| 353 | } |
| 354 | |
| 355 | inline void transpose_4lane_inline( |
| 356 | const u8* const lanes[4], |
| 357 | u8* output, |
| 358 | size_t num_bytes |
| 359 | ) FL_NOEXCEPT { |
| 360 | for (size_t byte_idx = 0; byte_idx < num_bytes; byte_idx++) { |
| 361 | u8 a = lanes[0][byte_idx]; |
| 362 | u8 b = lanes[1][byte_idx]; |
| 363 | u8 c = lanes[2][byte_idx]; |
| 364 | u8 d = lanes[3][byte_idx]; |
| 365 | |
| 366 | u8* dest = &output[byte_idx * 4]; |
| 367 | |
| 368 | dest[0] = ((a >> 7) & 0x01) << 0 | ((b >> 7) & 0x01) << 1 | ((c >> 7) & 0x01) << 2 | ((d >> 7) & 0x01) << 3 | |
| 369 | ((a >> 6) & 0x01) << 4 | ((b >> 6) & 0x01) << 5 | ((c >> 6) & 0x01) << 6 | ((d >> 6) & 0x01) << 7; |
| 370 | |
| 371 | dest[1] = ((a >> 5) & 0x01) << 0 | ((b >> 5) & 0x01) << 1 | ((c >> 5) & 0x01) << 2 | ((d >> 5) & 0x01) << 3 | |
| 372 | ((a >> 4) & 0x01) << 4 | ((b >> 4) & 0x01) << 5 | ((c >> 4) & 0x01) << 6 | ((d >> 4) & 0x01) << 7; |
| 373 | |
| 374 | dest[2] = ((a >> 3) & 0x01) << 0 | ((b >> 3) & 0x01) << 1 | ((c >> 3) & 0x01) << 2 | ((d >> 3) & 0x01) << 3 | |
| 375 | ((a >> 2) & 0x01) << 4 | ((b >> 2) & 0x01) << 5 | ((c >> 2) & 0x01) << 6 | ((d >> 2) & 0x01) << 7; |
| 376 | |
| 377 | dest[3] = ((a >> 1) & 0x01) << 0 | ((b >> 1) & 0x01) << 1 | ((c >> 1) & 0x01) << 2 | ((d >> 1) & 0x01) << 3 | |
| 378 | ((a >> 0) & 0x01) << 4 | ((b >> 0) & 0x01) << 5 | ((c >> 0) & 0x01) << 6 | ((d >> 0) & 0x01) << 7; |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | inline void transpose_8lane_inline( |
| 383 | const u8* const lanes[8], |