Added groundwork for col2im algorithm plus first non-working version of kernel and test

1 files changed, 7 insertions, 0 deletions

diff --git a/include/clblast_cuda.h b/include/clblast_cuda.h
index 1bbd898e..58f9b74b 100644
--- a/include/clblast_cuda.h
+++ b/include/clblast_cuda.h
@@ -608,6 +608,13 @@ 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);
 
+// Col2im function (non-BLAS function): SCOL2IM/DCOL2IM/CCOL2IM/ZCOL2IM/HCOL2IM
+template <typename T>
+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 CUdeviceptr col_buffer, const size_t col_offset,
+                  CUdeviceptr im_buffer, const size_t im_offset,
+                  const CUcontext context, const CUdevice device);
+
 // Batched convolution as GEMM (non-BLAS function): SCONVGEMM/DCONVGEMM/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,