masahi commented on code in PR #11088: URL: https://github.com/apache/tvm/pull/11088#discussion_r855988757
########## src/meta_schedule/postproc/rewrite_tensorize.cc: ########## @@ -0,0 +1,104 @@ +/* + * 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. + */ +#include <tvm/runtime/container/base.h> + +#include <algorithm> + +#include "../utils.h" + +namespace tvm { +namespace meta_schedule { + +using tir::BlockRV; +using tir::LoopRV; + +void ApplyTensorization(const tir::Schedule& sch, const String& func_name, + const tir::PrimFuncNode* func, bool vectorize_init_loop) { + std::vector<std::pair<std::string, std::function<void(tir::BlockRV)>>> jobs; + + tir::PostOrderVisit(func->body, [=, &jobs](const ObjectRef& obj) { + if (const auto* block = obj.as<tir::BlockNode>()) { + tir::StmtSRef block_sref = sch->GetSRef(block); + if (Optional<String> intrin_name = + tir::GetAnn<String>(block_sref, tir::attr::meta_schedule_auto_tensorize)) { + std::string block_name = block_sref->StmtAs<tir::BlockNode>()->name_hint; + if (block_name.find("init") == std::string::npos) { + jobs.emplace_back(block_name, [sch, intrin_name](tir::BlockRV block) { + try { + sch->Tensorize(block, intrin_name.value()); + } catch (const std::exception& e) { + LOG(WARNING) << "Tensorize failed with error " << e.what(); + } + }); + } else if (vectorize_init_loop) { + jobs.emplace_back(block_name, [sch](tir::BlockRV block) { + Array<BlockRV> child_blocks = sch->GetChildBlocks(block); + ICHECK(child_blocks.size() == 1); + Array<LoopRV> init_loops = sch->GetLoops(child_blocks[0]); + ICHECK(init_loops.size() == 1); + sch->Vectorize(init_loops[0]); Review Comment: The issue in question is vectorization for CPU targets. I'm using the default postprocs in https://github.com/apache/tvm/blob/effc23df7cb4c392f6957c94157da33293793eb7/python/tvm/meta_schedule/tune.py#L96-L103 Since loop parallelization or vectorization checks for the "compact dataflow" constraint, https://github.com/apache/tvm/blob/0ddaaa6a7d1009ea7ca8313a51eb19abb8ae7699/src/tir/schedule/primitive/for_kind.cc#L160, they need to be applied before `DecomposeReduction` in `RewriteReductionBlock()`. So having `RewriteParallelVectorizeUnroll` before `RewriteReductionBlock()` in the default postprocs makes sense. However, this is not sufficient to vectorize the init loop of reduction block, since it is generated during `RewriteReductionBlock()`. I don't think we should run `RewriteParallelVectorizeUnroll` again after `RewriteReductionBlock()` (and it doesn't work anyway), so we need to manually vectorize the decomposed init loop in `RewriteReductionBlock` or the new `RewriteTensorize` postproc I added. I prefer the former. -- 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: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
