Added convgemm skeleton, test infrastructure, and first reference implementation

1 files changed, 70 insertions, 0 deletions

diff --git a/doc/api.md b/doc/api.md
index 2f861ce0..c72b18c6 100644
--- a/doc/api.md
+++ b/doc/api.md
@@ -3072,6 +3072,76 @@ Arguments to IM2COL:
 
 
 
+xCONVGEMM: Batched convolution as GEMM (non-BLAS function)
+-------------
+
+Integrates im2col and GEMM for batched 3D convolution, in which _im_ is the 4D input tensor, _kernel_ the 4D kernel weights tensor, and _result_ the 4D output tensor.
+
+C++ API:
+```
+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)
+```
+
+C API:
+```
+CLBlastStatusCode 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 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 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 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 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)
+```
+
+Arguments to CONVGEMM:
+
+* `const size_t channels`: Integer size argument. This value must be positive.
+* `const size_t height`: Integer size argument. This value must be positive.
+* `const size_t width`: Integer size argument. This value must be positive.
+* `const size_t kernel_h`: Integer size argument. This value must be positive.
+* `const size_t kernel_w`: Integer size argument. This value must be positive.
+* `const size_t pad_h`: Integer size argument. This value must be positive.
+* `const size_t pad_w`: Integer size argument. This value must be positive.
+* `const size_t stride_h`: Integer size argument. This value must be positive.
+* `const size_t stride_w`: Integer size argument. This value must be positive.
+* `const size_t dilation_h`: Integer size argument. This value must be positive.
+* `const size_t dilation_w`: Integer size argument. This value must be positive.
+* `const size_t num_kernels`: Integer size argument. This value must be positive.
+* `const size_t batch_count`: Integer size argument. This value must be positive.
+* `const cl_mem im_buffer`: OpenCL buffer to store the input im tensor.
+* `const size_t im_offset`: The offset in elements from the start of the input im tensor.
+* `const cl_mem kernel_buffer`: OpenCL buffer to store the input kernel tensor.
+* `const size_t kernel_offset`: The offset in elements from the start of the input kernel tensor.
+* `cl_mem result_buffer`: OpenCL buffer to store the output result tensor.
+* `const size_t result_offset`: The offset in elements from the start of the output result tensor.
+* `cl_command_queue* queue`: Pointer to an OpenCL command queue associated with a context and device to execute the routine on.
+* `cl_event* event`: Pointer to an OpenCL event to be able to wait for completion of the routine's OpenCL kernel(s). This is an optional argument.
+
+
+
 xAXPYBATCHED: Batched version of AXPY
 -------------