From d45911b61dedafcbd74f65df263b4197697d6a81 Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Tue, 23 Oct 2018 20:52:25 +0200 Subject: Added groundwork for col2im algorithm plus first non-working version of kernel and test --- include/clblast.h | 7 +++++++ 1 file changed, 7 insertions(+) (limited to 'include/clblast.h') diff --git a/include/clblast.h b/include/clblast.h index 9a8988e7..27adf7fa 100644 --- a/include/clblast.h +++ b/include/clblast.h @@ -636,6 +636,13 @@ 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); +// Col2im function (non-BLAS function): SCOL2IM/DCOL2IM/CCOL2IM/ZCOL2IM/HCOL2IM +template +StatusCode Col2im(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 cl_mem col_buffer, const size_t col_offset, + cl_mem im_buffer, const size_t im_offset, + cl_command_queue* queue, cl_event* event = nullptr); + // Batched convolution as GEMM (non-BLAS function): SCONVGEMM/DCONVGEMM/HCONVGEMM template 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, -- cgit v1.2.3