Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in; |
| 25 | #include "helpers.h" |
| 26 | |
| 27 | #ifdef DATA_TYPE_FP32 |
| 28 | precision highp float; |
| 29 | |
| 30 | BUFFER_DECLARATION(src, 1, float, readonly); |
| 31 | BUFFER_DECLARATION(dst, 2, float, writeonly); |
| 32 | |
| 33 | layout(std140) uniform shader_params |
| 34 | { |
| 35 | IMAGE_PARAM_DECLARATION(src); |
| 36 | IMAGE_PARAM_DECLARATION(dst); |
| 37 | }; |
| 38 | |
| 39 | #define LOAD16(r, name, offset) \ |
| 40 | r.x = LOAD4(name, offset); \ |
| 41 | r.y = LOAD4(name, offset + uint(1)); \ |
| 42 | r.z = LOAD4(name, offset + uint(2)); \ |
| 43 | r.w = LOAD4(name, offset + uint(3)) |
| 44 | |
| 45 | #define STORE16(name, offset, r) \ |
| 46 | STORE4(name, offset, r.x); \ |
| 47 | STORE4(name, offset + uint(1), r.y); \ |
| 48 | STORE4(name, offset + uint(2), r.z); \ |
| 49 | STORE4(name, offset + uint(3), r.w) |
| 50 | |
| 51 | /** This OpenGL ES kernel computes the matrix transposition of input matrix |
| 52 | * |
| 53 | * @param[in] src_ptr Pointer to the source matrix. Supported data types: F32 |
| 54 | * @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes) |
| 55 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 56 | * @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 57 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 58 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 59 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as src_ptr |
| 60 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 61 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 62 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 63 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 64 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 65 | */ |
| 66 | void main(void) |
| 67 | { |
| 68 | // Compute source address |
| 69 | Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| 70 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 71 | |
| 72 | // Load the NxN block at (x, y) |
| 73 | vec4 u0; |
| 74 | vec4 u1; |
| 75 | vec4 u2; |
| 76 | vec4 u3; |
| 77 | LOAD16(u0, src, offset(src, 0, 0)); |
| 78 | LOAD16(u1, src, offset(src, 0, 1)); |
| 79 | LOAD16(u2, src, offset(src, 0, 2)); |
| 80 | LOAD16(u3, src, offset(src, 0, 3)); |
| 81 | |
| 82 | // Transpose the block |
| 83 | vec4 tmp; |
| 84 | tmp.xyz = u0.yzw; |
| 85 | u0.y = u1.x; |
| 86 | u0.z = u2.x; |
| 87 | u0.w = u3.x; |
| 88 | u1.x = tmp.x; |
| 89 | u2.x = tmp.y; |
| 90 | u3.x = tmp.z; |
| 91 | tmp.xy = u1.zw; |
| 92 | u1.z = u2.y; |
| 93 | u1.w = u3.y; |
| 94 | u2.y = tmp.x; |
| 95 | u3.y = tmp.y; |
| 96 | tmp.x = u2.w; |
| 97 | u2.w = u3.z; |
| 98 | u3.z = tmp.x; |
| 99 | |
| 100 | // Store the block at (y, x) |
| 101 | uint dst_offset_in_bytes = uint(16) * uint(gl_GlobalInvocationID.y) + uint(4) * uint(gl_GlobalInvocationID.x) * (dst.stride_y) + (dst.offset_first_element_in_bytes); |
| 102 | |
| 103 | STORE16(dst, uint((dst_offset_in_bytes + uint(0) * dst.stride_y) >> 2), u0); |
| 104 | STORE16(dst, uint((dst_offset_in_bytes + uint(1) * dst.stride_y) >> 2), u1); |
| 105 | STORE16(dst, uint((dst_offset_in_bytes + uint(2) * dst.stride_y) >> 2), u2); |
| 106 | STORE16(dst, uint((dst_offset_in_bytes + uint(3) * dst.stride_y) >> 2), u3); |
| 107 | } |
| 108 | |
| 109 | #elif defined(DATA_TYPE_FP16) |
| 110 | precision mediump float; |
| 111 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 112 | layout(std140) uniform shader_params |
| 113 | { |
| 114 | IMAGE_PARAM_DECLARATION(src); |
| 115 | IMAGE_PARAM_DECLARATION(dst); |
| 116 | }; |
| 117 | |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame^] | 118 | #if defined(TRANSPOSE_4X4) |
| 119 | BUFFER_DECLARATION(src, 1, uvec2, readonly); |
| 120 | BUFFER_DECLARATION(dst, 2, uvec2, writeonly); |
| 121 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 122 | /** This OpenGL ES kernel computes the matrix transposition of input matrix |
| 123 | * |
| 124 | * @param[in] src_ptr Pointer to the source matrix. Supported data types: F16 |
| 125 | * @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes) |
| 126 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 127 | * @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 128 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 129 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 130 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as src_ptr |
| 131 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 132 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 133 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 134 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 135 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 136 | */ |
| 137 | void main(void) |
| 138 | { |
| 139 | // Compute source address |
| 140 | Image src = GC_CONVERT_TO_IMAGE_STRUCT(src); |
| 141 | Image dst = GC_CONVERT_TO_IMAGE_STRUCT(dst); |
| 142 | |
| 143 | // Load the NxN block at (x, y) |
| 144 | vec4 u0; |
| 145 | vec4 u1; |
| 146 | vec4 u2; |
| 147 | vec4 u3; |
| 148 | uvec2 packed_s[4]; |
| 149 | GC_LOAD1_2D_OFFSET(packed_s[0], src, 0, 0); |
| 150 | GC_LOAD1_2D_OFFSET(packed_s[1], src, 0, 1); |
| 151 | GC_LOAD1_2D_OFFSET(packed_s[2], src, 0, 2); |
| 152 | GC_LOAD1_2D_OFFSET(packed_s[3], src, 0, 3); |
| 153 | u0 = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y)); |
| 154 | u1 = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y)); |
| 155 | u2 = vec4(unpackHalf2x16(packed_s[2].x), unpackHalf2x16(packed_s[2].y)); |
| 156 | u3 = vec4(unpackHalf2x16(packed_s[3].x), unpackHalf2x16(packed_s[3].y)); |
| 157 | |
| 158 | // Transpose the block |
| 159 | vec4 tmp; |
| 160 | tmp.xyz = u0.yzw; |
| 161 | u0.y = u1.x; |
| 162 | u0.z = u2.x; |
| 163 | u0.w = u3.x; |
| 164 | u1.x = tmp.x; |
| 165 | u2.x = tmp.y; |
| 166 | u3.x = tmp.z; |
| 167 | tmp.xy = u1.zw; |
| 168 | u1.z = u2.y; |
| 169 | u1.w = u3.y; |
| 170 | u2.y = tmp.x; |
| 171 | u3.y = tmp.y; |
| 172 | tmp.x = u2.w; |
| 173 | u2.w = u3.z; |
| 174 | u3.z = tmp.x; |
| 175 | |
| 176 | // Store the block at (y, x) |
| 177 | uint dst_offset_in_bytes = uint(8) * uint(gl_GlobalInvocationID.y) + uint(gl_GlobalInvocationID.x) * (dst_step_y) + (dst.offset_first_element_in_bytes); |
| 178 | |
| 179 | packed_s[0] = uvec2(packHalf2x16(u0.xy), packHalf2x16(u0.zw)); |
| 180 | packed_s[1] = uvec2(packHalf2x16(u1.xy), packHalf2x16(u1.zw)); |
| 181 | packed_s[2] = uvec2(packHalf2x16(u2.xy), packHalf2x16(u2.zw)); |
| 182 | packed_s[3] = uvec2(packHalf2x16(u3.xy), packHalf2x16(u3.zw)); |
| 183 | GC_STORE1(packed_s[0], dst, uint((dst_offset_in_bytes + uint(0) * dst_stride_y) >> 3)); |
| 184 | GC_STORE1(packed_s[1], dst, uint((dst_offset_in_bytes + uint(1) * dst_stride_y) >> 3)); |
| 185 | GC_STORE1(packed_s[2], dst, uint((dst_offset_in_bytes + uint(2) * dst_stride_y) >> 3)); |
| 186 | GC_STORE1(packed_s[3], dst, uint((dst_offset_in_bytes + uint(3) * dst_stride_y) >> 3)); |
| 187 | } |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame^] | 188 | #elif defined(TRANSPOSE_8X8) /* TRANSPOSE_4X4 */ |
| 189 | BUFFER_DECLARATION(src, 1, uvec4, readonly); |
| 190 | BUFFER_DECLARATION(dst, 2, uvec4, writeonly); |
| 191 | |
| 192 | #define SWAP_ROW(u0, l0) \ |
| 193 | { \ |
| 194 | tmp_swap = u0; \ |
| 195 | u0 = l0; \ |
| 196 | l0 = tmp_swap; \ |
| 197 | } |
| 198 | |
| 199 | #define SWAP_4x4(u0, u1, u2, u3, l0, l1, l2, l3) \ |
| 200 | { \ |
| 201 | vec4 tmp_swap; \ |
| 202 | SWAP_ROW(u0, l0); \ |
| 203 | SWAP_ROW(u1, l1); \ |
| 204 | SWAP_ROW(u2, l2); \ |
| 205 | SWAP_ROW(u3, l3); \ |
| 206 | } |
| 207 | |
| 208 | #define TRANSPOSE_4x4(u0, u1, u2, u3) \ |
| 209 | { \ |
| 210 | vec4 tmp; \ |
| 211 | tmp.xyz = u0.yzw; \ |
| 212 | u0.y = u1.x; \ |
| 213 | u0.z = u2.x; \ |
| 214 | u0.w = u3.x; \ |
| 215 | u1.x = tmp.x; \ |
| 216 | u2.x = tmp.y; \ |
| 217 | u3.x = tmp.z; \ |
| 218 | tmp.xy = u1.zw; \ |
| 219 | u1.z = u2.y; \ |
| 220 | u1.w = u3.y; \ |
| 221 | u2.y = tmp.x; \ |
| 222 | u3.y = tmp.y; \ |
| 223 | tmp.x = u2.w; \ |
| 224 | u2.w = u3.z; \ |
| 225 | u3.z = tmp.x; \ |
| 226 | } |
| 227 | |
| 228 | /** This OpenGL ES kernel computes the matrix transposition of input matrix |
| 229 | * |
| 230 | * @param[in] src_ptr Pointer to the source matrix. Supported data types:F16 |
| 231 | * @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes) |
| 232 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 233 | * @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes) |
| 234 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 235 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix |
| 236 | * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as src_ptr |
| 237 | * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) |
| 238 | * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) |
| 239 | * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) |
| 240 | * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) |
| 241 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix |
| 242 | */ |
| 243 | void main(void) |
| 244 | { |
| 245 | // Compute source address |
| 246 | Image src = GC_CONVERT_TO_IMAGE_STRUCT(src); |
| 247 | Image dst = GC_CONVERT_TO_IMAGE_STRUCT(dst); |
| 248 | |
| 249 | vec4 u[8][2]; |
| 250 | |
| 251 | uvec4 packed_s[8]; |
| 252 | |
| 253 | for(int i = 0; i < 8; i++) |
| 254 | { |
| 255 | GC_LOAD1_2D_OFFSET(packed_s[i], src, 0, i); |
| 256 | u[i][0] = vec4(unpackHalf2x16(packed_s[i].x), unpackHalf2x16(packed_s[i].y)); |
| 257 | u[i][1] = vec4(unpackHalf2x16(packed_s[i].z), unpackHalf2x16(packed_s[i].w)); |
| 258 | } |
| 259 | |
| 260 | // Transpose the block |
| 261 | TRANSPOSE_4x4(u[0][0], u[1][0], u[2][0], u[3][0]); |
| 262 | TRANSPOSE_4x4(u[0][1], u[1][1], u[2][1], u[3][1]); |
| 263 | TRANSPOSE_4x4(u[4][0], u[5][0], u[6][0], u[7][0]); |
| 264 | TRANSPOSE_4x4(u[4][1], u[5][1], u[6][1], u[7][1]); |
| 265 | SWAP_4x4(u[0][1], u[1][1], u[2][1], u[3][1], u[4][0], u[5][0], u[6][0], u[7][0]); |
| 266 | |
| 267 | // Store the block at (y, x) |
| 268 | uint dst_offset_in_bytes = uint(16) * uint(gl_GlobalInvocationID.y) + uint(gl_GlobalInvocationID.x) * (dst_step_y) + (dst.offset_first_element_in_bytes); |
| 269 | |
| 270 | for(int i = 0; i < 8; i++) |
| 271 | { |
| 272 | packed_s[i] = uvec4(packHalf2x16(u[i][0].xy), packHalf2x16(u[i][0].zw), packHalf2x16(u[i][1].xy), packHalf2x16(u[i][1].zw)); |
| 273 | GC_STORE1(packed_s[i], dst, uint((dst_offset_in_bytes + uint(i) * dst_stride_y) >> 4)); |
| 274 | } |
| 275 | } |
| 276 | #endif /* TRANSPOSE_4X4 */ |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 277 | #endif /*ARM_COMPUTE_ENABLE_FP16*/ |