feat: broadcast tiling 模板支持 PcieThrough 场景识别与 UB BRC 强制调度-opbase-AtomGit

已合并

feat: broadcast tiling 模板支持 PcieThrough 场景识别与 UB BRC 强制调度 #788

hahaha22创建于 25 天前从hahaha22/opbase： pcie_through 合入到 cann/opbase： master

feat: broadcast tiling 模板支持 PcieThrough 场景识别与 UB BRC 强制调度 #788

已合并

hahaha22创建于 25 天前

从hahaha22/opbase pcie_through合入到cann/opbase master

共 2 个文件变更+41-4

MCMakeLists.txt+1-0

	@@ -131,6 +131,7 @@ if(BUILD_WITH_INSTALLED_DEPENDENCY_CANN_PKG)
131	$<BUILD_INTERFACE:error_manager_headers>	131	$<BUILD_INTERFACE:error_manager_headers>
132	$<BUILD_INTERFACE:c_sec_headers>	132	$<BUILD_INTERFACE:c_sec_headers>
133	$<BUILD_INTERFACE:asc_host_headers>	133	$<BUILD_INTERFACE:asc_host_headers>
		134	+ $<BUILD_INTERFACE:cann_version_headers>
134	unified_dlog	135	unified_dlog
135	)	136	)
136		137

Mpkg_inc/op_common/atvoss/broadcast/broadcast_tiling.h+40-4

	@@ -17,6 +17,7 @@
17		17
18	#include "broadcast_tiling_base.h"	18	#include "broadcast_tiling_base.h"
19	#include "broadcast_tiling_noncontiguous.h"	19	#include "broadcast_tiling_noncontiguous.h"
		20	+#include "version/metadef_version.h"
20		21
21	namespace Ops {	22	namespace Ops {
22	namespace Base {	23	namespace Base {
	@@ -107,6 +108,13 @@ public:
107	return ge::GRAPH_FAILED;	108	return ge::GRAPH_FAILED;
108	}	109	}
109		110
		111	+ // 3. 获取PcieThrough信息
		112	+ status = GetPcieThroughInfo();
		113	+ if (status != ge::GRAPH_SUCCESS) {
		114	+ OP_LOGE(context_, "Get PcieThrough info failed.");
		115	+ return ge::GRAPH_FAILED;
		116	+ }
		117	+
110	BrcPrintShapes(inShapes, "inShapes");	118	BrcPrintShapes(inShapes, "inShapes");
111	BrcPrintShape(outShape, "outShape");	119	BrcPrintShape(outShape, "outShape");
112	BrcPrintStrides(inStrides, "inStrides");	120	BrcPrintStrides(inStrides, "inStrides");
	@@ -122,7 +130,7 @@ public:
122	OP_LOGI(context_->GetNodeName(), "coreNum: %llu, ubSize: %llu, inputAllContiguous:%d", coreNum, ubSize,	130	OP_LOGI(context_->GetNodeName(), "coreNum: %llu, ubSize: %llu, inputAllContiguous:%d", coreNum, ubSize,
123	inputAllContiguous);	131	inputAllContiguous);
124		132
125	- // 3. 合轴	133	+ // 4. 合轴
126	status = DoDimensionCollapse(broadcastTilingParams, tilingData);	134	status = DoDimensionCollapse(broadcastTilingParams, tilingData);
127	if (status != ge::GRAPH_SUCCESS) {	135	if (status != ge::GRAPH_SUCCESS) {
128	OP_LOGE(context_, "dimension collapse failed.");	136	OP_LOGE(context_, "dimension collapse failed.");
	@@ -132,10 +140,10 @@ public:
132	BrcPrintVectors(tilingData.dims, "after DoDimensionCollapse in&out dims");	140	BrcPrintVectors(tilingData.dims, "after DoDimensionCollapse in&out dims");
133	BrcPrintVectors(tilingData.strides, "after DoDimensionCollapse in&out strides");	141	BrcPrintVectors(tilingData.strides, "after DoDimensionCollapse in&out strides");
134		142
135	- // 4. 根据轴信息判断是否走onedim分支	143	+ // 5. 根据轴信息判断是否走onedim分支
136	OP_CHECK_IF((tilingData.dims.back().size() == 0),	144	OP_CHECK_IF((tilingData.dims.back().size() == 0),
137	OP_LOGE(context_->GetNodeName(), "tensor check is empty, check failed"), return ge::GRAPH_FAILED);	145	OP_LOGE(context_->GetNodeName(), "tensor check is empty, check failed"), return ge::GRAPH_FAILED);
138	- if (tilingData.dims.back().size() == 1 && broadcastTilingParams.inputAllContiguous) {	146	+ if (tilingData.dims.back().size() == 1 && broadcastTilingParams.inputAllContiguous && !isPcieThrough) {
139	if constexpr ((OpDag::InputSize + OpDag::OutputSize <= MAX_IN_OUT_ARGS_NUMBER) && OpDag::VarSize <= 0) {	147	if constexpr ((OpDag::InputSize + OpDag::OutputSize <= MAX_IN_OUT_ARGS_NUMBER) && OpDag::VarSize <= 0) {
140	status = DoOneDimOpTilingAdvance();	148	status = DoOneDimOpTilingAdvance();
141	} else {	149	} else {
	@@ -247,6 +255,27 @@ private:
247	return ge::GRAPH_SUCCESS;	255	return ge::GRAPH_SUCCESS;
248	}	256	}
249		257
		258	+ /**
		259	+ * 判断是否PcieThrough
		260	+ * @return
		261	+ */
		262	+ ge::graphStatus GetPcieThroughInfo()
		263	+ {
		264	+#if defined(METADEF_VERSION_NUM) && METADEF_VERSION_NUM >= 90200000
		265	+ isPcieThrough = context_->GetPcieThroughFlag();
		266	+ OP_LOGD(context_->GetNodeName(), "IsPcieThrough: %s", (isPcieThrough ? "true" : "false"));
		267	+#else
		268	+ isPcieThrough = false;
		269	+ OP_LOGD(context_->GetNodeName(), "IsPcieThrough: false");
		270	+#endif
		271	+ // 非连续模板不支持PCIE Through场景
		272	+ if (isPcieThrough && !inputAllContiguous) {
		273	+ OP_LOGE(context_, "PcieThrough not support noncontiguous.");
		274	+ return ge::GRAPH_FAILED;
		275	+ }
		276	+ return ge::GRAPH_SUCCESS;
		277	+ }
		278	+
250	/**	279	/**
251	* 获取输入输出shape	280	* 获取输入输出shape
252	* @return	281	* @return
	@@ -536,7 +565,13 @@ private:
536	// 1、判断ubbrc用到 2、设置copyIn参数用到	565	// 1、判断ubbrc用到 2、设置copyIn参数用到
537	GetCopyInBrcPosList();	566	GetCopyInBrcPosList();
538		567
539	- bool isUbBroadcast = IsUbBroadcast();	568	+ bool isUbBroadcast;
		569	+ if (isPcieThrough) {
		570	+ isUbBroadcast = true;
		571	+ } else {
		572	+ isUbBroadcast = IsUbBroadcast();
		573	+ }
		574	+
540	int64_t bufferNum = GetBufferNum(isUbBroadcast);	575	int64_t bufferNum = GetBufferNum(isUbBroadcast);
541	OP_LOGI(context_->GetNodeName(), "Broadcast DoBroadcastOpTiling. isUbBroadcast: %d bufferNum: %ld ",	576	OP_LOGI(context_->GetNodeName(), "Broadcast DoBroadcastOpTiling. isUbBroadcast: %d bufferNum: %ld ",
542	isUbBroadcast, bufferNum);	577	isUbBroadcast, bufferNum);
	@@ -1103,6 +1138,7 @@ private:
1103	std::vector<gert::Shape> inShapes;	1138	std::vector<gert::Shape> inShapes;
1104	std::vector<bool> inputIsContiguous;	1139	std::vector<bool> inputIsContiguous;
1105	std::vector<gert::Stride> inStrides;	1140	std::vector<gert::Stride> inStrides;
		1141	+ bool isPcieThrough = false;
1106	bool inputAllContiguous = true;	1142	bool inputAllContiguous = true;
1107	gert::Shape outShape;	1143	gert::Shape outShape;
1108	gert::Stride outStride;	1144	gert::Stride outStride;