Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016, 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 | #include "arm_compute/core/Error.h" |
| 25 | #include "arm_compute/core/Helpers.h" |
| 26 | #include "arm_compute/core/IMultiImage.h" |
| 27 | #include "arm_compute/core/Utils.h" |
| 28 | |
| 29 | #include <arm_neon.h> |
| 30 | |
| 31 | namespace |
| 32 | { |
| 33 | constexpr float red_coef_bt709 = 1.5748F; |
| 34 | constexpr float green_coef_bt709 = -0.1873f; |
| 35 | constexpr float green_coef2_bt709 = -0.4681f; |
| 36 | constexpr float blue_coef_bt709 = 1.8556f; |
| 37 | |
| 38 | constexpr float rgb2yuv_bt709_kr = 0.2126f; |
| 39 | constexpr float rgb2yuv_bt709_kb = 0.0722f; |
| 40 | // K_g = 1 - K_r - K_b |
| 41 | constexpr float rgb2yuv_bt709_kg = 0.7152f; |
| 42 | // C_u = 1 / (2 * (1 - K_b)) |
| 43 | constexpr float rgb2yuv_bt709_cu = 0.5389f; |
| 44 | // C_v = 1 / (2 * (1 - K_r)) |
| 45 | constexpr float rgb2yuv_bt709_cv = 0.6350f; |
| 46 | |
| 47 | inline void convert_uint8x16_to_float32x4x4(const uint8x16_t &in, float32x4x4_t &out) |
| 48 | { |
| 49 | const auto tmp1 = vmovl_u8(vget_low_u8(in)); |
| 50 | out.val[0] = vcvtq_f32_u32(vmovl_u16(vget_low_u16(tmp1))); |
| 51 | out.val[1] = vcvtq_f32_u32(vmovl_u16(vget_high_u16(tmp1))); |
| 52 | const auto tmp2 = vmovl_u8(vget_high_u8(in)); |
| 53 | out.val[2] = vcvtq_f32_u32(vmovl_u16(vget_low_u16(tmp2))); |
| 54 | out.val[3] = vcvtq_f32_u32(vmovl_u16(vget_high_u16(tmp2))); |
| 55 | } |
| 56 | |
| 57 | inline void convert_float32x4x3_to_uint8x8x3(const float32x4x3_t &in1, const float32x4x3_t &in2, uint8x8x3_t &out) |
| 58 | { |
| 59 | out.val[0] = vqmovn_u16(vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in1.val[0])), |
| 60 | vqmovn_u32(vcvtq_u32_f32(in2.val[0])))); |
| 61 | out.val[1] = vqmovn_u16(vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in1.val[1])), |
| 62 | vqmovn_u32(vcvtq_u32_f32(in2.val[1])))); |
| 63 | out.val[2] = vqmovn_u16(vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in1.val[2])), |
| 64 | vqmovn_u32(vcvtq_u32_f32(in2.val[2])))); |
| 65 | } |
| 66 | |
| 67 | inline void convert_float32x4x4_to_unit8x16(const float32x4x4_t &in, uint8x16_t &out) |
| 68 | { |
| 69 | const auto low = vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in.val[0])), |
| 70 | vqmovn_u32(vcvtq_u32_f32(in.val[1]))); |
| 71 | const auto high = vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in.val[2])), |
| 72 | vqmovn_u32(vcvtq_u32_f32(in.val[3]))); |
| 73 | out = vcombine_u8(vqmovn_u16(low), vqmovn_u16(high)); |
| 74 | } |
| 75 | |
| 76 | inline void rgb_to_yuv_calculation(const float32x4_t &rvec, const float32x4_t &gvec, const float32x4_t &bvec, |
| 77 | float32x4_t &yvec, float32x4_t &uvec, float32x4_t &vvec) |
| 78 | { |
| 79 | /* |
| 80 | Y'= 0.2126*R' + 0.7152*G' + 0.0722*B' |
| 81 | U'=-0.1146*R' - 0.3854*G' + 0.5000*B' |
| 82 | V'= 0.5000*R' - 0.4542*G' - 0.0458*B' |
| 83 | */ |
| 84 | const auto c128 = vdupq_n_f32(128.f); |
| 85 | |
| 86 | // Y = R * K_r + G * (1 - K_r - K_b) * B * K_b |
| 87 | yvec = vmulq_n_f32(rvec, rgb2yuv_bt709_kr); |
| 88 | yvec = vmlaq_n_f32(yvec, gvec, rgb2yuv_bt709_kg); |
| 89 | yvec = vmlaq_n_f32(yvec, bvec, rgb2yuv_bt709_kb); |
| 90 | |
| 91 | // U = (B - Y) / (2 * (1 - K_b)) |
| 92 | uvec = vsubq_f32(bvec, yvec); |
| 93 | uvec = vmlaq_n_f32(c128, uvec, rgb2yuv_bt709_cu); |
| 94 | |
| 95 | // V = (R - Y) / (2 * (1 - K_r)) |
| 96 | vvec = vsubq_f32(rvec, yvec); |
| 97 | vvec = vmlaq_n_f32(c128, vvec, rgb2yuv_bt709_cv); |
| 98 | } |
| 99 | |
| 100 | inline void yuyv_to_rgb_calculation(const float32x4_t &yvec_val, float32x4_t uvec_val, const float32x4_t &yyvec_val, |
| 101 | float32x4_t vvec_val, unsigned char *output_ptr, const bool alpha) |
| 102 | { |
| 103 | float32x4x3_t rgb1, rgb2; |
| 104 | |
| 105 | // Compute: cb - 128 and cr - 128; |
| 106 | const auto c128 = vdupq_n_f32(128.f); |
| 107 | uvec_val = vsubq_f32(uvec_val, c128); |
| 108 | vvec_val = vsubq_f32(vvec_val, c128); |
| 109 | |
| 110 | // Compute: |
| 111 | // r = 0.0000f*f_u + 1.5748f*f_v; |
| 112 | // g = 0.1873f*f_u - 0.4681f*f_v; |
| 113 | // b = 1.8556f*f_u + 0.0000f*f_v; |
| 114 | const auto red = vmulq_n_f32(vvec_val, red_coef_bt709); |
| 115 | const auto blue = vmulq_n_f32(uvec_val, blue_coef_bt709); |
| 116 | const auto green = vaddq_f32(vmulq_n_f32(uvec_val, green_coef_bt709), |
| 117 | vmulq_n_f32(vvec_val, green_coef2_bt709)); |
| 118 | |
| 119 | // Compute the final r,g,b values using y1 for the first texel and y2 for the second one. |
| 120 | // the result is stored in two float32x4x3_t which then are converted to one uint8x8x3_t |
| 121 | // and written back to memory using vst3 instruction |
| 122 | |
| 123 | rgb1.val[0] = vaddq_f32(yvec_val, red); |
| 124 | rgb1.val[1] = vaddq_f32(yvec_val, green); |
| 125 | rgb1.val[2] = vaddq_f32(yvec_val, blue); |
| 126 | |
| 127 | rgb2.val[0] = vaddq_f32(yyvec_val, red); |
| 128 | rgb2.val[1] = vaddq_f32(yyvec_val, green); |
| 129 | rgb2.val[2] = vaddq_f32(yyvec_val, blue); |
| 130 | |
| 131 | uint8x8x3_t u8_rgb; |
| 132 | convert_float32x4x3_to_uint8x8x3(rgb1, rgb2, u8_rgb); |
| 133 | |
| 134 | if(!alpha) |
| 135 | { |
| 136 | vst3_lane_u8(&output_ptr[0], u8_rgb, 0); |
| 137 | vst3_lane_u8(&output_ptr[3], u8_rgb, 4); |
| 138 | vst3_lane_u8(&output_ptr[6], u8_rgb, 1); |
| 139 | vst3_lane_u8(&output_ptr[9], u8_rgb, 5); |
| 140 | vst3_lane_u8(&output_ptr[12], u8_rgb, 2); |
| 141 | vst3_lane_u8(&output_ptr[15], u8_rgb, 6); |
| 142 | vst3_lane_u8(&output_ptr[18], u8_rgb, 3); |
| 143 | vst3_lane_u8(&output_ptr[21], u8_rgb, 7); |
| 144 | } |
| 145 | else |
| 146 | { |
| 147 | uint8x8x4_t u8_rgba; |
| 148 | u8_rgba.val[0] = u8_rgb.val[0]; |
| 149 | u8_rgba.val[1] = u8_rgb.val[1]; |
| 150 | u8_rgba.val[2] = u8_rgb.val[2]; |
| 151 | u8_rgba.val[3] = vdup_n_u8(255); |
| 152 | vst4_lane_u8(&output_ptr[0], u8_rgba, 0); |
| 153 | vst4_lane_u8(&output_ptr[4], u8_rgba, 4); |
| 154 | vst4_lane_u8(&output_ptr[8], u8_rgba, 1); |
| 155 | vst4_lane_u8(&output_ptr[12], u8_rgba, 5); |
| 156 | vst4_lane_u8(&output_ptr[16], u8_rgba, 2); |
| 157 | vst4_lane_u8(&output_ptr[20], u8_rgba, 6); |
| 158 | vst4_lane_u8(&output_ptr[24], u8_rgba, 3); |
| 159 | vst4_lane_u8(&output_ptr[28], u8_rgba, 7); |
| 160 | } |
| 161 | } |
| 162 | |
| 163 | inline uint8x16x3_t load_rgb(const unsigned char *const ptr, const bool alpha) |
| 164 | { |
| 165 | uint8x16x3_t rgb; |
| 166 | |
| 167 | if(alpha) |
| 168 | { |
| 169 | const auto tmp = vld4q_u8(ptr); |
| 170 | rgb.val[0] = tmp.val[0]; |
| 171 | rgb.val[1] = tmp.val[1]; |
| 172 | rgb.val[2] = tmp.val[2]; |
| 173 | } |
| 174 | else |
| 175 | { |
| 176 | rgb = vld3q_u8(ptr); |
| 177 | } |
| 178 | |
| 179 | return rgb; |
| 180 | } |
| 181 | |
| 182 | inline void rgb_to_yuv_conversion(uint8x16x3_t &vec_top, uint8x16x3_t &vec_bottom) |
| 183 | { |
| 184 | // Convert the uint8x16_t to float32x4x4_t |
| 185 | float32x4x4_t frvec_top, fgvec_top, fbvec_top; |
| 186 | convert_uint8x16_to_float32x4x4(vec_top.val[0], frvec_top); |
| 187 | convert_uint8x16_to_float32x4x4(vec_top.val[1], fgvec_top); |
| 188 | convert_uint8x16_to_float32x4x4(vec_top.val[2], fbvec_top); |
| 189 | |
| 190 | float32x4x4_t frvec_bottom, fgvec_bottom, fbvec_bottom; |
| 191 | convert_uint8x16_to_float32x4x4(vec_bottom.val[0], frvec_bottom); |
| 192 | convert_uint8x16_to_float32x4x4(vec_bottom.val[1], fgvec_bottom); |
| 193 | convert_uint8x16_to_float32x4x4(vec_bottom.val[2], fbvec_bottom); |
| 194 | |
| 195 | float32x4x4_t fyvec_top, fuvec_top, fvvec_top; |
| 196 | float32x4x4_t fyvec_bottom, fuvec_bottom, fvvec_bottom; |
| 197 | |
| 198 | for(auto i = 0; i < 4; ++i) |
| 199 | { |
| 200 | rgb_to_yuv_calculation(frvec_top.val[i], fgvec_top.val[i], fbvec_top.val[i], |
| 201 | fyvec_top.val[i], fuvec_top.val[i], fvvec_top.val[i]); |
| 202 | rgb_to_yuv_calculation(frvec_bottom.val[i], fgvec_bottom.val[i], fbvec_bottom.val[i], |
| 203 | fyvec_bottom.val[i], fuvec_bottom.val[i], fvvec_bottom.val[i]); |
| 204 | } |
| 205 | |
| 206 | convert_float32x4x4_to_unit8x16(fyvec_top, vec_top.val[0]); |
| 207 | convert_float32x4x4_to_unit8x16(fuvec_top, vec_top.val[1]); |
| 208 | convert_float32x4x4_to_unit8x16(fvvec_top, vec_top.val[2]); |
| 209 | convert_float32x4x4_to_unit8x16(fyvec_bottom, vec_bottom.val[0]); |
| 210 | convert_float32x4x4_to_unit8x16(fuvec_bottom, vec_bottom.val[1]); |
| 211 | convert_float32x4x4_to_unit8x16(fvvec_bottom, vec_bottom.val[2]); |
| 212 | } |
| 213 | |
| 214 | inline void store_rgb_to_nv12(const uint8x16_t &rvec_top, const uint8x16_t &gvec_top, const uint8x16_t &bvec_top, |
| 215 | const uint8x16_t &rvec_bottom, const uint8x16_t &gvec_bottom, const uint8x16_t &bvec_bottom, |
| 216 | unsigned char *const __restrict out_y_top, unsigned char *const __restrict out_y_bottom, |
| 217 | unsigned char *const __restrict out_uv) |
| 218 | { |
| 219 | uint8x16x3_t vec_top, vec_bottom; |
| 220 | vec_top.val[0] = rvec_top; |
| 221 | vec_top.val[1] = gvec_top; |
| 222 | vec_top.val[2] = bvec_top; |
| 223 | vec_bottom.val[0] = rvec_bottom; |
| 224 | vec_bottom.val[1] = gvec_bottom; |
| 225 | vec_bottom.val[2] = bvec_bottom; |
| 226 | |
| 227 | rgb_to_yuv_conversion(vec_top, vec_bottom); |
| 228 | |
| 229 | vst1q_u8(out_y_top, vec_top.val[0]); |
| 230 | vst1q_u8(out_y_bottom, vec_bottom.val[0]); |
| 231 | |
| 232 | const auto uvec = vuzpq_u8(vec_top.val[1], vec_bottom.val[1]); |
| 233 | const auto vvec = vuzpq_u8(vec_top.val[2], vec_bottom.val[2]); |
| 234 | const auto utmp = vrhaddq_u8(uvec.val[0], uvec.val[1]); |
| 235 | const auto vtmp = vrhaddq_u8(vvec.val[0], vvec.val[1]); |
| 236 | |
| 237 | uint8x8x2_t uvvec; |
| 238 | uvvec.val[0] = vhadd_u8(vget_low_u8(utmp), vget_high_u8(utmp)); |
| 239 | uvvec.val[1] = vhadd_u8(vget_low_u8(vtmp), vget_high_u8(vtmp)); |
| 240 | |
| 241 | vst2_u8(out_uv, uvvec); |
| 242 | } |
| 243 | |
| 244 | inline void store_rgb_to_iyuv(const uint8x16_t &rvec_top, const uint8x16_t &gvec_top, const uint8x16_t &bvec_top, |
| 245 | const uint8x16_t &rvec_bottom, const uint8x16_t &gvec_bottom, const uint8x16_t &bvec_bottom, |
| 246 | unsigned char *const __restrict out_y_top, unsigned char *const __restrict out_y_bottom, |
| 247 | unsigned char *const __restrict out_u, |
| 248 | unsigned char *const __restrict out_v) |
| 249 | { |
| 250 | uint8x16x3_t vec_top, vec_bottom; |
| 251 | vec_top.val[0] = rvec_top; |
| 252 | vec_top.val[1] = gvec_top; |
| 253 | vec_top.val[2] = bvec_top; |
| 254 | vec_bottom.val[0] = rvec_bottom; |
| 255 | vec_bottom.val[1] = gvec_bottom; |
| 256 | vec_bottom.val[2] = bvec_bottom; |
| 257 | |
| 258 | rgb_to_yuv_conversion(vec_top, vec_bottom); |
| 259 | |
| 260 | vst1q_u8(out_y_top, vec_top.val[0]); |
| 261 | vst1q_u8(out_y_bottom, vec_bottom.val[0]); |
| 262 | |
| 263 | const auto uvvec_top = vuzpq_u8(vec_top.val[1], vec_top.val[2]); |
| 264 | const auto uvvec_bottom = vuzpq_u8(vec_bottom.val[1], vec_bottom.val[2]); |
| 265 | const auto uvvec = vhaddq_u8(vrhaddq_u8(uvvec_top.val[0], uvvec_top.val[1]), |
| 266 | vrhaddq_u8(uvvec_bottom.val[0], uvvec_bottom.val[1])); |
| 267 | |
| 268 | vst1_u8(out_u, vget_low_u8(uvvec)); |
| 269 | vst1_u8(out_v, vget_high_u8(uvvec)); |
| 270 | } |
| 271 | |
| 272 | inline void store_rgb_to_yuv4(const uint8x16_t &rvec, const uint8x16_t &gvec, const uint8x16_t &bvec, |
| 273 | unsigned char *const __restrict out_y, |
| 274 | unsigned char *const __restrict out_u, |
| 275 | unsigned char *const __restrict out_v) |
| 276 | { |
| 277 | // Convert the uint8x16_t to float32x4x4_t |
| 278 | float32x4x4_t frvec, fgvec, fbvec; |
| 279 | convert_uint8x16_to_float32x4x4(rvec, frvec); |
| 280 | convert_uint8x16_to_float32x4x4(gvec, fgvec); |
| 281 | convert_uint8x16_to_float32x4x4(bvec, fbvec); |
| 282 | |
| 283 | float32x4x4_t fyvec, fuvec, fvvec; |
| 284 | for(auto i = 0; i < 4; ++i) |
| 285 | { |
| 286 | rgb_to_yuv_calculation(frvec.val[i], fgvec.val[i], fbvec.val[i], |
| 287 | fyvec.val[i], fuvec.val[i], fvvec.val[i]); |
| 288 | } |
| 289 | |
| 290 | uint8x16_t yvec, uvec, vvec; |
| 291 | convert_float32x4x4_to_unit8x16(fyvec, yvec); |
| 292 | convert_float32x4x4_to_unit8x16(fuvec, uvec); |
| 293 | convert_float32x4x4_to_unit8x16(fvvec, vvec); |
| 294 | |
| 295 | vst1q_u8(out_y, yvec); |
| 296 | vst1q_u8(out_u, uvec); |
| 297 | vst1q_u8(out_v, vvec); |
| 298 | } |
| 299 | } |
| 300 | |
| 301 | namespace arm_compute |
| 302 | { |
| 303 | void colorconvert_rgb_to_rgbx(const void *__restrict input, void *__restrict output, const Window &win) |
| 304 | { |
| 305 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 306 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 307 | |
| 308 | const auto input_ptr = static_cast<const IImage *__restrict>(input); |
| 309 | const auto output_ptr = static_cast<IImage *__restrict>(output); |
| 310 | |
| 311 | Iterator in(input_ptr, win); |
| 312 | Iterator out(output_ptr, win); |
| 313 | |
| 314 | execute_window_loop(win, [&](const Coordinates & id) |
| 315 | { |
| 316 | const auto ta1 = vld3q_u8(in.ptr()); |
| 317 | uint8x16x4_t ta2; |
| 318 | ta2.val[0] = ta1.val[0]; |
| 319 | ta2.val[1] = ta1.val[1]; |
| 320 | ta2.val[2] = ta1.val[2]; |
| 321 | ta2.val[3] = vdupq_n_u8(255); |
| 322 | vst4q_u8(out.ptr(), ta2); |
| 323 | }, |
| 324 | in, out); |
| 325 | } |
| 326 | |
| 327 | void colorconvert_rgbx_to_rgb(const void *input, void *output, const Window &win) |
| 328 | { |
| 329 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 330 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 331 | |
| 332 | const auto input_ptr = static_cast<const IImage *__restrict>(input); |
| 333 | const auto output_ptr = static_cast<IImage *__restrict>(output); |
| 334 | |
| 335 | Iterator in(input_ptr, win); |
| 336 | Iterator out(output_ptr, win); |
| 337 | |
| 338 | execute_window_loop(win, [&](const Coordinates & id) |
| 339 | { |
| 340 | const auto ta1 = vld4q_u8(in.ptr()); |
| 341 | uint8x16x3_t ta2; |
| 342 | ta2.val[0] = ta1.val[0]; |
| 343 | ta2.val[1] = ta1.val[1]; |
| 344 | ta2.val[2] = ta1.val[2]; |
| 345 | vst3q_u8(out.ptr(), ta2); |
| 346 | }, |
| 347 | in, out); |
| 348 | } |
| 349 | |
| 350 | template <bool yuyv, bool alpha> |
| 351 | void colorconvert_yuyv_to_rgb(const void *__restrict input, void *__restrict output, const Window &win) |
| 352 | { |
| 353 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 354 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 355 | |
| 356 | const auto input_ptr = static_cast<const IImage *__restrict>(input); |
| 357 | const auto output_ptr = static_cast<IImage *__restrict>(output); |
| 358 | |
| 359 | constexpr auto element_size = alpha ? 32 : 24; |
| 360 | constexpr auto shift = yuyv ? 0 : 1; |
| 361 | |
| 362 | Iterator in(input_ptr, win); |
| 363 | Iterator out(output_ptr, win); |
| 364 | |
| 365 | execute_window_loop(win, [&](const Coordinates & id) |
| 366 | { |
| 367 | float32x4x4_t uvec, yvec, vvec, yyvec; |
| 368 | const auto ta = vld4q_u8(in.ptr()); |
| 369 | //ta.val[0] = Y0 Y2 Y4 Y6 ... |
| 370 | //ta.val[1] = U0 U2 U4 U6 ... |
| 371 | //ta.val[2] = Y1 Y3 Y5 Y7 ... |
| 372 | //ta.val[3] = V0 V2 V4 V7 ... |
| 373 | |
| 374 | // Convert the uint8x16x4_t to float32x4x4_t |
| 375 | convert_uint8x16_to_float32x4x4(ta.val[0 + shift], yvec); |
| 376 | convert_uint8x16_to_float32x4x4(ta.val[1 - shift], uvec); |
| 377 | convert_uint8x16_to_float32x4x4(ta.val[2 + shift], yyvec); |
| 378 | convert_uint8x16_to_float32x4x4(ta.val[3 - shift], vvec); |
| 379 | |
| 380 | yuyv_to_rgb_calculation(yvec.val[0], uvec.val[0], yyvec.val[0], vvec.val[0], out.ptr() + 0 * element_size, alpha); |
| 381 | yuyv_to_rgb_calculation(yvec.val[1], uvec.val[1], yyvec.val[1], vvec.val[1], out.ptr() + 1 * element_size, alpha); |
| 382 | yuyv_to_rgb_calculation(yvec.val[2], uvec.val[2], yyvec.val[2], vvec.val[2], out.ptr() + 2 * element_size, alpha); |
| 383 | yuyv_to_rgb_calculation(yvec.val[3], uvec.val[3], yyvec.val[3], vvec.val[3], out.ptr() + 3 * element_size, alpha); |
| 384 | }, |
| 385 | in, out); |
| 386 | } |
| 387 | |
| 388 | template <bool uv, bool alpha> |
| 389 | void colorconvert_nv12_to_rgb(const void *__restrict input, void *__restrict output, const Window &win) |
| 390 | { |
| 391 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 392 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 393 | win.validate(); |
| 394 | |
| 395 | const auto input_ptr = static_cast<const IMultiImage *__restrict>(input); |
| 396 | const auto output_ptr = static_cast<IImage *__restrict>(output); |
| 397 | |
| 398 | constexpr auto element_size = alpha ? 32 : 24; |
| 399 | const auto out_stride = output_ptr->info()->strides_in_bytes().y(); |
| 400 | constexpr auto shift = uv ? 0 : 1; |
| 401 | |
| 402 | // UV's width and height are subsampled |
| 403 | Window win_uv(win); |
| 404 | win_uv.set(Window::DimX, Window::Dimension(win_uv.x().start() / 2, win_uv.x().end() / 2, win.x().step() / 2)); |
| 405 | win_uv.set(Window::DimY, Window::Dimension(win_uv.y().start() / 2, win_uv.y().end() / 2, 1)); |
| 406 | win_uv.validate(); |
| 407 | |
| 408 | Iterator in_y(input_ptr->plane(0), win); |
| 409 | Iterator in_uv(input_ptr->plane(1), win_uv); |
| 410 | Iterator out(output_ptr, win); |
| 411 | |
| 412 | execute_window_loop(win, [&](const Coordinates & id) |
| 413 | { |
| 414 | const auto ta_y_top = vld2q_u8(in_y.ptr()); |
| 415 | const auto ta_y_bottom = vld2q_u8(in_y.ptr() + input_ptr->plane(0)->info()->strides_in_bytes().y()); |
| 416 | const auto ta_uv = vld2q_u8(in_uv.ptr()); |
| 417 | //ta_y.val[0] = Y0 Y2 Y4 Y6 ... |
| 418 | //ta_y.val[1] = Y1 Y3 Y5 Y7 ... |
| 419 | //ta_uv.val[0] = U0 U2 U4 U6 ... |
| 420 | //ta_uv.val[1] = V0 V2 V4 V6 ... |
| 421 | |
| 422 | // Convert the uint8x16x4_t to float32x4x4_t |
| 423 | float32x4x4_t yvec_top, yyvec_top, yvec_bottom, yyvec_bottom, uvec, vvec; |
| 424 | convert_uint8x16_to_float32x4x4(ta_y_top.val[0], yvec_top); |
| 425 | convert_uint8x16_to_float32x4x4(ta_y_top.val[1], yyvec_top); |
| 426 | convert_uint8x16_to_float32x4x4(ta_y_bottom.val[0], yvec_bottom); |
| 427 | convert_uint8x16_to_float32x4x4(ta_y_bottom.val[1], yyvec_bottom); |
| 428 | convert_uint8x16_to_float32x4x4(ta_uv.val[0 + shift], uvec); |
| 429 | convert_uint8x16_to_float32x4x4(ta_uv.val[1 - shift], vvec); |
| 430 | |
| 431 | yuyv_to_rgb_calculation(yvec_top.val[0], uvec.val[0], yyvec_top.val[0], vvec.val[0], out.ptr() + 0 * element_size, alpha); |
| 432 | yuyv_to_rgb_calculation(yvec_top.val[1], uvec.val[1], yyvec_top.val[1], vvec.val[1], out.ptr() + 1 * element_size, alpha); |
| 433 | yuyv_to_rgb_calculation(yvec_top.val[2], uvec.val[2], yyvec_top.val[2], vvec.val[2], out.ptr() + 2 * element_size, alpha); |
| 434 | yuyv_to_rgb_calculation(yvec_top.val[3], uvec.val[3], yyvec_top.val[3], vvec.val[3], out.ptr() + 3 * element_size, alpha); |
| 435 | |
| 436 | yuyv_to_rgb_calculation(yvec_bottom.val[0], uvec.val[0], yyvec_bottom.val[0], vvec.val[0], out.ptr() + out_stride + 0 * element_size, alpha); |
| 437 | yuyv_to_rgb_calculation(yvec_bottom.val[1], uvec.val[1], yyvec_bottom.val[1], vvec.val[1], out.ptr() + out_stride + 1 * element_size, alpha); |
| 438 | yuyv_to_rgb_calculation(yvec_bottom.val[2], uvec.val[2], yyvec_bottom.val[2], vvec.val[2], out.ptr() + out_stride + 2 * element_size, alpha); |
| 439 | yuyv_to_rgb_calculation(yvec_bottom.val[3], uvec.val[3], yyvec_bottom.val[3], vvec.val[3], out.ptr() + out_stride + 3 * element_size, alpha); |
| 440 | }, |
| 441 | in_y, in_uv, out); |
| 442 | } |
| 443 | |
| 444 | template <bool alpha> |
| 445 | void colorconvert_iyuv_to_rgb(const void *__restrict input, void *__restrict output, const Window &win) |
| 446 | { |
| 447 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 448 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 449 | win.validate(); |
| 450 | |
| 451 | const auto input_ptr = static_cast<const IMultiImage *__restrict>(input); |
| 452 | const auto output_ptr = static_cast<IImage *__restrict>(output); |
| 453 | |
| 454 | constexpr auto element_size = alpha ? 32 : 24; |
| 455 | const auto out_stride = output_ptr->info()->strides_in_bytes().y(); |
| 456 | |
| 457 | // UV's width and height are subsampled |
| 458 | Window win_uv(win); |
| 459 | win_uv.set(Window::DimX, Window::Dimension(win_uv.x().start() / 2, win_uv.x().end() / 2, win_uv.x().step() / 2)); |
| 460 | win_uv.set(Window::DimY, Window::Dimension(win_uv.y().start() / 2, win_uv.y().end() / 2, 1)); |
| 461 | win_uv.validate(); |
| 462 | |
| 463 | Iterator in_y(input_ptr->plane(0), win); |
| 464 | Iterator in_u(input_ptr->plane(1), win_uv); |
| 465 | Iterator in_v(input_ptr->plane(2), win_uv); |
| 466 | Iterator out(output_ptr, win); |
| 467 | |
| 468 | execute_window_loop(win, [&](const Coordinates & id) |
| 469 | { |
| 470 | const auto ta_y_top = vld2q_u8(in_y.ptr()); |
| 471 | const auto ta_y_bottom = vld2q_u8(in_y.ptr() + input_ptr->plane(0)->info()->strides_in_bytes().y()); |
| 472 | const auto ta_u = vld1q_u8(in_u.ptr()); |
| 473 | const auto ta_v = vld1q_u8(in_v.ptr()); |
| 474 | //ta_y.val[0] = Y0 Y2 Y4 Y6 ... |
| 475 | //ta_y.val[1] = Y1 Y3 Y5 Y7 ... |
| 476 | //ta_u.val[0] = U0 U2 U4 U6 ... |
| 477 | //ta_v.val[0] = V0 V2 V4 V6 ... |
| 478 | |
| 479 | // Convert the uint8x16x4_t to float32x4x4_t |
| 480 | float32x4x4_t yvec_top, yyvec_top, yvec_bottom, yyvec_bottom, uvec, vvec; |
| 481 | convert_uint8x16_to_float32x4x4(ta_y_top.val[0], yvec_top); |
| 482 | convert_uint8x16_to_float32x4x4(ta_y_top.val[1], yyvec_top); |
| 483 | convert_uint8x16_to_float32x4x4(ta_y_bottom.val[0], yvec_bottom); |
| 484 | convert_uint8x16_to_float32x4x4(ta_y_bottom.val[1], yyvec_bottom); |
| 485 | convert_uint8x16_to_float32x4x4(ta_u, uvec); |
| 486 | convert_uint8x16_to_float32x4x4(ta_v, vvec); |
| 487 | |
| 488 | yuyv_to_rgb_calculation(yvec_top.val[0], uvec.val[0], yyvec_top.val[0], vvec.val[0], out.ptr() + 0 * element_size, alpha); |
| 489 | yuyv_to_rgb_calculation(yvec_top.val[1], uvec.val[1], yyvec_top.val[1], vvec.val[1], out.ptr() + 1 * element_size, alpha); |
| 490 | yuyv_to_rgb_calculation(yvec_top.val[2], uvec.val[2], yyvec_top.val[2], vvec.val[2], out.ptr() + 2 * element_size, alpha); |
| 491 | yuyv_to_rgb_calculation(yvec_top.val[3], uvec.val[3], yyvec_top.val[3], vvec.val[3], out.ptr() + 3 * element_size, alpha); |
| 492 | |
| 493 | yuyv_to_rgb_calculation(yvec_bottom.val[0], uvec.val[0], yyvec_bottom.val[0], vvec.val[0], out.ptr() + out_stride + 0 * element_size, alpha); |
| 494 | yuyv_to_rgb_calculation(yvec_bottom.val[1], uvec.val[1], yyvec_bottom.val[1], vvec.val[1], out.ptr() + out_stride + 1 * element_size, alpha); |
| 495 | yuyv_to_rgb_calculation(yvec_bottom.val[2], uvec.val[2], yyvec_bottom.val[2], vvec.val[2], out.ptr() + out_stride + 2 * element_size, alpha); |
| 496 | yuyv_to_rgb_calculation(yvec_bottom.val[3], uvec.val[3], yyvec_bottom.val[3], vvec.val[3], out.ptr() + out_stride + 3 * element_size, alpha); |
| 497 | }, |
| 498 | in_y, in_u, in_v, out); |
| 499 | } |
| 500 | |
| 501 | template <bool yuyv> |
| 502 | void colorconvert_yuyv_to_nv12(const void *__restrict input, void *__restrict output, const Window &win) |
| 503 | { |
| 504 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 505 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 506 | win.validate(); |
| 507 | |
| 508 | const auto input_ptr = static_cast<const IImage *__restrict>(input); |
| 509 | const auto output_ptr = static_cast<IMultiImage *__restrict>(output); |
| 510 | |
| 511 | constexpr auto shift = yuyv ? 0 : 1; |
| 512 | |
| 513 | // NV12's UV's width and height are subsampled |
| 514 | Window win_uv(win); |
| 515 | win_uv.set(Window::DimX, Window::Dimension(win_uv.x().start() / 2, win_uv.x().end() / 2, win_uv.x().step() / 2)); |
| 516 | win_uv.set(Window::DimY, Window::Dimension(win_uv.y().start() / 2, win_uv.y().end() / 2, 1)); |
| 517 | win_uv.validate(); |
| 518 | |
| 519 | Iterator in(input_ptr, win); |
| 520 | Iterator out_y(output_ptr->plane(0), win); |
| 521 | Iterator out_uv(output_ptr->plane(1), win_uv); |
| 522 | |
| 523 | execute_window_loop(win, [&](const Coordinates & id) |
| 524 | { |
| 525 | const auto ta_top = vld4q_u8(in.ptr()); |
| 526 | const auto ta_bottom = vld4q_u8(in.ptr() + input_ptr->info()->strides_in_bytes().y()); |
| 527 | //ta.val[0] = Y0 Y2 Y4 Y6 ... |
| 528 | //ta.val[1] = U0 U2 U4 U6 ... |
| 529 | //ta.val[2] = Y1 Y3 Y5 Y7 ... |
| 530 | //ta.val[3] = V0 V2 V4 V7 ... |
| 531 | |
| 532 | uint8x16x2_t yvec; |
| 533 | yvec.val[0] = ta_top.val[0 + shift]; |
| 534 | yvec.val[1] = ta_top.val[2 + shift]; |
| 535 | vst2q_u8(out_y.ptr(), yvec); |
| 536 | |
| 537 | uint8x16x2_t yyvec; |
| 538 | yyvec.val[0] = ta_bottom.val[0 + shift]; |
| 539 | yyvec.val[1] = ta_bottom.val[2 + shift]; |
| 540 | vst2q_u8(out_y.ptr() + output_ptr->plane(0)->info()->strides_in_bytes().y(), yyvec); |
| 541 | |
| 542 | uint8x16x2_t uvvec; |
| 543 | uvvec.val[0] = vhaddq_u8(ta_top.val[1 - shift], ta_bottom.val[1 - shift]); |
| 544 | uvvec.val[1] = vhaddq_u8(ta_top.val[3 - shift], ta_bottom.val[3 - shift]); |
| 545 | vst2q_u8(out_uv.ptr(), uvvec); |
| 546 | }, |
| 547 | in, out_y, out_uv); |
| 548 | } |
| 549 | |
| 550 | void colorconvert_iyuv_to_nv12(const void *__restrict input, void *__restrict output, const Window &win) |
| 551 | { |
| 552 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 553 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 554 | win.validate(); |
| 555 | |
| 556 | const auto input_ptr = static_cast<const IMultiImage *__restrict>(input); |
| 557 | const auto output_ptr = static_cast<IMultiImage *__restrict>(output); |
| 558 | |
| 559 | // UV's width and height are subsampled |
| 560 | Window win_uv(win); |
| 561 | win_uv.set(Window::DimX, Window::Dimension(win_uv.x().start() / 2, win_uv.x().end() / 2, win_uv.x().step() / 2)); |
| 562 | win_uv.set(Window::DimY, Window::Dimension(win_uv.y().start() / 2, win_uv.y().end() / 2, 1)); |
| 563 | win_uv.validate(); |
| 564 | |
| 565 | Iterator in_y(input_ptr->plane(0), win); |
| 566 | Iterator in_u(input_ptr->plane(1), win_uv); |
| 567 | Iterator in_v(input_ptr->plane(2), win_uv); |
| 568 | Iterator out_y(output_ptr->plane(0), win); |
| 569 | Iterator out_uv(output_ptr->plane(1), win_uv); |
| 570 | |
| 571 | execute_window_loop(win, [&](const Coordinates & id) |
| 572 | { |
| 573 | const auto ta_y_top = vld2q_u8(in_y.ptr()); |
| 574 | const auto ta_y_bottom = vld2q_u8(in_y.ptr() + input_ptr->plane(0)->info()->strides_in_bytes().y()); |
| 575 | uint8x16x2_t ta_uv; |
| 576 | ta_uv.val[0] = vld1q_u8(in_u.ptr()); |
| 577 | ta_uv.val[1] = vld1q_u8(in_v.ptr()); |
| 578 | //ta_y.val[0] = Y0 Y2 Y4 Y6 ... |
| 579 | //ta_y.val[1] = Y1 Y3 Y5 Y7 ... |
| 580 | //ta_uv.val[0] = U0 U2 U4 U6 ... |
| 581 | //ta_uv.val[1] = V0 V2 V4 V6 ... |
| 582 | |
| 583 | vst2q_u8(out_y.ptr(), ta_y_top); |
| 584 | vst2q_u8(out_y.ptr() + output_ptr->plane(0)->info()->strides_in_bytes().y(), ta_y_bottom); |
| 585 | vst2q_u8(out_uv.ptr(), ta_uv); |
| 586 | }, |
| 587 | in_y, in_u, in_v, out_y, out_uv); |
| 588 | } |
| 589 | |
| 590 | template <bool uv> |
| 591 | void colorconvert_nv12_to_iyuv(const void *__restrict input, void *__restrict output, const Window &win) |
| 592 | { |
| 593 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 594 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 595 | win.validate(); |
| 596 | |
| 597 | const auto input_ptr = static_cast<const IMultiImage *__restrict>(input); |
| 598 | const auto output_ptr = static_cast<IMultiImage *__restrict>(output); |
| 599 | |
| 600 | constexpr auto shift = uv ? 0 : 1; |
| 601 | |
| 602 | // UV's width and height are subsampled |
| 603 | Window win_uv(win); |
| 604 | win_uv.set(Window::DimX, Window::Dimension(win_uv.x().start() / 2, win_uv.x().end() / 2, win_uv.x().step() / 2)); |
| 605 | win_uv.set(Window::DimY, Window::Dimension(win_uv.y().start() / 2, win_uv.y().end() / 2, 1)); |
| 606 | win_uv.validate(); |
| 607 | |
| 608 | Iterator in_y(input_ptr->plane(0), win); |
| 609 | Iterator in_uv(input_ptr->plane(1), win_uv); |
| 610 | Iterator out_y(output_ptr->plane(0), win); |
| 611 | Iterator out_u(output_ptr->plane(1), win_uv); |
| 612 | Iterator out_v(output_ptr->plane(2), win_uv); |
| 613 | |
| 614 | execute_window_loop(win, [&](const Coordinates & id) |
| 615 | { |
| 616 | const auto ta_y_top = vld2q_u8(in_y.ptr()); |
| 617 | const auto ta_y_bottom = vld2q_u8(in_y.ptr() + input_ptr->plane(0)->info()->strides_in_bytes().y()); |
| 618 | const auto ta_uv = vld2q_u8(in_uv.ptr()); |
| 619 | //ta_y.val[0] = Y0 Y2 Y4 Y6 ... |
| 620 | //ta_y.val[1] = Y1 Y3 Y5 Y7 ... |
| 621 | //ta_uv.val[0] = U0 U2 U4 U6 ... |
| 622 | //ta_uv.val[1] = V0 V2 V4 V6 ... |
| 623 | |
| 624 | vst2q_u8(out_y.ptr(), ta_y_top); |
| 625 | vst2q_u8(out_y.ptr() + output_ptr->plane(0)->info()->strides_in_bytes().y(), ta_y_bottom); |
| 626 | vst1q_u8(out_u.ptr(), ta_uv.val[0 + shift]); |
| 627 | vst1q_u8(out_v.ptr(), ta_uv.val[1 - shift]); |
| 628 | }, |
| 629 | in_y, in_uv, out_y, out_u, out_v); |
| 630 | } |
| 631 | |
| 632 | template <bool yuyv> |
| 633 | void colorconvert_yuyv_to_iyuv(const void *__restrict input, void *__restrict output, const Window &win) |
| 634 | { |
| 635 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 636 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 637 | win.validate(); |
| 638 | |
| 639 | const auto input_ptr = static_cast<const IImage *__restrict>(input); |
| 640 | const auto output_ptr = static_cast<IMultiImage *__restrict>(output); |
| 641 | |
| 642 | constexpr auto shift = yuyv ? 0 : 1; |
| 643 | |
| 644 | // Destination's UV's width and height are subsampled |
| 645 | Window win_uv(win); |
| 646 | win_uv.set(Window::DimX, Window::Dimension(win_uv.x().start() / 2, win_uv.x().end() / 2, win_uv.x().step() / 2)); |
| 647 | win_uv.set(Window::DimY, Window::Dimension(win_uv.y().start() / 2, win_uv.y().end() / 2, 1)); |
| 648 | win_uv.validate(); |
| 649 | |
| 650 | Iterator in(input_ptr, win); |
| 651 | Iterator out_y(output_ptr->plane(0), win); |
| 652 | Iterator out_u(output_ptr->plane(1), win_uv); |
| 653 | Iterator out_v(output_ptr->plane(2), win_uv); |
| 654 | |
| 655 | execute_window_loop(win, [&](const Coordinates & id) |
| 656 | { |
| 657 | const auto ta_top = vld4q_u8(in.ptr()); |
| 658 | const auto ta_bottom = vld4q_u8(in.ptr() + input_ptr->info()->strides_in_bytes().y()); |
| 659 | //ta.val[0] = Y0 Y2 Y4 Y6 ... |
| 660 | //ta.val[1] = U0 U2 U4 U6 ... |
| 661 | //ta.val[2] = Y1 Y3 Y5 Y7 ... |
| 662 | //ta.val[3] = V0 V2 V4 V7 ... |
| 663 | |
| 664 | uint8x16x2_t yvec; |
| 665 | yvec.val[0] = ta_top.val[0 + shift]; |
| 666 | yvec.val[1] = ta_top.val[2 + shift]; |
| 667 | vst2q_u8(out_y.ptr(), yvec); |
| 668 | |
| 669 | uint8x16x2_t yyvec; |
| 670 | yyvec.val[0] = ta_bottom.val[0 + shift]; |
| 671 | yyvec.val[1] = ta_bottom.val[2 + shift]; |
| 672 | vst2q_u8(out_y.ptr() + output_ptr->plane(0)->info()->strides_in_bytes().y(), yyvec); |
| 673 | |
| 674 | uint8x16_t uvec; |
| 675 | uvec = vhaddq_u8(ta_top.val[1 - shift], ta_bottom.val[1 - shift]); |
| 676 | vst1q_u8(out_u.ptr(), uvec); |
| 677 | |
| 678 | uint8x16_t vvec; |
| 679 | vvec = vhaddq_u8(ta_top.val[3 - shift], ta_bottom.val[3 - shift]); |
| 680 | vst1q_u8(out_v.ptr(), vvec); |
| 681 | }, |
| 682 | in, out_y, out_u, out_v); |
| 683 | } |
| 684 | |
| 685 | template <bool uv> |
| 686 | void colorconvert_nv12_to_yuv4(const void *__restrict input, void *__restrict output, const Window &win) |
| 687 | { |
| 688 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 689 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 690 | win.validate(); |
| 691 | |
| 692 | const auto input_ptr = static_cast<const IMultiImage *__restrict>(input); |
| 693 | const auto output_ptr = static_cast<IMultiImage *__restrict>(output); |
| 694 | |
| 695 | constexpr auto shift = uv ? 0 : 1; |
| 696 | |
| 697 | // UV's width and height are subsampled |
| 698 | Window win_uv(win); |
| 699 | win_uv.set(Window::DimX, Window::Dimension(win_uv.x().start() / 2, win_uv.x().end() / 2, win_uv.x().step() / 2)); |
| 700 | win_uv.set(Window::DimY, Window::Dimension(win_uv.y().start() / 2, win_uv.y().end() / 2, 1)); |
| 701 | win_uv.validate(); |
| 702 | |
| 703 | Iterator in_y(input_ptr->plane(0), win); |
| 704 | Iterator in_uv(input_ptr->plane(1), win_uv); |
| 705 | Iterator out_y(output_ptr->plane(0), win); |
| 706 | Iterator out_u(output_ptr->plane(1), win); |
| 707 | Iterator out_v(output_ptr->plane(2), win); |
| 708 | |
| 709 | execute_window_loop(win, [&](const Coordinates & id) |
| 710 | { |
| 711 | const auto ta_y_top = vld2q_u8(in_y.ptr()); |
| 712 | const auto ta_y_bottom = vld2q_u8(in_y.ptr() + input_ptr->plane(0)->info()->strides_in_bytes().y()); |
| 713 | const auto ta_uv = vld2q_u8(in_uv.ptr()); |
| 714 | //ta_y.val[0] = Y0 Y2 Y4 Y6 ... |
| 715 | //ta_y.val[1] = Y1 Y3 Y5 Y7 ... |
| 716 | //ta_uv.val[0] = U0 U2 U4 U6 ... |
| 717 | //ta_uv.val[1] = V0 V2 V4 V6 ... |
| 718 | |
| 719 | vst2q_u8(out_y.ptr(), ta_y_top); |
| 720 | vst2q_u8(out_y.ptr() + output_ptr->plane(0)->info()->strides_in_bytes().y(), ta_y_bottom); |
| 721 | |
| 722 | uint8x16x2_t uvec; |
| 723 | uvec.val[0] = ta_uv.val[0 + shift]; |
| 724 | uvec.val[1] = ta_uv.val[0 + shift]; |
| 725 | vst2q_u8(out_u.ptr(), uvec); |
| 726 | vst2q_u8(out_u.ptr() + output_ptr->plane(1)->info()->strides_in_bytes().y(), uvec); |
| 727 | |
| 728 | uint8x16x2_t vvec; |
| 729 | vvec.val[0] = ta_uv.val[1 - shift]; |
| 730 | vvec.val[1] = ta_uv.val[1 - shift]; |
| 731 | vst2q_u8(out_v.ptr(), vvec); |
| 732 | vst2q_u8(out_v.ptr() + output_ptr->plane(2)->info()->strides_in_bytes().y(), vvec); |
| 733 | }, |
| 734 | in_y, in_uv, out_y, out_u, out_v); |
| 735 | } |
| 736 | |
| 737 | void colorconvert_iyuv_to_yuv4(const void *__restrict input, void *__restrict output, const Window &win) |
| 738 | { |
| 739 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 740 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 741 | win.validate(); |
| 742 | |
| 743 | const auto input_ptr = static_cast<const IMultiImage *__restrict>(input); |
| 744 | const auto output_ptr = static_cast<IMultiImage *__restrict>(output); |
| 745 | |
| 746 | // UV's width and height are subsampled |
| 747 | Window win_uv(win); |
| 748 | win_uv.set(Window::DimX, Window::Dimension(win_uv.x().start() / 2, win_uv.x().end() / 2, win_uv.x().step() / 2)); |
| 749 | win_uv.set(Window::DimY, Window::Dimension(win_uv.y().start() / 2, win_uv.y().end() / 2, 1)); |
| 750 | win_uv.validate(); |
| 751 | |
| 752 | Iterator in_y(input_ptr->plane(0), win); |
| 753 | Iterator in_u(input_ptr->plane(1), win_uv); |
| 754 | Iterator in_v(input_ptr->plane(2), win_uv); |
| 755 | Iterator out_y(output_ptr->plane(0), win); |
| 756 | Iterator out_u(output_ptr->plane(1), win); |
| 757 | Iterator out_v(output_ptr->plane(2), win); |
| 758 | |
| 759 | execute_window_loop(win, [&](const Coordinates & id) |
| 760 | { |
| 761 | const auto ta_y_top = vld2q_u8(in_y.ptr()); |
| 762 | const auto ta_y_bottom = vld2q_u8(in_y.ptr() + input_ptr->plane(0)->info()->strides_in_bytes().y()); |
| 763 | const auto ta_u = vld1q_u8(in_u.ptr()); |
| 764 | const auto ta_v = vld1q_u8(in_v.ptr()); |
| 765 | //ta_y.val[0] = Y0 Y2 Y4 Y6 ... |
| 766 | //ta_y.val[1] = Y1 Y3 Y5 Y7 ... |
| 767 | //ta_u = U0 U2 U4 U6 ... |
| 768 | //ta_v = V0 V2 V4 V6 ... |
| 769 | |
| 770 | vst2q_u8(out_y.ptr(), ta_y_top); |
| 771 | vst2q_u8(out_y.ptr() + output_ptr->plane(0)->info()->strides_in_bytes().y(), ta_y_bottom); |
| 772 | |
| 773 | uint8x16x2_t uvec; |
| 774 | uvec.val[0] = ta_u; |
| 775 | uvec.val[1] = ta_u; |
| 776 | vst2q_u8(out_u.ptr(), uvec); |
| 777 | vst2q_u8(out_u.ptr() + output_ptr->plane(1)->info()->strides_in_bytes().y(), uvec); |
| 778 | |
| 779 | uint8x16x2_t vvec; |
| 780 | vvec.val[0] = ta_v; |
| 781 | vvec.val[1] = ta_v; |
| 782 | vst2q_u8(out_v.ptr(), vvec); |
| 783 | vst2q_u8(out_v.ptr() + output_ptr->plane(2)->info()->strides_in_bytes().y(), vvec); |
| 784 | }, |
| 785 | in_y, in_u, in_v, out_y, out_u, out_v); |
| 786 | } |
| 787 | |
| 788 | template <bool alpha> |
| 789 | void colorconvert_rgb_to_nv12(const void *__restrict input, void *__restrict output, const Window &win) |
| 790 | { |
| 791 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 792 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 793 | win.validate(); |
| 794 | |
| 795 | const auto input_ptr = static_cast<const IImage *__restrict>(input); |
| 796 | const auto output_ptr = static_cast<IMultiImage *__restrict>(output); |
| 797 | |
| 798 | // UV's width and height are subsampled |
| 799 | Window win_uv(win); |
| 800 | win_uv.set(Window::DimX, Window::Dimension(win_uv.x().start() / 2, win_uv.x().end() / 2, win_uv.x().step() / 2)); |
| 801 | win_uv.set(Window::DimY, Window::Dimension(win_uv.y().start() / 2, win_uv.y().end() / 2, 1)); |
| 802 | win_uv.validate(); |
| 803 | |
| 804 | Iterator in(input_ptr, win); |
| 805 | Iterator out_y(output_ptr->plane(0), win); |
| 806 | Iterator out_uv(output_ptr->plane(1), win_uv); |
| 807 | |
| 808 | execute_window_loop(win, [&](const Coordinates & id) |
| 809 | { |
| 810 | const auto ta_rgb_top = load_rgb(in.ptr(), alpha); |
| 811 | const auto ta_rgb_bottom = load_rgb(in.ptr() + input_ptr->info()->strides_in_bytes().y(), alpha); |
| 812 | //ta_rgb.val[0] = R0 R1 R2 R3 ... |
| 813 | //ta_rgb.val[1] = G0 G1 G2 G3 ... |
| 814 | //ta_rgb.val[2] = B0 B1 B2 B3 ... |
| 815 | |
| 816 | store_rgb_to_nv12(ta_rgb_top.val[0], ta_rgb_top.val[1], ta_rgb_top.val[2], |
| 817 | ta_rgb_bottom.val[0], ta_rgb_bottom.val[1], ta_rgb_bottom.val[2], |
| 818 | out_y.ptr(), out_y.ptr() + output_ptr->plane(0)->info()->strides_in_bytes().y(), |
| 819 | out_uv.ptr()); |
| 820 | }, |
| 821 | in, out_y, out_uv); |
| 822 | } |
| 823 | |
| 824 | template <bool alpha> |
| 825 | void colorconvert_rgb_to_iyuv(const void *__restrict input, void *__restrict output, const Window &win) |
| 826 | { |
| 827 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 828 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 829 | win.validate(); |
| 830 | |
| 831 | const auto input_ptr = static_cast<const IImage *__restrict>(input); |
| 832 | const auto output_ptr = static_cast<IMultiImage *__restrict>(output); |
| 833 | |
| 834 | // UV's width and height are subsampled |
| 835 | Window win_uv(win); |
| 836 | win_uv.set(Window::DimX, Window::Dimension(win_uv.x().start() / 2, win_uv.x().end() / 2, win_uv.x().step() / 2)); |
| 837 | win_uv.set(Window::DimY, Window::Dimension(win_uv.y().start() / 2, win_uv.y().end() / 2, 1)); |
| 838 | win_uv.validate(); |
| 839 | |
| 840 | Iterator in(input_ptr, win); |
| 841 | Iterator out_y(output_ptr->plane(0), win); |
| 842 | Iterator out_u(output_ptr->plane(1), win_uv); |
| 843 | Iterator out_v(output_ptr->plane(2), win_uv); |
| 844 | |
| 845 | execute_window_loop(win, [&](const Coordinates & id) |
| 846 | { |
| 847 | const auto ta_rgb_top = load_rgb(in.ptr(), alpha); |
| 848 | const auto ta_rgb_bottom = load_rgb(in.ptr() + input_ptr->info()->strides_in_bytes().y(), alpha); |
| 849 | //ta_rgb.val[0] = R0 R1 R2 R3 ... |
| 850 | //ta_rgb.val[1] = G0 G1 G2 G3 ... |
| 851 | //ta_rgb.val[2] = B0 B1 B2 B3 ... |
| 852 | |
| 853 | store_rgb_to_iyuv(ta_rgb_top.val[0], ta_rgb_top.val[1], ta_rgb_top.val[2], |
| 854 | ta_rgb_bottom.val[0], ta_rgb_bottom.val[1], ta_rgb_bottom.val[2], |
| 855 | out_y.ptr(), out_y.ptr() + output_ptr->plane(0)->info()->strides_in_bytes().y(), |
| 856 | out_u.ptr(), out_v.ptr()); |
| 857 | }, |
| 858 | in, out_y, out_u, out_v); |
| 859 | } |
| 860 | |
| 861 | template <bool alpha> |
| 862 | void colorconvert_rgb_to_yuv4(const void *__restrict input, void *__restrict output, const Window &win) |
| 863 | { |
| 864 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 865 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 866 | win.validate(); |
| 867 | |
| 868 | const auto input_ptr = static_cast<const IImage *__restrict>(input); |
| 869 | const auto output_ptr = static_cast<IMultiImage *__restrict>(output); |
| 870 | |
| 871 | Iterator in(input_ptr, win); |
| 872 | Iterator out_y(output_ptr->plane(0), win); |
| 873 | Iterator out_u(output_ptr->plane(1), win); |
| 874 | Iterator out_v(output_ptr->plane(2), win); |
| 875 | |
| 876 | execute_window_loop(win, [&](const Coordinates & id) |
| 877 | { |
| 878 | const auto ta_rgb = load_rgb(in.ptr(), alpha); |
| 879 | //ta_rgb.val[0] = R0 R1 R2 R3 ... |
| 880 | //ta_rgb.val[1] = G0 G1 G2 G3 ... |
| 881 | //ta_rgb.val[2] = B0 B1 B2 B3 ... |
| 882 | |
| 883 | store_rgb_to_yuv4(ta_rgb.val[0], ta_rgb.val[1], ta_rgb.val[2], |
| 884 | out_y.ptr(), out_u.ptr(), out_v.ptr()); |
| 885 | }, |
| 886 | in, out_y, out_u, out_v); |
| 887 | } |
Gian Marco Iodice | 356f643 | 2017-09-22 11:32:21 +0100 | [diff] [blame] | 888 | } // namespace arm_compute |