}
#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
pa_volume_s16ne_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
{
* we overread the volume array, which should have enough padding. */
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__ (
" 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" /* do odd samples */
- " movw (%0), %4 \n\t"
+ " movd (%1, %3, 4), %%mm0 \n\t" /* | v0h | v0l | */
+ " movw (%0), %4 \n\t" /* .. | p0 | */
" movd %4, %%mm1 \n\t"
VOLUME_32x16 (%%mm1, %%mm0)
- " movd %%mm0, %4 \n\t"
+ " movd %%mm0, %4 \n\t" /* .. | p0*v0 | */
" movw %4, (%0) \n\t"
" add $2, %0 \n\t"
MOD_ADD ($1, %5)
" test $1, %2 \n\t" /* check for odd samples */
" je 4f \n\t"
- "3: \n\t" /* do samples in pairs of 2 */
- " movq (%1, %3, 4), %%mm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */
- " movd (%0), %%mm1 \n\t" /* X | X | p1 | p0 */
+ "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 | */
VOLUME_32x16 (%%mm1, %%mm0)
- " movd %%mm0, (%0) \n\t"
+ " movd %%mm0, (%0) \n\t" /* | p1*v1 | p0*v0 | */
" add $4, %0 \n\t"
MOD_ADD ($2, %5)
" cmp $0, %2 \n\t"
" je 6f \n\t"
- "5: \n\t" /* do samples in pairs of 4 */
- " movq (%1, %3, 4), %%mm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */
- " movq 8(%1, %3, 4), %%mm2 \n\t" /* v3_h | v3_l | v2_h | v2_l */
- " movd (%0), %%mm1 \n\t" /* X | X | p1 | p0 */
- " movd 4(%0), %%mm3 \n\t" /* X | X | p3 | p2 */
+ "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 | */
VOLUME_32x16 (%%mm1, %%mm0)
VOLUME_32x16 (%%mm3, %%mm2)
- " movd %%mm0, (%0) \n\t"
- " movd %%mm2, 4(%0) \n\t"
+ " movd %%mm0, (%0) \n\t" /* | p1*v1 | p0*v0 | */
+ " movd %%mm2, 4(%0) \n\t" /* | p3*v3 | p2*v2 | */
" add $8, %0 \n\t"
MOD_ADD ($4, %5)
" dec %2 \n\t"
);
}
-#if 0
static void
pa_volume_s16re_mmx (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
{
- unsigned channel;
+ int64_t channel, temp;
- length /= sizeof (int16_t);
+ /* 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);
- for (channel = 0; length; length--) {
- int32_t t, hi, lo;
+ __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 | */
- hi = volumes[channel] >> 16;
- lo = volumes[channel] & 0xFFFF;
+ " test $1, %2 \n\t" /* check for odd samples */
+ " je 2f \n\t"
- 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);
+ " 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)
- if (PA_UNLIKELY(++channel >= channels))
- channel = 0;
- }
+ "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"
+ 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"
+ );
}
+#if 0
static void
pa_volume_float32ne_mmx (float *samples, float *volumes, unsigned channels, unsigned length)
{
}
#endif
-#undef RUN_TEST
+#define RUN_TEST
#ifdef RUN_TEST
#define CHANNELS 2
#define SAMPLES 1021
#define TIMES 1000
+#define PADDING 16
static void run_test (void) {
int16_t samples[SAMPLES];
int16_t samples_ref[SAMPLES];
int16_t samples_orig[SAMPLES];
- int32_t volumes[CHANNELS + 16];
+ int32_t volumes[CHANNELS + PADDING];
int i, j, padding;
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 (i = 0; i < SAMPLES; i++) {
- samples[i] samples_ref[i] = samples_orig[i] = rand() >> 16;
- }
- */
-
pa_random (samples, sizeof (samples));
memcpy (samples_ref, samples, sizeof (samples));
memcpy (samples_orig, samples, sizeof (samples));
for (i = 0; i < CHANNELS; i++)
- volumes[i] = rand() >> 15;
- for (padding = 0; padding < 16; padding++, i++)
+ volumes[i] = rand() >> 1;
+ for (padding = 0; padding < PADDING; padding++, i++)
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));
for (i = 0; i < SAMPLES; i++) {
printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
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
}
}
}
#endif
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);
}
}
#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
pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
{
* we overread the volume array, which should have enough padding. */
channels = MAX (8, channels);
-#define VOLUME_32x16(s,v) /* v1_h | v1_l | v0_h | v0_l */ \
- " pxor %%xmm4, %%xmm4 \n\t" \
- " punpcklwd %%xmm4, "#s" \n\t" /* 0 | p1 | 0 | p0 */ \
- " 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 */
+ __asm__ __volatile__ (
+ " xor %3, %3 \n\t"
+ " sar $1, %2 \n\t" /* length /= sizeof (int16_t) */
-#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"
+ " test $1, %2 \n\t" /* check for odd samples */
+ " je 2f \n\t"
+
+ " movd (%1, %3, 4), %%xmm0 \n\t" /* | v0h | v0l | */
+ " movw (%0), %4 \n\t" /* .. | p0 | */
+ " 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__ (
" xor %3, %3 \n\t"
" movd (%1, %3, 4), %%xmm0 \n\t" /* do odd sample */
" movw (%0), %4 \n\t"
+ " rorw $8, %4 \n\t"
" movd %4, %%xmm1 \n\t"
VOLUME_32x16 (%%xmm1, %%xmm0)
" movd %%xmm0, %4 \n\t"
+ " rorw $8, %4 \n\t"
" movw %4, (%0) \n\t"
" add $2, %0 \n\t"
MOD_ADD ($1, %5)
"3: \n\t" /* do samples in pairs of 2 */
" movq (%1, %3, 4), %%xmm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */
" movd (%0), %%xmm1 \n\t" /* X | X | p1 | p0 */
+ SWAP_16 (%%xmm1)
VOLUME_32x16 (%%xmm1, %%xmm0)
+ SWAP_16 (%%xmm0)
" movd %%xmm0, (%0) \n\t"
" add $4, %0 \n\t"
MOD_ADD ($2, %5)
" je 6f \n\t"
"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 */
+ SWAP_16 (%%xmm1)
VOLUME_32x16 (%%xmm1, %%xmm0)
+ SWAP_16 (%%xmm0)
" movq %%xmm0, (%0) \n\t"
" add $8, %0 \n\t"
MOD_ADD ($4, %5)
" je 8f \n\t"
"7: \n\t" /* do samples in pairs of 8 */
- " movdqa (%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 (%1, %3, 4), %%xmm0 \n\t" /* v1_h | v1_l | v0_h | v0_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 8(%0), %%xmm3 \n\t" /* X | X | p3 | p2 */
+ SWAP_16_2 (%%xmm1, %%xmm3)
VOLUME_32x16 (%%xmm1, %%xmm0)
VOLUME_32x16 (%%xmm3, %%xmm2)
+ SWAP_16_2 (%%xmm0, %%xmm2)
" movq %%xmm0, (%0) \n\t"
" movq %%xmm2, 8(%0) \n\t"
" add $16, %0 \n\t"
}
#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
pa_volume_float32ne_sse (float *samples, float *volumes, unsigned channels, unsigned length)
{
}
#endif
-#undef RUN_TEST
+#define RUN_TEST
#ifdef RUN_TEST
#define CHANNELS 2
#define SAMPLES 1021
#define TIMES 1000
+#define PADDING 16
static void run_test (void) {
int16_t samples[SAMPLES];
int16_t samples_ref[SAMPLES];
int16_t samples_orig[SAMPLES];
- int32_t volumes[CHANNELS + 16];
+ int32_t volumes[CHANNELS + PADDING];
int i, j, padding;
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++) {
pa_random (samples, sizeof (samples));
memcpy (samples_orig, samples, sizeof (samples));
for (i = 0; i < CHANNELS; i++)
- volumes[i] = rand() >> 15;
-
- for (padding = 0; padding < 16; padding++, i++)
+ volumes[i] = rand() >> 1;
+ for (padding = 0; padding < PADDING; padding++, i++)
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));
for (i = 0; i < SAMPLES; i++) {
printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
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
}
}
}
#endif
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);
}