From 76faef88284e6fd51f53b23063374d3d3a884e4f Mon Sep 17 00:00:00 2001 From: Gian Marco Date: Mon, 29 Jan 2018 12:15:32 +0000 Subject: COMPMID-855 - Optimizing im2col on OpenCL (DCHW) Introduced optimizations for 1x1, 3x3, 5x5 and 11x11 Change-Id: Ibb7f7a9fbec01a7684746ed8513634078126e452 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118107 Tested-by: Jenkins Reviewed-by: Michalis Spyrou --- src/core/CL/cl_kernels/col2im.cl | 68 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 68 insertions(+) create mode 100644 src/core/CL/cl_kernels/col2im.cl (limited to 'src/core/CL/cl_kernels/col2im.cl') diff --git a/src/core/CL/cl_kernels/col2im.cl b/src/core/CL/cl_kernels/col2im.cl new file mode 100644 index 0000000000..58fb80a416 --- /dev/null +++ b/src/core/CL/cl_kernels/col2im.cl @@ -0,0 +1,68 @@ +/* + * Copyright (c) 2017-2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "helpers.h" + +#if defined(FIXED_POINT_POSITION) +#include "fixed_point.h" +#endif // FIXED_POINT_POSITION + +#if defined(DATA_TYPE) && defined(WIDTH_OUTPUT) +/** This kernel performs a reshaping of the output of the convolution layer. + * + * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float + * @note The width of the output tensor must be passed at compile time using -DWIDTH_OUTPUT: e.g. -DWIDTH_OUTPUT=320 + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/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 Z 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. Supported data types: 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_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_step_z dst_stride_z * number of elements along Z 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] dst_stride_w Stride of the destination tensor in W dimension (in bytes) + */ +__kernel void col2im( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst), + uint dst_stride_w) +{ + Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); + Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(dst); + + // Compute output offset + int idx = get_global_id(0) * dst.stride_z + (get_global_id(1) / WIDTH_OUTPUT) * dst_stride_y + (get_global_id(1) % WIDTH_OUTPUT) * dst_stride_x + get_global_id(2) * dst_stride_w; + + // Store value + *((__global DATA_TYPE *)(dst.ptr + idx)) = *((__global DATA_TYPE *)(src.ptr)); +} +#endif // defined(DATA_TYPE) && defined(WIDTH_OUTPUT) \ No newline at end of file -- cgit v1.2.1