Mercurial > mplayer.hg
changeset 3677:2a36c3d11dbd
mix5to1 & mix4to1 in SSE
| author | michael |
|---|---|
| date | Sun, 23 Dec 2001 17:16:07 +0000 |
| parents | 45ecea1c23f3 |
| children | 265680bbdcfd |
| files | liba52/downmix.c |
| diffstat | 1 files changed, 41 insertions(+), 0 deletions(-) [+] |
line wrap: on
line diff
--- a/liba52/downmix.c Sun Dec 23 17:10:27 2001 +0000 +++ b/liba52/downmix.c Sun Dec 23 17:16:07 2001 +0000 @@ -342,18 +342,59 @@ { int i; +#ifdef HAVE_SSE // FIXME untested + asm volatile( + "movlps %1, %%xmm7 \n\t" + "shufps $0x00, %%xmm7, %%xmm7 \n\t" + "movl $-1024, %%esi \n\t" + "1: \n\t" + "movaps (%0, %%esi), %%xmm0 \n\t" + "movaps 1024(%0, %%esi), %%xmm1 \n\t" + "addps 2048(%0, %%esi), %%xmm0 \n\t" + "addps 3072(%0, %%esi), %%xmm1 \n\t" + "addps %%xmm7, %%xmm0 \n\t" + "addps %%xmm1, %%xmm0 \n\t" + "movaps %%xmm0, (%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++) samples[i] += (samples[i + 256] + samples[i + 512] + samples[i + 768] + bias); +#endif } static void mix5to1 (sample_t * samples, sample_t bias) { int i; +#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 (%0, %%esi), %%xmm0 \n\t" + "movaps 1024(%0, %%esi), %%xmm1 \n\t" + "addps 2048(%0, %%esi), %%xmm0 \n\t" + "addps 3072(%0, %%esi), %%xmm1 \n\t" + "addps %%xmm7, %%xmm0 \n\t" + "addps 4096(%0, %%esi), %%xmm1 \n\t" + "addps %%xmm1, %%xmm0 \n\t" + "movaps %%xmm0, (%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++) samples[i] += (samples[i + 256] + samples[i + 512] + samples[i + 768] + samples[i + 1024] + bias); +#endif } static void mix3to2 (sample_t * samples, sample_t bias)