diff options
Diffstat (limited to 'include/clblast.h')
-rw-r--r-- | include/clblast.h | 8 |
1 files changed, 8 insertions, 0 deletions
diff --git a/include/clblast.h b/include/clblast.h index ce64b37a..3e65f52a 100644 --- a/include/clblast.h +++ b/include/clblast.h @@ -636,6 +636,14 @@ StatusCode Im2col(const size_t channels, const size_t height, const size_t width cl_mem col_buffer, const size_t col_offset, cl_command_queue* queue, cl_event* event = nullptr); +// 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 cl_mem im_buffer, const size_t im_offset, + const cl_mem kernel_buffer, const size_t kernel_offset, + cl_mem result_buffer, const size_t result_offset, + cl_command_queue* queue, cl_event* event = nullptr); + // Batched version of AXPY: SAXPYBATCHED/DAXPYBATCHED/CAXPYBATCHED/ZAXPYBATCHED/HAXPYBATCHED template <typename T> StatusCode AxpyBatched(const size_t n, |