Mercurial > libavcodec.hg
comparison simple_idct.c @ 215:1fe6b64feefb libavcodec
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
| author | nickols_k |
|---|---|
| date | Mon, 21 Jan 2002 17:46:42 +0000 |
| parents | 73df666cacc7 |
| children | 5a8eb5cf9f92 |
comparison
equal
deleted
inserted
replaced
| 214:73df666cacc7 | 215:1fe6b64feefb |
|---|---|
| 102 | 102 |
| 103 return 1; | 103 return 1; |
| 104 } | 104 } |
| 105 | 105 |
| 106 #ifdef ARCH_ALPHA | 106 #ifdef ARCH_ALPHA |
| 107 /* 0: all entries 0, 1: only first entry nonzero, 2: otherwise */ | |
| 107 static int inline idctRowCondDC(int16_t *row) | 108 static int inline idctRowCondDC(int16_t *row) |
| 108 { | 109 { |
| 109 int_fast32_t a0, a1, a2, a3, b0, b1, b2, b3; | 110 int_fast32_t a0, a1, a2, a3, b0, b1, b2, b3; |
| 110 uint64_t *lrow = (uint64_t *) row; | 111 uint64_t *lrow = (uint64_t *) row; |
| 111 | 112 |
| 199 row[4] = (a3 - b3) >> ROW_SHIFT; | 200 row[4] = (a3 - b3) >> ROW_SHIFT; |
| 200 row[5] = (a2 - b2) >> ROW_SHIFT; | 201 row[5] = (a2 - b2) >> ROW_SHIFT; |
| 201 row[6] = (a1 - b1) >> ROW_SHIFT; | 202 row[6] = (a1 - b1) >> ROW_SHIFT; |
| 202 row[7] = (a0 - b0) >> ROW_SHIFT; | 203 row[7] = (a0 - b0) >> ROW_SHIFT; |
| 203 | 204 |
| 204 return 1; | 205 return 2; |
| 205 } | 206 } |
| 206 #else /* not ARCH_ALPHA */ | 207 #else /* not ARCH_ALPHA */ |
| 207 static int inline idctRowCondDC (int16_t * row) | 208 static int inline idctRowCondDC (int16_t * row) |
| 208 { | 209 { |
| 209 int a0, a1, a2, a3, b0, b1, b2, b3; | 210 int a0, a1, a2, a3, b0, b1, b2, b3; |
| 545 { | 546 { |
| 546 for(i=0; i<8; i++) | 547 for(i=0; i<8; i++) |
| 547 idctSparse2Col(block + i); | 548 idctSparse2Col(block + i); |
| 548 } | 549 } |
| 549 #elif defined(ARCH_ALPHA) | 550 #elif defined(ARCH_ALPHA) |
| 550 int shortcut = 1; | 551 int rowsZero = 1; /* all rows except row 0 zero */ |
| 552 int rowsConstant = 1; /* all rows consist of a constant value */ | |
| 551 | 553 |
| 552 for (i = 0; i < 8; i++) { | 554 for (i = 0; i < 8; i++) { |
| 553 int anynonzero = idctRowCondDC(block + 8 * i); | 555 int sparseness = idctRowCondDC(block + 8 * i); |
| 554 if (i > 0 && anynonzero) | 556 |
| 555 shortcut = 0; | 557 if (i > 0 && sparseness > 0) |
| 556 } | 558 rowsZero = 0; |
| 557 | 559 if (sparseness == 2) |
| 558 if (shortcut) { | 560 rowsConstant = 0; |
| 559 idctCol2(block); | 561 } |
| 562 | |
| 563 if (rowsZero) { | |
| 564 idctCol2(block); | |
| 565 } else if (rowsConstant) { | |
| 566 uint64_t *lblock = (uint64_t *) block; | |
| 567 | |
| 568 idctSparseCol(block); | |
| 569 for (i = 0; i < 8; i++) { | |
| 570 uint64_t v = (uint16_t) block[i * 8]; | |
| 571 | |
| 572 v += v << 16; | |
| 573 v += v << 32; | |
| 574 lblock[0] = v; | |
| 575 lblock[1] = v; | |
| 576 lblock += 2; | |
| 577 } | |
| 560 } else { | 578 } else { |
| 561 for (i = 0; i < 8; i++) | 579 for (i = 0; i < 8; i++) |
| 562 idctSparseCol(block + i); | 580 idctSparseCol(block + i); |
| 563 } | 581 } |
| 564 #else | 582 #else |