mirrors/pulseaudio
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/pulseaudio/pulseaudio.git synced 2025-11-07 13:30:03 -05:00
Code Issues Projects Releases Packages Wiki Activity Actions

svolume: improve SSE and MMX code

Browse source
This commit is contained in:
Wim Taymans 2009-08-13 13:45:01 +02:00
parent 7086784573
commit 5998cf99b0
2 changed files with 277 additions and 132 deletions
Download patch file Download diff file Expand all files Collapse all files

218
src/pulsecore/svolume_mmx.c
View file

@ -99,6 +99,46 @@ pa_volume_ulaw_mmx (uint8_t *samples, int32_t *volumes, unsigned channels, unsig
} }
#endif #endif
#define VOLUME_32x16(s,v) /* .. | vh | vl | */ \
" pxor %%mm4, %%mm4 \n\t" /* .. | 0 | 0 | */ \
" 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 | */ \
" packssdw "#v", "#v" \n\t" /* .. | p1*v1 | p0*v0 | */
#define MOD_ADD(a,b) \
" add "#a", %3 \n\t" \
" mov %3, %4 \n\t" \
" sub "#b", %4 \n\t" \
" cmp "#b", %3 \n\t" \
" cmovae %4, %3 \n\t"
/* swap 16 bits */
#define SWAP_16(s) \
" movq "#s", %%mm4 \n\t" /* .. | h l | */ \
" psrlw $8, %%mm4 \n\t" /* .. | 0 h | */ \
" psllw $8, "#s" \n\t" /* .. | l 0 | */ \
" por %%mm4, "#s" \n\t" /* .. | l h | */
/* swap 2 registers 16 bits for better pairing */
#define SWAP_16_2(s1,s2) \
" movq "#s1", %%mm4 \n\t" /* .. | h l | */ \
" movq "#s2", %%mm5 \n\t" \
" psrlw $8, %%mm4 \n\t" /* .. | 0 h | */ \
" psrlw $8, %%mm5 \n\t" \
" psllw $8, "#s1" \n\t" /* .. | l 0 | */ \
" psllw $8, "#s2" \n\t" \
" por %%mm4, "#s1" \n\t" /* .. | l h | */ \
" por %%mm5, "#s2" \n\t"
static void static void
pa_volume_s16ne_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) pa_volume_s16ne_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
{ {
@ -108,38 +148,22 @@ pa_volume_s16ne_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsi
* we overread the volume array, which should have enough padding. */ * we overread the volume array, which should have enough padding. */
channels = MAX (4, channels); channels = MAX (4, channels);
#define VOLUME_32x16(s,v) /* v1_h | v1_l | v0_h | v0_l */ \
" pxor %%mm4, %%mm4 \n\t" \
" punpcklwd %%mm4, "#s" \n\t" /* 0 | p1 | 0 | p0 */ \
" pcmpgtw "#s", %%mm4 \n\t" /* select sign from sample */ \
" pand "#v", %%mm4 \n\t" /* extract sign correction factors */ \
" movq "#s", %%mm5 \n\t" \
" pmulhuw "#v", "#s" \n\t" /* 0 | p1*v1lh | 0 | p0*v0lh */ \
" psubd %%mm4, "#s" \n\t" /* sign correction */ \
" psrld $16, "#v" \n\t" /* 0 | v1h | 0 | v0h */ \
" pmaddwd %%mm5, "#v" \n\t" /* p1 * v1h | p0 * v0h */ \
" paddd "#s", "#v" \n\t" /* p1 * v1 | p0 * v0 */ \
" packssdw "#v", "#v" \n\t" /* p0*v0 | p1*v1 | p0*v0 | p1*v1 */
#define MOD_ADD(a,b) \
" add "#a", %3 \n\t" \
" mov %3, %4 \n\t" \
" sub "#b", %4 \n\t" \
" cmp "#b", %3 \n\t" \
" cmovae %4, %3 \n\t"
__asm__ __volatile__ ( __asm__ __volatile__ (
" xor %3, %3 \n\t" " xor %3, %3 \n\t"
" sar $1, %2 \n\t" /* length /= sizeof (int16_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 */ " test $1, %2 \n\t" /* check for odd samples */
" je 2f \n\t" " je 2f \n\t"
" movd (%1, %3, 4), %%mm0 \n\t" /* do odd samples */ " movd (%1, %3, 4), %%mm0 \n\t" /* | v0h | v0l | */
" movw (%0), %4 \n\t" " movw (%0), %4 \n\t" /* .. | p0 | */
" movd %4, %%mm1 \n\t" " movd %4, %%mm1 \n\t"
VOLUME_32x16 (%%mm1, %%mm0) VOLUME_32x16 (%%mm1, %%mm0)
" movd %%mm0, %4 \n\t" " movd %%mm0, %4 \n\t" /* .. | p0*v0 | */
" movw %4, (%0) \n\t" " movw %4, (%0) \n\t"
" add $2, %0 \n\t" " add $2, %0 \n\t"
MOD_ADD ($1, %5) MOD_ADD ($1, %5)
@ -149,11 +173,11 @@ pa_volume_s16ne_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsi
" test $1, %2 \n\t" /* check for odd samples */ " test $1, %2 \n\t" /* check for odd samples */
" je 4f \n\t" " je 4f \n\t"
"3: \n\t" /* do samples in pairs of 2 */ "3: \n\t" /* do samples in groups of 2 */
" movq (%1, %3, 4), %%mm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */ " movq (%1, %3, 4), %%mm0 \n\t" /* | v1h | v1l | v0h | v0l | */
" movd (%0), %%mm1 \n\t" /* X | X | p1 | p0 */ " movd (%0), %%mm1 \n\t" /* .. | p1 | p0 | */
VOLUME_32x16 (%%mm1, %%mm0) VOLUME_32x16 (%%mm1, %%mm0)
" movd %%mm0, (%0) \n\t" " movd %%mm0, (%0) \n\t" /* | p1*v1 | p0*v0 | */
" add $4, %0 \n\t" " add $4, %0 \n\t"
MOD_ADD ($2, %5) MOD_ADD ($2, %5)
@ -162,15 +186,91 @@ pa_volume_s16ne_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsi
" cmp $0, %2 \n\t" " cmp $0, %2 \n\t"
" je 6f \n\t" " je 6f \n\t"
"5: \n\t" /* do samples in pairs of 4 */ "5: \n\t" /* do samples in groups of 4 */
" movq (%1, %3, 4), %%mm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */ " movq (%1, %3, 4), %%mm0 \n\t" /* | v1h | v1l | v0h | v0l | */
" movq 8(%1, %3, 4), %%mm2 \n\t" /* v3_h | v3_l | v2_h | v2_l */ " movq 8(%1, %3, 4), %%mm2 \n\t" /* | v3h | v3l | v2h | v2l | */
" movd (%0), %%mm1 \n\t" /* X | X | p1 | p0 */ " movd (%0), %%mm1 \n\t" /* .. | p1 | p0 | */
" movd 4(%0), %%mm3 \n\t" /* X | X | p3 | p2 */ " movd 4(%0), %%mm3 \n\t" /* .. | p3 | p2 | */
VOLUME_32x16 (%%mm1, %%mm0) VOLUME_32x16 (%%mm1, %%mm0)
VOLUME_32x16 (%%mm3, %%mm2) VOLUME_32x16 (%%mm3, %%mm2)
" movd %%mm0, (%0) \n\t" " movd %%mm0, (%0) \n\t" /* | p1*v1 | p0*v0 | */
" movd %%mm2, 4(%0) \n\t" " movd %%mm2, 4(%0) \n\t" /* | p3*v3 | p2*v2 | */
" add $8, %0 \n\t"
MOD_ADD ($4, %5)
" dec %2 \n\t"
" jne 5b \n\t"
"6: \n\t"
" emms \n\t"
: "+r" (samples), "+r" (volumes), "+r" (length), "=D" ((int64_t)channel), "=&r" (temp)
: "r" ((int64_t)channels)
: "cc"
);
}
static void
pa_volume_s16re_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
{
int64_t channel, temp;
/* the max number of samples we process at a time, this is also the max amount
* we overread the volume array, which should have enough padding. */
channels = MAX (4, channels);
__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"
" movd (%1, %3, 4), %%mm0 \n\t" /* | v0h | v0l | */
" movw (%0), %4 \n\t" /* .. | p0 | */
" rorw $8, %4 \n\t"
" movd %4, %%mm1 \n\t"
VOLUME_32x16 (%%mm1, %%mm0)
" movd %%mm0, %4 \n\t" /* .. | p0*v0 | */
" rorw $8, %4 \n\t"
" movw %4, (%0) \n\t"
" add $2, %0 \n\t"
MOD_ADD ($1, %5)
"2: \n\t"
" sar $1, %2 \n\t" /* prepare for processing 2 samples at a time */
" test $1, %2 \n\t" /* check for odd samples */
" je 4f \n\t"
"3: \n\t" /* do samples in groups of 2 */
" movq (%1, %3, 4), %%mm0 \n\t" /* | v1h | v1l | v0h | v0l | */
" movd (%0), %%mm1 \n\t" /* .. | p1 | p0 | */
SWAP_16 (%%mm1)
VOLUME_32x16 (%%mm1, %%mm0)
SWAP_16 (%%mm0)
" movd %%mm0, (%0) \n\t" /* | p1*v1 | p0*v0 | */
" add $4, %0 \n\t"
MOD_ADD ($2, %5)
"4: \n\t"
" sar $1, %2 \n\t" /* prepare for processing 4 samples at a time */
" cmp $0, %2 \n\t"
" je 6f \n\t"
"5: \n\t" /* do samples in groups of 4 */
" movq (%1, %3, 4), %%mm0 \n\t" /* | v1h | v1l | v0h | v0l | */
" movq 8(%1, %3, 4), %%mm2 \n\t" /* | v3h | v3l | v2h | v2l | */
" movd (%0), %%mm1 \n\t" /* .. | p1 | p0 | */
" movd 4(%0), %%mm3 \n\t" /* .. | p3 | p2 | */
SWAP_16_2 (%%mm1, %%mm3)
VOLUME_32x16 (%%mm1, %%mm0)
VOLUME_32x16 (%%mm3, %%mm2)
SWAP_16_2 (%%mm0, %%mm2)
" movd %%mm0, (%0) \n\t" /* | p1*v1 | p0*v0 | */
" movd %%mm2, 4(%0) \n\t" /* | p3*v3 | p2*v2 | */
" add $8, %0 \n\t" " add $8, %0 \n\t"
MOD_ADD ($4, %5) MOD_ADD ($4, %5)
" dec %2 \n\t" " dec %2 \n\t"
@ -186,29 +286,6 @@ pa_volume_s16ne_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsi
} }
#if 0 #if 0
static void
pa_volume_s16re_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
{
unsigned channel;
length /= sizeof (int16_t);
for (channel = 0; length; length--) {
int32_t t, hi, lo;
hi = volumes[channel] >> 16;
lo = volumes[channel] & 0xFFFF;
t = (int32_t) PA_INT16_SWAP(*samples);
t = ((t * lo) >> 16) + (t * hi);
t = PA_CLAMP_UNLIKELY(t, -0x8000, 0x7FFF);
*samples++ = PA_INT16_SWAP((int16_t) t);
if (PA_UNLIKELY(++channel >= channels))
channel = 0;
}
}
static void static void
pa_volume_float32ne_mmx (float *samples, float *volumes, unsigned channels, unsigned length) pa_volume_float32ne_mmx (float *samples, float *volumes, unsigned channels, unsigned length)
{ {
@ -366,42 +443,37 @@ pa_volume_s24_32re_mmx (uint32_t *samples, int32_t *volumes, unsigned channels,
} }
#endif #endif
#undef RUN_TEST #define RUN_TEST
#ifdef RUN_TEST #ifdef RUN_TEST
#define CHANNELS 2 #define CHANNELS 2
#define SAMPLES 1021 #define SAMPLES 1021
#define TIMES 1000 #define TIMES 1000
#define PADDING 16
static void run_test (void) { static void run_test (void) {
int16_t samples[SAMPLES]; int16_t samples[SAMPLES];
int16_t samples_ref[SAMPLES]; int16_t samples_ref[SAMPLES];
int16_t samples_orig[SAMPLES]; int16_t samples_orig[SAMPLES];
int32_t volumes[CHANNELS + 16]; int32_t volumes[CHANNELS + PADDING];
int i, j, padding; int i, j, padding;
pa_do_volume_func_t func; pa_do_volume_func_t func;
func = pa_get_volume_func (PA_SAMPLE_S16NE); func = pa_get_volume_func (PA_SAMPLE_S16RE);
printf ("checking %d\n", sizeof (samples)); printf ("checking MMX %d\n", sizeof (samples));
for (j = 0; j < TIMES; j++) { for (j = 0; j < TIMES; j++) {
/*
for (i = 0; i < SAMPLES; i++) {
samples[i] samples_ref[i] = samples_orig[i] = rand() >> 16;
}
*/
pa_random (samples, sizeof (samples)); pa_random (samples, sizeof (samples));
memcpy (samples_ref, samples, sizeof (samples)); memcpy (samples_ref, samples, sizeof (samples));
memcpy (samples_orig, samples, sizeof (samples)); memcpy (samples_orig, samples, sizeof (samples));
for (i = 0; i < CHANNELS; i++) for (i = 0; i < CHANNELS; i++)
volumes[i] = rand() >> 15; volumes[i] = rand() >> 1;
for (padding = 0; padding < 16; padding++, i++) for (padding = 0; padding < PADDING; padding++, i++)
volumes[i] = volumes[padding]; volumes[i] = volumes[padding];
pa_volume_s16ne_mmx (samples, volumes, CHANNELS, sizeof (samples)); pa_volume_s16re_mmx (samples, volumes, CHANNELS, sizeof (samples));
func (samples_ref, volumes, CHANNELS, sizeof (samples)); func (samples_ref, volumes, CHANNELS, sizeof (samples));
for (i = 0; i < SAMPLES; i++) { for (i = 0; i < SAMPLES; i++) {
@ -409,11 +481,6 @@ static void run_test (void) {
printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i], printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
samples_orig[i], volumes[i % CHANNELS]); samples_orig[i], volumes[i % CHANNELS]);
} }
#if 0
else
printf ("%d: %04x == %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
samples_orig[i], volumes[i % CHANNELS]);
#endif
} }
} }
} }
@ -427,4 +494,5 @@ void pa_volume_func_init_mmx (void) {
#endif #endif
pa_set_volume_func (PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_mmx); pa_set_volume_func (PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_mmx);
pa_set_volume_func (PA_SAMPLE_S16RE, (pa_do_volume_func_t) pa_volume_s16re_mmx);
} }

191
src/pulsecore/svolume_sse.c
View file

@ -99,6 +99,44 @@ pa_volume_ulaw_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsig
} }
#endif #endif
#define VOLUME_32x16(s,v) /* .. | vh | vl | */ \
" pxor %%xmm4, %%xmm4 \n\t" /* .. | 0 | 0 | */ \
" punpcklwd %%xmm4, "#s" \n\t" /* .. | 0 | p0 | */ \
" pcmpgtw "#s", %%xmm4 \n\t" /* .. | 0 | s(p0) | */ \
" pand "#v", %%xmm4 \n\t" /* .. | 0 | (vl) | */ \
" movdqa "#s", %%xmm5 \n\t" \
" pmulhuw "#v", "#s" \n\t" /* .. | 0 | vl*p0 | */ \
" psubd %%xmm4, "#s" \n\t" /* .. | 0 | vl*p0 | + sign correct */ \
" psrld $16, "#v" \n\t" /* .. | p0 | 0 | */ \
" pmaddwd %%xmm5, "#v" \n\t" /* .. | p0 * vh | */ \
" paddd "#s", "#v" \n\t" /* .. | p0 * v0 | */ \
" packssdw "#v", "#v" \n\t" /* .. | p1*v1 | p0*v0 | */
#define MOD_ADD(a,b) \
" add "#a", %3 \n\t" /* channel += inc */ \
" mov %3, %4 \n\t" \
" sub "#b", %4 \n\t" /* tmp = channel - channels */ \
" cmp "#b", %3 \n\t" /* if (channel >= channels) */ \
" cmovae %4, %3 \n\t" /* channel = tmp */
/* swap 16 bits */
#define SWAP_16(s) \
" movdqa "#s", %%xmm4 \n\t" /* .. | h l | */ \
" psrlw $8, %%xmm4 \n\t" /* .. | 0 h | */ \
" psllw $8, "#s" \n\t" /* .. | l 0 | */ \
" por %%xmm4, "#s" \n\t" /* .. | l h | */
/* swap 2 registers 16 bits for better pairing */
#define SWAP_16_2(s1,s2) \
" movdqa "#s1", %%xmm4 \n\t" /* .. | h l | */ \
" movdqa "#s2", %%xmm5 \n\t" \
" psrlw $8, %%xmm4 \n\t" /* .. | 0 h | */ \
" psrlw $8, %%xmm5 \n\t" \
" psllw $8, "#s1" \n\t" /* .. | l 0 | */ \
" psllw $8, "#s2" \n\t" \
" por %%xmm4, "#s1" \n\t" /* .. | l h | */ \
" por %%xmm5, "#s2" \n\t"
static void static void
pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
{ {
@ -109,25 +147,83 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi
* we overread the volume array, which should have enough padding. */ * we overread the volume array, which should have enough padding. */
channels = MAX (8, channels); channels = MAX (8, channels);
#define VOLUME_32x16(s,v) /* v1_h | v1_l | v0_h | v0_l */ \ __asm__ __volatile__ (
" pxor %%xmm4, %%xmm4 \n\t" \ " xor %3, %3 \n\t"
" punpcklwd %%xmm4, "#s" \n\t" /* 0 | p1 | 0 | p0 */ \ " sar $1, %2 \n\t" /* length /= sizeof (int16_t) */
" pcmpgtw "#s", %%xmm4 \n\t" /* select sign from sample */ \
" pand "#v", %%xmm4 \n\t" /* extract sign correction factors */ \
" movdqa "#s", %%xmm5 \n\t" \
" pmulhuw "#v", "#s" \n\t" /* 0 | p1*v1lh | 0 | p0*v0lh */ \
" psubd %%xmm4, "#s" \n\t" /* sign correction */ \
" psrld $16, "#v" \n\t" /* 0 | v1h | 0 | v0h */ \
" pmaddwd %%xmm5, "#v" \n\t" /* p1 * v1h | p0 * v0h */ \
" paddd "#s", "#v" \n\t" /* p1 * v1 | p0 * v0 */ \
" packssdw "#v", "#v" \n\t" /* p0*v0 | p1*v1 | p0*v0 | p1*v1 */
#define MOD_ADD(a,b) \ " test $1, %2 \n\t" /* check for odd samples */
" add "#a", %3 \n\t" \ " je 2f \n\t"
" mov %3, %4 \n\t" \
" sub "#b", %4 \n\t" \ " movd (%1, %3, 4), %%xmm0 \n\t" /* | v0h | v0l | */
" cmp "#b", %3 \n\t" \ " movw (%0), %4 \n\t" /* .. | p0 | */
" cmovae %4, %3 \n\t" " movd %4, %%xmm1 \n\t"
VOLUME_32x16 (%%xmm1, %%xmm0)
" movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */
" movw %4, (%0) \n\t"
" add $2, %0 \n\t"
MOD_ADD ($1, %5)
"2: \n\t"
" sar $1, %2 \n\t" /* prepare for processing 2 samples at a time */
" test $1, %2 \n\t"
" je 4f \n\t"
"3: \n\t" /* do samples in groups of 2 */
" movq (%1, %3, 4), %%xmm0 \n\t" /* | v1h | v1l | v0h | v0l | */
" movd (%0), %%xmm1 \n\t" /* .. | p1 | p0 | */
VOLUME_32x16 (%%xmm1, %%xmm0)
" movd %%xmm0, (%0) \n\t" /* .. | p1*v1 | p0*v0 | */
" add $4, %0 \n\t"
MOD_ADD ($2, %5)
"4: \n\t"
" sar $1, %2 \n\t" /* prepare for processing 4 samples at a time */
" test $1, %2 \n\t"
" je 6f \n\t"
"5: \n\t" /* do samples in groups of 4 */
" movdqu (%1, %3, 4), %%xmm0 \n\t" /* | v3h | v3l .. v0h | v0l | */
" movq (%0), %%xmm1 \n\t" /* .. | p3 .. p0 | */
VOLUME_32x16 (%%xmm1, %%xmm0)
" movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */
" add $8, %0 \n\t"
MOD_ADD ($4, %5)
"6: \n\t"
" sar $1, %2 \n\t" /* prepare for processing 8 samples at a time */
" cmp $0, %2 \n\t"
" je 8f \n\t"
"7: \n\t" /* do samples in groups of 8 */
" movdqu (%1, %3, 4), %%xmm0 \n\t" /* | v3h | v3l .. v0h | v0l | */
" movdqu 16(%1, %3, 4), %%xmm2 \n\t" /* | v7h | v7l .. v4h | v4l | */
" movq (%0), %%xmm1 \n\t" /* .. | p3 .. p0 | */
" movq 8(%0), %%xmm3 \n\t" /* .. | p7 .. p4 | */
VOLUME_32x16 (%%xmm1, %%xmm0)
VOLUME_32x16 (%%xmm3, %%xmm2)
" movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */
" movq %%xmm2, 8(%0) \n\t" /* .. | p7*v7 .. p4*v4 | */
" add $16, %0 \n\t"
MOD_ADD ($8, %5)
" dec %2 \n\t"
" jne 7b \n\t"
"8: \n\t"
: "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp)
: "r" ((int64_t)channels)
: "cc"
);
}
static void
pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
{
int64_t channel;
int64_t temp;
/* the max number of samples we process at a time, this is also the max amount
* we overread the volume array, which should have enough padding. */
channels = MAX (8, channels);
__asm__ __volatile__ ( __asm__ __volatile__ (
" xor %3, %3 \n\t" " xor %3, %3 \n\t"
@ -138,9 +234,11 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi
" movd (%1, %3, 4), %%xmm0 \n\t" /* do odd sample */ " movd (%1, %3, 4), %%xmm0 \n\t" /* do odd sample */
" movw (%0), %4 \n\t" " movw (%0), %4 \n\t"
" rorw $8, %4 \n\t"
" movd %4, %%xmm1 \n\t" " movd %4, %%xmm1 \n\t"
VOLUME_32x16 (%%xmm1, %%xmm0) VOLUME_32x16 (%%xmm1, %%xmm0)
" movd %%xmm0, %4 \n\t" " movd %%xmm0, %4 \n\t"
" rorw $8, %4 \n\t"
" movw %4, (%0) \n\t" " movw %4, (%0) \n\t"
" add $2, %0 \n\t" " add $2, %0 \n\t"
MOD_ADD ($1, %5) MOD_ADD ($1, %5)
@ -153,7 +251,9 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi
"3: \n\t" /* do samples in pairs of 2 */ "3: \n\t" /* do samples in pairs of 2 */
" movq (%1, %3, 4), %%xmm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */ " movq (%1, %3, 4), %%xmm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */
" movd (%0), %%xmm1 \n\t" /* X | X | p1 | p0 */ " movd (%0), %%xmm1 \n\t" /* X | X | p1 | p0 */
SWAP_16 (%%xmm1)
VOLUME_32x16 (%%xmm1, %%xmm0) VOLUME_32x16 (%%xmm1, %%xmm0)
SWAP_16 (%%xmm0)
" movd %%xmm0, (%0) \n\t" " movd %%xmm0, (%0) \n\t"
" add $4, %0 \n\t" " add $4, %0 \n\t"
MOD_ADD ($2, %5) MOD_ADD ($2, %5)
@ -164,9 +264,11 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi
" je 6f \n\t" " je 6f \n\t"
"5: \n\t" /* do samples in pairs of 4 */ "5: \n\t" /* do samples in pairs of 4 */
" movdqa (%1, %3, 4), %%xmm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */ " movdqu (%1, %3, 4), %%xmm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */
" movq (%0), %%xmm1 \n\t" /* X | X | p1 | p0 */ " movq (%0), %%xmm1 \n\t" /* X | X | p1 | p0 */
SWAP_16 (%%xmm1)
VOLUME_32x16 (%%xmm1, %%xmm0) VOLUME_32x16 (%%xmm1, %%xmm0)
SWAP_16 (%%xmm0)
" movq %%xmm0, (%0) \n\t" " movq %%xmm0, (%0) \n\t"
" add $8, %0 \n\t" " add $8, %0 \n\t"
MOD_ADD ($4, %5) MOD_ADD ($4, %5)
@ -177,12 +279,14 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi
" je 8f \n\t" " je 8f \n\t"
"7: \n\t" /* do samples in pairs of 8 */ "7: \n\t" /* do samples in pairs of 8 */
" movdqa (%1, %3, 4), %%xmm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */ " movdqu (%1, %3, 4), %%xmm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */
" movdqa 16(%1, %3, 4), %%xmm2 \n\t" /* v3_h | v3_l | v2_h | v2_l */ " movdqu 16(%1, %3, 4), %%xmm2 \n\t" /* v3_h | v3_l | v2_h | v2_l */
" movq (%0), %%xmm1 \n\t" /* X | X | p1 | p0 */ " movq (%0), %%xmm1 \n\t" /* X | X | p1 | p0 */
" movq 8(%0), %%xmm3 \n\t" /* X | X | p3 | p2 */ " movq 8(%0), %%xmm3 \n\t" /* X | X | p3 | p2 */
SWAP_16_2 (%%xmm1, %%xmm3)
VOLUME_32x16 (%%xmm1, %%xmm0) VOLUME_32x16 (%%xmm1, %%xmm0)
VOLUME_32x16 (%%xmm3, %%xmm2) VOLUME_32x16 (%%xmm3, %%xmm2)
SWAP_16_2 (%%xmm0, %%xmm2)
" movq %%xmm0, (%0) \n\t" " movq %%xmm0, (%0) \n\t"
" movq %%xmm2, 8(%0) \n\t" " movq %%xmm2, 8(%0) \n\t"
" add $16, %0 \n\t" " add $16, %0 \n\t"
@ -198,29 +302,6 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi
} }
#if 0 #if 0
static void
pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
{
unsigned channel;
length /= sizeof (int16_t);
for (channel = 0; length; length--) {
int32_t t, hi, lo;
hi = volumes[channel] >> 16;
lo = volumes[channel] & 0xFFFF;
t = (int32_t) PA_INT16_SWAP(*samples);
t = ((t * lo) >> 16) + (t * hi);
t = PA_CLAMP_UNLIKELY(t, -0x8000, 0x7FFF);
*samples++ = PA_INT16_SWAP((int16_t) t);
if (PA_UNLIKELY(++channel >= channels))
channel = 0;
}
}
static void static void
pa_volume_float32ne_sse (float *samples, float *volumes, unsigned channels, unsigned length) pa_volume_float32ne_sse (float *samples, float *volumes, unsigned channels, unsigned length)
{ {
@ -378,24 +459,25 @@ pa_volume_s24_32re_sse (uint32_t *samples, int32_t *volumes, unsigned channels,
} }
#endif #endif
#undef RUN_TEST #define RUN_TEST
#ifdef RUN_TEST #ifdef RUN_TEST
#define CHANNELS 2 #define CHANNELS 2
#define SAMPLES 1021 #define SAMPLES 1021
#define TIMES 1000 #define TIMES 1000
#define PADDING 16
static void run_test (void) { static void run_test (void) {
int16_t samples[SAMPLES]; int16_t samples[SAMPLES];
int16_t samples_ref[SAMPLES]; int16_t samples_ref[SAMPLES];
int16_t samples_orig[SAMPLES]; int16_t samples_orig[SAMPLES];
int32_t volumes[CHANNELS + 16]; int32_t volumes[CHANNELS + PADDING];
int i, j, padding; int i, j, padding;
pa_do_volume_func_t func; pa_do_volume_func_t func;
func = pa_get_volume_func (PA_SAMPLE_S16NE); func = pa_get_volume_func (PA_SAMPLE_S16RE);
printf ("checking %d\n", sizeof (samples)); printf ("checking SSE %d\n", sizeof (samples));
for (j = 0; j < TIMES; j++) { for (j = 0; j < TIMES; j++) {
pa_random (samples, sizeof (samples)); pa_random (samples, sizeof (samples));
@ -403,12 +485,11 @@ static void run_test (void) {
memcpy (samples_orig, samples, sizeof (samples)); memcpy (samples_orig, samples, sizeof (samples));
for (i = 0; i < CHANNELS; i++) for (i = 0; i < CHANNELS; i++)
volumes[i] = rand() >> 15; volumes[i] = rand() >> 1;
for (padding = 0; padding < PADDING; padding++, i++)
for (padding = 0; padding < 16; padding++, i++)
volumes[i] = volumes[padding]; volumes[i] = volumes[padding];
pa_volume_s16ne_sse (samples, volumes, CHANNELS, SAMPLES * sizeof (int16_t)); pa_volume_s16re_sse (samples, volumes, CHANNELS, SAMPLES * sizeof (int16_t));
func (samples_ref, volumes, CHANNELS, SAMPLES * sizeof (int16_t)); func (samples_ref, volumes, CHANNELS, SAMPLES * sizeof (int16_t));
for (i = 0; i < SAMPLES; i++) { for (i = 0; i < SAMPLES; i++) {
@ -416,11 +497,6 @@ static void run_test (void) {
printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i], printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
samples_orig[i], volumes[i % CHANNELS]); samples_orig[i], volumes[i % CHANNELS]);
} }
#if 0
else
printf ("%d: %04x == %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
samples_orig[i], volumes[i % CHANNELS]);
#endif
} }
} }
} }
@ -434,4 +510,5 @@ void pa_volume_func_init_sse (void) {
#endif #endif
pa_set_volume_func (PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_sse); pa_set_volume_func (PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_sse);
pa_set_volume_func (PA_SAMPLE_S16RE, (pa_do_volume_func_t) pa_volume_s16re_sse);
} }