Added interface of batched convolution as GEMM

3 files changed, 43 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,
diff --git a/include/clblast_c.h b/include/clblast_c.h
index 23a3afcc..918c25f6 100644
--- a/include/clblast_c.h
+++ b/include/clblast_c.h
@@ -1410,6 +1410,33 @@ CLBlastStatusCode PUBLIC_API CLBlastHim2col(const size_t channels, const size_t
                                             cl_mem col_buffer, const size_t col_offset,
                                             cl_command_queue* queue, cl_event* event);
 
+// Batched convolution as GEMM (non-BLAS function): SCONVGEMM/DCONVGEMM/CCONVGEMM/ZCONVGEMM/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,
+                                              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,
+                                              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 CLBlastCconvgemm(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 CLBlastZconvgemm(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,
+                                              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);
+
 // Batched version of AXPY: SAXPYBATCHED/DAXPYBATCHED/CAXPYBATCHED/ZAXPYBATCHED/HAXPYBATCHED
 CLBlastStatusCode PUBLIC_API CLBlastSaxpyBatched(const size_t n,
                                                  const float *alphas,
diff --git a/include/clblast_cuda.h b/include/clblast_cuda.h
index d82ee331..01044037 100644
--- a/include/clblast_cuda.h
+++ b/include/clblast_cuda.h
@@ -608,6 +608,14 @@ 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);
 
+// 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 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,
+                    const CUcontext context, const CUdevice device);
+
 // Batched version of AXPY: SAXPYBATCHED/DAXPYBATCHED/CAXPYBATCHED/ZAXPYBATCHED/HAXPYBATCHED
 template <typename T>
 StatusCode AxpyBatched(const size_t n,