Add kernel_mode option to im2col, col2im, and convgemm functions

4 files changed, 58 insertions, 27 deletions

diff --git a/include/clblast.h b/include/clblast.h
index 27adf7fa..7a82361c 100644
--- a/include/clblast.h
+++ b/include/clblast.h
@@ -117,6 +117,7 @@ enum class Transpose { kNo = 111, kYes = 112, kConjugate = 113 };
 enum class Triangle { kUpper = 121, kLower = 122 };
 enum class Diagonal { kNonUnit = 131, kUnit = 132 };
 enum class Side { kLeft = 141, kRight = 142 };
+enum class KernelMode { kCrossCorrelation = 151, kConvolution = 152 };
 
 // Precision scoped enum (values in bits)
 enum class Precision { kHalf = 16, kSingle = 32, kDouble = 64,
@@ -631,21 +632,24 @@ StatusCode Omatcopy(const Layout layout, const Transpose a_transpose,
 
 // Im2col function (non-BLAS function): SIM2COL/DIM2COL/CIM2COL/ZIM2COL/HIM2COL
 template <typename T>
-StatusCode Im2col(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,
+StatusCode Im2col(const KernelMode kernel_mode,
+                  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 im_buffer, const size_t im_offset,
                   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 <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,
+StatusCode Col2im(const KernelMode kernel_mode,
+                  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 <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,
+StatusCode Convgemm(const KernelMode kernel_mode,
+                    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,
diff --git a/include/clblast_c.h b/include/clblast_c.h
index 1c681bfe..2ba6375a 100644
--- a/include/clblast_c.h
+++ b/include/clblast_c.h
@@ -120,6 +120,7 @@ typedef enum CLBlastTriangle_ { CLBlastTriangleUpper = 121,
 typedef enum CLBlastDiagonal_ { CLBlastDiagonalNonUnit = 131,
                                 CLBlastDiagonalUnit = 132 } CLBlastDiagonal;
 typedef enum CLBlastSide_ { CLBlastSideLeft = 141, CLBlastSideRight = 142 } CLBlastSide;
+typedef enum CLBlastKernelMode_ { CLBlastKernelModeCrossCorrelation = 151, CLBlastKernelModeConvolution = 152 } CLBlastKernelMode;
 
 // Precision enum (values in bits)
 typedef enum CLBlastPrecision_ { CLBlastPrecisionHalf = 16, CLBlastPrecisionSingle = 32,
@@ -1389,61 +1390,74 @@ CLBlastStatusCode PUBLIC_API CLBlastHomatcopy(const CLBlastLayout layout, const
                                               cl_command_queue* queue, cl_event* event);
 
 // Im2col function (non-BLAS function): SIM2COL/DIM2COL/CIM2COL/ZIM2COL/HIM2COL
-CLBlastStatusCode PUBLIC_API CLBlastSim2col(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,
+CLBlastStatusCode PUBLIC_API CLBlastSim2col(const CLBlastKernelMode kernel_mode,
+                                            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 im_buffer, const size_t im_offset,
                                             cl_mem col_buffer, const size_t col_offset,
                                             cl_command_queue* queue, cl_event* event);
-CLBlastStatusCode PUBLIC_API CLBlastDim2col(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,
+CLBlastStatusCode PUBLIC_API CLBlastDim2col(const CLBlastKernelMode kernel_mode,
+                                            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 im_buffer, const size_t im_offset,
                                             cl_mem col_buffer, const size_t col_offset,
                                             cl_command_queue* queue, cl_event* event);
-CLBlastStatusCode PUBLIC_API CLBlastCim2col(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,
+CLBlastStatusCode PUBLIC_API CLBlastCim2col(const CLBlastKernelMode kernel_mode,
+                                            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 im_buffer, const size_t im_offset,
                                             cl_mem col_buffer, const size_t col_offset,
                                             cl_command_queue* queue, cl_event* event);
-CLBlastStatusCode PUBLIC_API CLBlastZim2col(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,
+CLBlastStatusCode PUBLIC_API CLBlastZim2col(const CLBlastKernelMode kernel_mode,
+                                            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 im_buffer, const size_t im_offset,
                                             cl_mem col_buffer, const size_t col_offset,
                                             cl_command_queue* queue, cl_event* event);
-CLBlastStatusCode PUBLIC_API CLBlastHim2col(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,
+CLBlastStatusCode PUBLIC_API CLBlastHim2col(const CLBlastKernelMode kernel_mode,
+                                            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 im_buffer, const size_t im_offset,
                                             cl_mem col_buffer, const size_t col_offset,
                                             cl_command_queue* queue, cl_event* event);
 
 // Col2im function (non-BLAS function): SCOL2IM/DCOL2IM/CCOL2IM/ZCOL2IM/HCOL2IM
-CLBlastStatusCode PUBLIC_API CLBlastScol2im(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,
+CLBlastStatusCode PUBLIC_API CLBlastScol2im(const CLBlastKernelMode kernel_mode,
+                                            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);
-CLBlastStatusCode PUBLIC_API CLBlastDcol2im(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,
+CLBlastStatusCode PUBLIC_API CLBlastDcol2im(const CLBlastKernelMode kernel_mode,
+                                            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);
-CLBlastStatusCode PUBLIC_API CLBlastCcol2im(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,
+CLBlastStatusCode PUBLIC_API CLBlastCcol2im(const CLBlastKernelMode kernel_mode,
+                                            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);
-CLBlastStatusCode PUBLIC_API CLBlastZcol2im(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,
+CLBlastStatusCode PUBLIC_API CLBlastZcol2im(const CLBlastKernelMode kernel_mode,
+                                            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);
-CLBlastStatusCode PUBLIC_API CLBlastHcol2im(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,
+CLBlastStatusCode PUBLIC_API CLBlastHcol2im(const CLBlastKernelMode kernel_mode,
+                                            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);
 
 // Batched convolution as GEMM (non-BLAS function): SCONVGEMM/DCONVGEMM/HCONVGEMM
-CLBlastStatusCode PUBLIC_API CLBlastSconvgemm(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,
+CLBlastStatusCode PUBLIC_API CLBlastSconvgemm(const CLBlastKernelMode kernel_mode,
+                                              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);
-CLBlastStatusCode PUBLIC_API CLBlastDconvgemm(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,
+CLBlastStatusCode PUBLIC_API CLBlastDconvgemm(const CLBlastKernelMode kernel_mode,
+                                              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);
-CLBlastStatusCode PUBLIC_API CLBlastHconvgemm(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,
+CLBlastStatusCode PUBLIC_API CLBlastHconvgemm(const CLBlastKernelMode kernel_mode,
+                                              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,
diff --git a/include/clblast_cuda.h b/include/clblast_cuda.h
index 58f9b74b..f6d6372d 100644
--- a/include/clblast_cuda.h
+++ b/include/clblast_cuda.h
@@ -89,6 +89,7 @@ enum class Transpose { kNo = 111, kYes = 112, kConjugate = 113 };
 enum class Triangle { kUpper = 121, kLower = 122 };
 enum class Diagonal { kNonUnit = 131, kUnit = 132 };
 enum class Side { kLeft = 141, kRight = 142 };
+enum class KernelMode { kCrossCorrelation = 151, kConvolution = 152 };
 
 // Precision scoped enum (values in bits)
 enum class Precision { kHalf = 16, kSingle = 32, kDouble = 64,
@@ -603,21 +604,24 @@ StatusCode Omatcopy(const Layout layout, const Transpose a_transpose,
 
 // Im2col function (non-BLAS function): SIM2COL/DIM2COL/CIM2COL/ZIM2COL/HIM2COL
 template <typename T>
-StatusCode Im2col(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,
+StatusCode Im2col(const KernelMode kernel_mode,
+                  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 im_buffer, const size_t im_offset,
                   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,
+StatusCode Col2im(const KernelMode kernel_mode,
+                  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,
+StatusCode Convgemm(const KernelMode kernel_mode,
+                    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,
diff --git a/include/clblast_netlib_c.h b/include/clblast_netlib_c.h
index 65545bfb..4c54fb18 100644
--- a/include/clblast_netlib_c.h
+++ b/include/clblast_netlib_c.h
@@ -45,6 +45,7 @@ typedef enum CLBlastTriangle_ { CLBlastTriangleUpper = 121,
 typedef enum CLBlastDiagonal_ { CLBlastDiagonalNonUnit = 131,
                                 CLBlastDiagonalUnit = 132 } CLBlastDiagonal;
 typedef enum CLBlastSide_ { CLBlastSideLeft = 141, CLBlastSideRight = 142 } CLBlastSide;
+typedef enum CLBlastKernelMode_ { CLBlastKernelModeCrossCorrelation = 141, CLBlastKernelModeConvolution = 152 } CLBlastKernelMode;
 
 // For full compatibility with CBLAS
 typedef CLBlastLayout CBLAS_ORDER;
@@ -947,30 +948,38 @@ void PUBLIC_API cblas_zomatcopy(const CLBlastLayout layout, const CLBlastTranspo
                                 void* b, const int b_ld);
 
 // Im2col function (non-BLAS function): SIM2COL/DIM2COL/CIM2COL/ZIM2COL/HIM2COL
-void PUBLIC_API cblas_sim2col(const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
+void PUBLIC_API cblas_sim2col(const CLBlastKernelMode kernel_mode,
+                              const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
                               const float* im,
                               float* col);
-void PUBLIC_API cblas_dim2col(const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
+void PUBLIC_API cblas_dim2col(const CLBlastKernelMode kernel_mode,
+                              const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
                               const double* im,
                               double* col);
-void PUBLIC_API cblas_cim2col(const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
+void PUBLIC_API cblas_cim2col(const CLBlastKernelMode kernel_mode,
+                              const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
                               const void* im,
                               void* col);
-void PUBLIC_API cblas_zim2col(const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
+void PUBLIC_API cblas_zim2col(const CLBlastKernelMode kernel_mode,
+                              const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
                               const void* im,
                               void* col);
 
 // Col2im function (non-BLAS function): SCOL2IM/DCOL2IM/CCOL2IM/ZCOL2IM/HCOL2IM
-void PUBLIC_API cblas_scol2im(const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
+void PUBLIC_API cblas_scol2im(const CLBlastKernelMode kernel_mode,
+                              const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
                               const float* col,
                               float* im);
-void PUBLIC_API cblas_dcol2im(const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
+void PUBLIC_API cblas_dcol2im(const CLBlastKernelMode kernel_mode,
+                              const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
                               const double* col,
                               double* im);
-void PUBLIC_API cblas_ccol2im(const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
+void PUBLIC_API cblas_ccol2im(const CLBlastKernelMode kernel_mode,
+                              const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
                               const void* col,
                               void* im);
-void PUBLIC_API cblas_zcol2im(const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
+void PUBLIC_API cblas_zcol2im(const CLBlastKernelMode kernel_mode,
+                              const int channels, const int height, const int width, const int kernel_h, const int kernel_w, const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int dilation_h, const int dilation_w,
                               const void* col,
                               void* im);