# HG changeset patch # User michael # Date 1009302325 0 # Node ID 99ac808d97c0a4fd351ceda4ee88f6e8bbedc78b # Parent dcbdad3fc5cd472635ab1884d5e189f86582ee9b mix22toS & move2to1 in SSE diff -r dcbdad3fc5cd -r 99ac808d97c0 liba52/downmix.c --- a/liba52/downmix.c Tue Dec 25 17:36:54 2001 +0000 +++ b/liba52/downmix.c Tue Dec 25 17:45:25 2001 +0000 @@ -592,11 +592,34 @@ int i; sample_t surround; +#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 2048(%0, %%esi), %%xmm0 \n\t" + "addps 3072(%0, %%esi), %%xmm0 \n\t" // surround + "movaps (%0, %%esi), %%xmm1 \n\t" + "movaps 1024(%0, %%esi), %%xmm2 \n\t" + "addps %%xmm7, %%xmm1 \n\t" + "addps %%xmm7, %%xmm2 \n\t" + "subps %%xmm0, %%xmm1 \n\t" + "addps %%xmm0, %%xmm2 \n\t" + "movaps %%xmm1, (%0, %%esi) \n\t" + "movaps %%xmm2, 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++) { surround = samples[i + 512] + samples[i + 768]; samples[i] += bias - surround; samples[i + 256] += bias + surround; } +#endif } static void mix32to2 (sample_t * samples, sample_t bias) @@ -675,8 +698,29 @@ { int i; +#ifdef HAVE_SSE + asm volatile( + "movlps %2, %%xmm7 \n\t" + "shufps $0x00, %%xmm7, %%xmm7 \n\t" + "movl $-1024, %%esi \n\t" + "1: \n\t" + "movaps (%0, %%esi), %%xmm0 \n\t" + "movaps 16(%0, %%esi), %%xmm1 \n\t" + "addps 1024(%0, %%esi), %%xmm0 \n\t" + "addps 1040(%0, %%esi), %%xmm1 \n\t" + "addps %%xmm7, %%xmm0 \n\t" + "addps %%xmm7, %%xmm1 \n\t" + "movaps %%xmm0, (%1, %%esi) \n\t" + "movaps %%xmm1, 16(%1, %%esi) \n\t" + "addl $32, %%esi \n\t" + " jnz 1b \n\t" + :: "r" (src+256), "r" (dest+256), "m" (bias) + : "%esi" + ); +#else for (i = 0; i < 256; i++) dest[i] = src[i] + src[i + 256] + bias; +#endif } static void zero (sample_t * samples)