax::ObjectTypeNode, ReprPrintRelax);
-TVM_SCRIPT_REPR(relax::DynTensorTypeNode, ReprPrintRelax);
+TVM_SCRIPT_REPR(relax::TensorTypeNode, ReprPrintRelax);
TVM_SCRIPT_REPR(relax::PackedFuncTypeNode, ReprPrintRelax);
TVM_REGISTER_GLOBAL("script.printer.ReprPrintRelax").set_body_typed(ReprPrintRelax);
diff --git a/src/support/libinfo.cc b/src/support/libinfo.cc
index fd77427c70..b981fcd6d7 100644
--- a/src/support/libinfo.cc
+++ b/src/support/libinfo.cc
@@ -231,10 +231,6 @@
#define TVM_INFO_USE_COREML "NOT-FOUND"
#endif
-#ifndef TVM_INFO_USE_TARGET_ONNX
-#define TVM_INFO_USE_TARGET_ONNX "NOT-FOUND"
-#endif
-
#ifndef TVM_INFO_USE_ARM_COMPUTE_LIB
#define TVM_INFO_USE_ARM_COMPUTE_LIB "NOT-FOUND"
#endif
@@ -351,7 +347,6 @@ TVM_DLL Map<String, String> GetLibInfo() {
{"USE_SORT", TVM_INFO_USE_SORT},
{"USE_SPIRV_KHR_INTEGER_DOT_PRODUCT", TVM_INFO_USE_SPIRV_KHR_INTEGER_DOT_PRODUCT},
{"USE_STACKVM_RUNTIME", TVM_INFO_USE_STACKVM_RUNTIME},
- {"USE_TARGET_ONNX", TVM_INFO_USE_TARGET_ONNX},
{"USE_TENSORFLOW_PATH", TVM_INFO_USE_TENSORFLOW_PATH},
{"USE_TENSORRT_CODEGEN", TVM_INFO_USE_TENSORRT_CODEGEN},
{"USE_TENSORRT_RUNTIME", TVM_INFO_USE_TENSORRT_RUNTIME},
diff --git a/src/target/generic_func.cc b/src/target/generic_func.cc
deleted file mode 100644
index 3135f6a9f2..0000000000
--- a/src/target/generic_func.cc
+++ /dev/null
@@ -1,182 +0,0 @@
-/*
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements. See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership. The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * "License"); you may not use this file except in compliance
- * with the License. You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations
- * under the License.
- */
-/*!
- * \file src/target/generic_func.cc
- */
-#include <dmlc/thread_local.h>
-#include <tvm/node/node.h>
-#include <tvm/node/repr_printer.h>
-#include <tvm/runtime/registry.h>
-#include <tvm/target/generic_func.h>
-#include <tvm/target/target.h>
-#include <tvm/tir/expr.h>
-
-#include <algorithm>
-#include <mutex>
-#include <stack>
-
-#include "../runtime/object_internal.h"
-
-namespace tvm {
-
-TVM_REGISTER_NODE_TYPE(GenericFuncNode);
-
-struct GenericFunc::Manager {
- std::unordered_map<std::string, GenericFunc> fmap;
- // mutex
- std::mutex mutex;
-
- Manager() {}
-
- static Manager* Global() {
- static Manager inst;
- return &inst;
- }
-};
-
-GenericFunc GenericFunc::Get(const std::string& name) {
- Manager* m = Manager::Global();
- std::lock_guard<std::mutex> lock(m->mutex);
- auto it = m->fmap.find(name);
- if (it == m->fmap.end()) {
- auto f = make_object<GenericFuncNode>();
- f->name_ = name;
- auto gf = GenericFunc(f);
- m->fmap[name] = gf;
- return gf;
- } else {
- return it->second;
- }
-}
-
-void GenericFunc::RegisterGenericFunc(GenericFunc func, const std::string& name) {
- Manager* m = Manager::Global();
- std::lock_guard<std::mutex> lock(m->mutex);
- auto it = m->fmap.find(name);
- ICHECK(it == m->fmap.end()) << "GenericFunc already registered " << name;
- func->name_ = name;
- m->fmap[name] = func;
-}
-
-GenericFunc& GenericFunc::set_default(const PackedFunc value, bool allow_override) {
- auto node = static_cast<GenericFuncNode*>(operator->());
- if (!allow_override) {
- ICHECK(node->generic_func_ == nullptr)
- << "Generic function already registered for " << node->name_;
- }
- node->generic_func_ = value;
- return *this;
-}
-
-GenericFunc& GenericFunc::register_func(const std::vector<std::string>& tags,
- const PackedFunc value, bool allow_override) {
- for (auto& t : tags) {
- if (!allow_override) {
- auto iter = (*this)->dispatch_dict_.find(t);
- ICHECK(iter == (*this)->dispatch_dict_.end())
- << "Tag " << t << " already registered for schedule factory " << (*this)->name_;
- }
- (*this)->dispatch_dict_[t] = value;
- }
- return *this;
-}
-
-void GenericFunc::CallPacked(TVMArgs args, TVMRetValue* ret) const {
- auto node = static_cast<const GenericFuncNode*>(get());
- auto target = Target::Current(true);
- PackedFunc func;
-
- if (target.defined()) {
- for (auto& k : target->GetKeys()) {
- auto iter = node->dispatch_dict_.find(k);
- if (iter != node->dispatch_dict_.end()) {
- func = iter->second;
- break;
- }
- }
- }
-
- if (func == nullptr) {
- ICHECK(node->generic_func_ != nullptr) << "No generic function registered for " << node->name_;
- func = node->generic_func_;
- }
-
- func.CallPacked(args, ret);
-}
-
-PackedFunc GenericFunc::GetPacked() const {
- auto node = static_cast<const GenericFuncNode*>(get());
- auto target = Target::Current(true);
- if (target.defined()) {
- for (auto& k : target->GetKeys()) {
- auto iter = node->dispatch_dict_.find(k);
- if (iter != node->dispatch_dict_.end()) {
- return iter->second;
- }
- }
- }
- return node->generic_func_;
-}
-
-TVM_REGISTER_GLOBAL("target.GenericFuncCreate").set_body([](TVMArgs args, TVMRetValue* ret) {
- *ret = GenericFunc(make_object<GenericFuncNode>());
-});
-
-TVM_REGISTER_GLOBAL("target.GenericFuncGetGlobal").set_body([](TVMArgs args, TVMRetValue* ret) {
- std::string func_name = args[0];
- *ret = GenericFunc::Get(func_name);
-});
-
-TVM_REGISTER_GLOBAL("target.GenericFuncSetDefault").set_body([](TVMArgs args, TVMRetValue* ret) {
- GenericFunc generic_func = args[0];
- PackedFunc func = args[1];
- bool allow_override = args[2];
- // Intentionally copy and not de-allocate it, to avoid free pyobject during shutdown
- runtime::ObjectInternal::ObjectRetain((TVMObjectHandle)(func.get()));
- generic_func.set_default(func, allow_override);
-});
-
-TVM_REGISTER_GLOBAL("target.GenericFuncRegisterFunc").set_body([](TVMArgs args, TVMRetValue* ret) {
- GenericFunc generic_func = args[0];
- PackedFunc func = args[1];
- Array<runtime::String> tags = args[2];
- bool allow_override = args[3];
- // Intentionally copy and not de-allocate it, to avoid free pyobject during shutdown
- runtime::ObjectInternal::ObjectRetain((TVMObjectHandle)(func.get()));
- std::vector<std::string> tags_vector;
- for (auto& tag : tags) {
- tags_vector.push_back(tag);
- }
-
- generic_func.register_func(tags_vector, func, allow_override);
-});
-
-TVM_REGISTER_GLOBAL("target.GenericFuncCallFunc").set_body([](TVMArgs args, TVMRetValue* ret) {
- GenericFunc generic_func = args[0];
- TVMArgs func_args(&args.values[1], &args.type_codes[1], args.num_args - 1);
-
- generic_func.CallPacked(func_args, ret);
-});
-
-TVM_REGISTER_GLOBAL("target.GenericFuncGetPackedFunc").set_body([](TVMArgs args, TVMRetValue* ret) {
- GenericFunc generic_func = args[0];
- *ret = generic_func.GetPacked();
-});
-
-} // namespace tvm
diff --git a/src/te/operation/create_primfunc.h b/src/te/operation/create_primfunc.h
index 496ee45ba4..dc045156d1 100644
--- a/src/te/operation/create_primfunc.h
+++ b/src/te/operation/create_primfunc.h
@@ -42,8 +42,6 @@ PrimFunc CreatePrimFuncWithConstants(const Array<te::Tensor>& arg_list,
const Array<runtime::NDArray>& constants,
std::optional<DataType> index_dtype_override = std::nullopt);
-// Relax version
-// TODO(relax-team) combine with the relay version
/*! \brief Use Tensor Expression to create a schedulable TensorIR func. */
PrimFunc CreatePrimFunc(const Array<ObjectRef>& arg_list,
std::optional<DataType> index_dtype_override);
diff --git a/src/tir/ir/expr.cc b/src/tir/ir/expr.cc
index c38237a664..ca28520f8f 100644
--- a/src/tir/ir/expr.cc
+++ b/src/tir/ir/expr.cc
@@ -544,7 +544,7 @@ TVM_REGISTER_GLOBAL("tir.Let").set_body_typed([](Var var, PrimExpr value, PrimEx
TVM_REGISTER_NODE_TYPE(LetNode);
// Call
-Call::Call(DataType dtype, RelayExpr op, Array<PrimExpr> args, Span span) {
+Call::Call(DataType dtype, RelaxExpr op, Array<PrimExpr> args, Span span) {
for (size_t i = 0; i < args.size(); ++i) {
ICHECK(args[i].defined()) << "arg " << i << " is not defined()";
}
@@ -558,7 +558,7 @@ Call::Call(DataType dtype, RelayExpr op, Array<PrimExpr> args, Span span) {
}
TVM_REGISTER_GLOBAL("tir.Call")
- .set_body_typed([](DataType type, RelayExpr op,
+ .set_body_typed([](DataType type, RelaxExpr op,
Array<Variant<runtime::String, IterVar, BufferRegion, PrimExpr>> args,
Span span) {
Array<PrimExpr> prim_expr_args;
diff --git a/tests/lint/check_file_type.py b/tests/lint/check_file_type.py
index c27bd410f4..119cc2ea78 100644
--- a/tests/lint/check_file_type.py
+++ b/tests/lint/check_file_type.py
@@ -48,8 +48,6 @@ ALLOW_EXTENSION = {
"cu",
"cuh",
"bat",
- # relay text format
- "rly",
# configurations
"mk",
"in",
diff --git a/tests/python/contrib/test_hexagon/infrastructure.py b/tests/python/contrib/test_hexagon/infrastructure.py
index c728ae842c..0fbe92f508 100644
--- a/tests/python/contrib/test_hexagon/infrastructure.py
+++ b/tests/python/contrib/test_hexagon/infrastructure.py
@@ -21,7 +21,6 @@
import numpy
import tvm
from tvm import te
-from tvm.relay.backend import Executor
def ceildiv(o, d):
@@ -113,19 +112,6 @@ def build_and_run(inputs, func, target: str, target_host: str, *args, **kwargs):
return tensors[-1].asnumpy()
-def build_module(relay_mod, target):
- """builds a relay module for a specified target"""
- params = {}
- executor = Executor("aot", {"link-params": True})
- lowered = tvm.relay.build(
- relay_mod,
- tvm.target.Target(target, host=target),
- executor=executor,
- params=params,
- )
- return lowered
-
-
def run_module(mod, inputs):
"""invokes run function of specified module with inputs provided"""
mod.set_input(**inputs)
diff --git a/tests/python/contrib/test_hexagon/test_relax_integration.py b/tests/python/contrib/test_hexagon/test_relax_integration.py
index 89539b7951..7d90adbc95 100644
--- a/tests/python/contrib/test_hexagon/test_relax_integration.py
+++ b/tests/python/contrib/test_hexagon/test_relax_integration.py
@@ -19,102 +19,9 @@
import numpy as np
import pytest
import tvm.testing
-from tvm import relay, relax, runtime
+from tvm import relax, runtime
from tvm.relax.testing import relay_translator
from tvm.contrib.hexagon.session import Session
-from tvm.relay import testing
-
-
-class TestConv2d:
- """Test conv2d op"""
-
- n_batch = tvm.testing.parameter(1, relay.Any())
-
- @tvm.testing.requires_hexagon
- def test_conv2d(self, hexagon_session: Session, n_batch):
- """Test Relax conv2d op and compare with Relay"""
- dtype = "float32"
- data = relay.var("data", relay.TensorType((n_batch, 64, 64, 3), dtype))
- weight = relay.var("weight", relay.TensorType((5, 5, 3, 8), dtype))
- y = relay.nn.conv2d(
- data,
- weight,
- padding=(2, 2),
- kernel_size=(5, 5),
- data_layout="NHWC",
- kernel_layout="HWIO",
- out_dtype="float32",
- )
- f = relay.Function([data, weight], y)
- relay_mod = tvm.IRModule.from_expr(f)
-
- target_hexagon = tvm.target.hexagon("v68")
- target = tvm.target.Target(target_hexagon, host=target_hexagon)
- relax_mod = relay_translator.from_relay(relay_mod["main"], target)
-
- exe = relax.build(relax_mod, target)
- dev = hexagon_session.device
- vm_mod = hexagon_session.get_executor_from_factory(exe)
- vm_rt = relax.VirtualMachine(vm_mod, dev)
-
- data_np = np.random.rand(1, 64, 64, 3).astype(np.float32)
- weight_np = np.random.rand(5, 5, 3, 8).astype(np.float32)
-
- # Run on hexagon and get result
- data = tvm.nd.array(data_np, dev)
- weight = tvm.nd.array(weight_np, dev)
- vm_rt.set_input("main", data, weight)
- vm_rt.invoke_stateful("main")
- hexagon_res = vm_rt.get_outputs("main")
-
- # Compile and run on Relay for comparison.
- dev = tvm.cpu()
- data = tvm.nd.array(data_np, dev)
- weight = tvm.nd.array(weight_np, dev)
-
- target = tvm.target.Target("llvm", host="llvm")
- vm_exec = relay.vm.compile(relay_mod, target=target)
- vm_factory = runtime.vm.VirtualMachine(vm_exec, tvm.cpu())
- relay_res = vm_factory.invoke("main", data, weight)
- tvm.testing.assert_allclose(hexagon_res.numpy(), relay_res.numpy(), rtol=1e-3)
-
-
-class TestMLP:
- """Test MLP"""
-
- n_batch = tvm.testing.parameter(1, relay.Any())
-
- @tvm.testing.requires_hexagon
- def test_mlp(self, hexagon_session: Session, n_batch):
- """Test Relax MLP and compare with Relay"""
- relay_mod, params = testing.mlp.get_workload(batch_size=n_batch, dtype="float32")
-
- target_hexagon = tvm.target.hexagon("v68")
- target = tvm.target.Target(target_hexagon, host=target_hexagon)
- relax_mod = relay_translator.from_relay(relay_mod["main"], target, params)
-
- exe = relax.build(relax_mod, target)
- hexagon_device = hexagon_session.device
-
- vm_mod = hexagon_session.get_executor_from_factory(exe)
- vm_rt = relax.VirtualMachine(vm_mod, hexagon_device)
-
- shape = (1, 1, 28, 28)
- data_np = np.random.rand(*shape).astype("float32")
- data = tvm.nd.array(data_np, hexagon_device)
- vm_rt.set_input("main", data)
- vm_rt.invoke_stateful("main")
- hexagon_res = vm_rt.get_outputs("main")
-
- # Compile and run on Relay for comparison.
- cpu_dev = tvm.cpu()
- data = tvm.nd.array(data_np, cpu_dev)
-
- target = tvm.target.Target("llvm", host="llvm")
- vm_exec = relay.vm.compile(relay_mod, target=target)
- vm_factory = runtime.vm.VirtualMachine(vm_exec, cpu_dev)
- relay_res = vm_factory.invoke("main", data, **params)
- tvm.testing.assert_allclose(hexagon_res.numpy(), relay_res.numpy(), rtol=1e-3)
def get_onnx_mobilenet():
@@ -197,40 +104,5 @@ def test_mobilenet(hexagon_session: Session):
tvm.testing.assert_allclose(hexagon_res.numpy(), llvm_res.numpy(), rtol=1e-3)
-@pytest.mark.skip("takes too long (~20min)")
-@tvm.testing.requires_hexagon
-def test_mobilenet_dyn(hexagon_session: Session):
- """Test MobileNet workload with dynamic batch size"""
- relay_mod, params = testing.mobilenet.get_workload(batch_size=relay.Any(), dtype="float32")
- data_np = np.random.rand(1, 3, 224, 224).astype("float32")
-
- target_hexagon = tvm.target.hexagon("v68")
- target = tvm.target.Target(target_hexagon, host=target_hexagon)
-
- # translate the relay mobilenet and bind params
- relax_mod = relay_translator.from_relay(relay_mod["main"], target, params)
-
- # Compile and run on Hexagon.
- exe = relax.build(relax_mod, target)
- dev = hexagon_session.device
-
- vm_mod = hexagon_session.get_executor_from_factory(exe)
- vm_rt = relax.VirtualMachine(vm_mod, dev)
- data = tvm.nd.array(data_np, dev)
- vm_rt.set_input("main", data)
- vm_rt.invoke_stateful("main")
- hexagon_res = vm_rt.get_outputs("main")
-
- # Compile and run on Relay for comparison.
- dev = tvm.cpu()
- data = tvm.nd.array(data_np, dev)
-
- target = tvm.target.Target("llvm", host="llvm")
- vm_exec = relay.vm.compile(relay_mod, target=target)
- vm_factory = runtime.vm.VirtualMachine(vm_exec, tvm.cpu())
- relay_res = vm_factory.invoke("main", data, **params)
- tvm.testing.assert_allclose(hexagon_res.numpy(), relay_res.numpy(), rtol=1e-3)
-
-
if __name__ == "__main__":
tvm.testing.main()
diff --git a/tests/python/ir/test_ir_type.py b/tests/python/ir/test_ir_type.py
index 0c95bc4554..d1f9ab3100 100644
--- a/tests/python/ir/test_ir_type.py
+++ b/tests/python/ir/test_ir_type.py
@@ -31,16 +31,9 @@ def test_prim_type():
assert x.dtype == "int32"
-def test_tensor_type_bad_constructor():
- try:
- x = tvm.ir.TensorType("xx", "xx")
- except tvm.error.TVMError:
- pass
-
-
def test_func_type():
arg_types = tvm.runtime.convert([])
- ret_type = tvm.ir.TensorType((1, 2, 3), "float32")
+ ret_type = tvm.ir.PrimType("float32")
tf = tvm.ir.FuncType(arg_types, ret_type)
assert tf.arg_types == arg_types
assert tf.ret_type == ret_type
@@ -52,7 +45,7 @@ def test_func_type():
def test_tuple_type():
tf = tvm.ir.FuncType([], tvm.ir.TupleType([]))
- tt = tvm.ir.TensorType(tvm.runtime.convert([1, 2, 3]), "float32")
+ tt = tvm.ir.PrimType("float32")
fields = tvm.runtime.convert([tf, tt])
tup_ty = tvm.ir.TupleType(fields)
diff --git a/tests/python/nightly/test_nnapi/infrastructure.py b/tests/python/nightly/test_nnapi/infrastructure.py
index a86c681f0b..9f3813ef61 100644
--- a/tests/python/nightly/test_nnapi/infrastructure.py
+++ b/tests/python/nightly/test_nnapi/infrastructure.py
@@ -79,7 +79,7 @@ def decompose_clip(mod: tvm.IRModule) -> tvm.IRModule:
def _build(mod, enable_nnapi):
- if isinstance(mod, tvm.relay.expr.Call):
+ if isinstance(mod, tvm.relax.expr.Call):
mod = tvm.IRModule.from_expr(mod)
if enable_nnapi:
diff --git a/tests/python/relax/test_analysis_struct_info_analysis.py b/tests/python/relax/test_analysis_struct_info_analysis.py
index b2931549e9..2d5afdc73f 100644
--- a/tests/python/relax/test_analysis_struct_info_analysis.py
+++ b/tests/python/relax/test_analysis_struct_info_analysis.py
@@ -54,7 +54,7 @@ def test_get_static_type_tensor():
s4 = rx.TensorStructInfo([1, n + 1, m], "int64")
tvm.ir.assert_structural_equal(
- rx.analysis.get_static_type(s4), rx.DynTensorType(ndim=3, dtype="int64")
+ rx.analysis.get_static_type(s4), rx.TensorType(ndim=3, dtype="int64")
)
@@ -71,7 +71,7 @@ def test_get_static_type_tuple():
rx.analysis.get_static_type(t1),
rx.TupleType(
[
- rx.TupleType([rx.DynTensorType(ndim=3, dtype="int64"), rx.ObjectType()]),
+ rx.TupleType([rx.TensorType(ndim=3, dtype="int64"), rx.ObjectType()]),
rx.ShapeType(ndim=3),
]
),
@@ -88,9 +88,9 @@ def test_get_static_type_func():
return rx.FuncStructInfo([x, y], z)
def fn_type():
- x = rx.DynTensorType(ndim=3, dtype="float32")
- y = rx.DynTensorType(ndim=3, dtype="float32")
- z = rx.DynTensorType(ndim=2, dtype="float32")
+ x = rx.TensorType(ndim=3, dtype="float32")
+ y = rx.TensorType(ndim=3, dtype="float32")
+ z = rx.TensorType(ndim=2, dtype="float32")
return rx.FuncType([x, y], z)
f0 = fn_info(1)
diff --git a/tests/python/relax/test_analysis_suggest_layout_transforms.py b/tests/python/relax/test_analysis_suggest_layout_transforms.py
index 6e47c1d681..03eaef0267 100644
--- a/tests/python/relax/test_analysis_suggest_layout_transforms.py
+++ b/tests/python/relax/test_analysis_suggest_layout_transforms.py
@@ -476,7 +476,7 @@ def test_op_reduce():
def test_op_upsampling():
- # relay materializes the layout if H, W or D dimensions are moved or tiled.
+ # relax materializes the layout if H, W or D dimensions are moved or tiled.
@T.prim_func(private=True)
def before(
arg: T.Buffer((32, 64, 224, 224), "float32"),
diff --git a/tests/python/relax/test_ast_printer.py b/tests/python/relax/test_ast_printer.py
index 1df7dcf36f..6f55c542ea 100644
--- a/tests/python/relax/test_ast_printer.py
+++ b/tests/python/relax/test_ast_printer.py
@@ -270,8 +270,8 @@ def test_types():
assert strip_whitespace(printer.visit_type_(object_type)) == "ObjectType()"
packed_type = rx.PackedFuncType()
assert strip_whitespace(printer.visit_type_(packed_type)) == "PackedFuncType()"
- tensor_type = rx.DynTensorType(ndim=2, dtype="int32")
- assert strip_whitespace(printer.visit_type_(tensor_type)) == "DynTensorType(ndim=2,dtype=int32)"
+ tensor_type = rx.TensorType(ndim=2, dtype="int32")
+ assert strip_whitespace(printer.visit_type_(tensor_type)) == "TensorType(ndim=2,dtype=int32)"
unit_type = rx.TupleType([])
assert strip_whitespace(printer.visit_type_(unit_type)) == "TupleType(fields=[])"
tuple_type = rx.TupleType([rx.ShapeType(), object_type])
@@ -284,7 +284,7 @@ def test_types():
func_type = rx.FuncType([tensor_type], unit_type)
assert_fields(
"FuncType",
- {"arg_types": "[DynTensorType(ndim=2, dtype=int32)]", "ret_type": "TupleType(fields=[])"},
+ {"arg_types": "[TensorType(ndim=2, dtype=int32)]", "ret_type": "TupleType(fields=[])"},
printer.visit_type_(func_type),
)
@@ -622,7 +622,7 @@ def test_print_type_annotation_non_var():
arg = call.args[0]
arg_str = strip_whitespace(dump_ast(arg))
# the constant should have a tensor type
- assert "checked_type_=DynTensorType(ndim=0" in arg_str
+ assert "checked_type_=TensorType(ndim=0" in arg_str
call_str = strip_whitespace(dump_ast(call))
# we expect the shape_of call to have a checked_type_ of ShapeType
diff --git a/tests/python/relax/test_blockbuilder_core.py b/tests/python/relax/test_blockbuilder_core.py
index da030935c3..efab59a0e6 100644
--- a/tests/python/relax/test_blockbuilder_core.py
+++ b/tests/python/relax/test_blockbuilder_core.py
@@ -206,7 +206,7 @@ def test_binary_shape_type_deduction():
gv0 = bb.emit_output(lv3)
bb.emit_func_output(gv0)
- assert isinstance(gv0.checked_type, rx.DynTensorType)
+ assert isinstance(gv0.checked_type, rx.TensorType)
assert gv0.checked_type.ndim == 1
assert gv0.checked_type.dtype == "float16"
diff --git a/tests/python/relax/test_dataflow_pattern.py b/tests/python/relax/test_dataflow_pattern.py
index a534b7c0c7..ab6c6df31b 100644
--- a/tests/python/relax/test_dataflow_pattern.py
+++ b/tests/python/relax/test_dataflow_pattern.py
@@ -227,7 +227,7 @@ def test_not_pattern():
def test_type_pattern():
- assert wildcard().has_type(rx.DynTensorType(2, "float32")).match(bindings[0].var)
+ assert wildcard().has_type(rx.TensorType(2, "float32")).match(bindings[0].var)
def test_dtype_pattern():
@@ -486,7 +486,7 @@ class SmallParallel:
def test_distinguish_diamond_and_parallel():
- # relay pattern lang cannot distinguish the two cases above.
+ # pattern lang cannot distinguish the two cases above.
diamond = SmallDiamond["main"].body.blocks[0]
parallel = SmallParallel["main"].body.blocks[0]
diff --git a/tests/python/relax/test_expr.py b/tests/python/relax/test_expr.py
index b20c9ef2d9..c20f0b2681 100644
--- a/tests/python/relax/test_expr.py
+++ b/tests/python/relax/test_expr.py
@@ -48,7 +48,7 @@ def test_var() -> None:
assert v1.name_hint == "v1"
for s0, s1 in zip(v1.struct_info.shape, shape):
assert s0 == s1
- assert v1.checked_type == rx.DynTensorType(2, "float32")
+ assert v1.checked_type == rx.TensorType(2, "float32")
tvm.ir.assert_structural_equal(v1.struct_info, rx.TensorStructInfo(shape, "float32"))
@@ -62,7 +62,7 @@ def test_dataflow_var() -> None:
v1 = rx.DataflowVar("v1", R.Tensor(shape, "float16"))
assert v1.name_hint == "v1"
- assert v1._checked_type_ == rx.DynTensorType(2, "float16")
+ assert v1._checked_type_ == rx.TensorType(2, "float16")
assert isinstance(v1, rx.DataflowVar)
tvm.ir.assert_structural_equal(v1.struct_info, rx.TensorStructInfo(shape, "float16"))
@@ -128,7 +128,7 @@ def test_match_cast() -> None:
assert b1.pattern[0] == m
assert b1.pattern[1] == n
assert b1.var is not None
- assert b1.var.checked_type == rx.DynTensorType(2, "float32")
+ assert b1.var.checked_type == rx.TensorType(2, "float32")
def test_match_cast() -> None:
diff --git a/tests/python/relax/test_struct_info.py b/tests/python/relax/test_struct_info.py
index 33dcd7e9d7..71f9827a14 100644
--- a/tests/python/relax/test_struct_info.py
+++ b/tests/python/relax/test_struct_info.py
@@ -57,9 +57,9 @@ def test_shape_type():
def test_dyn_tensor_type():
- t0 = rx.DynTensorType()
+ t0 = rx.TensorType()
assert t0.ndim == -1
- t1 = rx.DynTensorType(3, "int32")
+ t1 = rx.TensorType(3, "int32")
assert t1.ndim == 3
assert t1.dtype == "int32"
diff --git a/tests/python/relax/test_transform_merge_composite_functions.py b/tests/python/relax/test_transform_merge_composite_functions.py
index 27537edd9e..bbdf49a5cb 100644
--- a/tests/python/relax/test_transform_merge_composite_functions.py
+++ b/tests/python/relax/test_transform_merge_composite_functions.py
@@ -995,7 +995,6 @@ def test_merge_producers_cyclic_dep():
def test_merge_compiler_regions_example():
"""
A tricky example from https://discuss.tvm.apache.org/t/relay-improved-graph-partitioning-algorithm/5830
- See also the corresponding test case for Relay MergeCompilerRegions in relay/test_pass_merge_compiler_regions.py.
"""
check(
MergeCompilerRegionsExample,
diff --git a/tests/python/relax/test_transform_normalize.py b/tests/python/relax/test_transform_normalize.py
index 335ca7c70a..b5559b4972 100644
--- a/tests/python/relax/test_transform_normalize.py
+++ b/tests/python/relax/test_transform_normalize.py
@@ -190,7 +190,7 @@ def test_normalize_if_branches():
# an if node's branches must be seq exprs
x = relax.Var("x", R.Tensor([], "int32"))
y = relax.Var("y", R.Tensor([], "int32"))
- # TODO(@relax-team): z has a shape of () and type of DynTensorType(ndim=0),
+ # TODO(@relax-team): z has a shape of () and type of TensorType(ndim=0),
# but normalization fails to infer these even though it should
z = relax.Var("z")
cond = relax.Var("cond", R.Tensor([], "bool"))
diff --git a/tests/scripts/task_mypy.sh b/tests/scripts/task_mypy.sh
index a8a788f4ce..c8d29237bb 100755
--- a/tests/scripts/task_mypy.sh
+++ b/tests/scripts/task_mypy.sh
@@ -38,15 +38,5 @@ mypy --check-untyped-defs python/tvm/script/printer
echo "Checking MyPy Type defs in the TIR package with unittest"
MYPYPATH=$TVM_PATH/python mypy --check-untyped-defs tests/python/tvmscript/test_tvmscript_type.py
-echo "Checking MyPy Type defs in tvm.relay.op.contrib"
-mypy --disallow-untyped-defs python/tvm/relay/op/contrib/cublas.py
-mypy --disallow-untyped-defs python/tvm/relay/op/contrib/cudnn.py
-mypy --disallow-untyped-defs python/tvm/relay/op/contrib/te_target.py
-mypy --disallow-untyped-defs python/tvm/relay/op/contrib/tensorrt.py
-
-#TODO(@mikepapadim): This is failing atm
-# echo "Checking MyPy Type defs in the tvm.relay.backend.contrib.ethosu package."
-# mypy --check-untyped-defs python/tvm/relay/backend/contrib/ethosu/
-
echo "Checking MyPy Type defs in the tvmscript IRBuilder package."
mypy --check-untyped-defs python/tvm/script/ir_builder