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 2021/02/23 00:16:18 UTC
[GitHub] [tvm] mbrookhart commented on a change in pull request #7477: SparseReshape Op
mbrookhart commented on a change in pull request #7477:
URL: https://github.com/apache/tvm/pull/7477#discussion_r580685169
##########
File path: python/tvm/topi/cuda/sparse_reshape.py
##########
@@ -0,0 +1,232 @@
+# 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, too-many-arguments, too-many-nested-blocks
+"""Sparse_Reshape operator"""
+import tvm
+from tvm import te
+from ...tir import decl_buffer, ir_builder, Cast
+from ...te import extern, div, floordiv, floormod
+from ..utils import ceil_div
+
+
+def sparse_reshape(
+ sparse_indices,
+ prev_shape,
+ new_shape,
+ new_sparse_indices_shape,
+ new_shape_shape,
+):
+ """
+ Reshape a Sparse Tensor
+ Parameters
+ ----------
+ sparse_indices : relay.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
+ A 1-D tensor containing the previous shape of the dense tensor
+ new_shape : relay.Expr
+ A 1-D tensor containing the new shape of the dense tensor
+ Returns
+ -------
+ result: relay.Expr
+ Output tensor.
+ Examples
+ --------
+ .. code-block:: python
+ sparse_indices = [[0, 0, 0],
+ [0, 0, 1],
+ [0, 1, 0],
+ [1, 0, 0],
+ [1, 2, 3]]
+ prev_shape = [2, 3, 4]
+ new_shape = [9, -1]
+ new_sparse_indices, new_shape = relay.sparse_reshape(sparse_indices,
+ prev_shape,
+ new_shape)
+ new_sparse_indices = [[0, 0],
+ [0, 1],
+ [1, 2],
+ [4, 2],
+ [8, 1]]
+ new_shape = [9, 4]
+ """
+
+ def gen_ir(
+ sparse_indices_ptr,
+ prev_shape_ptr,
+ new_shape_ptr,
+ new_sparse_indices_ptr,
+ out_new_shape_ptr,
+ ):
+ ib = ir_builder.create()
+
+ sparse_indices = ib.buffer_ptr(sparse_indices_ptr)
+ prev_shape = ib.buffer_ptr(prev_shape_ptr)
+
+ new_shape = ib.buffer_ptr(new_shape_ptr)
+ out_new_shape = ib.buffer_ptr(out_new_shape_ptr)
+ new_sparse_indices = ib.buffer_ptr(new_sparse_indices_ptr)
+ out_new_shape = ib.buffer_ptr(out_new_shape_ptr)
+
+ prev_shape_size = prev_shape_ptr.shape[0]
+ new_shape_size = new_shape_ptr.shape[0]
+
+ multipliers = ib.allocate("int64", (prev_shape_size,), name="multipliers", scope="local")
+ dividers = ib.allocate("int64", (new_shape_size,), name="dividers", scope="local")
+ flattened_indices = ib.allocate(
+ "int64", (sparse_indices_ptr.shape[0],), name="flattened_indices", scope="local"
+ )
+ total_ele = ib.allocate("int64", (1,), name="total_ele", scope="local")
+ division_total_ele = ib.allocate("int64", (1,), name="division_total_ele", scope="local")
+ equal_shape = ib.allocate("bool", (1,), name="equal_shape", scope="local")
+ max_threads = int(tvm.target.Target.current(allow_none=False).max_num_threads)
+ with ib.new_scope():
+ # The computation in this block is very very miniscule since we are just iterating over
+ # shape tensors which are very small (< 10) and there is no need of parallelization
+ nthread_tx = 1
+ nthread_bx = 1
+ tx = te.thread_axis("threadIdx.x")
+ bx = te.thread_axis("blockIdx.x")
+ ib.scope_attr(tx, "thread_extent", nthread_tx)
+ ib.scope_attr(bx, "thread_extent", nthread_bx)
+
+ total_ele[0] = prev_shape[0]
+
+ # Cumulative Reverse Exclusive Multiply
+ multipliers[prev_shape_size - 1] = Cast("int64", 1)
+ with ib.for_range(0, prev_shape_size - 1) as i_:
+ i = i_ + 1
+ multipliers[prev_shape_size - 1 - i] = (
+ prev_shape[prev_shape_size - i] * multipliers[prev_shape_size - i]
+ )
+ total_ele[0] *= prev_shape[prev_shape_size - i]
+
+ division_total_ele[0] = Cast("int64", 1)
+ with ib.for_range(0, new_shape_size) as i:
+ with ib.if_scope(new_shape[i] != -1):
+ division_total_ele[0] *= new_shape[i]
+
+ # Compute true output shape (replace negative ones)
+ with ib.for_range(0, new_shape_size) as i:
+ with ib.if_scope(new_shape[i] == -1):
+ out_new_shape[i] = Cast("int64", div(total_ele[0], division_total_ele[0]))
+ with ib.else_scope():
+ out_new_shape[i] = new_shape[i]
+
+ # Check if prev_shape and new_shape are equal
+ equal_shape[0] = True
+ with ib.if_scope(prev_shape_size == new_shape_size):
+ with ib.for_range(0, prev_shape_size) as i:
+ with ib.if_scope(prev_shape[i] != out_new_shape[i]):
+ equal_shape[0] = False
+ with ib.else_scope():
+ equal_shape[0] = False
+
Review comment:
There is no global sync, you need to break the cuda kernel, go back to host, and then launch a new cuda kernel to get a global sync
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org