feat: mul/select 算子支持 PcieThrough 场景识别与回退-ops-math-AtomGit

Mmath/mul/op_host/mul_infershape.cpp+28-21

	@@ -16,33 +16,40 @@
16	#include "log/log.h"	16	#include "log/log.h"
17	#include "infershape_broadcast_util.h"	17	#include "infershape_broadcast_util.h"
18	#include "register/op_impl_registry.h"	18	#include "register/op_impl_registry.h"
		19	+#include "version/metadef_version.h"
19		20
20	using namespace ge;	21	using namespace ge;
21	namespace ops {	22	namespace ops {
22	-static std::string ShapeCannotBroadcastMsg(const gert::Shape& shape1, const gert::Shape& shape2) {	23	+static std::string ShapeCannotBroadcastMsg(const gert::Shape& shape1, const gert::Shape& shape2)
23	- std::string res = "shape ";	24	+{
24	- res += Ops::Base::ToString(shape1);	25	+ std::string res = "shape ";
25	- res += " and ";	26	+ res += Ops::Base::ToString(shape1);
26	- res += Ops::Base::ToString(shape2);	27	+ res += " and ";
27	- res += " cannot broadcast!";	28	+ res += Ops::Base::ToString(shape2);
28	- return res;	29	+ res += " cannot broadcast!";
		30	+ return res;
29	}	31	}
30		32
31	-static ge::graphStatus InferShape4Broadcast(gert::InferShapeContext* context) {	33	+static ge::graphStatus InferShape4Broadcast(gert::InferShapeContext* context)
32	- auto in_shape1 = context->GetInputShape(0);	34	+{
33	- OP_CHECK_NULL_WITH_CONTEXT(context, in_shape1);	35	+ auto in_shape1 = context->GetInputShape(0);
34	- auto in_shape2 = context->GetInputShape(1);	36	+ OP_CHECK_NULL_WITH_CONTEXT(context, in_shape1);
35	- OP_CHECK_NULL_WITH_CONTEXT(context, in_shape2);	37	+ auto in_shape2 = context->GetInputShape(1);
36	- auto out_shape = context->GetOutputShape(0);	38	+ OP_CHECK_NULL_WITH_CONTEXT(context, in_shape2);
37	- OP_CHECK_NULL_WITH_CONTEXT(context, out_shape);	39	+ auto out_shape = context->GetOutputShape(0);
		40	+ OP_CHECK_NULL_WITH_CONTEXT(context, out_shape);
38		41
39	- OP_CHECK_IF((!Ops::Base::BroadcastShape(in_shape1, in_shape2, out_shape)),	42	+ OP_CHECK_IF((!Ops::Base::BroadcastShape(in_shape1, in_shape2, out_shape)),
40	- OP_LOGE(context->GetNodeName(), "%s", ShapeCannotBroadcastMsg(in_shape2, in_shape1).c_str()),	43	+ OP_LOGE(context->GetNodeName(), "%s", ShapeCannotBroadcastMsg(in_shape2, in_shape1).c_str()),
41	- return ge::GRAPH_FAILED);	44	+ return ge::GRAPH_FAILED);
42		45
43	- return ge::GRAPH_SUCCESS;	46	+ return ge::GRAPH_SUCCESS;
44	}	47	}
45		48
46	-IMPL_OP_INFERSHAPE(Mul).InferShape(InferShape4Broadcast);	49	+IMPL_OP_INFERSHAPE(Mul)
47	-	50	+ .InferShape(InferShape4Broadcast)
48	-}	51	+#if defined(METADEF_VERSION_NUM) && METADEF_VERSION_NUM >= 90200000
		52	+ .SetSupportPcieThrough()
		53	+#endif
		54	+ ;
		55	+} // namespace ops

Mmath/select/op_host/arch35/select_simt_tiling.cpp+16-0

	@@ -21,6 +21,7 @@
21	#include "register/op_def_registry.h"	21	#include "register/op_def_registry.h"
22	#include "util/math_util.h"	22	#include "util/math_util.h"
23	#include "op_host/math_tiling_templates_registry.h"	23	#include "op_host/math_tiling_templates_registry.h"
		24	+#include "version/metadef_version.h"
24		25
25	using namespace AscendC;	26	using namespace AscendC;
26	using namespace ge;	27	using namespace ge;
	@@ -104,8 +105,23 @@ bool SelectSimtTiling::XDtypeImprove()
104	return false;	105	return false;
105	}	106	}
106		107
		108	+bool SelectSimtTiling::IsPcieThrough()
		109	+{
		110	+#if defined(METADEF_VERSION_NUM) && METADEF_VERSION_NUM >= 90200000
		111	+ bool isPcieThrough = context_->GetPcieThroughFlag();
		112	+ OP_LOGD(context_->GetNodeName(), "IsPcieThrough: %s", (isPcieThrough ? "true" : "false"));
		113	+ return isPcieThrough;
		114	+#else
		115	+ OP_LOGD(context_->GetNodeName(), "IsPcieThrough: false");
		116	+ return false;
		117	+#endif
		118	+}
		119	+
107	bool SelectSimtTiling::IsCapable()	120	bool SelectSimtTiling::IsCapable()
108	{	121	{
		122	+ if (IsPcieThrough()) {
		123	+ return false;
		124	+ }
109	if (!IsMatchAB()) {	125	if (!IsMatchAB()) {
110	return false;	126	return false;
111	}	127	}

Mmath/select/op_host/arch35/select_simt_tiling.h+29-28

	@@ -26,40 +26,41 @@
26	namespace optiling {	26	namespace optiling {
27		27
28	class SelectSimtTiling : public Ops::Base::TilingBaseClass {	28	class SelectSimtTiling : public Ops::Base::TilingBaseClass {
29	- public:	29	+public:
30	- explicit SelectSimtTiling(gert::TilingContext* context) : Ops::Base::TilingBaseClass(context) {}	30	+ explicit SelectSimtTiling(gert::TilingContext* context) : Ops::Base::TilingBaseClass(context) {}
31		31
32	- protected:	32	+protected:
33	- bool IsCapable() override;	33	+ bool IsCapable() override;
34	- ge::graphStatus GetPlatformInfo() override;	34	+ ge::graphStatus GetPlatformInfo() override;
35	- ge::graphStatus GetShapeAttrsInfo() override;	35	+ ge::graphStatus GetShapeAttrsInfo() override;
36	- ge::graphStatus DoOpTiling() override;	36	+ ge::graphStatus DoOpTiling() override;
37	- ge::graphStatus DoLibApiTiling() override;	37	+ ge::graphStatus DoLibApiTiling() override;
38	- uint64_t GetTilingKey() const override;	38	+ uint64_t GetTilingKey() const override;
39	- ge::graphStatus GetWorkspaceSize() override;	39	+ ge::graphStatus GetWorkspaceSize() override;
40	- ge::graphStatus PostTiling() override;	40	+ ge::graphStatus PostTiling() override;
41		41
42	- private:	42	+private:
43	- bool IsMatchAB();	43	+ bool IsPcieThrough();
44	- bool XDtypeImprove();	44	+ bool IsMatchAB();
		45	+ bool XDtypeImprove();
45		46
46	- uint64_t tilingKey = 0;	47	+ uint64_t tilingKey = 0;
47	- int64_t aivNum_;	48	+ int64_t aivNum_;
48	- int64_t ubSize_ = 0;	49	+ int64_t ubSize_ = 0;
49	- int64_t needCoreNum_ = 0;	50	+ int64_t needCoreNum_ = 0;
50	- gert::Shape conditionShape_;	51	+ gert::Shape conditionShape_;
51	- gert::Shape x1Shape_;	52	+ gert::Shape x1Shape_;
52	- gert::Shape x2Shape_;	53	+ gert::Shape x2Shape_;
53	- int64_t aSize_ = 1;	54	+ int64_t aSize_ = 1;
54	- int64_t bSize_ = 1;	55	+ int64_t bSize_ = 1;
55	- int64_t ySize_ = 1;	56	+ int64_t ySize_ = 1;
56	#ifdef DAVID_FPGA	57	#ifdef DAVID_FPGA
57	- int64_t threadNum_ = 128;	58	+ int64_t threadNum_ = 128;
58	#else	59	#else
59	- int64_t threadNum_ = 2048;	60	+ int64_t threadNum_ = 2048;
60	#endif	61	#endif
61	};	62	};
62		63
63	-} // namespace optiling	64	+} // namespace optiling
64		65
65	-#endif // OPS_BUILD_IN_OP_TILING_RUNTIME_SELECT_SIMT_TILING_H	66	+#endif // OPS_BUILD_IN_OP_TILING_RUNTIME_SELECT_SIMT_TILING_H

Mmath/select/op_host/select_infershape.cpp+16-9

	@@ -14,18 +14,25 @@
14	*/	14	*/
15	#include "log/log.h"	15	#include "log/log.h"
16	#include "register/op_impl_registry.h"	16	#include "register/op_impl_registry.h"
		17	+#include "version/metadef_version.h"
17		18
18	using namespace ge;	19	using namespace ge;
19	namespace ops {	20	namespace ops {
20	-static graphStatus InferShape4Select(gert::InferShapeContext* context) {	21	+static graphStatus InferShape4Select(gert::InferShapeContext* context)
21	- const gert::Shape* x1_shape = context->GetInputShape(1);	22	+{
22	- OP_CHECK_NULL_WITH_CONTEXT(context, x1_shape);	23	+ const gert::Shape* x1_shape = context->GetInputShape(1);
		24	+ OP_CHECK_NULL_WITH_CONTEXT(context, x1_shape);
23		25
24	- gert::Shape* y_shape = context->GetOutputShape(0);	26	+ gert::Shape* y_shape = context->GetOutputShape(0);
25	- OP_CHECK_NULL_WITH_CONTEXT(context, y_shape);	27	+ OP_CHECK_NULL_WITH_CONTEXT(context, y_shape);
26	- y_shape = x1_shape;	28	+ y_shape = x1_shape;
27		29
28	- return GRAPH_SUCCESS;	30	+ return GRAPH_SUCCESS;
29	-}
30	-IMPL_OP_INFERSHAPE(Select).InferShape(InferShape4Select);
31	}	31	}
		32	+IMPL_OP_INFERSHAPE(Select)
		33	+ .InferShape(InferShape4Select)
		34	+#if defined(METADEF_VERSION_NUM) && METADEF_VERSION_NUM >= 90200000
		35	+ .SetSupportPcieThrough()
		36	+#endif
		37	+ ;
		38	+} // namespace ops