diff options
| author | Wim Taymans <wim.taymans@collabora.co.uk> | 2009-09-12 12:02:59 +0200 | 
|---|---|---|
| committer | Wim Taymans <wim@metal.(none)> | 2009-09-16 17:14:12 +0200 | 
| commit | 3d5a572694388008fd4cf522699fc6c431f97325 (patch) | |
| tree | b8d172492a3ce81b57dcfafaeb3c74fa38fe89dc /src | |
| parent | d397a82e14da192995ba8caee5c95e70e545df29 (diff) | |
svolume_mmx: optimize some more
We can reorder the algortihm a bit like we do for sse so that we
don't need the contants and masking instructions. Saves 2 instructions
for the mmx code.
Diffstat (limited to 'src')
| -rw-r--r-- | src/pulsecore/svolume_mmx.c | 20 | 
1 files changed, 7 insertions, 13 deletions
| diff --git a/src/pulsecore/svolume_mmx.c b/src/pulsecore/svolume_mmx.c index 74918e78..170f01d4 100644 --- a/src/pulsecore/svolume_mmx.c +++ b/src/pulsecore/svolume_mmx.c @@ -57,14 +57,12 @@        " punpcklwd %%mm4, "#s"        \n\t" /* .. |    0  |   p0  | */                   \        " pcmpgtw "#v", %%mm4          \n\t" /* .. |    0  | s(vl) | */                   \        " pand "#s", %%mm4             \n\t" /* .. |    0  |  (p0) |  (vl >> 15) & p */   \ -      " movq %%mm6, %%mm5            \n\t" /* .. |  ffff |   0   | */                   \ -      " pand "#v", %%mm5             \n\t" /* .. |   vh  |   0   | */                   \ -      " por %%mm5, %%mm4             \n\t" /* .. |   vh  |  (p0) | */                   \ -      " pmulhw "#s", "#v"            \n\t" /* .. |    0  | vl*p0 | */                   \ -      " paddw %%mm4, "#v"            \n\t" /* .. |   vh  | vl*p0 | vh + sign correct */ \ -      " pslld $16, "#s"              \n\t" /* .. |   p0  |    0  | */                   \ -      " por %%mm7, "#s"              \n\t" /* .. |   p0  |    1  | */                   \ -      " pmaddwd "#s", "#v"           \n\t" /* .. |    p0 * v0    | */                   \ +      " movq "#s", %%mm5             \n\t"                                              \ +      " pmulhw "#v", "#s"            \n\t" /* .. |    0  | vl*p0 | */                   \ +      " paddw %%mm4, "#s"            \n\t" /* .. |    0  | vl*p0 | + sign correct */    \ +      " psrld $16, "#v"              \n\t" /* .. |    0  |   vh  | */                   \ +      " pmaddwd %%mm5, "#v"          \n\t" /* .. |    p0 * vh    | */                   \ +      " paddd "#s", "#v"             \n\t" /* .. |    p0 * v0    | */                   \        " packssdw "#v", "#v"          \n\t" /* .. | p1*v1 | p0*v0 | */  /* approximately advances %3 = (%3 + a) % b. This function requires that @@ -105,10 +103,6 @@ pa_volume_s16ne_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsi      __asm__ __volatile__ (          " xor %3, %3                    \n\t"          " sar $1, %2                    \n\t" /* length /= sizeof (int16_t) */ -        " pcmpeqw %%mm6, %%mm6          \n\t" /* .. |  ffff |  ffff | */ -        " pcmpeqw %%mm7, %%mm7          \n\t" /* .. |  ffff |  ffff | */ -        " pslld  $16, %%mm6             \n\t" /* .. |  ffff |     0 | */ -        " psrld  $31, %%mm7             \n\t" /* .. |     0 |     1 | */          " test $1, %2                   \n\t" /* check for odd samples */          " je 2f                         \n\t" @@ -239,7 +233,7 @@ pa_volume_s16re_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsi      );  } -#undef RUN_TEST +#define RUN_TEST  #ifdef RUN_TEST  #define CHANNELS 2 | 
