You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2022/06/02 00:46:21 UTC
[GitHub] [tvm] mbs-octoml commented on a diff in pull request #11474: [BYOC] Two helper passes for external codegen using RelayToTIR custom pass machinery
mbs-octoml commented on code in PR #11474:
URL: https://github.com/apache/tvm/pull/11474#discussion_r887406423
##########
tests/python/relay/transform/test_compiler_function_utils.py:
##########
@@ -0,0 +1,162 @@
+# 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
+"""Unit tests for the OutlineCompilerFunctionsWithExistingGlobalSymbols and
+ MarkCompilerFunctionsAsExtern external codegen helper passes."""
+
+import tvm
+import tvm.testing
+import numpy as np
+
+
+def make_const(dtype, shape):
+ return tvm.relay.const(np.random.rand(*shape).astype(dtype))
+
+
+def make_consts(dtype, shapes):
+ return [make_const(dtype, shape) for shape in shapes]
+
+
+metatable = {
+ "relay.Constant": make_consts(
+ "float16",
+ [
+ (2304, 768), # 0
+ (2304,), # 1
+ (600, 32, 64), # 2
+ ],
+ ),
+ "attributes": [{"relay_attrs": None}],
+}
+
+
+def inlined_mod():
+ return tvm.parser.parse(
+ """
+ #[version = "0.0.5"]
+ def @main(%x0 : Tensor[(1600, 768), float16], %x3 : Tensor[(600, 32, 64), float16]) -> (Tensor[(1600, 2304), float16], Tensor[(600, 32, 32), float16]) {
+ %0 = fn(%y_0_i0: Tensor[(1600, 768), float16], %y_0_i1: Tensor[(2304, 768), float16], %y_0_i2: Tensor[(2304), float16],
+ Inline=1, Compiler="cutlass", global_symbol="tvmgen_default_cutlass_main_0", Primitive=1) -> Tensor[(1600, 2304), float16] {
+ %4 = fn (%FunctionVar_0_0: Tensor[(1600, 768), float16], %FunctionVar_0_1: Tensor[(2304, 768), float16], %FunctionVar_0_2: Tensor[(2304), float16],
+ PartitionedFromPattern="nn.dense_add_", Composite="cutlass.dense_bias") -> Tensor[(1600, 2304), float16] {
+ %5 = nn.dense(%FunctionVar_0_0, %FunctionVar_0_1, units=2304);
+ add(%5, %FunctionVar_0_2)
+ };
+ %4(%y_0_i0, %y_0_i1, %y_0_i2)
+ };
+ %1 = %0(%x0, meta[relay.Constant][0], meta[relay.Constant][1]);
+ %2 = fn(%y_3_i0: Tensor[(600, 32, 64), float16], %y_3_i1: Tensor[(600, 32, 64), float16],
+ Inline=1, Compiler="cublas", global_symbol="tvmgen_default_cublas_main_3", Primitive=1) -> Tensor[(600, 32, 32), float16] {
+ %6 = fn (%FunctionVar_0_01: Tensor[(600, 32, 64), float16], %FunctionVar_0_11: Tensor[(600, 32, 64), float16],
+ PartitionedFromPattern="nn.batch_matmul_", Composite="cublas.batch_matmul") -> Tensor[(600, 32, 32), float16] {
+ nn.batch_matmul(%FunctionVar_0_01, %FunctionVar_0_11, out_dtype="float16", transpose_b=True)
+ };
+ %6(%y_3_i0, %y_3_i1)
+ };
+ %3 = %2(%x3, meta[relay.Constant][2]);
+ (%1, %3)
+ }
+ """,
+ "from_string",
+ None,
+ metatable,
+ )
+
+
+def expected_outlined_mod():
+ return tvm.parser.parse(
+ """
+ #[version = "0.0.5"]
+ def @main(%x0 : Tensor[(1600, 768), float16], %x3 : Tensor[(600, 32, 64), float16]) -> (Tensor[(1600, 2304), float16], Tensor[(600, 32, 32), float16]) {
+ %1 = @tvmgen_default_cutlass_main_0(%x0, meta[relay.Constant][0], meta[relay.Constant][1]);
+ %2 = fn(%y_3_i0: Tensor[(600, 32, 64), float16], %y_3_i1: Tensor[(600, 32, 64), float16],
+ Inline=1, Compiler="cublas", global_symbol="tvmgen_default_cublas_main_3", Primitive=1) -> Tensor[(600, 32, 32), float16] {
+ %6 = fn (%FunctionVar_0_01: Tensor[(600, 32, 64), float16], %FunctionVar_0_11: Tensor[(600, 32, 64), float16],
+ PartitionedFromPattern="nn.batch_matmul_", Composite="cublas.batch_matmul") -> Tensor[(600, 32, 32), float16] {
+ nn.batch_matmul(%FunctionVar_0_01, %FunctionVar_0_11, out_dtype="float16", transpose_b=True)
+ };
+ %6(%y_3_i0, %y_3_i1)
+ };
+ %3 = %2(%x3, meta[relay.Constant][2]);
+ (%1, %3)
+ }
+
+ def @tvmgen_default_cutlass_main_0(%y_0_i0: Tensor[(1600, 768), float16], %y_0_i1: Tensor[(2304, 768), float16], %y_0_i2: Tensor[(2304), float16],
+ Inline=1, Compiler="cutlass", global_symbol="tvmgen_default_cutlass_main_0", Primitive=1) -> Tensor[(1600, 2304), float16] {
Review Comment:
AFAIKT only PartitionGraph sets it, and it triggers Inline to actually do the inline. I guess at the time folks were thinking of supporting a kind of inline profit analysis separate from the inline pass itself?
The 'outlining' pass will preserve the attribute, along with all the other attributes. So, nothing stopping someone running Inline to push them back in again if that's what they really want!
The 'make extern' pass however intentionally drops all attributes since these functions are now just stubs and the body is no longer meaningful.
In an ideal world I think:
- "Compiler" only makes sense on global functions
- The graph and aot flows would respect the entire IRModule and not be restricted to seeing just 'main', hence there'd never be a need to inline anything other than for perf.
- IRModule would support the notion of extern functions (along with extern/defined constants, and runtime modules)
But I'm trying to work with what we have :-)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org