fix: 拦截不支持ScalarData的codegen输入(#161)-graph-autofusion-AtomGit

已合并

fix: 拦截不支持ScalarData的codegen输入(#161) #1209

zqmin创建于 7月6日从zqmin/graph-autofusion： fix/bitwiseand-scalardata-guard 合入到 cann/graph-autofusion： develop

fix: 拦截不支持ScalarData的codegen输入(#161) #1209

已合并

zqmin创建于 7月6日

从zqmin/graph-autofusion fix/bitwiseand-scalardata-guard合入到cann/graph-autofusion develop

共 2 个文件变更+103-2

Mautofuse/ascir/generator/ascir_common.cpp+7-2

	@@ -14,6 +14,11 @@
14		14
15	namespace af {	15	namespace af {
16	namespace ascir {	16	namespace ascir {
		17	+namespace {
		18	+bool IsScalarInputType(const std::string &node_type) {
		19	+ return node_type == "Scalar" \|\| node_type == "ScalarData";
		20	+}
		21	+} // namespace
17		22
18	bool OnlySecondInputSupportScalar(const std::vector<bool> &is_scalar_list) {	23	bool OnlySecondInputSupportScalar(const std::vector<bool> &is_scalar_list) {
19	GE_ASSERT_EQ(is_scalar_list.size(), 2UL);	24	GE_ASSERT_EQ(is_scalar_list.size(), 2UL);
	@@ -213,7 +218,7 @@ Status ValidateShapeConsistencyWithSingleOutput(const AscNode &node, const Broad
213	bool IsNodeHasScalarInput(const AscNode &node) {	218	bool IsNodeHasScalarInput(const AscNode &node) {
214	AscNodeInputs node_inputs = node.inputs;	219	AscNodeInputs node_inputs = node.inputs;
215	for (size_t i = 0; i < node_inputs().size(); i++) {	220	for (size_t i = 0; i < node_inputs().size(); i++) {
216	- if (node.GetInDataNodes().at(i)->GetType() == "Scalar") {	221	+ if (IsScalarInputType(node.GetInDataNodes().at(i)->GetType())) {
217	return true;	222	return true;
218	}	223	}
219	}	224	}
	@@ -221,7 +226,7 @@ bool IsNodeHasScalarInput(const AscNode &node) {
221	}	226	}
222		227
223	bool IsNodeFirstInputScalar(const AscNode &node) {	228	bool IsNodeFirstInputScalar(const AscNode &node) {
224	- return node.GetInDataNodes().at(0)->GetType() == "Scalar";	229	+ return IsScalarInputType(node.GetInDataNodes().at(0)->GetType());
225	}	230	}
226	} // namespace ascir	231	} // namespace ascir
227	} // namespace af	232	} // namespace af

Mautofuse/tests/ut/codegen/test_codegen_graph_check.cpp+96-0

	@@ -32,6 +32,51 @@ namespace {
32	std::string ToString(const af::Expression &e) {	32	std::string ToString(const af::Expression &e) {
33	return std::string(e.Serialize().get());	33	return std::string(e.Serialize().get());
34	}	34	}
		35	+
		36	+void SetGlobalOutput(const af::AscNodePtr &node, uint32_t tensor_id) {
		37	+ node->outputs[0].attr.mem.alloc_type = af::AllocType::kAllocTypeGlobal;
		38	+ node->outputs[0].attr.mem.tensor_id = tensor_id;
		39	+}
		40	+
		41	+void SetScalarLikeQueueOutput(const af::AscNodePtr &node, const af::Axis &axis, const af::Expression &repeat,
		42	+ uint32_t tensor_id, uint32_t queue_id) {
		43	+ auto &attr = node->outputs[0].attr;
		44	+ attr.axis = {axis.id};
		45	+ attr.vectorized_axis = {axis.id};
		46	+ attr.vectorized_strides = {af::ops::Zero};
		47	+ attr.repeats = {repeat};
		48	+ attr.strides = {af::ops::Zero};
		49	+ attr.mem.alloc_type = af::AllocType::kAllocTypeQueue;
		50	+ attr.mem.tensor_id = tensor_id;
		51	+ attr.que.id = queue_id;
		52	+ attr.mem.reuse_id = 0;
		53	+ attr.que.depth = 2;
		54	+ attr.que.buf_num = 2;
		55	+ attr.opt.merge_scope = af::kIdNone;
		56	+}
		57	+
		58	+void SetVectorQueueOutput(const af::AscNodePtr &node, const af::Axis &axis, const af::Expression &repeat,
		59	+ af::Position position, uint32_t tensor_id, uint32_t queue_id) {
		60	+ auto &attr = node->outputs[0].attr;
		61	+ attr.axis = {axis.id};
		62	+ attr.vectorized_axis = {axis.id};
		63	+ attr.vectorized_strides = {af::ops::One};
		64	+ attr.repeats = {repeat};
		65	+ attr.strides = {af::ops::One};
		66	+ attr.mem.position = position;
		67	+ attr.mem.alloc_type = af::AllocType::kAllocTypeQueue;
		68	+ attr.mem.tensor_id = tensor_id;
		69	+ attr.que.id = queue_id;
		70	+ attr.opt.merge_scope = af::kIdNone;
		71	+}
		72	+
		73	+void SetVectorInputQueueOutput(const af::AscNodePtr &node, const af::Axis &axis, const af::Expression &repeat,
		74	+ uint32_t tensor_id, uint32_t queue_id) {
		75	+ SetVectorQueueOutput(node, axis, repeat, af::Position::kPositionVecIn, tensor_id, queue_id);
		76	+ node->outputs[0].attr.mem.reuse_id = 0;
		77	+ node->outputs[0].attr.que.depth = 2;
		78	+ node->outputs[0].attr.que.buf_num = 2;
		79	+}
35	} // namespace	80	} // namespace
36		81
37	TEST(CodegenKernel, Kernel_DynamicInputDtypeCheck) {	82	TEST(CodegenKernel, Kernel_DynamicInputDtypeCheck) {
	@@ -443,6 +488,57 @@ TEST(CodegenKernel, Kernel_DynamicShapeConsistencyCheckValid) {
443	EXPECT_EQ(ret, af::SUCCESS);	488	EXPECT_EQ(ret, af::SUCCESS);
444	}	489	}
445		490
		491	+TEST(CodegenKernel, Kernel_BitwiseAndScalarDataInputInvalidCheck) {
		492	+ af::AscGraph graph("test_graph");
		493	+ const af::Expression s0 = graph.CreateSizeVar(8);
		494	+ auto z0 = graph.CreateAxis("z0", s0);
		495	+
		496	+ af::ascir_op::Data x_op("x", graph);
		497	+ x_op.ir_attr.SetIndex(0);
		498	+ af::ascir_op::ScalarData scalar_op("scalar", graph);
		499	+ scalar_op.ir_attr.SetIndex(1);
		500	+
		501	+ af::ascir_op::Load load_op("load");
		502	+ af::ascir_op::BitwiseAnd bitwise_and_op("bitwise_and");
		503	+ af::ascir_op::Store store_op("store");
		504	+ af::ascir_op::Output y_op("y");
		505	+ y_op.ir_attr.SetIndex(0);
		506	+
		507	+ x_op.y.dtype = ge::DT_UINT8;
		508	+ scalar_op.y.dtype = ge::DT_UINT8;
		509	+ load_op.x = x_op.y;
		510	+ load_op.y.dtype = ge::DT_UINT8;
		511	+ bitwise_and_op.x1 = load_op.y;
		512	+ bitwise_and_op.x2 = scalar_op.y;
		513	+ bitwise_and_op.y.dtype = ge::DT_UINT8;
		514	+ store_op.x = bitwise_and_op.y;
		515	+ store_op.y.dtype = ge::DT_UINT8;
		516	+ y_op.x = store_op.y;
		517	+ y_op.y.dtype = ge::DT_UINT8;
		518	+
		519	+ auto x = graph.FindNode("x");
		520	+ auto scalar = graph.FindNode("scalar");
		521	+ auto load = graph.FindNode("load");
		522	+ auto bitwise_and = graph.FindNode("bitwise_and");
		523	+ auto store = graph.FindNode("store");
		524	+ ASSERT_NE(x, nullptr);
		525	+ ASSERT_NE(scalar, nullptr);
		526	+ ASSERT_NE(load, nullptr);
		527	+ ASSERT_NE(bitwise_and, nullptr);
		528	+ ASSERT_NE(store, nullptr);
		529	+
		530	+ SetGlobalOutput(x, 0);
		531	+ SetScalarLikeQueueOutput(scalar, z0, s0, 1, 0);
		532	+ SetVectorInputQueueOutput(load, z0, s0, 2, 1);
		533	+
		534	+ bitwise_and->attr.api.unit = af::ComputeUnit::kUnitVector;
		535	+ SetVectorQueueOutput(bitwise_and, z0, s0, af::Position::kPositionVecOut, 3, 2);
		536	+ SetGlobalOutput(store, 4);
		537	+
		538	+ auto ret = CheckGraphValidity(graph);
		539	+ EXPECT_NE(ret, af::SUCCESS);
		540	+}
		541	+
446	TEST(CodegenKernel, Kernel_StaticShapeVecAxisConsistencyInValidCheck) {	542	TEST(CodegenKernel, Kernel_StaticShapeVecAxisConsistencyInValidCheck) {
447	af::AscGraph graph("test_graph");	543	af::AscGraph graph("test_graph");
448	const af::Expression s0 = graph.CreateSizeVar(3);	544	const af::Expression s0 = graph.CreateSizeVar(3);