why 7234 和 8406 diffs in layer's output dimensions when constructing network

[zhoutianzi666]

Description

Environment

TensorRT Version: NVIDIA GPU: NVIDIA Driver Version: CUDA Version: CUDNN Version: Operating System: Python Version (if applicable): Tensorflow Version (if applicable): PyTorch Version (if applicable): Baremetal or Container (if so, version):

Relevant Files

Steps To Reproduce

#include "/usr/local/tensorrt/include/NvInfer.h"
#include <vector>
#include <iostream>
// g++ A.cc -lnvinfer -lcudart  -L /usr/local/tensorrt/lib/ -L/usr/local/cuda/lib64 -I /usr/local/cuda/include
// export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/tensorrt/lib:/usr/local/cuda/lib64

#define TRT_VERSION                                    \
  NV_TENSORRT_MAJOR * 1000 + NV_TENSORRT_MINOR * 100 + \
      NV_TENSORRT_PATCH * 10 + NV_TENSORRT_BUILD

#include "/usr/local/cuda/include/cuda_runtime.h"
class TensorrtLogger : public nvinfer1::ILogger {
  nvinfer1::ILogger::Severity verbosity_;

 public:
  TensorrtLogger(Severity verbosity = Severity::kWARNING)
      : verbosity_(verbosity) {}
  void log(Severity severity, const char* msg) noexcept override {
    if (severity <= verbosity_) {
        printf("%s\n", msg);
  }
  }
};

int main()
{
    std::cout << TRT_VERSION  << std::endl;
    static TensorrtLogger trt_logger(nvinfer1::ILogger::Severity::kWARNING);
    auto trt_builder = nvinfer1::createInferBuilder(trt_logger);
    const auto explicitBatch = 1U << static_cast<uint32_t>(nvinfer1::NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);
    auto trt_network = trt_builder->createNetworkV2(explicitBatch);
    auto trt_config = trt_builder->createBuilderConfig();
    trt_config->setMaxWorkspaceSize(1<<30);

    nvinfer1::Dims a;
    a.nbDims = 4;
    a.d[0] = -1;
    a.d[1] = -1;
    a.d[2] = -1;
    a.d[3] = 32;
    trt_network->addInput("foo0", nvinfer1::DataType::kFLOAT, a);

nvinfer1::IOptimizationProfile* profile1 = trt_builder->createOptimizationProfile();
profile1->setDimensions("foo0", nvinfer1::OptProfileSelector::kMIN, nvinfer1::Dims4(2, 4, 4, 32));
profile1->setDimensions("foo0", nvinfer1::OptProfileSelector::kOPT, nvinfer1::Dims4(16, 64, 16, 32));
profile1->setDimensions("foo0", nvinfer1::OptProfileSelector::kMAX, nvinfer1::Dims4(128, 256, 64, 32));
trt_config->addOptimizationProfile(profile1);


    auto x = trt_network->getInput(0);
    auto layer = trt_network->addShuffle(*x);
    layer->setInput(1, *(trt_network->addShape(*x)->getOutput(0)));

    for (int i = 0 ; i < layer->getOutput(0)->getDimensions().nbDims; i++)
    {
        std::cout << layer->getOutput(0)->getDimensions().d[i] << std::endl;
    }

    trt_network->markOutput(*layer->getOutput(0));


    auto engine = trt_builder->buildEngineWithConfig(*trt_network, *trt_config);
    auto engine_out_dims = engine->getBindingDimensions(0);

std::cout << "after build" << std::endl;
    for (int i = 0 ; i < engine_out_dims.nbDims; i++)
    {
        std::cout << engine_out_dims.d[i] << std::endl;
    }
}

• in 7234

7234
-1
-1
-1
-1
after build
-1
-1
-1
32

• in 8406

8406
-1
-1
-1
32
after build
-1
-1
-1
32

I think the second is good ! It seems like this: in 7234, shape infer is carried out when building engine. In 8034, it seems that when every layer is added, shape infer will be carried out. Is that right？

[nvpohanh]

Yes, that was a bug already fixed in TRT 8.4.

[zhoutianzi666]

Yes, that was a bug already fixed in TRT 8.4.

I think is was not a bug. It seems like this: in 7234, shape infer is carried out when building engine. In 8034, it seems that when every layer is added, shape infer will be carried out. Is that right？

[zhoutianzi666]

Yes, that was a bug already fixed in TRT 8.4.

I think is was not a bug. It seems like this: in 7234, shape infer is carried out when building engine. In 8034, it seems that when every layer is added, shape infer will be carried out. Is that right？

please reply me