Merged
fix(example): preserve ND outputs for INTrainingReduceV2 on Ascend950 #14400
fix(example): preserve ND outputs for INTrainingReduceV2 on Ascend950 #14400
Merged
zsq_888Add on 9 days ago
Total: 1 files changed+32-1
@@ -33,6 +33,9 @@ using namespace ge;
33using std::string;33using std::string;
34using std::vector;34using std::vector;
35 35 
36+// Both outputs are always FP32, including when the input is FP16.
37+constexpr DataType OUTPUT_DTYPE = DT_FLOAT;
38+ 
36string GetTime()39string GetTime()
37{40{
38 time_t timep;41 time_t timep;
@@ -145,7 +148,7 @@ bool CheckOutputs(const ExampleOptions& options, const vector<Tensor>& outputs)
145 for (size_t outputIndex = 0; outputIndex < outputs.size(); ++outputIndex) {148 for (size_t outputIndex = 0; outputIndex < outputs.size(); ++outputIndex) {
146 const auto desc = outputs[outputIndex].GetTensorDesc();149 const auto desc = outputs[outputIndex].GetTensorDesc();
147 const auto& expected = outputIndex == 0 ? sums : squares;150 const auto& expected = outputIndex == 0 ? sums : squares;
148- if (desc.GetDataType() != DT_FLOAT || desc.GetFormat() != FORMAT_ND ||151+ if (desc.GetDataType() != OUTPUT_DTYPE || desc.GetFormat() != FORMAT_ND ||
149 desc.GetShape().GetDims() != expectedShape || outputs[outputIndex].GetData() == nullptr ||152 desc.GetShape().GetDims() != expectedShape || outputs[outputIndex].GetData() == nullptr ||
150 outputs[outputIndex].GetSize() != expected.size() * sizeof(float)) {153 outputs[outputIndex].GetSize() != expected.size() * sizeof(float)) {
151 printf("Output %zu has an unexpected descriptor or buffer size\n", outputIndex);154 printf("Output %zu has an unexpected descriptor or buffer size\n", outputIndex);
@@ -165,8 +168,23 @@ bool CheckOutputs(const ExampleOptions& options, const vector<Tensor>& outputs)
165 return true;168 return true;
166}169}
167 170 
171+ge::graphStatus InferExampleOutputFormats(ge::Operator& op)
172+{
173+ // Keep the declared ND outputs through GE's format propagation stage.
174+ for (uint32_t index = 0; index < 2; ++index) {
175+ auto desc = op.GetOutputDesc(index);
176+ desc.SetFormat(FORMAT_ND);
177+ desc.SetOriginFormat(FORMAT_ND);
178+ if (op.UpdateOutputDesc(index, desc) != ge::GRAPH_SUCCESS) {
179+ return ge::GRAPH_FAILED;
180+ }
181+ }
182+ return ge::GRAPH_SUCCESS;
183+}
184+ 
168int32_t RunExample(const ExampleOptions& options)185int32_t RunExample(const ExampleOptions& options)
169{186{
187+ const ge::InferFormatFuncRegister outputFormatRegistration("INTrainingReduceV2", InferExampleOutputFormats);
170 Graph graph("tc_ge_irrun_in_training_reduce_v2");188 Graph graph("tc_ge_irrun_in_training_reduce_v2");
171 auto xData = op::Data("placeholder_x").set_attr_index(0);189 auto xData = op::Data("placeholder_x").set_attr_index(0);
172 TensorDesc xDesc(ge::Shape(options.shape), options.format, options.dtype);190 TensorDesc xDesc(ge::Shape(options.shape), options.format, options.dtype);
@@ -179,6 +197,19 @@ int32_t RunExample(const ExampleOptions& options)
179 }197 }
180 auto reduceOp = op::INTrainingReduceV2("in_training_reduce_v2");198 auto reduceOp = op::INTrainingReduceV2("in_training_reduce_v2");
181 reduceOp.set_input_x(xData);199 reduceOp.set_input_x(xData);
200+ // Both outputs use ND storage, independently of the input's channel layout.
201+ const bool channelLast = options.format == FORMAT_NHWC || options.format == FORMAT_NDHWC;
202+ const size_t channelAxis = channelLast ? options.shape.size() - 1 : 1;
203+ vector<int64_t> outputShape(options.shape.size(), 1);
204+ outputShape[0] = options.shape[0];
205+ outputShape[channelAxis] = options.shape[channelAxis];
206+ TensorDesc outputDesc(ge::Shape(outputShape), FORMAT_ND, OUTPUT_DTYPE);
207+ outputDesc.SetOriginFormat(FORMAT_ND);
atomgit-bot
atomgit-botatomgit-bot9 days ago

🟡 Medium Priority

变更行 203 用 TensorDesc(ge::Shape(outputShape), FORMAT_ND, DT_FLOAT) 把 sum/square_sum 两个输出的 dtype 硬编码为 DT_FLOAT,而同一函数第 187-188 行创建输入描述符时按 options.dtype 走,且 main/ParseOptions(56-62 行)明确支持 fp16 参数、GenFloatData 也真的按 DT_FLOAT16 填数据。

改动前后行为差异:改动前输出 dtype 由 GE 的 infer-dtype 阶段推导;改动后被显式覆盖为 DT_FLOAT,与输入 dtype 解耦。

触发条件:以 ./test_geir_in_training_reduce_v2 fp16 运行。失败模式(取决于算子 fp16 情况下输出 dtype 的真实语义):

  1. 若算子对 fp16 输入的输出为 DT_FLOAT16,则描述符声明 float32,GE 按 float32 分配输出缓冲,CheckOutputs 第 148 行 dtype 校验与第 150 行 expected.size() * sizeof(float) 长度校验都会“通过”,但 156-158 行按 float32 解释 float16 的原始字节,必然与 expected 不等,用例稳定报 FAIL(错值被当成真实测试失败,掩盖真实原因);
  2. 或在 AddGraph/InferDataType 阶段因手工 desc 与算子推导的 dtype 冲突而编译失败,只留下 "AddGraph failed" 这类无信息日志。

建议:输出 TensorDesc 的 dtype 改为按约定推导(例如 const DataType outDtype = DT_FLOAT; 并注释“该算子输出恒为 fp32,与输入 dtype 无关”),或直接使用 GE 推导的 dtype;同时与 CheckOutputs 的 dtype 校验保持同一来源,避免两处硬编码各写一份。

likedislike
Incorrect?
208+ if (reduceOp.update_output_desc_sum(outputDesc) != ge::GRAPH_SUCCESS ||
209+ reduceOp.update_output_desc_square_sum(outputDesc) != ge::GRAPH_SUCCESS) {
210+ printf("Set output descriptors failed\n");
211+ return FAILED;
212+ }
182 const vector<Operator> inputs = {xData};213 const vector<Operator> inputs = {xData};
183 const vector<Operator> outputs = {reduceOp};214 const vector<Operator> outputs = {reduceOp};
184 graph.SetInputs(inputs).SetOutputs(outputs);215 graph.SetInputs(inputs).SetOutputs(outputs);