adaptivemaxpool3dgradxA5上性能问题修复-ops-nn-AtomGit

已合并

adaptivemaxpool3dgradxA5上性能问题修复 #6799

adaptivemaxpool3dgradxA5上性能问题修复 #6799

已合并

胡一航创建于 6月30日

共 1 个文件变更+2-4

	@@ -19,7 +19,6 @@ using namespace AdaptiveMaxPool3dGradOp;
19	namespace optiling {	19	namespace optiling {
20		20
21	static constexpr uint64_t DCACHE_SIZE = 128 * 1024UL;	21	static constexpr uint64_t DCACHE_SIZE = 128 * 1024UL;
22	-static constexpr int64_t MAX_THREAD_NUM = 1024;
23		22
24	bool AdaptiveMaxPool3dGradTilingSimt::IsCapable()	23	bool AdaptiveMaxPool3dGradTilingSimt::IsCapable()
25	{	24	{
	@@ -51,11 +50,10 @@ ge::graphStatus AdaptiveMaxPool3dGradTilingSimt::PostTiling()
51	blockNum = Ops::Base::CeilDiv(totalNC, detThreadDim);	50	blockNum = Ops::Base::CeilDiv(totalNC, detThreadDim);
52	} else {	51	} else {
53	int64_t outDataCount = inputData.nX * inputData.cX * inputData.dGrad * inputData.hGrad * inputData.wGrad;	52	int64_t outDataCount = inputData.nX * inputData.cX * inputData.dGrad * inputData.hGrad * inputData.wGrad;
54	- int64_t threads = std::min(outDataCount, MAX_THREAD_NUM);	53	+ blockNum = (outDataCount == 0) ? 1 : std::min(outDataCount, static_cast<int64_t>(coreNum_));
55	- blockNum = Ops::Base::CeilDiv(outDataCount, threads);
56	}	54	}
57	- blockNum = std::max(blockNum, static_cast<int64_t>(1));
58	blockNum = std::min(blockNum, static_cast<int64_t>(coreNum_));	55	blockNum = std::min(blockNum, static_cast<int64_t>(coreNum_));
		56	+ blockNum = std::max(blockNum, static_cast<int64_t>(1));
59	context_->SetBlockDim(blockNum);	57	context_->SetBlockDim(blockNum);
60	context_->SetLocalMemorySize(ubSize_ - DCACHE_SIZE);	58	context_->SetLocalMemorySize(ubSize_ - DCACHE_SIZE);
61	return ge::GRAPH_SUCCESS;	59	return ge::GRAPH_SUCCESS;