deepcore_source_code

This iterm is subpart source of deepcore v0.7, include fftconv, cellconvand and gemm operation; remove the conv, reduce ,batch-normalization, pooling and activation operation.

deepcore is based on CUDA and very fast, if want to use it please do with below: 0 compile and build the device kernels in deepcore_device,copy the fatBinary to the deepcore/include/dev//kbin_sm.h 1 compile and build the deepcore to generate deepcore.dll or libdeepcore.so

Note: The data struct of deepcore is CNHW(Not NCHW), please see the deepcore doc for more info.# deepcore_source_code

build from source

currently only the fftconv related kernel is built.

nvidia/cuda

for nvidia platform, make sure installed cuda 9.0+, and use build_cuda.sh to build. the output will in build/ directory

NOTE: tested under V100(sm70). you need to change to other sm targeting other nv arch.

# build device code, host code, example code
B0=1 B1=1 sh build_cuda.sh
# only need to build host code, example
B1=1 sh build_cuda.sh
# only need to build example
sh build_cuda.sh

# then run example code like:
./build/fftconv

amd/rocm

for amd platform, make sure rocm sw stack is installed properly. Then use build_rocm.sh to build.

NOTE: tested on gfx906. change to other gfx*** to target other amd gpu arch.

# build device code, host code, example code
B0=1 B1=1 sh build_rocm.sh
# only need to build host code, example
B1=1 sh build_rocm.sh
# only need to build example
sh build_rocm.sh

# then run example code like:
./build/fftconv