Star1324
1438
代码
代码
Issues114
Pull Requests238
Actions
项目讨论
Wiki
分析
Star1324
1438
已合并
fix: 修复可选输入索引处理问题，区分必选和可选输入校验 #3075
liangtongxue创建于 6月2日
fix: 修复可选输入索引处理问题，区分必选和可选输入校验 #3075
已合并
liangtongxue创建于 6月2日
从liangtongxue/ops-mathfix/dropoutv3-optional-input-9.1.0-beta.2合入到cann/ops-math9.1.0-beta.2
共 7 个文件变更+54-34
Mrandom/drop_out_v3/op_host/arch35/drop_out_v3_tiling_arch35.cpp+1-1
@@ -47,7 +47,7 @@ OpTilingConfig DropOutV3Tiling::BuildOpConfig()
47 {OUTPUT_IDX_Y, {{ge::DT_FLOAT, ge::DT_FLOAT16, ge::DT_BF16}, -1, {}, nullptr}}};47 {OUTPUT_IDX_Y, {{ge::DT_FLOAT, ge::DT_FLOAT16, ge::DT_BF16}, -1, {}, nullptr}}};
48 48 
49 config.getOutputSize = [](gert::TilingContext* ctx, int64_t& size) {49 config.getOutputSize = [](gert::TilingContext* ctx, int64_t& size) {
50- auto inputShape = ctx->GetInputShape(INPUT_IDX_X);50+ auto inputShape = ctx->GetRequiredInputShape(INPUT_IDX_X);
51 OP_CHECK_NULL_WITH_CONTEXT(ctx, inputShape);51 OP_CHECK_NULL_WITH_CONTEXT(ctx, inputShape);
52 auto storageShape = inputShape->GetStorageShape();52 auto storageShape = inputShape->GetStorageShape();
53 size = storageShape.IsScalar() ? 1 : storageShape.GetShapeSize();53 size = storageShape.IsScalar() ? 1 : storageShape.GetShapeSize();
Mrandom/drop_out_v3/op_host/drop_out_v3_infershape.cpp+4-8
@@ -30,17 +30,13 @@ static constexpr size_t MAX_DIM_NUM = 8;
30 30 
31static graphStatus InferShapeDropOutV3(gert::InferShapeContext* context)31static graphStatus InferShapeDropOutV3(gert::InferShapeContext* context)
32{32{
33- const std::unordered_map<std::string, size_t>& inputMap = {33+ const std::unordered_map<std::string, size_t>& requiredInputMap = {
34 {"x", DropOutV3_X}, {"p", DropOutV3_P}, {"seed", DropOutV3_SEED}, {"offset", DropOutV3_OFFSET}};34 {"x", DropOutV3_X}, {"p", DropOutV3_P}, {"seed", DropOutV3_SEED}, {"offset", DropOutV3_OFFSET}};
35+ const std::unordered_map<std::string, size_t>& optionalInputMap = {
36+ {"noise_shape", DropOutV3_NOISE}};
35 const std::unordered_map<std::string, size_t>& outputMap = {{"y", DropOutV3_Y}, {"mask", DropOutV3_MASK}};37 const std::unordered_map<std::string, size_t>& outputMap = {{"y", DropOutV3_Y}, {"mask", DropOutV3_MASK}};
36 int32_t mode = ops::randomCommon::MODE_NO_DEPENDENCY;38 int32_t mode = ops::randomCommon::MODE_NO_DEPENDENCY;
37- const gert::Shape* noiseInputShape = context->GetOptionalInputShape(DropOutV3_NOISE);39+ return ops::randomCommon::CommonInferShape(context, requiredInputMap, outputMap, mode, optionalInputMap);
38- if (noiseInputShape != nullptr) {
39- if (noiseInputShape->GetDimNum() > MAX_DIM_NUM) {
40- return ge::GRAPH_FAILED;
41- }
42- }
43- return ops::randomCommon::CommonInferShape(context, inputMap, outputMap, mode);
44}40}
45IMPL_OP_INFERSHAPE(DropOutV3).InferShape(InferShapeDropOutV3);41IMPL_OP_INFERSHAPE(DropOutV3).InferShape(InferShapeDropOutV3);
46 42 
Mrandom/random_common/op_host/arch35/random_tiling_arch35.cpp+18-8
@@ -149,29 +149,39 @@ ge::graphStatus RandomTilingArch35::DoTiling()
149ge::graphStatus RandomTilingArch35::CheckInputsOutputsAndAttrs()149ge::graphStatus RandomTilingArch35::CheckInputsOutputsAndAttrs()
150{150{
151 OP_LOGI(opName_, "TilingContext: %s", RandomUtils::GetTilingContext(context_).c_str());151 OP_LOGI(opName_, "TilingContext: %s", RandomUtils::GetTilingContext(context_).c_str());
152- // 校验输入tensor
153 for (const auto& [idx, rule] : config_.inputCheckRules) {152 for (const auto& [idx, rule] : config_.inputCheckRules) {
154- auto tensorDesc = context_->GetInputDesc(idx);153+ auto tensorDesc = context_->GetRequiredInputDesc(idx);
155 OP_CHECK_NULL_WITH_CONTEXT(context_, tensorDesc);154 OP_CHECK_NULL_WITH_CONTEXT(context_, tensorDesc);
156- auto inputShape = context_->GetInputShape(idx);155+ auto inputShape = context_->GetRequiredInputShape(idx);
157 OP_CHECK_NULL_WITH_CONTEXT(context_, inputShape);156 OP_CHECK_NULL_WITH_CONTEXT(context_, inputShape);
158- auto inputTensor = inputShape->GetStorageShape();157+ auto storageShape = inputShape->GetStorageShape();
159 158 
160- auto ret = CheckTensor(tensorDesc, inputTensor, rule, "input_" + std::to_string(idx));159+ auto ret = CheckTensor(tensorDesc, storageShape, rule, "input_" + std::to_string(idx));
161 if (ret != ge::GRAPH_SUCCESS) {160 if (ret != ge::GRAPH_SUCCESS) {
162 return ret;161 return ret;
163 }162 }
164 }163 }
165 164 
166- // 校验输出tensor165+ for (const auto& [idx, rule] : config_.optionalInputCheckRules) {
166+ auto tensorDesc = context_->GetOptionalInputDesc(idx);
167+ auto inputShape = context_->GetOptionalInputShape(idx);
168+ if (tensorDesc != nullptr && inputShape != nullptr) {
169+ auto storageShape = inputShape->GetStorageShape();
170+ auto ret = CheckTensor(tensorDesc, storageShape, rule, "optional_input_" + std::to_string(idx));
171+ if (ret != ge::GRAPH_SUCCESS) {
172+ return ret;
173+ }
174+ }
175+ }
176+ 
167 for (const auto& [idx, rule] : config_.outputCheckRules) {177 for (const auto& [idx, rule] : config_.outputCheckRules) {
168 auto tensorDesc = context_->GetOutputDesc(idx);178 auto tensorDesc = context_->GetOutputDesc(idx);
169 OP_CHECK_NULL_WITH_CONTEXT(context_, tensorDesc);179 OP_CHECK_NULL_WITH_CONTEXT(context_, tensorDesc);
170 auto outputShape = context_->GetOutputShape(idx);180 auto outputShape = context_->GetOutputShape(idx);
171 OP_CHECK_NULL_WITH_CONTEXT(context_, outputShape);181 OP_CHECK_NULL_WITH_CONTEXT(context_, outputShape);
172- auto outputTensor = outputShape->GetStorageShape();182+ auto storageShape = outputShape->GetStorageShape();
173 183 
174- auto ret = CheckTensor(tensorDesc, outputTensor, rule, "output_" + std::to_string(idx));184+ auto ret = CheckTensor(tensorDesc, storageShape, rule, "output_" + std::to_string(idx));
175 if (ret != ge::GRAPH_SUCCESS) {185 if (ret != ge::GRAPH_SUCCESS) {
176 return ret;186 return ret;
177 }187 }
Mrandom/random_common/op_host/arch35/random_tiling_arch35.h+1-2
@@ -45,9 +45,8 @@ struct TensorCheckRule {
45enum class RandomKernelMode {SIMD, SIMT};45enum class RandomKernelMode {SIMD, SIMT};
46 46 
47struct OpTilingConfig {47struct OpTilingConfig {
48- // 输入tensor校验规则（key: 输入索引，value: 校验规则）
49 std::unordered_map<int32_t, TensorCheckRule> inputCheckRules;48 std::unordered_map<int32_t, TensorCheckRule> inputCheckRules;
50- // 输出tensor校验规则（key: 输出索引，value: 校验规则）49+ std::unordered_map<int32_t, TensorCheckRule> optionalInputCheckRules;
51 std::unordered_map<int32_t, TensorCheckRule> outputCheckRules;50 std::unordered_map<int32_t, TensorCheckRule> outputCheckRules;
52 // // 属性校验规则（key: 属性名，value: 自定义校验函数）51 // // 属性校验规则（key: 属性名，value: 自定义校验函数）
53 std::unordered_map<int32_t, std::function<bool(gert::TilingContext*)>> attrCheckRules;52 std::unordered_map<int32_t, std::function<bool(gert::TilingContext*)>> attrCheckRules;
Mrandom/random_common/op_host/random_infershape_base.cpp+18-8
@@ -88,14 +88,23 @@ bool DependencyMode(const gert::Tensor* inTensor, gert::Shape& outShape, size_t
88}88}
89 89 
90bool InputAndOutputCheck(90bool InputAndOutputCheck(
91- gert::InferShapeContext* context, const std::unordered_map<std::string, size_t>& inputMap,91+ gert::InferShapeContext* context, const std::unordered_map<std::string, size_t>& requiredInputMap,
92- const std::unordered_map<std::string, size_t>& outputMap, int64_t& maskIndex, int64_t& offsetIndex)92+ const std::unordered_map<std::string, size_t>& outputMap, int64_t& maskIndex, int64_t& offsetIndex,
93+ const std::unordered_map<std::string, size_t>& optionalInputMap)
93{94{
94 OP_LOGD(context->GetNodeName(), "InputAndOutputCheck start");95 OP_LOGD(context->GetNodeName(), "InputAndOutputCheck start");
95- for (const auto& item : inputMap) {96+ for (const auto& item : requiredInputMap) {
96 size_t inputIndex = item.second;97 size_t inputIndex = item.second;
97- auto input = context->GetInputTensor(inputIndex);98+ auto inputShape = context->GetRequiredInputShape(inputIndex);
98- OP_CHECK_NULL_WITH_CONTEXT(context, input);99+ OP_CHECK_NULL_WITH_CONTEXT(context, inputShape);
100+ }
101+ 
102+ for (const auto& item : optionalInputMap) {
103+ size_t inputIndex = item.second;
104+ auto inputShape = context->GetOptionalInputShape(inputIndex);
105+ if (inputShape != nullptr) {
106+ OP_LOGD(context->GetNodeName(), "Optional input %zu is provided", inputIndex);
107+ }
99 }108 }
100 109 
101 for (const auto& item : outputMap) {110 for (const auto& item : outputMap) {
@@ -116,15 +125,16 @@ bool InputAndOutputCheck(
116}125}
117 126 
118ge::graphStatus CommonInferShape(127ge::graphStatus CommonInferShape(
119- gert::InferShapeContext* context, const std::unordered_map<std::string, size_t>& inputMap,128+ gert::InferShapeContext* context, const std::unordered_map<std::string, size_t>& requiredInputMap,
120- const std::unordered_map<std::string, size_t>& outputMap, int32_t mode)129+ const std::unordered_map<std::string, size_t>& outputMap, int32_t mode,
130+ const std::unordered_map<std::string, size_t>& optionalInputMap)
121{131{
122 if (context == nullptr) {132 if (context == nullptr) {
123 return ge::GRAPH_FAILED;133 return ge::GRAPH_FAILED;
124 }134 }
125 int64_t maskIndex = -1;135 int64_t maskIndex = -1;
126 int64_t offsetIndex = -1;136 int64_t offsetIndex = -1;
127- if (!InputAndOutputCheck(context, inputMap, outputMap, maskIndex, offsetIndex)) {137+ if (!InputAndOutputCheck(context, requiredInputMap, outputMap, maskIndex, offsetIndex, optionalInputMap)) {
128 return ge::GRAPH_FAILED;138 return ge::GRAPH_FAILED;
129 }139 }
130 const gert::Shape* inShape = context->GetInputShape(0);140 const gert::Shape* inShape = context->GetInputShape(0);
Mrandom/random_common/op_host/random_infershape_base.h+6-4
@@ -33,8 +33,9 @@ static constexpr int32_t MODE_NO_DEPENDENCY = 1;
33// 使用时默认输入输出的位置为0。mode = 0:只支持int32和int64，输入输出的dim相等,使用时必须有值依赖。33// 使用时默认输入输出的位置为0。mode = 0:只支持int32和int64，输入输出的dim相等,使用时必须有值依赖。
34// mode = 1:只支持float,float16和bf16，输出shape直接使用输入shape，可以无需值依赖。34// mode = 1:只支持float,float16和bf16，输出shape直接使用输入shape，可以无需值依赖。
35ge::graphStatus CommonInferShape(35ge::graphStatus CommonInferShape(
36- gert::InferShapeContext* context, const std::unordered_map<std::string, size_t>& inputMap,36+ gert::InferShapeContext* context, const std::unordered_map<std::string, size_t>& requiredInputMap,
37- const std::unordered_map<std::string, size_t>& outputMap, int32_t mode);37+ const std::unordered_map<std::string, size_t>& outputMap, int32_t mode,
38+ const std::unordered_map<std::string, size_t>& optionalInputMap = {});
38 39 
39template <typename T>40template <typename T>
40ge::graphStatus HandleShapeTensor(gert::Shape& outputShape, size_t shapeSize, const T* shapeData);41ge::graphStatus HandleShapeTensor(gert::Shape& outputShape, size_t shapeSize, const T* shapeData);
@@ -46,8 +47,9 @@ bool InferShapeForUnknow(
46bool DependencyMode(const gert::Tensor* inTensor, gert::Shape& outShape, size_t xShapeSize);47bool DependencyMode(const gert::Tensor* inTensor, gert::Shape& outShape, size_t xShapeSize);
47 48 
48bool InputAndOutputCheck(49bool InputAndOutputCheck(
49- gert::InferShapeContext* context, const std::unordered_map<std::string, size_t>& inputMap,50+ gert::InferShapeContext* context, const std::unordered_map<std::string, size_t>& requiredInputMap,
50- const std::unordered_map<std::string, size_t>& outputMap, int64_t& maskIndex, int64_t& offsetIndex);51+ const std::unordered_map<std::string, size_t>& outputMap, int64_t& maskIndex, int64_t& offsetIndex,
52+ const std::unordered_map<std::string, size_t>& optionalInputMap = {});
51 53 
52} // namespace randomCommon54} // namespace randomCommon
53} // namespace ops55} // namespace ops
Mrandom/stateless_drop_out_gen_mask/op_host/arch35/stateless_drop_out_gen_mask_tiling_arch35.cpp+6-3
@@ -49,7 +49,10 @@ OpTilingConfig StatelessDropOutGenMaskTiling::BuildOpConfig()
49 {1, {{ge::DT_FLOAT16, ge::DT_FLOAT, ge::DT_BF16}, 1, {}, nullptr}}, // prob49 {1, {{ge::DT_FLOAT16, ge::DT_FLOAT, ge::DT_BF16}, 1, {}, nullptr}}, // prob
50 {2, {{ge::DT_INT32, ge::DT_INT64}, 1, {}, nullptr}}, // seed50 {2, {{ge::DT_INT32, ge::DT_INT64}, 1, {}, nullptr}}, // seed
51 {3, {{ge::DT_INT32, ge::DT_INT64}, 1, {}, nullptr}}, // seed151 {3, {{ge::DT_INT32, ge::DT_INT64}, 1, {}, nullptr}}, // seed1
52- {4, {{ge::DT_INT64}, -1, {}, nullptr}}, // offset52+ };
53+ 
54+ config.optionalInputCheckRules = {
55+ {4, {{ge::DT_INT64}, -1, {}, nullptr}}, // offset (可选输入)
53 };56 };
54 57 
55 config.outputCheckRules = {58 config.outputCheckRules = {
@@ -84,7 +87,7 @@ OpTilingConfig StatelessDropOutGenMaskTiling::BuildOpConfig()
84 counter[0] = 0; counter[1] = 0; counter[2] = 0; counter[3] = 0;87 counter[0] = 0; counter[1] = 0; counter[2] = 0; counter[3] = 0;
85 88 
86 // offset element count (shape info, no D2H needed)89 // offset element count (shape info, no D2H needed)
87- auto offsetTensor = ctx->GetRequiredInputTensor(IN_OFFSET_IDX);90+ auto offsetTensor = ctx->GetOptionalInputTensor(IN_OFFSET_IDX);
88 uint32_t offsetElemCount = 2;91 uint32_t offsetElemCount = 2;
89 if (offsetTensor != nullptr) {92 if (offsetTensor != nullptr) {
90 offsetElemCount = static_cast<uint32_t>(offsetTensor->GetShapeSize());93 offsetElemCount = static_cast<uint32_t>(offsetTensor->GetShapeSize());
@@ -92,7 +95,7 @@ OpTilingConfig StatelessDropOutGenMaskTiling::BuildOpConfig()
92 key[0] = offsetElemCount;95 key[0] = offsetElemCount;
93 96 
94 // seed byte size: 4 for INT32, 8 for INT64 (dtype info, no D2H needed)97 // seed byte size: 4 for INT32, 8 for INT64 (dtype info, no D2H needed)
95- auto seedDesc = ctx->GetInputDesc(IN_SEED_IDX);98+ auto seedDesc = ctx->GetRequiredInputDesc(IN_SEED_IDX);
96 uint32_t seedByteSize = 8;99 uint32_t seedByteSize = 8;
97 if (seedDesc != nullptr) {100 if (seedDesc != nullptr) {
98 auto seedDtype = seedDesc->GetDataType();101 auto seedDtype = seedDesc->GetDataType();