【fix】: 修复Rand构图问题-graph-autofusion-AtomGit

已合并

【fix】: 修复Rand构图问题 #1964

【fix】: 修复Rand构图问题 #1964

已合并

liyuewei创建于 9月4日

共 4 个文件变更+17-11

	@@ -1028,7 +1028,8 @@ af::Status DefaultMTE3Api([[maybe_unused]] const std::vector<TensorShapeInfo> &i
1028	af::Status DefaultVECApi([[maybe_unused]] const std::vector<TensorShapeInfo> &input_shapes,	1028	af::Status DefaultVECApi([[maybe_unused]] const std::vector<TensorShapeInfo> &input_shapes,
1029	[[maybe_unused]] const std::vector<TensorShapeInfo> &output_shapes,	1029	[[maybe_unused]] const std::vector<TensorShapeInfo> &output_shapes,
1030	[[maybe_unused]] const NodeInfo &node, PerfOutputInfo &perf_res) {	1030	[[maybe_unused]] const NodeInfo &node, PerfOutputInfo &perf_res) {
1031	- GE_ASSERT_TRUE(!input_shapes.empty() && !output_shapes.empty());	1031	+ // rand/randn have no input shape, so check output shape only
		1032	+ GE_ASSERT_TRUE(!output_shapes.empty());
1032	perf_res.pipe_res[PipeType::AIV_VEC] = af::sym::kSymbolOne;	1033	perf_res.pipe_res[PipeType::AIV_VEC] = af::sym::kSymbolOne;
1033	return af::SUCCESS;	1034	return af::SUCCESS;
1034	}	1035	}

	@@ -133,8 +133,8 @@ static const std::map<std::string, af::ComputeType> kOpTypeToComputeType = {
133	{Scalar::Type, af::ComputeType::kComputeInvalid},	133	{Scalar::Type, af::ComputeType::kComputeInvalid},
134	{Output::Type, af::ComputeType::kComputeInvalid},	134	{Output::Type, af::ComputeType::kComputeInvalid},
135	{IndexExpr::Type, af::ComputeType::kComputeInvalid},	135	{IndexExpr::Type, af::ComputeType::kComputeInvalid},
136	- {"Rand", af::ComputeType::kComputeInvalid},	136	+ {Rand::Type, af::ComputeType::kComputeElewise},
137	- {"Randn", af::ComputeType::kComputeInvalid},	137	+ {Randn::Type, af::ComputeType::kComputeElewise},
138		138
139	{Load::Type, af::ComputeType::kComputeLoad},	139	{Load::Type, af::ComputeType::kComputeLoad},
140	{Store::Type, af::ComputeType::kComputeStore},	140	{Store::Type, af::ComputeType::kComputeStore},

	@@ -13346,13 +13346,22 @@ static void CreateRandStoreAscGraph(af::AscGraph &graph, size_t dims_size) {
13346	*rand_op.y.repeats = {s0};	13346	*rand_op.y.repeats = {s0};
13347	*rand_op.y.strides = {One};	13347	*rand_op.y.strides = {One};
13348		13348
13349	- rand_op.attr.api.compute_type = af::ComputeType::kComputeElewise;
13350	- rand_op.attr.api.type = af::ApiType::kAPITypeCompute;
13351	- rand_op.attr.api.unit = af::ComputeUnit::kUnitVector;
13352	rand_op.attr.sched.axis = {z0.id};	13349	rand_op.attr.sched.axis = {z0.id};
13353		13350
		13351	+ af::ascir_op::Abs abs_op("abs");
		13352	+ graph.AddNode(abs_op);
		13353	+ abs_op.x = rand_op.y;
		13354	+ *abs_op.y.axis = {z0.id};
		13355	+ *abs_op.y.repeats = {s0};
		13356	+ *abs_op.y.strides = {One};
		13357	+
		13358	+ abs_op.attr.api.compute_type = af::ComputeType::kComputeElewise;
		13359	+ abs_op.attr.api.type = af::ApiType::kAPITypeCompute;
		13360	+ abs_op.attr.api.unit = af::ComputeUnit::kUnitVector;
		13361	+ abs_op.attr.sched.axis = {z0.id};
		13362	+
13354	af::ascir_op::Store store("store");	13363	af::ascir_op::Store store("store");
13355	- store.x = rand_op.y;	13364	+ store.x = abs_op.y;
13356	store.attr.api.compute_type = ComputeType::kComputeStore;	13365	store.attr.api.compute_type = ComputeType::kComputeStore;
13357	store.attr.api.type = af::ApiType::kAPITypeCompute;	13366	store.attr.api.type = af::ApiType::kAPITypeCompute;
13358	store.attr.api.unit = ComputeUnit::kUnitMTE3;	13367	store.attr.api.unit = ComputeUnit::kUnitMTE3;

	@@ -56,10 +56,6 @@ TEST_P(E2E_FusedBackendElewise_Code, CalculateCorrect) {
56	// Launch	56	// Launch
57	uint32_t ws_size = 0;	57	uint32_t ws_size = 0;
58	AutofuseTiling(test_shape[0], test_shape[1], &tiling_data, &ws_size, &block_dim, 48, 192 * 1024);	58	AutofuseTiling(test_shape[0], test_shape[1], &tiling_data, &ws_size, &block_dim, 48, 192 * 1024);
59	- printf("g0_tiling_key: %d\n", tiling_data.graph0_tiling_key);
60	- printf("g1_tiling_key: %d\n", tiling_data.graph1_tiling_key);
61	- printf("g2_tiling_key: %d\n", tiling_data.graph2_tiling_key);
62	- printf("ws_size: %d, ws3_size: %d, ws5_size: %d\n", ws_size, tiling_data.workspace3, tiling_data.workspace5);
63	float workspace = (float )AscendC::GmAlloc(ws_size + 32);	59	float workspace = (float )AscendC::GmAlloc(ws_size + 32);
64		60
65	AscendC::SetKernelMode(KernelMode::AIV_MODE);	61	AscendC::SetKernelMode(KernelMode::AIV_MODE);