miopen_cudnn_ops's Introduction

miopen_cudnn_ops

compare cudnn/miopen ops

build on amd rocm platform

please install rocm following this link, and make sure miopen is installed properly

sh build.sh miopen

build on nvidia cuda platform

please install cuda to /usr/local/cuda, as well as cudnn

sh build.sh cudnn

compare op with op_driver

# after above build step, binary will result in ./build/ directory

# run alexnet 1st conv layer command:
./build/op_driver  conv -k 64 -W 227 -H 227 -c 3 -x 11 -y 11 -u 4 -v 4 -p 1 -q 1 -n 512
# run alexnet 2nd conv layer command:
./build/op_driver  conv -k 192 -W 55 -H 55 -c 64 -x 5 -y 5 -u 1 -v 1 -p 2 -q 2 -n 512

# more detail about conv parameters:
./build/op_driver  conv -h

miopen_cudnn_ops's People

Contributors

Stargazers

Watchers

miopen_cudnn_ops's Issues

Build fails on ROCm 6.0

Branch: 4-build-fail-using-rocm-60
PR: fix the build on ROCm 6.0

Tracking/Plan of op_driver

Add support of enabling/disabling Tensor Core in convolution on NVIDIA/cuDNN backend.
Add support of measure wall-time of both Find/Get interface on both NVIDIA/AMD backend.
- Print out both kernel execution time and wall-time.
- For NVIDIA/cuDNN backend, add both paths of cudnnFindConvolution[Direction]AlgorithmEx/cudnnGetConvolution[Direction]Algorithm. Direction can be Forward/BackwardData/BackwardFilter.
- For AMD/MIOpen backend, add intermediate mode support, see below APIs for reference: miopenConvolutionForwardCompileSolution/miopenConvolutionForwardGetSolution/miopenConvolutionForwardImmediate
Group/depthwise convolution support both in CPU/GPU side.
Add support of tensor format on NVIDIA/cuDNN other than CUDNN_TENSOR_NCHW, such as CUDNN_TENSOR_NHWC and CUDNN_TENSOR_NCHW_VECT_C.

Recommend Projects