fatal error: cudnn.h: No such file or directory

[Alex-fishred]

I'm trying to Build and Install Custom TensorRT Plugins
When I execute make -j$(nproc)
fatal error: cudnn.h: No such file or directory
#include <cudnn.h>
^
compilation terminated.
make[2]: *** [CMakeFiles/tensorrt_ops.dir/build.make:206: CMakeFiles/tensorrt_ops.dir/plugin/multi_head_attn/multiHeadAttnPlugin.cpp.o] Error 1
make[2]: *** Waiting for unfinished jobs....
/mnt/f/BEVFormer_tensorrt-main/TensorRT/plugin/multi_head_attn/multiHeadAttnKernel.cu:8:26: fatal error: cuda/std/cmath: No such file or directory
#include <cuda/std/cmath>
Any solution?

[Alex-fishred]

Update
I tried to copy the downloaded include file under Tensorrt8.x.x.x to TensorRT’s common
Solved the above problem
But then a lot of warnings and error appeared
warning: ‘virtual void nvinfer1::IGpuAllocator::free(void*)’ is deprecated [-Wdeprecated-declarations]
this->free(memory);
error: ‘cudaMallocAsync’ was not declared in this scope
cudaMallocAsync(&A, batch * q_len * embed_dim * sizeof(int8_t), stream);
^
/mnt/f/BEVFormer_tensorrt-main/TensorRT/plugin/multi_head_attn/multiHeadAttnPlugin.cpp:488:26: error: ‘cudaFreeAsync’ was not declared in this scope
cudaFreeAsync(A, stream);
^
/mnt/f/BEVFormer_tensorrt-main/TensorRT/plugin/multi_head_attn/multiHeadAttnPlugin.cpp: In instantiation of ‘void trt_plugin::QKVPlugin::chooseAlgo(const nvinfer1::DynamicPluginTensorDesc*, const nvinfer1::DynamicPluginTensorDesc*) [with T = float]’:
/mnt/f/BEVFormer_tensorrt-main/TensorRT/plugin/multi_head_attn/multiHeadAttnPlugin.cpp:249:32: required from here
/mnt/f/BEVFormer_tensorrt-main/TensorRT/plugin/multi_head_attn/multiHeadAttnPlugin.cpp:291:18: error: ‘cudaMallocAsync’ was not declared in this scope
cudaMallocAsync(&A, batch * q_len * embed_dim * sizeof(T), stream);

/mnt/f/BEVFormer_tensorrt-main/TensorRT/plugin/multi_head_attn/multiHeadAttnPlugin.cpp:292:18: error: ‘cudaMallocAsync’ was not declared in this scope, and no declarations were found by argument-dependent lookup at the point of instantiation [-fpermissive]
cudaMallocAsync(&B, batch * kv_len * embed_dim * sizeof(T), stream);
~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

[Alex-fishred]

I successfully solved the problem by copying the cudnn file to cuda in usr/local. You can refer to the cudnn installation method. I think I missed the step of copying the file.