| OLD | NEW |
| 1 /* | 1 /* |
| 2 * Copyright 2014 The LibYuv Project Authors. All rights reserved. | 2 * Copyright 2014 The LibYuv Project Authors. All rights reserved. |
| 3 * | 3 * |
| 4 * Use of this source code is governed by a BSD-style license | 4 * Use of this source code is governed by a BSD-style license |
| 5 * that can be found in the LICENSE file in the root of the source | 5 * that can be found in the LICENSE file in the root of the source |
| 6 * tree. An additional intellectual property rights grant can be found | 6 * tree. An additional intellectual property rights grant can be found |
| 7 * in the file PATENTS. All contributing project authors may | 7 * in the file PATENTS. All contributing project authors may |
| 8 * be found in the AUTHORS file in the root of the source tree. | 8 * be found in the AUTHORS file in the root of the source tree. |
| 9 */ | 9 */ |
| 10 | 10 |
| (...skipping 123 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 134 "movi v22.8h, #19, lsl #0 \n" /* UR coefficient (-0.2969) / 2 */ \ | 134 "movi v22.8h, #19, lsl #0 \n" /* UR coefficient (-0.2969) / 2 */ \ |
| 135 "movi v23.8h, #9, lsl #0 \n" /* VB coefficient (-0.1406) / 2 */ \ | 135 "movi v23.8h, #9, lsl #0 \n" /* VB coefficient (-0.1406) / 2 */ \ |
| 136 "movi v24.8h, #47, lsl #0 \n" /* VG coefficient (-0.7344) / 2 */ \ | 136 "movi v24.8h, #47, lsl #0 \n" /* VG coefficient (-0.7344) / 2 */ \ |
| 137 "movi v25.16b, #0x80 \n" /* 128.5 (0x8080 in 16-bit) */ | 137 "movi v25.16b, #0x80 \n" /* 128.5 (0x8080 in 16-bit) */ |
| 138 | 138 |
| 139 #ifdef HAS_I444TOARGBROW_NEON | 139 #ifdef HAS_I444TOARGBROW_NEON |
| 140 void I444ToARGBRow_NEON(const uint8* src_y, | 140 void I444ToARGBRow_NEON(const uint8* src_y, |
| 141 const uint8* src_u, | 141 const uint8* src_u, |
| 142 const uint8* src_v, | 142 const uint8* src_v, |
| 143 uint8* dst_argb, | 143 uint8* dst_argb, |
| 144 struct YuvConstants* yuvconstants, | 144 const struct YuvConstants* yuvconstants, |
| 145 int width) { | 145 int width) { |
| 146 asm volatile ( | 146 asm volatile ( |
| 147 YUVTORGB_SETUP | 147 YUVTORGB_SETUP |
| 148 "1: \n" | 148 "1: \n" |
| 149 READYUV444 | 149 READYUV444 |
| 150 YUVTORGB(v22, v21, v20) | 150 YUVTORGB(v22, v21, v20) |
| 151 "subs %w4, %w4, #8 \n" | 151 "subs %w4, %w4, #8 \n" |
| 152 "movi v23.8b, #255 \n" /* A */ | 152 "movi v23.8b, #255 \n" /* A */ |
| 153 MEMACCESS(3) | 153 MEMACCESS(3) |
| 154 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" | 154 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" |
| (...skipping 12 matching lines...) Expand all Loading... |
| 167 ); | 167 ); |
| 168 } | 168 } |
| 169 #endif // HAS_I444TOARGBROW_NEON | 169 #endif // HAS_I444TOARGBROW_NEON |
| 170 | 170 |
| 171 // TODO(fbarchard): Switch to Matrix version of this function. | 171 // TODO(fbarchard): Switch to Matrix version of this function. |
| 172 #ifdef HAS_I422TOARGBROW_NEON | 172 #ifdef HAS_I422TOARGBROW_NEON |
| 173 void I422ToARGBRow_NEON(const uint8* src_y, | 173 void I422ToARGBRow_NEON(const uint8* src_y, |
| 174 const uint8* src_u, | 174 const uint8* src_u, |
| 175 const uint8* src_v, | 175 const uint8* src_v, |
| 176 uint8* dst_argb, | 176 uint8* dst_argb, |
| 177 struct YuvConstants* yuvconstants, | 177 const struct YuvConstants* yuvconstants, |
| 178 int width) { | 178 int width) { |
| 179 asm volatile ( | 179 asm volatile ( |
| 180 YUVTORGB_SETUP | 180 YUVTORGB_SETUP |
| 181 "1: \n" | 181 "1: \n" |
| 182 READYUV422 | 182 READYUV422 |
| 183 YUVTORGB(v22, v21, v20) | 183 YUVTORGB(v22, v21, v20) |
| 184 "subs %w4, %w4, #8 \n" | 184 "subs %w4, %w4, #8 \n" |
| 185 "movi v23.8b, #255 \n" /* A */ | 185 "movi v23.8b, #255 \n" /* A */ |
| 186 MEMACCESS(3) | 186 MEMACCESS(3) |
| 187 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" | 187 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" |
| (...skipping 11 matching lines...) Expand all Loading... |
| 199 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" | 199 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" |
| 200 ); | 200 ); |
| 201 } | 201 } |
| 202 #endif // HAS_I422TOARGBROW_NEON | 202 #endif // HAS_I422TOARGBROW_NEON |
| 203 | 203 |
| 204 #ifdef HAS_I411TOARGBROW_NEON | 204 #ifdef HAS_I411TOARGBROW_NEON |
| 205 void I411ToARGBRow_NEON(const uint8* src_y, | 205 void I411ToARGBRow_NEON(const uint8* src_y, |
| 206 const uint8* src_u, | 206 const uint8* src_u, |
| 207 const uint8* src_v, | 207 const uint8* src_v, |
| 208 uint8* dst_argb, | 208 uint8* dst_argb, |
| 209 struct YuvConstants* yuvconstants, | 209 const struct YuvConstants* yuvconstants, |
| 210 int width) { | 210 int width) { |
| 211 asm volatile ( | 211 asm volatile ( |
| 212 YUVTORGB_SETUP | 212 YUVTORGB_SETUP |
| 213 "1: \n" | 213 "1: \n" |
| 214 READYUV411 | 214 READYUV411 |
| 215 YUVTORGB(v22, v21, v20) | 215 YUVTORGB(v22, v21, v20) |
| 216 "subs %w4, %w4, #8 \n" | 216 "subs %w4, %w4, #8 \n" |
| 217 "movi v23.8b, #255 \n" /* A */ | 217 "movi v23.8b, #255 \n" /* A */ |
| 218 MEMACCESS(3) | 218 MEMACCESS(3) |
| 219 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" | 219 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" |
| (...skipping 11 matching lines...) Expand all Loading... |
| 231 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" | 231 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" |
| 232 ); | 232 ); |
| 233 } | 233 } |
| 234 #endif // HAS_I411TOARGBROW_NEON | 234 #endif // HAS_I411TOARGBROW_NEON |
| 235 | 235 |
| 236 #ifdef HAS_I422TOBGRAROW_NEON | 236 #ifdef HAS_I422TOBGRAROW_NEON |
| 237 void I422ToBGRARow_NEON(const uint8* src_y, | 237 void I422ToBGRARow_NEON(const uint8* src_y, |
| 238 const uint8* src_u, | 238 const uint8* src_u, |
| 239 const uint8* src_v, | 239 const uint8* src_v, |
| 240 uint8* dst_bgra, | 240 uint8* dst_bgra, |
| 241 struct YuvConstants* yuvconstants, | 241 const struct YuvConstants* yuvconstants, |
| 242 int width) { | 242 int width) { |
| 243 asm volatile ( | 243 asm volatile ( |
| 244 YUVTORGB_SETUP | 244 YUVTORGB_SETUP |
| 245 "1: \n" | 245 "1: \n" |
| 246 READYUV422 | 246 READYUV422 |
| 247 YUVTORGB(v21, v22, v23) | 247 YUVTORGB(v21, v22, v23) |
| 248 "subs %w4, %w4, #8 \n" | 248 "subs %w4, %w4, #8 \n" |
| 249 "movi v20.8b, #255 \n" /* A */ | 249 "movi v20.8b, #255 \n" /* A */ |
| 250 MEMACCESS(3) | 250 MEMACCESS(3) |
| 251 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" | 251 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" |
| (...skipping 12 matching lines...) Expand all Loading... |
| 264 ); | 264 ); |
| 265 } | 265 } |
| 266 #endif // HAS_I422TOBGRAROW_NEON | 266 #endif // HAS_I422TOBGRAROW_NEON |
| 267 | 267 |
| 268 // TODO(fbarchard): Switch to Matrix version of this function. | 268 // TODO(fbarchard): Switch to Matrix version of this function. |
| 269 #ifdef HAS_I422TOABGRROW_NEON | 269 #ifdef HAS_I422TOABGRROW_NEON |
| 270 void I422ToABGRRow_NEON(const uint8* src_y, | 270 void I422ToABGRRow_NEON(const uint8* src_y, |
| 271 const uint8* src_u, | 271 const uint8* src_u, |
| 272 const uint8* src_v, | 272 const uint8* src_v, |
| 273 uint8* dst_abgr, | 273 uint8* dst_abgr, |
| 274 struct YuvConstants* yuvconstants, | 274 const struct YuvConstants* yuvconstants, |
| 275 int width) { | 275 int width) { |
| 276 asm volatile ( | 276 asm volatile ( |
| 277 YUVTORGB_SETUP | 277 YUVTORGB_SETUP |
| 278 "1: \n" | 278 "1: \n" |
| 279 READYUV422 | 279 READYUV422 |
| 280 YUVTORGB(v20, v21, v22) | 280 YUVTORGB(v20, v21, v22) |
| 281 "subs %w4, %w4, #8 \n" | 281 "subs %w4, %w4, #8 \n" |
| 282 "movi v23.8b, #255 \n" /* A */ | 282 "movi v23.8b, #255 \n" /* A */ |
| 283 MEMACCESS(3) | 283 MEMACCESS(3) |
| 284 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" | 284 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" |
| (...skipping 11 matching lines...) Expand all Loading... |
| 296 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" | 296 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" |
| 297 ); | 297 ); |
| 298 } | 298 } |
| 299 #endif // HAS_I422TOABGRROW_NEON | 299 #endif // HAS_I422TOABGRROW_NEON |
| 300 | 300 |
| 301 #ifdef HAS_I422TORGBAROW_NEON | 301 #ifdef HAS_I422TORGBAROW_NEON |
| 302 void I422ToRGBARow_NEON(const uint8* src_y, | 302 void I422ToRGBARow_NEON(const uint8* src_y, |
| 303 const uint8* src_u, | 303 const uint8* src_u, |
| 304 const uint8* src_v, | 304 const uint8* src_v, |
| 305 uint8* dst_rgba, | 305 uint8* dst_rgba, |
| 306 struct YuvConstants* yuvconstants, | 306 const struct YuvConstants* yuvconstants, |
| 307 int width) { | 307 int width) { |
| 308 asm volatile ( | 308 asm volatile ( |
| 309 YUVTORGB_SETUP | 309 YUVTORGB_SETUP |
| 310 "1: \n" | 310 "1: \n" |
| 311 READYUV422 | 311 READYUV422 |
| 312 YUVTORGB(v23, v22, v21) | 312 YUVTORGB(v23, v22, v21) |
| 313 "subs %w4, %w4, #8 \n" | 313 "subs %w4, %w4, #8 \n" |
| 314 "movi v20.8b, #255 \n" /* A */ | 314 "movi v20.8b, #255 \n" /* A */ |
| 315 MEMACCESS(3) | 315 MEMACCESS(3) |
| 316 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" | 316 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" |
| (...skipping 11 matching lines...) Expand all Loading... |
| 328 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" | 328 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" |
| 329 ); | 329 ); |
| 330 } | 330 } |
| 331 #endif // HAS_I422TORGBAROW_NEON | 331 #endif // HAS_I422TORGBAROW_NEON |
| 332 | 332 |
| 333 #ifdef HAS_I422TORGB24ROW_NEON | 333 #ifdef HAS_I422TORGB24ROW_NEON |
| 334 void I422ToRGB24Row_NEON(const uint8* src_y, | 334 void I422ToRGB24Row_NEON(const uint8* src_y, |
| 335 const uint8* src_u, | 335 const uint8* src_u, |
| 336 const uint8* src_v, | 336 const uint8* src_v, |
| 337 uint8* dst_rgb24, | 337 uint8* dst_rgb24, |
| 338 struct YuvConstants* yuvconstants, | 338 const struct YuvConstants* yuvconstants, |
| 339 int width) { | 339 int width) { |
| 340 asm volatile ( | 340 asm volatile ( |
| 341 YUVTORGB_SETUP | 341 YUVTORGB_SETUP |
| 342 "1: \n" | 342 "1: \n" |
| 343 READYUV422 | 343 READYUV422 |
| 344 YUVTORGB(v22, v21, v20) | 344 YUVTORGB(v22, v21, v20) |
| 345 "subs %w4, %w4, #8 \n" | 345 "subs %w4, %w4, #8 \n" |
| 346 MEMACCESS(3) | 346 MEMACCESS(3) |
| 347 "st3 {v20.8b,v21.8b,v22.8b}, [%3], #24 \n" | 347 "st3 {v20.8b,v21.8b,v22.8b}, [%3], #24 \n" |
| 348 "b.gt 1b \n" | 348 "b.gt 1b \n" |
| (...skipping 10 matching lines...) Expand all Loading... |
| 359 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" | 359 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" |
| 360 ); | 360 ); |
| 361 } | 361 } |
| 362 #endif // HAS_I422TORGB24ROW_NEON | 362 #endif // HAS_I422TORGB24ROW_NEON |
| 363 | 363 |
| 364 #ifdef HAS_I422TORAWROW_NEON | 364 #ifdef HAS_I422TORAWROW_NEON |
| 365 void I422ToRAWRow_NEON(const uint8* src_y, | 365 void I422ToRAWRow_NEON(const uint8* src_y, |
| 366 const uint8* src_u, | 366 const uint8* src_u, |
| 367 const uint8* src_v, | 367 const uint8* src_v, |
| 368 uint8* dst_raw, | 368 uint8* dst_raw, |
| 369 struct YuvConstants* yuvconstants, | 369 const struct YuvConstants* yuvconstants, |
| 370 int width) { | 370 int width) { |
| 371 asm volatile ( | 371 asm volatile ( |
| 372 YUVTORGB_SETUP | 372 YUVTORGB_SETUP |
| 373 "1: \n" | 373 "1: \n" |
| 374 READYUV422 | 374 READYUV422 |
| 375 YUVTORGB(v20, v21, v22) | 375 YUVTORGB(v20, v21, v22) |
| 376 "subs %w4, %w4, #8 \n" | 376 "subs %w4, %w4, #8 \n" |
| 377 MEMACCESS(3) | 377 MEMACCESS(3) |
| 378 "st3 {v20.8b,v21.8b,v22.8b}, [%3], #24 \n" | 378 "st3 {v20.8b,v21.8b,v22.8b}, [%3], #24 \n" |
| 379 "b.gt 1b \n" | 379 "b.gt 1b \n" |
| (...skipping 17 matching lines...) Expand all Loading... |
| 397 "shll v20.8h, v20.8b, #8 \n" /* B */ \ | 397 "shll v20.8h, v20.8b, #8 \n" /* B */ \ |
| 398 "shll v21.8h, v21.8b, #8 \n" /* G */ \ | 398 "shll v21.8h, v21.8b, #8 \n" /* G */ \ |
| 399 "sri v0.8h, v21.8h, #5 \n" /* RG */ \ | 399 "sri v0.8h, v21.8h, #5 \n" /* RG */ \ |
| 400 "sri v0.8h, v20.8h, #11 \n" /* RGB */ | 400 "sri v0.8h, v20.8h, #11 \n" /* RGB */ |
| 401 | 401 |
| 402 #ifdef HAS_I422TORGB565ROW_NEON | 402 #ifdef HAS_I422TORGB565ROW_NEON |
| 403 void I422ToRGB565Row_NEON(const uint8* src_y, | 403 void I422ToRGB565Row_NEON(const uint8* src_y, |
| 404 const uint8* src_u, | 404 const uint8* src_u, |
| 405 const uint8* src_v, | 405 const uint8* src_v, |
| 406 uint8* dst_rgb565, | 406 uint8* dst_rgb565, |
| 407 struct YuvConstants* yuvconstants, | 407 const struct YuvConstants* yuvconstants, |
| 408 int width) { | 408 int width) { |
| 409 asm volatile ( | 409 asm volatile ( |
| 410 YUVTORGB_SETUP | 410 YUVTORGB_SETUP |
| 411 "1: \n" | 411 "1: \n" |
| 412 READYUV422 | 412 READYUV422 |
| 413 YUVTORGB(v22, v21, v20) | 413 YUVTORGB(v22, v21, v20) |
| 414 "subs %w4, %w4, #8 \n" | 414 "subs %w4, %w4, #8 \n" |
| 415 ARGBTORGB565 | 415 ARGBTORGB565 |
| 416 MEMACCESS(3) | 416 MEMACCESS(3) |
| 417 "st1 {v0.8h}, [%3], #16 \n" // store 8 pixels RGB565. | 417 "st1 {v0.8h}, [%3], #16 \n" // store 8 pixels RGB565. |
| (...skipping 20 matching lines...) Expand all Loading... |
| 438 "shll v21.8h, v21.8b, #8 \n" /* G */ \ | 438 "shll v21.8h, v21.8b, #8 \n" /* G */ \ |
| 439 "sri v0.8h, v22.8h, #1 \n" /* AR */ \ | 439 "sri v0.8h, v22.8h, #1 \n" /* AR */ \ |
| 440 "sri v0.8h, v21.8h, #6 \n" /* ARG */ \ | 440 "sri v0.8h, v21.8h, #6 \n" /* ARG */ \ |
| 441 "sri v0.8h, v20.8h, #11 \n" /* ARGB */ | 441 "sri v0.8h, v20.8h, #11 \n" /* ARGB */ |
| 442 | 442 |
| 443 #ifdef HAS_I422TOARGB1555ROW_NEON | 443 #ifdef HAS_I422TOARGB1555ROW_NEON |
| 444 void I422ToARGB1555Row_NEON(const uint8* src_y, | 444 void I422ToARGB1555Row_NEON(const uint8* src_y, |
| 445 const uint8* src_u, | 445 const uint8* src_u, |
| 446 const uint8* src_v, | 446 const uint8* src_v, |
| 447 uint8* dst_argb1555, | 447 uint8* dst_argb1555, |
| 448 struct YuvConstants* yuvconstants, | 448 const struct YuvConstants* yuvconstants, |
| 449 int width) { | 449 int width) { |
| 450 asm volatile ( | 450 asm volatile ( |
| 451 YUVTORGB_SETUP | 451 YUVTORGB_SETUP |
| 452 "1: \n" | 452 "1: \n" |
| 453 READYUV422 | 453 READYUV422 |
| 454 YUVTORGB(v22, v21, v20) | 454 YUVTORGB(v22, v21, v20) |
| 455 "subs %w4, %w4, #8 \n" | 455 "subs %w4, %w4, #8 \n" |
| 456 "movi v23.8b, #255 \n" | 456 "movi v23.8b, #255 \n" |
| 457 ARGBTOARGB1555 | 457 ARGBTOARGB1555 |
| 458 MEMACCESS(3) | 458 MEMACCESS(3) |
| (...skipping 22 matching lines...) Expand all Loading... |
| 481 "bic v23.8b, v23.8b, v4.8b \n" /* A */ \ | 481 "bic v23.8b, v23.8b, v4.8b \n" /* A */ \ |
| 482 "orr v0.8b, v20.8b, v21.8b \n" /* BG */ \ | 482 "orr v0.8b, v20.8b, v21.8b \n" /* BG */ \ |
| 483 "orr v1.8b, v22.8b, v23.8b \n" /* RA */ \ | 483 "orr v1.8b, v22.8b, v23.8b \n" /* RA */ \ |
| 484 "zip1 v0.16b, v0.16b, v1.16b \n" /* BGRA */ | 484 "zip1 v0.16b, v0.16b, v1.16b \n" /* BGRA */ |
| 485 | 485 |
| 486 #ifdef HAS_I422TOARGB4444ROW_NEON | 486 #ifdef HAS_I422TOARGB4444ROW_NEON |
| 487 void I422ToARGB4444Row_NEON(const uint8* src_y, | 487 void I422ToARGB4444Row_NEON(const uint8* src_y, |
| 488 const uint8* src_u, | 488 const uint8* src_u, |
| 489 const uint8* src_v, | 489 const uint8* src_v, |
| 490 uint8* dst_argb4444, | 490 uint8* dst_argb4444, |
| 491 struct YuvConstants* yuvconstants, | 491 const struct YuvConstants* yuvconstants, |
| 492 int width) { | 492 int width) { |
| 493 asm volatile ( | 493 asm volatile ( |
| 494 YUVTORGB_SETUP | 494 YUVTORGB_SETUP |
| 495 "movi v4.16b, #0x0f \n" // bits to clear with vbic. | 495 "movi v4.16b, #0x0f \n" // bits to clear with vbic. |
| 496 "1: \n" | 496 "1: \n" |
| 497 READYUV422 | 497 READYUV422 |
| 498 YUVTORGB(v22, v21, v20) | 498 YUVTORGB(v22, v21, v20) |
| 499 "subs %w4, %w4, #8 \n" | 499 "subs %w4, %w4, #8 \n" |
| 500 "movi v23.8b, #255 \n" | 500 "movi v23.8b, #255 \n" |
| 501 ARGBTOARGB4444 | 501 ARGBTOARGB4444 |
| (...skipping 26 matching lines...) Expand all Loading... |
| 528 READYUV400 | 528 READYUV400 |
| 529 YUVTORGB(v22, v21, v20) | 529 YUVTORGB(v22, v21, v20) |
| 530 "subs %w2, %w2, #8 \n" | 530 "subs %w2, %w2, #8 \n" |
| 531 "movi v23.8b, #255 \n" | 531 "movi v23.8b, #255 \n" |
| 532 MEMACCESS(1) | 532 MEMACCESS(1) |
| 533 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n" | 533 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n" |
| 534 "b.gt 1b \n" | 534 "b.gt 1b \n" |
| 535 : "+r"(src_y), // %0 | 535 : "+r"(src_y), // %0 |
| 536 "+r"(dst_argb), // %1 | 536 "+r"(dst_argb), // %1 |
| 537 "+r"(width64) // %2 | 537 "+r"(width64) // %2 |
| 538 : [kUVToRB]"r"(&kYuvConstants.kUVToRB), | 538 : [kUVToRB]"r"(&kYuvIConstants.kUVToRB), |
| 539 [kUVToG]"r"(&kYuvConstants.kUVToG), | 539 [kUVToG]"r"(&kYuvIConstants.kUVToG), |
| 540 [kUVBiasBGR]"r"(&kYuvConstants.kUVBiasBGR), | 540 [kUVBiasBGR]"r"(&kYuvIConstants.kUVBiasBGR), |
| 541 [kYToRgb]"r"(&kYuvConstants.kYToRgb) | 541 [kYToRgb]"r"(&kYuvIConstants.kYToRgb) |
| 542 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", | 542 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", |
| 543 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" | 543 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" |
| 544 ); | 544 ); |
| 545 } | 545 } |
| 546 #endif // HAS_I400TOARGBROW_NEON | 546 #endif // HAS_I400TOARGBROW_NEON |
| 547 | 547 |
| 548 #ifdef HAS_J400TOARGBROW_NEON | 548 #ifdef HAS_J400TOARGBROW_NEON |
| 549 void J400ToARGBRow_NEON(const uint8* src_y, | 549 void J400ToARGBRow_NEON(const uint8* src_y, |
| 550 uint8* dst_argb, | 550 uint8* dst_argb, |
| 551 int width) { | 551 int width) { |
| (...skipping 14 matching lines...) Expand all Loading... |
| 566 : | 566 : |
| 567 : "cc", "memory", "v20", "v21", "v22", "v23" | 567 : "cc", "memory", "v20", "v21", "v22", "v23" |
| 568 ); | 568 ); |
| 569 } | 569 } |
| 570 #endif // HAS_J400TOARGBROW_NEON | 570 #endif // HAS_J400TOARGBROW_NEON |
| 571 | 571 |
| 572 #ifdef HAS_NV12TOARGBROW_NEON | 572 #ifdef HAS_NV12TOARGBROW_NEON |
| 573 void NV12ToARGBRow_NEON(const uint8* src_y, | 573 void NV12ToARGBRow_NEON(const uint8* src_y, |
| 574 const uint8* src_uv, | 574 const uint8* src_uv, |
| 575 uint8* dst_argb, | 575 uint8* dst_argb, |
| 576 struct YuvConstants* yuvconstants, | 576 const struct YuvConstants* yuvconstants, |
| 577 int width) { | 577 int width) { |
| 578 asm volatile ( | 578 asm volatile ( |
| 579 YUVTORGB_SETUP | 579 YUVTORGB_SETUP |
| 580 "1: \n" | 580 "1: \n" |
| 581 READNV12 | 581 READNV12 |
| 582 YUVTORGB(v22, v21, v20) | 582 YUVTORGB(v22, v21, v20) |
| 583 "subs %w3, %w3, #8 \n" | 583 "subs %w3, %w3, #8 \n" |
| 584 "movi v23.8b, #255 \n" | 584 "movi v23.8b, #255 \n" |
| 585 MEMACCESS(2) | 585 MEMACCESS(2) |
| 586 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32 \n" | 586 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32 \n" |
| 587 "b.gt 1b \n" | 587 "b.gt 1b \n" |
| 588 : "+r"(src_y), // %0 | 588 : "+r"(src_y), // %0 |
| 589 "+r"(src_uv), // %1 | 589 "+r"(src_uv), // %1 |
| 590 "+r"(dst_argb), // %2 | 590 "+r"(dst_argb), // %2 |
| 591 "+r"(width) // %3 | 591 "+r"(width) // %3 |
| 592 : [kUVToRB]"r"(&yuvconstants->kUVToRB), | 592 : [kUVToRB]"r"(&yuvconstants->kUVToRB), |
| 593 [kUVToG]"r"(&yuvconstants->kUVToG), | 593 [kUVToG]"r"(&yuvconstants->kUVToG), |
| 594 [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR), | 594 [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR), |
| 595 [kYToRgb]"r"(&yuvconstants->kYToRgb) | 595 [kYToRgb]"r"(&yuvconstants->kYToRgb) |
| 596 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", | 596 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", |
| 597 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" | 597 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" |
| 598 ); | 598 ); |
| 599 } | 599 } |
| 600 #endif // HAS_NV12TOARGBROW_NEON | 600 #endif // HAS_NV12TOARGBROW_NEON |
| 601 | 601 |
| 602 #ifdef HAS_NV12TOARGBROW_NEON | 602 #ifdef HAS_NV12TOARGBROW_NEON |
| 603 void NV21ToARGBRow_NEON(const uint8* src_y, | 603 void NV21ToARGBRow_NEON(const uint8* src_y, |
| 604 const uint8* src_vu, | 604 const uint8* src_vu, |
| 605 uint8* dst_argb, | 605 uint8* dst_argb, |
| 606 struct YuvConstants* yuvconstants, | 606 const struct YuvConstants* yuvconstants, |
| 607 int width) { | 607 int width) { |
| 608 asm volatile ( | 608 asm volatile ( |
| 609 YUVTORGB_SETUP | 609 YUVTORGB_SETUP |
| 610 "1: \n" | 610 "1: \n" |
| 611 READNV21 | 611 READNV21 |
| 612 YUVTORGB(v22, v21, v20) | 612 YUVTORGB(v22, v21, v20) |
| 613 "subs %w3, %w3, #8 \n" | 613 "subs %w3, %w3, #8 \n" |
| 614 "movi v23.8b, #255 \n" | 614 "movi v23.8b, #255 \n" |
| 615 MEMACCESS(2) | 615 MEMACCESS(2) |
| 616 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32 \n" | 616 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32 \n" |
| 617 "b.gt 1b \n" | 617 "b.gt 1b \n" |
| 618 : "+r"(src_y), // %0 | 618 : "+r"(src_y), // %0 |
| 619 "+r"(src_vu), // %1 | 619 "+r"(src_vu), // %1 |
| 620 "+r"(dst_argb), // %2 | 620 "+r"(dst_argb), // %2 |
| 621 "+r"(width) // %3 | 621 "+r"(width) // %3 |
| 622 : [kUVToRB]"r"(&yuvconstants->kUVToRB), | 622 : [kUVToRB]"r"(&yuvconstants->kUVToRB), |
| 623 [kUVToG]"r"(&yuvconstants->kUVToG), | 623 [kUVToG]"r"(&yuvconstants->kUVToG), |
| 624 [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR), | 624 [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR), |
| 625 [kYToRgb]"r"(&yuvconstants->kYToRgb) | 625 [kYToRgb]"r"(&yuvconstants->kYToRgb) |
| 626 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", | 626 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", |
| 627 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" | 627 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" |
| 628 ); | 628 ); |
| 629 } | 629 } |
| 630 #endif // HAS_NV12TOARGBROW_NEON | 630 #endif // HAS_NV12TOARGBROW_NEON |
| 631 | 631 |
| 632 #ifdef HAS_NV12TORGB565ROW_NEON | 632 #ifdef HAS_NV12TORGB565ROW_NEON |
| 633 void NV12ToRGB565Row_NEON(const uint8* src_y, | 633 void NV12ToRGB565Row_NEON(const uint8* src_y, |
| 634 const uint8* src_uv, | 634 const uint8* src_uv, |
| 635 uint8* dst_rgb565, | 635 uint8* dst_rgb565, |
| 636 struct YuvConstants* yuvconstants, | 636 const struct YuvConstants* yuvconstants, |
| 637 int width) { | 637 int width) { |
| 638 asm volatile ( | 638 asm volatile ( |
| 639 YUVTORGB_SETUP | 639 YUVTORGB_SETUP |
| 640 "1: \n" | 640 "1: \n" |
| 641 READNV12 | 641 READNV12 |
| 642 YUVTORGB(v22, v21, v20) | 642 YUVTORGB(v22, v21, v20) |
| 643 "subs %w3, %w3, #8 \n" | 643 "subs %w3, %w3, #8 \n" |
| 644 ARGBTORGB565 | 644 ARGBTORGB565 |
| 645 MEMACCESS(2) | 645 MEMACCESS(2) |
| 646 "st1 {v0.8h}, [%2], 16 \n" // store 8 pixels RGB565. | 646 "st1 {v0.8h}, [%2], 16 \n" // store 8 pixels RGB565. |
| 647 "b.gt 1b \n" | 647 "b.gt 1b \n" |
| 648 : "+r"(src_y), // %0 | 648 : "+r"(src_y), // %0 |
| 649 "+r"(src_uv), // %1 | 649 "+r"(src_uv), // %1 |
| 650 "+r"(dst_rgb565), // %2 | 650 "+r"(dst_rgb565), // %2 |
| 651 "+r"(width) // %3 | 651 "+r"(width) // %3 |
| 652 : [kUVToRB]"r"(&yuvconstants->kUVToRB), | 652 : [kUVToRB]"r"(&yuvconstants->kUVToRB), |
| 653 [kUVToG]"r"(&yuvconstants->kUVToG), | 653 [kUVToG]"r"(&yuvconstants->kUVToG), |
| 654 [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR), | 654 [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR), |
| 655 [kYToRgb]"r"(&yuvconstants->kYToRgb) | 655 [kYToRgb]"r"(&yuvconstants->kYToRgb) |
| 656 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", | 656 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", |
| 657 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" | 657 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" |
| 658 ); | 658 ); |
| 659 } | 659 } |
| 660 #endif // HAS_NV12TORGB565ROW_NEON | 660 #endif // HAS_NV12TORGB565ROW_NEON |
| 661 | 661 |
| 662 #ifdef HAS_YUY2TOARGBROW_NEON | 662 #ifdef HAS_YUY2TOARGBROW_NEON |
| 663 void YUY2ToARGBRow_NEON(const uint8* src_yuy2, | 663 void YUY2ToARGBRow_NEON(const uint8* src_yuy2, |
| 664 uint8* dst_argb, | 664 uint8* dst_argb, |
| 665 struct YuvConstants* yuvconstants, | 665 const struct YuvConstants* yuvconstants, |
| 666 int width) { | 666 int width) { |
| 667 int64 width64 = (int64)(width); | 667 int64 width64 = (int64)(width); |
| 668 asm volatile ( | 668 asm volatile ( |
| 669 YUVTORGB_SETUP | 669 YUVTORGB_SETUP |
| 670 "1: \n" | 670 "1: \n" |
| 671 READYUY2 | 671 READYUY2 |
| 672 YUVTORGB(v22, v21, v20) | 672 YUVTORGB(v22, v21, v20) |
| 673 "subs %w2, %w2, #8 \n" | 673 "subs %w2, %w2, #8 \n" |
| 674 "movi v23.8b, #255 \n" | 674 "movi v23.8b, #255 \n" |
| 675 MEMACCESS(1) | 675 MEMACCESS(1) |
| 676 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n" | 676 "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n" |
| 677 "b.gt 1b \n" | 677 "b.gt 1b \n" |
| 678 : "+r"(src_yuy2), // %0 | 678 : "+r"(src_yuy2), // %0 |
| 679 "+r"(dst_argb), // %1 | 679 "+r"(dst_argb), // %1 |
| 680 "+r"(width64) // %2 | 680 "+r"(width64) // %2 |
| 681 : [kUVToRB]"r"(&yuvconstants->kUVToRB), | 681 : [kUVToRB]"r"(&yuvconstants->kUVToRB), |
| 682 [kUVToG]"r"(&yuvconstants->kUVToG), | 682 [kUVToG]"r"(&yuvconstants->kUVToG), |
| 683 [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR), | 683 [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR), |
| 684 [kYToRgb]"r"(&yuvconstants->kYToRgb) | 684 [kYToRgb]"r"(&yuvconstants->kYToRgb) |
| 685 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", | 685 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", |
| 686 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" | 686 "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" |
| 687 ); | 687 ); |
| 688 } | 688 } |
| 689 #endif // HAS_YUY2TOARGBROW_NEON | 689 #endif // HAS_YUY2TOARGBROW_NEON |
| 690 | 690 |
| 691 #ifdef HAS_UYVYTOARGBROW_NEON | 691 #ifdef HAS_UYVYTOARGBROW_NEON |
| 692 void UYVYToARGBRow_NEON(const uint8* src_uyvy, | 692 void UYVYToARGBRow_NEON(const uint8* src_uyvy, |
| 693 uint8* dst_argb, | 693 uint8* dst_argb, |
| 694 struct YuvConstants* yuvconstants, | 694 const struct YuvConstants* yuvconstants, |
| 695 int width) { | 695 int width) { |
| 696 int64 width64 = (int64)(width); | 696 int64 width64 = (int64)(width); |
| 697 asm volatile ( | 697 asm volatile ( |
| 698 YUVTORGB_SETUP | 698 YUVTORGB_SETUP |
| 699 "1: \n" | 699 "1: \n" |
| 700 READUYVY | 700 READUYVY |
| 701 YUVTORGB(v22, v21, v20) | 701 YUVTORGB(v22, v21, v20) |
| 702 "subs %w2, %w2, #8 \n" | 702 "subs %w2, %w2, #8 \n" |
| 703 "movi v23.8b, #255 \n" | 703 "movi v23.8b, #255 \n" |
| 704 MEMACCESS(1) | 704 MEMACCESS(1) |
| (...skipping 2367 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 3072 : "cc", "memory", "v0", "v1", "v2", "v3" // Clobber List | 3072 : "cc", "memory", "v0", "v1", "v2", "v3" // Clobber List |
| 3073 ); | 3073 ); |
| 3074 } | 3074 } |
| 3075 #endif // HAS_SOBELYROW_NEON | 3075 #endif // HAS_SOBELYROW_NEON |
| 3076 #endif // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) | 3076 #endif // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) |
| 3077 | 3077 |
| 3078 #ifdef __cplusplus | 3078 #ifdef __cplusplus |
| 3079 } // extern "C" | 3079 } // extern "C" |
| 3080 } // namespace libyuv | 3080 } // namespace libyuv |
| 3081 #endif | 3081 #endif |
| OLD | NEW |