From 19ea419e7f14d02aeb208c2fbd5a4ac55f4cb101 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Tue, 19 Jun 2018 13:09:53 +0100 Subject: COMPMID-809: Add NHWC data format on CLGEMMConvolutionLayer. Change-Id: I50e4f5e7d47e21c300f754bee2c216863075b5cf Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/136191 Tested-by: Jenkins Reviewed-by: Giorgio Arena Reviewed-by: Gian Marco Iodice --- src/core/CL/cl_kernels/convolution_layer.cl | 72 ++++++++++++++++++++++++++++- 1 file changed, 71 insertions(+), 1 deletion(-) (limited to 'src/core/CL/cl_kernels/convolution_layer.cl') diff --git a/src/core/CL/cl_kernels/convolution_layer.cl b/src/core/CL/cl_kernels/convolution_layer.cl index f8e0c27724..6a70b009c8 100644 --- a/src/core/CL/cl_kernels/convolution_layer.cl +++ b/src/core/CL/cl_kernels/convolution_layer.cl @@ -55,7 +55,7 @@ * @param[in] depth The depth of the input tensor * @param[in] total_filters Total number of filters. 4th dimension of the weights matrix */ -__kernel void reshape_to_columns( +__kernel void reshape_to_columns_nchw( TENSOR3D_DECLARATION(src), IMAGE_DECLARATION(dst), #ifdef HAS_BIAS @@ -97,4 +97,74 @@ __kernel void reshape_to_columns( } } } + +/** This kernel reshapes the tensor's low three dimensions to single column + * + * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 + * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] dst_ptr Pointer to the destination tensor. Same as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] bias_ptr Pointer to the bias tensor. Same as @p src_ptr + * @param[in] bias_stride_x Stride of the bias tensor in X dimension (in bytes) + * @param[in] bias_step_x bias_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] bias_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] depth The depth of the input tensor + * @param[in] width The width of the input tensor + * @param[in] height The height of the input tensor + * @param[in] total_filters Total number of filters. 4th dimension of the weights matrix + */ +__kernel void reshape_to_columns_nhwc( + TENSOR3D_DECLARATION(src), + IMAGE_DECLARATION(dst), +#ifdef HAS_BIAS + VECTOR_DECLARATION(bias), +#endif /* HAS_BIAS */ + uint depth, uint width, uint height, uint total_filters) +{ + Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); + bool is_last_thread = (get_global_id(0) == (get_global_size(0) - 1) && get_global_id(1) == (get_global_size(1) - 1) && get_global_id(2) == (get_global_size(2) - 1)); + + __global uchar *tmp_src_ptr = src.ptr; + __global uchar *tmp_dst_ptr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(1) * dst_stride_y + get_global_id(2) * width * dst_stride_y + get_global_id( + 0) * width * height * dst_stride_y; +#ifdef HAS_BIAS + __global uchar *tmp_bias_ptr = bias_ptr + bias_offset_first_element_in_bytes; +#endif /* HAS_BIAS */ + + if(is_last_thread) + { + for(uint i = 0; i < total_filters; ++i) + { + *((__global DATA_TYPE *)tmp_dst_ptr) = *((__global DATA_TYPE *)tmp_src_ptr); + +#ifdef HAS_BIAS + *((__global DATA_TYPE *)(tmp_dst_ptr + dst_stride_y)) = *((__global DATA_TYPE *)(tmp_bias_ptr)); + tmp_bias_ptr += bias_stride_x; +#endif /* HAS_BIAS */ + tmp_src_ptr += height * src_stride_z; + tmp_dst_ptr += dst_stride_x; + } + } + else + { + for(uint i = 0; i < total_filters; ++i) + { + *((__global DATA_TYPE *)tmp_dst_ptr) = *((__global DATA_TYPE *)tmp_src_ptr); + tmp_src_ptr += height * src_stride_z; + tmp_dst_ptr += dst_stride_x; + } + } +} #endif // defined(DATA_TYPE) \ No newline at end of file -- cgit v1.2.1