Verify if the operator's calculation results are correct on other gpu products compatible with NVIDIA CUDA/cuDNN.
mkdir build
cd build
cmake ..
make
Note: The most recent cuDNN distribution will be obtained automatically by installing PyTorch into the build directory.
Run all correctness tests at once:
cd build
make
ctest
Run all correctness tests at once, with CUDA/cuDNN stack intercepted by GPUOcelot:
cd build
cmake -DENABLE_OCELOT=/home/marcusmae/gpuocelot/gpuocelot/ocelot/build/libgpuocelot.so ..
make
ctest
We generate input/weight/bias/outout data with TensorFlow API. For example, we use tf.nn.maxpool()
, and then compare against it the result of a call to cudnnMaxpoolForward
:
- Generate input/output data:
python tf-maxpooling.py
- Paste new input/output data into
float32.h
:
dtype input[IN_SIZE] = {...}
dtype output[OUT_SIZE] = {...}
- Re-run the test to verify
cudnnMaxpoolForward
.