README.md

cuDNN correctness tests

Verify if the operator's calculation results are correct on other gpu products compatible with NVIDIA CUDA/cuDNN.

Building

mkdir build
cd build
cmake ..
make

Note: The most recent cuDNN distribution will be obtained automatically by installing PyTorch into the build directory.

Testing

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

Generating the ground truth data

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:

1. Generate input/output data:

python tf-maxpooling.py

2. Paste new input/output data into float32.h:

dtype input[IN_SIZE] = {...}
dtype output[OUT_SIZE] = {...}

3. Re-run the test to verify cudnnMaxpoolForward.