diff options
author | Cedric Nugteren <web@cedricnugteren.nl> | 2018-05-05 14:06:33 +0200 |
---|---|---|
committer | Cedric Nugteren <web@cedricnugteren.nl> | 2018-05-05 14:06:33 +0200 |
commit | 2776d761768295b01a8be7c333dbb337805d7f77 (patch) | |
tree | 3fb955ea8e6d962e60dc6c8eba4b5e9800c3db91 /include/clblast_cuda.h | |
parent | 8b381480f8bd3b40cc030b07a599da10092b8117 (diff) |
Added interface of batched convolution as GEMM
Diffstat (limited to 'include/clblast_cuda.h')
-rw-r--r-- | include/clblast_cuda.h | 8 |
1 files changed, 8 insertions, 0 deletions
diff --git a/include/clblast_cuda.h b/include/clblast_cuda.h index d82ee331..01044037 100644 --- a/include/clblast_cuda.h +++ b/include/clblast_cuda.h @@ -608,6 +608,14 @@ StatusCode Im2col(const size_t channels, const size_t height, const size_t width CUdeviceptr col_buffer, const size_t col_offset, const CUcontext context, const CUdevice device); +// Batched convolution as GEMM (non-BLAS function): SCONVGEMM/DCONVGEMM/CCONVGEMM/ZCONVGEMM/HCONVGEMM +template <typename T> +StatusCode Convgemm(const size_t channels, const size_t height, const size_t width, const size_t kernel_h, const size_t kernel_w, const size_t pad_h, const size_t pad_w, const size_t stride_h, const size_t stride_w, const size_t dilation_h, const size_t dilation_w, const size_t num_kernels, const size_t batch_count, + const CUdeviceptr im_buffer, const size_t im_offset, + const CUdeviceptr kernel_buffer, const size_t kernel_offset, + CUdeviceptr result_buffer, const size_t result_offset, + const CUcontext context, const CUdevice device); + // Batched version of AXPY: SAXPYBATCHED/DAXPYBATCHED/CAXPYBATCHED/ZAXPYBATCHED/HAXPYBATCHED template <typename T> StatusCode AxpyBatched(const size_t n, |