diff options
author | Koichi Akabe <vbkaisetsu@gmail.com> | 2018-11-12 10:12:07 +0900 |
---|---|---|
committer | Koichi Akabe <vbkaisetsu@gmail.com> | 2018-11-12 10:12:07 +0900 |
commit | 032e3b0cc00a15dd2af8b4fb82d261eb7b086e26 (patch) | |
tree | cdcf4d0fc342c9ff92ee7ab3f75b0cdeced46e96 /include/clblast.h | |
parent | 90112618daa0d6b24ae3e53203a636d2e908dfba (diff) |
Add kernel_mode option to im2col, col2im, and convgemm functions
Diffstat (limited to 'include/clblast.h')
-rw-r--r-- | include/clblast.h | 10 |
1 files changed, 7 insertions, 3 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, |