From 0bd2483a7553bfd60af3d8f137cab02795d59771 Mon Sep 17 00:00:00 2001 From: Michael Niedermayer Date: Tue, 22 Aug 2006 06:24:59 +0000 Subject: [PATCH] replace MULL by MULH + shift in dct32() (code is very slightly faster, and the compiler should be able to optimize the shift away on 64bit archs) Originally committed as revision 6045 to svn://svn.ffmpeg.org/ffmpeg/trunk --- libavcodec/mpegaudiodec.c | 209 +++++++++++++++++----------------- tests/ffmpeg.regression.ref | 2 +- tests/libav.regression.ref | 22 ++-- tests/rotozoom.regression.ref | 2 +- 4 files changed, 117 insertions(+), 118 deletions(-) diff --git a/libavcodec/mpegaudiodec.c b/libavcodec/mpegaudiodec.c index d374031cea..53a3890667 100644 --- a/libavcodec/mpegaudiodec.c +++ b/libavcodec/mpegaudiodec.c @@ -513,62 +513,62 @@ static int decode_init(AVCodecContext * avctx) /* cos(i*pi/64) */ -#define COS0_0 FIXR(0.50060299823519630134) -#define COS0_1 FIXR(0.50547095989754365998) -#define COS0_2 FIXR(0.51544730992262454697) -#define COS0_3 FIXR(0.53104259108978417447) -#define COS0_4 FIXR(0.55310389603444452782) -#define COS0_5 FIXR(0.58293496820613387367) -#define COS0_6 FIXR(0.62250412303566481615) -#define COS0_7 FIXR(0.67480834145500574602) -#define COS0_8 FIXR(0.74453627100229844977) -#define COS0_9 FIXR(0.83934964541552703873) -#define COS0_10 FIXR(0.97256823786196069369) -#define COS0_11 FIXR(1.16943993343288495515) -#define COS0_12 FIXR(1.48416461631416627724) -#define COS0_13 FIXR(2.05778100995341155085) -#define COS0_14 FIXR(3.40760841846871878570) -#define COS0_15 FIXR(10.19000812354805681150) +#define COS0_0 FIXHR(0.50060299823519630134/2) +#define COS0_1 FIXHR(0.50547095989754365998/2) +#define COS0_2 FIXHR(0.51544730992262454697/2) +#define COS0_3 FIXHR(0.53104259108978417447/2) +#define COS0_4 FIXHR(0.55310389603444452782/2) +#define COS0_5 FIXHR(0.58293496820613387367/2) +#define COS0_6 FIXHR(0.62250412303566481615/2) +#define COS0_7 FIXHR(0.67480834145500574602/2) +#define COS0_8 FIXHR(0.74453627100229844977/2) +#define COS0_9 FIXHR(0.83934964541552703873/2) +#define COS0_10 FIXHR(0.97256823786196069369/2) +#define COS0_11 FIXHR(1.16943993343288495515/4) +#define COS0_12 FIXHR(1.48416461631416627724/4) +#define COS0_13 FIXHR(2.05778100995341155085/8) +#define COS0_14 FIXHR(3.40760841846871878570/8) +#define COS0_15 FIXHR(10.19000812354805681150/32) -#define COS1_0 FIXR(0.50241928618815570551) -#define COS1_1 FIXR(0.52249861493968888062) -#define COS1_2 FIXR(0.56694403481635770368) -#define COS1_3 FIXR(0.64682178335999012954) -#define COS1_4 FIXR(0.78815462345125022473) -#define COS1_5 FIXR(1.06067768599034747134) -#define COS1_6 FIXR(1.72244709823833392782) -#define COS1_7 FIXR(5.10114861868916385802) +#define COS1_0 FIXHR(0.50241928618815570551/2) +#define COS1_1 FIXHR(0.52249861493968888062/2) +#define COS1_2 FIXHR(0.56694403481635770368/2) +#define COS1_3 FIXHR(0.64682178335999012954/2) +#define COS1_4 FIXHR(0.78815462345125022473/2) +#define COS1_5 FIXHR(1.06067768599034747134/4) +#define COS1_6 FIXHR(1.72244709823833392782/4) +#define COS1_7 FIXHR(5.10114861868916385802/16) -#define COS2_0 FIXR(0.50979557910415916894) -#define COS2_1 FIXR(0.60134488693504528054) -#define COS2_2 FIXR(0.89997622313641570463) -#define COS2_3 FIXR(2.56291544774150617881) +#define COS2_0 FIXHR(0.50979557910415916894/2) +#define COS2_1 FIXHR(0.60134488693504528054/2) +#define COS2_2 FIXHR(0.89997622313641570463/2) +#define COS2_3 FIXHR(2.56291544774150617881/8) -#define COS3_0 FIXR(0.54119610014619698439) -#define COS3_1 FIXR(1.30656296487637652785) +#define COS3_0 FIXHR(0.54119610014619698439/2) +#define COS3_1 FIXHR(1.30656296487637652785/4) -#define COS4_0 FIXR(0.70710678118654752439) +#define COS4_0 FIXHR(0.70710678118654752439/2) /* butterfly operator */ -#define BF(a, b, c)\ +#define BF(a, b, c, s)\ {\ tmp0 = tab[a] + tab[b];\ tmp1 = tab[a] - tab[b];\ tab[a] = tmp0;\ - tab[b] = MULL(tmp1, c);\ + tab[b] = MULH(tmp1<<(s), c);\ } #define BF1(a, b, c, d)\ {\ - BF(a, b, COS4_0);\ - BF(c, d, -COS4_0);\ + BF(a, b, COS4_0, 1);\ + BF(c, d,-COS4_0, 1);\ tab[c] += tab[d];\ } #define BF2(a, b, c, d)\ {\ - BF(a, b, COS4_0);\ - BF(c, d, -COS4_0);\ + BF(a, b, COS4_0, 1);\ + BF(c, d,-COS4_0, 1);\ tab[c] += tab[d];\ tab[a] += tab[c];\ tab[c] += tab[b];\ @@ -582,102 +582,101 @@ static void dct32(int32_t *out, int32_t *tab) { int tmp0, tmp1; - /* pass 1 */ - BF(0, 31, COS0_0); - BF(15, 16, COS0_15); + BF( 0, 31, COS0_0 , 1); + BF(15, 16, COS0_15, 5); /* pass 2 */ - BF(0, 15, COS1_0); - BF(16, 31, -COS1_0); + BF( 0, 15, COS1_0 , 1); + BF(16, 31,-COS1_0 , 1); /* pass 1 */ - BF(7, 24, COS0_7); - BF(8, 23, COS0_8); + BF( 7, 24, COS0_7 , 1); + BF( 8, 23, COS0_8 , 1); /* pass 2 */ - BF(7, 8, COS1_7); - BF(23, 24, -COS1_7); + BF( 7, 8, COS1_7 , 4); + BF(23, 24,-COS1_7 , 4); /* pass 3 */ - BF(0, 7, COS2_0); - BF(8, 15, -COS2_0); - BF(16, 23, COS2_0); - BF(24, 31, -COS2_0); + BF( 0, 7, COS2_0 , 1); + BF( 8, 15,-COS2_0 , 1); + BF(16, 23, COS2_0 , 1); + BF(24, 31,-COS2_0 , 1); /* pass 1 */ - BF(3, 28, COS0_3); - BF(12, 19, COS0_12); + BF( 3, 28, COS0_3 , 1); + BF(12, 19, COS0_12, 2); /* pass 2 */ - BF(3, 12, COS1_3); - BF(19, 28, -COS1_3); + BF( 3, 12, COS1_3 , 1); + BF(19, 28,-COS1_3 , 1); /* pass 1 */ - BF(4, 27, COS0_4); - BF(11, 20, COS0_11); + BF( 4, 27, COS0_4 , 1); + BF(11, 20, COS0_11, 2); /* pass 2 */ - BF(4, 11, COS1_4); - BF(20, 27, -COS1_4); + BF( 4, 11, COS1_4 , 1); + BF(20, 27,-COS1_4 , 1); /* pass 3 */ - BF(3, 4, COS2_3); - BF(11, 12, -COS2_3); - BF(19, 20, COS2_3); - BF(27, 28, -COS2_3); + BF( 3, 4, COS2_3 , 3); + BF(11, 12,-COS2_3 , 3); + BF(19, 20, COS2_3 , 3); + BF(27, 28,-COS2_3 , 3); /* pass 4 */ - BF(0, 3, COS3_0); - BF(4, 7, -COS3_0); - BF(8, 11, COS3_0); - BF(12, 15, -COS3_0); - BF(16, 19, COS3_0); - BF(20, 23, -COS3_0); - BF(24, 27, COS3_0); - BF(28, 31, -COS3_0); + BF( 0, 3, COS3_0 , 1); + BF( 4, 7,-COS3_0 , 1); + BF( 8, 11, COS3_0 , 1); + BF(12, 15,-COS3_0 , 1); + BF(16, 19, COS3_0 , 1); + BF(20, 23,-COS3_0 , 1); + BF(24, 27, COS3_0 , 1); + BF(28, 31,-COS3_0 , 1); /* pass 1 */ - BF(1, 30, COS0_1); - BF(14, 17, COS0_14); + BF( 1, 30, COS0_1 , 1); + BF(14, 17, COS0_14, 3); /* pass 2 */ - BF(1, 14, COS1_1); - BF(17, 30, -COS1_1); + BF( 1, 14, COS1_1 , 1); + BF(17, 30,-COS1_1 , 1); /* pass 1 */ - BF(6, 25, COS0_6); - BF(9, 22, COS0_9); + BF( 6, 25, COS0_6 , 1); + BF( 9, 22, COS0_9 , 1); /* pass 2 */ - BF(6, 9, COS1_6); - BF(22, 25, -COS1_6); + BF( 6, 9, COS1_6 , 2); + BF(22, 25,-COS1_6 , 2); /* pass 3 */ - BF(1, 6, COS2_1); - BF(9, 14, -COS2_1); - BF(17, 22, COS2_1); - BF(25, 30, -COS2_1); + BF( 1, 6, COS2_1 , 1); + BF( 9, 14,-COS2_1 , 1); + BF(17, 22, COS2_1 , 1); + BF(25, 30,-COS2_1 , 1); /* pass 1 */ - BF(2, 29, COS0_2); - BF(13, 18, COS0_13); + BF( 2, 29, COS0_2 , 1); + BF(13, 18, COS0_13, 3); /* pass 2 */ - BF(2, 13, COS1_2); - BF(18, 29, -COS1_2); + BF( 2, 13, COS1_2 , 1); + BF(18, 29,-COS1_2 , 1); /* pass 1 */ - BF(5, 26, COS0_5); - BF(10, 21, COS0_10); + BF( 5, 26, COS0_5 , 1); + BF(10, 21, COS0_10, 1); /* pass 2 */ - BF(5, 10, COS1_5); - BF(21, 26, -COS1_5); + BF( 5, 10, COS1_5 , 2); + BF(21, 26,-COS1_5 , 2); /* pass 3 */ - BF(2, 5, COS2_2); - BF(10, 13, -COS2_2); - BF(18, 21, COS2_2); - BF(26, 29, -COS2_2); + BF( 2, 5, COS2_2 , 1); + BF(10, 13,-COS2_2 , 1); + BF(18, 21, COS2_2 , 1); + BF(26, 29,-COS2_2 , 1); /* pass 4 */ - BF(1, 2, COS3_1); - BF(5, 6, -COS3_1); - BF(9, 10, COS3_1); - BF(13, 14, -COS3_1); - BF(17, 18, COS3_1); - BF(21, 22, -COS3_1); - BF(25, 26, COS3_1); - BF(29, 30, -COS3_1); + BF( 1, 2, COS3_1 , 2); + BF( 5, 6,-COS3_1 , 2); + BF( 9, 10, COS3_1 , 2); + BF(13, 14,-COS3_1 , 2); + BF(17, 18, COS3_1 , 2); + BF(21, 22,-COS3_1 , 2); + BF(25, 26, COS3_1 , 2); + BF(29, 30,-COS3_1 , 2); /* pass 5 */ - BF1(0, 1, 2, 3); - BF2(4, 5, 6, 7); - BF1(8, 9, 10, 11); + BF1( 0, 1, 2, 3); + BF2( 4, 5, 6, 7); + BF1( 8, 9, 10, 11); BF2(12, 13, 14, 15); BF1(16, 17, 18, 19); BF2(20, 21, 22, 23); diff --git a/tests/ffmpeg.regression.ref b/tests/ffmpeg.regression.ref index 037c748e3c..cc44375b81 100644 --- a/tests/ffmpeg.regression.ref +++ b/tests/ffmpeg.regression.ref @@ -155,7 +155,7 @@ bbff871d1475e1eee4231a08e075de2c *./data/out.yuv stddev: 10.99 PSNR:27.30 bytes:7602176 21f8ff9f1daacd9133683bb4ea0f50a4 *./data/a-mp2.mp2 95712 ./data/a-mp2.mp2 -6956f2c5185c4edbae648c614992a14b *./data/out.wav +83f8df5d5f84480566af548bb037fceb *./data/out.wav stddev:9330.70 PSNR:16.92 bytes:1054720 stddev:4396.13 PSNR:23.46 bytes:1052672 aefe11ab5067621a1c674859d6413891 *./data/a-ac3.rm diff --git a/tests/libav.regression.ref b/tests/libav.regression.ref index 4b102091db..458ae427f4 100644 --- a/tests/libav.regression.ref +++ b/tests/libav.regression.ref @@ -1,24 +1,24 @@ ffmpeg regression test 5fb135845dbde3ce5208c468173ef2c2 *./data/b-libav.avi - 343318 ./data/b-libav.avi -./data/b-libav.avi CRC=0xccab3a27 +343318 ./data/b-libav.avi +./data/b-libav.avi CRC=0x1fdd352c 6f4dca897d9a009009798e434fe5f651 *./data/b-libav.asf - 342967 ./data/b-libav.asf -./data/b-libav.asf CRC=0x41af1ed9 +342967 ./data/b-libav.asf +./data/b-libav.asf CRC=0x94d219de 5275686385346789108ec4fdf06ecef8 *./data/b-libav.rm 360269 ./data/b-libav.rm 90784a1b9589095f20fc6bcc0cc23cc4 *./data/b-libav.mpg - 387072 ./data/b-libav.mpg -./data/b-libav.mpg CRC=0x8b644837 +387072 ./data/b-libav.mpg +./data/b-libav.mpg CRC=0xde87433c 2933a367c996b49f4b7e66a481fb334e *./data/b-libav.ts 476768 ./data/b-libav.ts -./data/b-libav.ts CRC=0x0dd7fc5f +./data/b-libav.ts CRC=0x60faf764 57a8dfc7926802bb337a9d8918de94a8 *./data/b-libav.swf - 41816 ./data/b-libav.swf +41816 ./data/b-libav.swf ./data/b-libav.swf CRC=0x2b273fea 0026c4e2528ff8f71b79ef53d38cc04d *./data/b-libav.ffm - 389120 ./data/b-libav.ffm -./data/b-libav.ffm CRC=0x688bd1eb +389120 ./data/b-libav.ffm +./data/b-libav.ffm CRC=0x1acfcaf1 1596b6d08a94e6b5519357ed23a48aa3 *./data/b-libav.flv 339476 ./data/b-libav.flv ./data/b-libav.flv CRC=0x7b9076f8 @@ -27,7 +27,7 @@ ffmpeg regression test ./data/b-libav.mov CRC=0x48f5a90d 8bf16d40a2ec19fa36b124a928e47e23 *./data/b-libav.nut 332358 ./data/b-libav.nut -./data/b-libav.nut CRC=0xccab3a27 +./data/b-libav.nut CRC=0x1fdd352c 2cc647904d3cca01703f36056b9d3f42 *./data/b-libav.dv 3600000 ./data/b-libav.dv ./data/b-libav.dv CRC=0xa6b8b635 diff --git a/tests/rotozoom.regression.ref b/tests/rotozoom.regression.ref index 441c951017..41209b5629 100644 --- a/tests/rotozoom.regression.ref +++ b/tests/rotozoom.regression.ref @@ -155,7 +155,7 @@ stddev: 2.98 PSNR:38.61 bytes:7602176 stddev: 3.44 PSNR:37.38 bytes:7602176 21f8ff9f1daacd9133683bb4ea0f50a4 *./data/a-mp2.mp2 95712 ./data/a-mp2.mp2 -6956f2c5185c4edbae648c614992a14b *./data/out.wav +83f8df5d5f84480566af548bb037fceb *./data/out.wav stddev:9330.70 PSNR:16.92 bytes:1054720 stddev:4396.13 PSNR:23.46 bytes:1052672 aefe11ab5067621a1c674859d6413891 *./data/a-ac3.rm