# HG changeset patch # User michael # Date 1008788285 0 # Node ID 0e4ff1cac6a25c60046f78f07a81c44647faf853 # Parent 3f1c2c06d0d808741352cd8fc6ff85b9812137d9 sse opt of mix32to2() diff -r 3f1c2c06d0d8 -r 0e4ff1cac6a2 liba52/downmix.c --- a/liba52/downmix.c Wed Dec 19 18:42:36 2001 +0000 +++ b/liba52/downmix.c Wed Dec 19 18:58:05 2001 +0000 @@ -411,11 +411,33 @@ int i; sample_t common; +#ifdef HAVE_SSE + asm volatile( + "movlps %1, %%xmm7 \n\t" + "shufps $0x00, %%xmm7, %%xmm7 \n\t" + "movl $-1024, %%esi \n\t" + "1: \n\t" + "movaps 1024(%0, %%esi), %%xmm0 \n\t" + "addps %%xmm7, %%xmm0 \n\t" // common + "movaps %%xmm0, %%xmm1 \n\t" // common + "addps (%0, %%esi), %%xmm0 \n\t" + "addps 2048(%0, %%esi), %%xmm1 \n\t" + "addps 3072(%0, %%esi), %%xmm0 \n\t" + "addps 4096(%0, %%esi), %%xmm1 \n\t" + "movaps %%xmm0, (%0, %%esi) \n\t" + "movaps %%xmm1, 1024(%0, %%esi) \n\t" + "addl $16, %%esi \n\t" + " jnz 1b \n\t" + :: "r" (samples+256), "m" (bias) + : "%esi" + ); +#else for (i = 0; i < 256; i++) { common = samples[i + 256] + bias; samples[i] += common + samples[i + 768]; samples[i + 256] = common + samples[i + 512] + samples[i + 1024]; } +#endif } static void mix32toS (sample_t * samples, sample_t bias)