diff options
Diffstat (limited to 'src/core/CL/cl_kernels/transpose.cl')
| -rw-r--r-- | src/core/CL/cl_kernels/transpose.cl | 217 |
1 files changed, 217 insertions, 0 deletions
diff --git a/src/core/CL/cl_kernels/transpose.cl b/src/core/CL/cl_kernels/transpose.cl new file mode 100644 index 0000000000..c30158f280 --- /dev/null +++ b/src/core/CL/cl_kernels/transpose.cl @@ -0,0 +1,217 @@ +/* + * Copyright (c) 2017 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" + +#define SWAP_ROW(u0, l0) \ + ({ \ + tmp_swap = u0; \ + u0 = l0; \ + l0 = tmp_swap; \ + }) + +#define SWAP_4x4(u0, u1, u2, u3, l0, l1, l2, l3) \ + ({ \ + VEC_DATA_TYPE(DATA_TYPE, 4) \ + tmp_swap; \ + SWAP_ROW(u0, l0); \ + SWAP_ROW(u1, l1); \ + SWAP_ROW(u2, l2); \ + SWAP_ROW(u3, l3); \ + }) + +#define SWAP_8x8(u0, u1, u2, u3, u4, u5, u6, u7, l0, l1, l2, l3, l4, l5, l6, l7) \ + ({ \ + VEC_DATA_TYPE(DATA_TYPE, 8) \ + tmp_swap; \ + SWAP_ROW(u0, l0); \ + SWAP_ROW(u1, l1); \ + SWAP_ROW(u2, l2); \ + SWAP_ROW(u3, l3); \ + SWAP_ROW(u4, l4); \ + SWAP_ROW(u5, l5); \ + SWAP_ROW(u6, l6); \ + SWAP_ROW(u7, l7); \ + }) + +#define TRANSPOSE_4x4(u0, u1, u2, u3) \ + ({ \ + VEC_DATA_TYPE(DATA_TYPE, 4) \ + tmp; \ + tmp.s012 = u0.s123; \ + u0.s1 = u1.s0; \ + u0.s2 = u2.s0; \ + u0.s3 = u3.s0; \ + u1.s0 = tmp.s0; \ + u2.s0 = tmp.s1; \ + u3.s0 = tmp.s2; \ + \ + tmp.s01 = u1.s23; \ + u1.s2 = u2.s1; \ + u1.s3 = u3.s1; \ + u2.s1 = tmp.s0; \ + u3.s1 = tmp.s1; \ + \ + tmp.s0 = u2.s3; \ + u2.s3 = u3.s2; \ + u3.s2 = tmp.s0; \ + }) + +#define TRANSPOSE_8x8(u0, u1, u2, u3, u4, u5, u6, u7) \ + ({ \ + TRANSPOSE_4x4(u0.s0123, u1.s0123, u2.s0123, u3.s0123); \ + TRANSPOSE_4x4(u0.s4567, u1.s4567, u2.s4567, u3.s4567); \ + TRANSPOSE_4x4(u4.s0123, u5.s0123, u6.s0123, u7.s0123); \ + TRANSPOSE_4x4(u4.s4567, u5.s4567, u6.s4567, u7.s4567); \ + SWAP_4x4(u0.s4567, u1.s4567, u2.s4567, u3.s4567, u4.s0123, u5.s0123, u6.s0123, u7.s0123); \ + }) + +#define TRANSPOSE_16x16(u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15) \ + ({ \ + TRANSPOSE_8x8(u0.s01234567, u1.s01234567, u2.s01234567, u3.s01234567, u4.s01234567, u5.s01234567, u6.s01234567, u7.s01234567); \ + TRANSPOSE_8x8(u0.s89ABCDEF, u1.s89ABCDEF, u2.s89ABCDEF, u3.s89ABCDEF, u4.s89ABCDEF, u5.s89ABCDEF, u6.s89ABCDEF, u7.s89ABCDEF); \ + TRANSPOSE_8x8(u8.s01234567, u9.s01234567, u10.s01234567, u11.s01234567, u12.s01234567, u13.s01234567, u14.s01234567, u15.s01234567); \ + TRANSPOSE_8x8(u8.s89ABCDEF, u9.s89ABCDEF, u10.s89ABCDEF, u11.s89ABCDEF, u12.s89ABCDEF, u13.s89ABCDEF, u14.s89ABCDEF, u15.s89ABCDEF); \ + SWAP_8x8(u0.s89ABCDEF, u1.s89ABCDEF, u2.s89ABCDEF, u3.s89ABCDEF, u4.s89ABCDEF, u5.s89ABCDEF, u6.s89ABCDEF, u7.s89ABCDEF, \ + u8.s01234567, u9.s01234567, u10.s01234567, u11.s01234567, u12.s01234567, u13.s01234567, u14.s01234567, u15.s01234567); \ + }) + +#ifndef DATA_TYPE_IN_BYTES +#error DATA_TYPE_IN_BYTES not set for the transpose OpenCL kernel +#endif + +#if DATA_TYPE_IN_BYTES == 4 +#define DATA_TYPE uint +#define TRANSPOSE() TRANSPOSE_4x4(u0, u1, u2, u3) +#define VLOAD(x, y) vload4(x, y) +#define VSTORE(x, y, z) vstore4(x, y, z) +#define BLOCK_SIZE 4 +#elif DATA_TYPE_IN_BYTES == 2 +#define DATA_TYPE ushort +#define TRANSPOSE() TRANSPOSE_8x8(u0, u1, u2, u3, u4, u5, u6, u7) +#define VLOAD(x, y) vload8(x, y) +#define VSTORE(x, y, z) vstore8(x, y, z) +#define BLOCK_SIZE 8 +#elif DATA_TYPE_IN_BYTES == 1 +#define DATA_TYPE uchar +#define TRANSPOSE() TRANSPOSE_16x16(u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15) +#define VLOAD(x, y) vload16(x, y) +#define VSTORE(x, y, z) vstore16(x, y, z) +#define BLOCK_SIZE 16 +#else +#error DATA_TYPE_IN_BYTES not supported for transpose +#endif + +/** This OpenCL kernel computes the matrix transposition of input matrix + * + * @attention The number of bytes of the data type need to be passed at compile time using -DDATA_TYPE_IN_BYTES. DATA_TYPE_IN_BYTES can be: + * -# -DDATA_TYPE_IN_BYTES=1 for transposing U8 or S8 matrices + * -# -DDATA_TYPE_IN_BYTES=2 for transposing U16, S16 or FP16 matrices + * -# -DDATA_TYPE_IN_BYTES=4 for transposing U32, S32 or FP32 matrices + * + * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/U16/S16/F16/U32/S32/F32 + * @param[in] src_stride_x Stride of the source matrix 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 matrix 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_offset_first_element_in_bytes The offset of the first element in the source matrix + * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as src_ptr + * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) + * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) + * @param[in] dst_step_y dst_gx_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 matrix + */ +__kernel void transpose(IMAGE_DECLARATION(src), + IMAGE_DECLARATION(dst)) +{ + uint x = get_global_id(0) * BLOCK_SIZE; + uint y = get_global_id(1) * BLOCK_SIZE; + + // Compute source address + Image src = CONVERT_TO_IMAGE_STRUCT(src); + + // Load the NxN block at (x, y) + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u0 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 0))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u1 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 1))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u2 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 2))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u3 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 3))); +#if BLOCK_SIZE > 4 + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u4 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 4))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u5 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 5))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u6 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 6))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u7 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 7))); +#if BLOCK_SIZE == 16 + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u8 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 8))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u9 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 9))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u10 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 10))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u11 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 11))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u12 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 12))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u13 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 13))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u14 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 14))); + VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) + u15 = VLOAD(0, (__global DATA_TYPE *)(offset(&src, 0, 15))); +#endif /* BLOCK_SIZE == 16 */ +#endif /* BLOCK_SIZE > 4 */ + + // Transpose the block + TRANSPOSE(); + + // Store the block at (y, x) + uint dst_offset_in_bytes = y * DATA_TYPE_IN_BYTES + x * dst_stride_y + dst_offset_first_element_in_bytes; + VSTORE(u0, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 0 * dst_stride_y)); + VSTORE(u1, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 1 * dst_stride_y)); + VSTORE(u2, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 2 * dst_stride_y)); + VSTORE(u3, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 3 * dst_stride_y)); +#if BLOCK_SIZE > 4 + VSTORE(u4, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 4 * dst_stride_y)); + VSTORE(u5, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 5 * dst_stride_y)); + VSTORE(u6, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 6 * dst_stride_y)); + VSTORE(u7, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 7 * dst_stride_y)); +#if BLOCK_SIZE == 16 + VSTORE(u8, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 8 * dst_stride_y)); + VSTORE(u9, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 9 * dst_stride_y)); + VSTORE(u10, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 10 * dst_stride_y)); + VSTORE(u11, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 11 * dst_stride_y)); + VSTORE(u12, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 12 * dst_stride_y)); + VSTORE(u13, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 13 * dst_stride_y)); + VSTORE(u14, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 14 * dst_stride_y)); + VSTORE(u15, 0, (__global DATA_TYPE *)(dst_ptr + dst_offset_in_bytes + 15 * dst_stride_y)); +#endif /* BLOCK_SIZE == 16 */ +#endif /* BLOCK_SIZE > 4 */ +} |