From 7086784573e9e6c92d4c34404f18891c2d19872a Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Wed, 12 Aug 2009 20:44:12 +0200 Subject: volume_sse: add sse optimisations --- src/pulsecore/svolume_sse.c | 437 ++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 437 insertions(+) create mode 100644 src/pulsecore/svolume_sse.c (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c new file mode 100644 index 00000000..2d4c541b --- /dev/null +++ b/src/pulsecore/svolume_sse.c @@ -0,0 +1,437 @@ +/*** + This file is part of PulseAudio. + + Copyright 2004-2006 Lennart Poettering + Copyright 2009 Wim Taymans + + PulseAudio is free software; you can redistribute it and/or modify + it under the terms of the GNU Lesser General Public License as published + by the Free Software Foundation; either version 2.1 of the License, + or (at your option) any later version. + + PulseAudio is distributed in the hope that it will be useful, but + WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + General Public License for more details. + + You should have received a copy of the GNU Lesser General Public License + along with PulseAudio; if not, write to the Free Software + Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 + USA. +***/ + +#ifdef HAVE_CONFIG_H +#include +#endif + +#include + +#include +#include +#include +#include + +#include "sample-util.h" +#include "endianmacros.h" + +#if 0 +static void +pa_volume_u8_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsigned length) +{ + unsigned channel; + + for (channel = 0; length; length--) { + int32_t t, hi, lo; + + hi = volumes[channel] >> 16; + lo = volumes[channel] & 0xFFFF; + + t = (int32_t) *samples - 0x80; + t = ((t * lo) >> 16) + (t * hi); + t = PA_CLAMP_UNLIKELY(t, -0x80, 0x7F); + *samples++ = (uint8_t) (t + 0x80); + + if (PA_UNLIKELY(++channel >= channels)) + channel = 0; + } +} + +static void +pa_volume_alaw_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsigned length) +{ + unsigned channel; + + for (channel = 0; length; length--) { + int32_t t, hi, lo; + + hi = volumes[channel] >> 16; + lo = volumes[channel] & 0xFFFF; + + t = (int32_t) st_alaw2linear16(*samples); + t = ((t * lo) >> 16) + (t * hi); + t = PA_CLAMP_UNLIKELY(t, -0x8000, 0x7FFF); + *samples++ = (uint8_t) st_13linear2alaw((int16_t) t >> 3); + + if (PA_UNLIKELY(++channel >= channels)) + channel = 0; + } +} + +static void +pa_volume_ulaw_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsigned length) +{ + unsigned channel; + + for (channel = 0; length; length--) { + int32_t t, hi, lo; + + hi = volumes[channel] >> 16; + lo = volumes[channel] & 0xFFFF; + + t = (int32_t) st_ulaw2linear16(*samples); + t = ((t * lo) >> 16) + (t * hi); + t = PA_CLAMP_UNLIKELY(t, -0x8000, 0x7FFF); + *samples++ = (uint8_t) st_14linear2ulaw((int16_t) t >> 2); + + if (PA_UNLIKELY(++channel >= channels)) + channel = 0; + } +} +#endif + +static void +pa_volume_s16ne_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); + +#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 */ + +#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) */ + + " test $1, %2 \n\t" /* check for odd samples */ + " je 2f \n\t" + + " movd (%1, %3, 4), %%xmm0 \n\t" /* do odd sample */ + " movw (%0), %4 \n\t" + " movd %4, %%xmm1 \n\t" + VOLUME_32x16 (%%xmm1, %%xmm0) + " movd %%xmm0, %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 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 */ + VOLUME_32x16 (%%xmm1, %%xmm0) + " movd %%xmm0, (%0) \n\t" + " 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" /* check for odd samples */ + " 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 */ + " movq (%0), %%xmm1 \n\t" /* X | X | p1 | p0 */ + VOLUME_32x16 (%%xmm1, %%xmm0) + " movq %%xmm0, (%0) \n\t" + " 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 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 */ + " movq (%0), %%xmm1 \n\t" /* X | X | p1 | p0 */ + " movq 8(%0), %%xmm3 \n\t" /* X | X | p3 | p2 */ + VOLUME_32x16 (%%xmm1, %%xmm0) + VOLUME_32x16 (%%xmm3, %%xmm2) + " movq %%xmm0, (%0) \n\t" + " movq %%xmm2, 8(%0) \n\t" + " 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" + ); +} + +#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) +{ + unsigned channel; + + length /= sizeof (float); + + for (channel = 0; length; length--) { + *samples++ *= volumes[channel]; + + if (PA_UNLIKELY(++channel >= channels)) + channel = 0; + } +} + +static void +pa_volume_float32re_sse (float *samples, float *volumes, unsigned channels, unsigned length) +{ + unsigned channel; + + length /= sizeof (float); + + for (channel = 0; length; length--) { + float t; + + t = PA_FLOAT32_SWAP(*samples); + t *= volumes[channel]; + *samples++ = PA_FLOAT32_SWAP(t); + + if (PA_UNLIKELY(++channel >= channels)) + channel = 0; + } +} + +static void +pa_volume_s32ne_sse (int32_t *samples, int32_t *volumes, unsigned channels, unsigned length) +{ + unsigned channel; + + length /= sizeof (int32_t); + + for (channel = 0; length; length--) { + int64_t t; + + t = (int64_t)(*samples); + t = (t * volumes[channel]) >> 16; + t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); + *samples++ = (int32_t) t; + + if (PA_UNLIKELY(++channel >= channels)) + channel = 0; + } +} + +static void +pa_volume_s32re_sse (int32_t *samples, int32_t *volumes, unsigned channels, unsigned length) +{ + unsigned channel; + + length /= sizeof (int32_t); + + for (channel = 0; length; length--) { + int64_t t; + + t = (int64_t) PA_INT32_SWAP(*samples); + t = (t * volumes[channel]) >> 16; + t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); + *samples++ = PA_INT32_SWAP((int32_t) t); + + if (PA_UNLIKELY(++channel >= channels)) + channel = 0; + } +} + +static void +pa_volume_s24ne_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsigned length) +{ + unsigned channel; + uint8_t *e; + + e = samples + length; + + for (channel = 0; samples < e; samples += 3) { + int64_t t; + + t = (int64_t)((int32_t) (PA_READ24NE(samples) << 8)); + t = (t * volumes[channel]) >> 16; + t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); + PA_WRITE24NE(samples, ((uint32_t) (int32_t) t) >> 8); + + if (PA_UNLIKELY(++channel >= channels)) + channel = 0; + } +} + +static void +pa_volume_s24re_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsigned length) +{ + unsigned channel; + uint8_t *e; + + e = samples + length; + + for (channel = 0; samples < e; samples += 3) { + int64_t t; + + t = (int64_t)((int32_t) (PA_READ24RE(samples) << 8)); + t = (t * volumes[channel]) >> 16; + t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); + PA_WRITE24RE(samples, ((uint32_t) (int32_t) t) >> 8); + + if (PA_UNLIKELY(++channel >= channels)) + channel = 0; + } +} + +static void +pa_volume_s24_32ne_sse (uint32_t *samples, int32_t *volumes, unsigned channels, unsigned length) +{ + unsigned channel; + + length /= sizeof (uint32_t); + + for (channel = 0; length; length--) { + int64_t t; + + t = (int64_t) ((int32_t) (*samples << 8)); + t = (t * volumes[channel]) >> 16; + t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); + *samples++ = ((uint32_t) ((int32_t) t)) >> 8; + + if (PA_UNLIKELY(++channel >= channels)) + channel = 0; + } +} + +static void +pa_volume_s24_32re_sse (uint32_t *samples, int32_t *volumes, unsigned channels, unsigned length) +{ + unsigned channel; + + length /= sizeof (uint32_t); + + for (channel = 0; length; length--) { + int64_t t; + + t = (int64_t) ((int32_t) (PA_UINT32_SWAP(*samples) << 8)); + t = (t * volumes[channel]) >> 16; + t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); + *samples++ = PA_UINT32_SWAP(((uint32_t) ((int32_t) t)) >> 8); + + if (PA_UNLIKELY(++channel >= channels)) + channel = 0; + } +} +#endif + +#undef RUN_TEST + +#ifdef RUN_TEST +#define CHANNELS 2 +#define SAMPLES 1021 +#define TIMES 1000 + +static void run_test (void) { + int16_t samples[SAMPLES]; + int16_t samples_ref[SAMPLES]; + int16_t samples_orig[SAMPLES]; + int32_t volumes[CHANNELS + 16]; + int i, j, padding; + pa_do_volume_func_t func; + + func = pa_get_volume_func (PA_SAMPLE_S16NE); + + printf ("checking %d\n", sizeof (samples)); + + for (j = 0; j < TIMES; j++) { + 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] = volumes[padding]; + + pa_volume_s16ne_sse (samples, volumes, CHANNELS, SAMPLES * sizeof (int16_t)); + func (samples_ref, volumes, CHANNELS, SAMPLES * sizeof (int16_t)); + + for (i = 0; i < SAMPLES; i++) { + if (samples[i] != samples_ref[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 + +void pa_volume_func_init_sse (void) { + pa_log_info("Initialising SSE optimized functions."); + +#ifdef RUN_TEST + run_test (); +#endif + + pa_set_volume_func (PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_sse); +} -- cgit From 5998cf99b08d448dd5158ed6229262aa67ea4a66 Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Thu, 13 Aug 2009 13:45:01 +0200 Subject: svolume: improve SSE and MMX code --- src/pulsecore/svolume_sse.c | 191 +++++++++++++++++++++++++++++++------------- 1 file changed, 134 insertions(+), 57 deletions(-) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index 2d4c541b..ff583a06 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -99,6 +99,44 @@ pa_volume_ulaw_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsig } #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) { @@ -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. */ 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" @@ -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 */ " 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) @@ -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 */ " 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) @@ -164,9 +264,11 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi " 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) @@ -177,12 +279,14 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi " 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" @@ -198,29 +302,6 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi } #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) { @@ -378,24 +459,25 @@ pa_volume_s24_32re_sse (uint32_t *samples, int32_t *volumes, unsigned channels, } #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)); @@ -403,12 +485,11 @@ static void run_test (void) { 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++) { @@ -416,11 +497,6 @@ static void run_test (void) { 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 } } } @@ -434,4 +510,5 @@ void pa_volume_func_init_sse (void) { #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); } -- cgit From 563cb2dea9f7f73180e2b8cc8d45b0df9358c936 Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Thu, 13 Aug 2009 17:12:44 +0200 Subject: main: hook up cpu detection code Add CPU detection code to activate the various optimisations. Move some method definitions around. Use compatibility macros when we can. --- src/pulsecore/svolume_sse.c | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index ff583a06..b60471a7 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -31,6 +31,8 @@ #include #include +#include "cpu-x86.h" + #include "sample-util.h" #include "endianmacros.h" @@ -140,8 +142,7 @@ pa_volume_ulaw_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsig static void pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) { - int64_t channel; - int64_t temp; + pa_reg_x86 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. */ @@ -210,7 +211,7 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi "8: \n\t" : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp) - : "r" ((int64_t)channels) + : "r" ((pa_reg_x86)channels) : "cc" ); } @@ -218,8 +219,7 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi static void pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) { - int64_t channel; - int64_t temp; + pa_reg_x86 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. */ @@ -296,7 +296,7 @@ pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi "8: \n\t" : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp) - : "r" ((int64_t)channels) + : "r" ((pa_reg_x86)channels) : "cc" ); } @@ -459,7 +459,7 @@ pa_volume_s24_32re_sse (uint32_t *samples, int32_t *volumes, unsigned channels, } #endif -#define RUN_TEST +#undef RUN_TEST #ifdef RUN_TEST #define CHANNELS 2 @@ -502,7 +502,7 @@ static void run_test (void) { } #endif -void pa_volume_func_init_sse (void) { +void pa_volume_func_init_sse (pa_cpu_x86_flag_t flags) { pa_log_info("Initialising SSE optimized functions."); #ifdef RUN_TEST -- cgit From a1235446a733164f00a96688784913172456a34e Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Fri, 14 Aug 2009 15:19:26 +0200 Subject: volume: make the benchmark more meaningfull MMX is about 6x faster, SSE around 15x on my machine. --- src/pulsecore/svolume_sse.c | 59 +++++++++++++++++++++++++++------------------ 1 file changed, 36 insertions(+), 23 deletions(-) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index b60471a7..8138c6c1 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -24,8 +24,7 @@ #include #endif -#include - +#include #include #include #include @@ -459,7 +458,7 @@ pa_volume_s24_32re_sse (uint32_t *samples, int32_t *volumes, unsigned channels, } #endif -#undef RUN_TEST +#define RUN_TEST #ifdef RUN_TEST #define CHANNELS 2 @@ -474,31 +473,45 @@ static void run_test (void) { int32_t volumes[CHANNELS + PADDING]; int i, j, padding; pa_do_volume_func_t func; + struct timeval start, stop; - func = pa_get_volume_func (PA_SAMPLE_S16RE); + func = pa_get_volume_func (PA_SAMPLE_S16NE); - printf ("checking SSE %d\n", sizeof (samples)); + printf ("checking SSE %zd\n", sizeof (samples)); - for (j = 0; j < TIMES; j++) { - 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() >> 1; - for (padding = 0; padding < PADDING; padding++, i++) - volumes[i] = volumes[padding]; - - 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++) { - if (samples[i] != samples_ref[i]) { - printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i], - samples_orig[i], volumes[i % CHANNELS]); - } + 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() >> 1; + for (padding = 0; padding < PADDING; padding++, i++) + volumes[i] = volumes[padding]; + + func (samples_ref, volumes, CHANNELS, sizeof (samples)); + pa_volume_s16ne_sse (samples, volumes, CHANNELS, sizeof (samples)); + for (i = 0; i < SAMPLES; i++) { + if (samples[i] != samples_ref[i]) { + printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i], + samples_orig[i], volumes[i % CHANNELS]); } } + + pa_gettimeofday(&start); + for (j = 0; j < TIMES; j++) { + memcpy (samples, samples_orig, sizeof (samples)); + pa_volume_s16ne_sse (samples, volumes, CHANNELS, sizeof (samples)); + } + pa_gettimeofday(&stop); + pa_log_info("SSE: %llu usec.", (long long unsigned int)pa_timeval_diff (&stop, &start)); + + pa_gettimeofday(&start); + for (j = 0; j < TIMES; j++) { + memcpy (samples_ref, samples_orig, sizeof (samples)); + func (samples_ref, volumes, CHANNELS, sizeof (samples)); + } + pa_gettimeofday(&stop); + pa_log_info("ref: %llu usec.", (long long unsigned int)pa_timeval_diff (&stop, &start)); } #endif -- cgit From f24c24c14b6614cf19ee916886c8b02384bac435 Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Fri, 14 Aug 2009 15:41:32 +0200 Subject: volume: improved comments --- src/pulsecore/svolume_sse.c | 45 ++++++++++++++++++++++++--------------------- 1 file changed, 24 insertions(+), 21 deletions(-) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index 8138c6c1..d95fa9d9 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -231,12 +231,12 @@ pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi " test $1, %2 \n\t" /* check for odd samples */ " je 2f \n\t" - " movd (%1, %3, 4), %%xmm0 \n\t" /* do odd sample */ - " movw (%0), %4 \n\t" + " movd (%1, %3, 4), %%xmm0 \n\t" /* | v0h | v0l | */ + " movw (%0), %4 \n\t" /* .. | p0 | */ " rorw $8, %4 \n\t" " movd %4, %%xmm1 \n\t" VOLUME_32x16 (%%xmm1, %%xmm0) - " movd %%xmm0, %4 \n\t" + " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */ " rorw $8, %4 \n\t" " movw %4, (%0) \n\t" " add $2, %0 \n\t" @@ -244,31 +244,34 @@ pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi "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 */ + " test $1, %2 \n\t" " je 4f \n\t" - "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 */ + "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 | */ SWAP_16 (%%xmm1) VOLUME_32x16 (%%xmm1, %%xmm0) SWAP_16 (%%xmm0) - " movd %%xmm0, (%0) \n\t" + " 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" /* check for odd samples */ + " test $1, %2 \n\t" " je 6f \n\t" - "5: \n\t" /* do samples in pairs of 4 */ - " movdqu (%1, %3, 4), %%xmm0 \n\t" /* v1_h | v1_l | v0_h | v0_l */ - " movq (%0), %%xmm1 \n\t" /* X | X | p1 | p0 */ + /* FIXME, we can do aligned access of the volume values if we can guarantee + * that the array is 16 bytes aligned, we probably have to do the odd values + * after this then. */ + "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 | */ SWAP_16 (%%xmm1) VOLUME_32x16 (%%xmm1, %%xmm0) SWAP_16 (%%xmm0) - " movq %%xmm0, (%0) \n\t" + " movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */ " add $8, %0 \n\t" MOD_ADD ($4, %5) @@ -277,17 +280,17 @@ pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi " cmp $0, %2 \n\t" " je 8f \n\t" - "7: \n\t" /* do samples in pairs of 8 */ - " 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 */ + "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 | */ 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" + " 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" @@ -458,7 +461,7 @@ pa_volume_s24_32re_sse (uint32_t *samples, int32_t *volumes, unsigned channels, } #endif -#define RUN_TEST +#undef RUN_TEST #ifdef RUN_TEST #define CHANNELS 2 -- cgit From 591baacba5913de32e6556a71a8300d25addbec4 Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Fri, 14 Aug 2009 15:48:10 +0200 Subject: volume: remove ref functions --- src/pulsecore/svolume_sse.c | 223 -------------------------------------------- 1 file changed, 223 deletions(-) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index d95fa9d9..b0a6e0dd 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -35,71 +35,6 @@ #include "sample-util.h" #include "endianmacros.h" -#if 0 -static void -pa_volume_u8_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsigned length) -{ - unsigned channel; - - for (channel = 0; length; length--) { - int32_t t, hi, lo; - - hi = volumes[channel] >> 16; - lo = volumes[channel] & 0xFFFF; - - t = (int32_t) *samples - 0x80; - t = ((t * lo) >> 16) + (t * hi); - t = PA_CLAMP_UNLIKELY(t, -0x80, 0x7F); - *samples++ = (uint8_t) (t + 0x80); - - if (PA_UNLIKELY(++channel >= channels)) - channel = 0; - } -} - -static void -pa_volume_alaw_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsigned length) -{ - unsigned channel; - - for (channel = 0; length; length--) { - int32_t t, hi, lo; - - hi = volumes[channel] >> 16; - lo = volumes[channel] & 0xFFFF; - - t = (int32_t) st_alaw2linear16(*samples); - t = ((t * lo) >> 16) + (t * hi); - t = PA_CLAMP_UNLIKELY(t, -0x8000, 0x7FFF); - *samples++ = (uint8_t) st_13linear2alaw((int16_t) t >> 3); - - if (PA_UNLIKELY(++channel >= channels)) - channel = 0; - } -} - -static void -pa_volume_ulaw_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsigned length) -{ - unsigned channel; - - for (channel = 0; length; length--) { - int32_t t, hi, lo; - - hi = volumes[channel] >> 16; - lo = volumes[channel] & 0xFFFF; - - t = (int32_t) st_ulaw2linear16(*samples); - t = ((t * lo) >> 16) + (t * hi); - t = PA_CLAMP_UNLIKELY(t, -0x8000, 0x7FFF); - *samples++ = (uint8_t) st_14linear2ulaw((int16_t) t >> 2); - - if (PA_UNLIKELY(++channel >= channels)) - channel = 0; - } -} -#endif - #define VOLUME_32x16(s,v) /* .. | vh | vl | */ \ " pxor %%xmm4, %%xmm4 \n\t" /* .. | 0 | 0 | */ \ " punpcklwd %%xmm4, "#s" \n\t" /* .. | 0 | p0 | */ \ @@ -303,164 +238,6 @@ pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi ); } -#if 0 -static void -pa_volume_float32ne_sse (float *samples, float *volumes, unsigned channels, unsigned length) -{ - unsigned channel; - - length /= sizeof (float); - - for (channel = 0; length; length--) { - *samples++ *= volumes[channel]; - - if (PA_UNLIKELY(++channel >= channels)) - channel = 0; - } -} - -static void -pa_volume_float32re_sse (float *samples, float *volumes, unsigned channels, unsigned length) -{ - unsigned channel; - - length /= sizeof (float); - - for (channel = 0; length; length--) { - float t; - - t = PA_FLOAT32_SWAP(*samples); - t *= volumes[channel]; - *samples++ = PA_FLOAT32_SWAP(t); - - if (PA_UNLIKELY(++channel >= channels)) - channel = 0; - } -} - -static void -pa_volume_s32ne_sse (int32_t *samples, int32_t *volumes, unsigned channels, unsigned length) -{ - unsigned channel; - - length /= sizeof (int32_t); - - for (channel = 0; length; length--) { - int64_t t; - - t = (int64_t)(*samples); - t = (t * volumes[channel]) >> 16; - t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); - *samples++ = (int32_t) t; - - if (PA_UNLIKELY(++channel >= channels)) - channel = 0; - } -} - -static void -pa_volume_s32re_sse (int32_t *samples, int32_t *volumes, unsigned channels, unsigned length) -{ - unsigned channel; - - length /= sizeof (int32_t); - - for (channel = 0; length; length--) { - int64_t t; - - t = (int64_t) PA_INT32_SWAP(*samples); - t = (t * volumes[channel]) >> 16; - t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); - *samples++ = PA_INT32_SWAP((int32_t) t); - - if (PA_UNLIKELY(++channel >= channels)) - channel = 0; - } -} - -static void -pa_volume_s24ne_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsigned length) -{ - unsigned channel; - uint8_t *e; - - e = samples + length; - - for (channel = 0; samples < e; samples += 3) { - int64_t t; - - t = (int64_t)((int32_t) (PA_READ24NE(samples) << 8)); - t = (t * volumes[channel]) >> 16; - t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); - PA_WRITE24NE(samples, ((uint32_t) (int32_t) t) >> 8); - - if (PA_UNLIKELY(++channel >= channels)) - channel = 0; - } -} - -static void -pa_volume_s24re_sse (uint8_t *samples, int32_t *volumes, unsigned channels, unsigned length) -{ - unsigned channel; - uint8_t *e; - - e = samples + length; - - for (channel = 0; samples < e; samples += 3) { - int64_t t; - - t = (int64_t)((int32_t) (PA_READ24RE(samples) << 8)); - t = (t * volumes[channel]) >> 16; - t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); - PA_WRITE24RE(samples, ((uint32_t) (int32_t) t) >> 8); - - if (PA_UNLIKELY(++channel >= channels)) - channel = 0; - } -} - -static void -pa_volume_s24_32ne_sse (uint32_t *samples, int32_t *volumes, unsigned channels, unsigned length) -{ - unsigned channel; - - length /= sizeof (uint32_t); - - for (channel = 0; length; length--) { - int64_t t; - - t = (int64_t) ((int32_t) (*samples << 8)); - t = (t * volumes[channel]) >> 16; - t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); - *samples++ = ((uint32_t) ((int32_t) t)) >> 8; - - if (PA_UNLIKELY(++channel >= channels)) - channel = 0; - } -} - -static void -pa_volume_s24_32re_sse (uint32_t *samples, int32_t *volumes, unsigned channels, unsigned length) -{ - unsigned channel; - - length /= sizeof (uint32_t); - - for (channel = 0; length; length--) { - int64_t t; - - t = (int64_t) ((int32_t) (PA_UINT32_SWAP(*samples) << 8)); - t = (t * volumes[channel]) >> 16; - t = PA_CLAMP_UNLIKELY(t, -0x80000000LL, 0x7FFFFFFFLL); - *samples++ = PA_UINT32_SWAP(((uint32_t) ((int32_t) t)) >> 8); - - if (PA_UNLIKELY(++channel >= channels)) - channel = 0; - } -} -#endif - #undef RUN_TEST #ifdef RUN_TEST -- cgit From a98fa950d2f04e2ba4d4a470296a081e1050f76d Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Wed, 19 Aug 2009 15:56:44 +0200 Subject: svolume: remove unneeded compare We don't need the compare because the sub operation already set the right flags for us. --- src/pulsecore/svolume_sse.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index b0a6e0dd..98f828c0 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -52,8 +52,7 @@ " 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 */ + " cmovae %4, %3 \n\t" /* if (tmp >= 0) channel = tmp */ /* swap 16 bits */ #define SWAP_16(s) \ -- cgit From bd49d43bd387758f151c56b7ed1643ecb72c0258 Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Wed, 19 Aug 2009 16:09:48 +0200 Subject: svolume: add CPU guards around code Mark code that should only be compiled on x86 CPUs with proper defines. --- src/pulsecore/svolume_sse.c | 5 +++++ 1 file changed, 5 insertions(+) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index 98f828c0..141c466e 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -35,6 +35,8 @@ #include "sample-util.h" #include "endianmacros.h" +#if defined (__i386__) || defined (__amd64__) + #define VOLUME_32x16(s,v) /* .. | vh | vl | */ \ " pxor %%xmm4, %%xmm4 \n\t" /* .. | 0 | 0 | */ \ " punpcklwd %%xmm4, "#s" \n\t" /* .. | 0 | p0 | */ \ @@ -293,8 +295,10 @@ static void run_test (void) { pa_log_info("ref: %llu usec.", (long long unsigned int)pa_timeval_diff (&stop, &start)); } #endif +#endif /* defined (__i386__) || defined (__amd64__) */ void pa_volume_func_init_sse (pa_cpu_x86_flag_t flags) { +#if defined (__i386__) || defined (__amd64__) pa_log_info("Initialising SSE optimized functions."); #ifdef RUN_TEST @@ -303,4 +307,5 @@ void pa_volume_func_init_sse (pa_cpu_x86_flag_t flags) { 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); +#endif /* defined (__i386__) || defined (__amd64__) */ } -- cgit From 370016c0e73236830513a9ea9c16366c15bd30a2 Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Wed, 19 Aug 2009 18:52:28 +0200 Subject: svolume: fix compilation in 32bits --- src/pulsecore/svolume_sse.c | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index 141c466e..0054d301 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -91,11 +91,11 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi " je 2f \n\t" " movd (%1, %3, 4), %%xmm0 \n\t" /* | v0h | v0l | */ - " movw (%0), %4 \n\t" /* .. | p0 | */ + " movw (%0), %w4 \n\t" /* .. | p0 | */ " movd %4, %%xmm1 \n\t" VOLUME_32x16 (%%xmm1, %%xmm0) " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */ - " movw %4, (%0) \n\t" + " movw %w4, (%0) \n\t" " add $2, %0 \n\t" MOD_ADD ($1, %5) @@ -168,13 +168,13 @@ pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi " je 2f \n\t" " movd (%1, %3, 4), %%xmm0 \n\t" /* | v0h | v0l | */ - " movw (%0), %4 \n\t" /* .. | p0 | */ - " rorw $8, %4 \n\t" + " movw (%0), %w4 \n\t" /* .. | p0 | */ + " rorw $8, %w4 \n\t" " movd %4, %%xmm1 \n\t" VOLUME_32x16 (%%xmm1, %%xmm0) " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */ - " rorw $8, %4 \n\t" - " movw %4, (%0) \n\t" + " rorw $8, %w4 \n\t" + " movw %w4, (%0) \n\t" " add $2, %0 \n\t" MOD_ADD ($1, %5) -- cgit From aeae567f8861d2f068ebd0f054cd9d0aa6a7fe95 Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Wed, 19 Aug 2009 20:00:28 +0200 Subject: svolume: add comment --- src/pulsecore/svolume_sse.c | 3 +++ 1 file changed, 3 insertions(+) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index 0054d301..5979f7c2 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -117,6 +117,9 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi " test $1, %2 \n\t" " je 6f \n\t" + /* FIXME, we can do aligned access of the volume values if we can guarantee + * that the array is 16 bytes aligned, we probably have to do the odd values + * after this then. */ "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 | */ -- cgit From f09b51198f43d79b22cb92b5223d01a7ab339d9f Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Thu, 20 Aug 2009 10:56:20 +0200 Subject: whitespace fixes --- src/pulsecore/svolume_sse.c | 410 ++++++++++++++++++++++---------------------- 1 file changed, 205 insertions(+), 205 deletions(-) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index 5979f7c2..b5e3687f 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -48,7 +48,7 @@ " 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 | */ + " packssdw "#v", "#v" \n\t" /* .. | p1*v1 | p0*v0 | */ #define MOD_ADD(a,b) \ " add "#a", %3 \n\t" /* channel += inc */ \ @@ -77,169 +77,169 @@ static void pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) { - pa_reg_x86 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 (8, channels); - - __asm__ __volatile__ ( - " xor %3, %3 \n\t" - " sar $1, %2 \n\t" /* length /= sizeof (int16_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), %w4 \n\t" /* .. | p0 | */ - " movd %4, %%xmm1 \n\t" - VOLUME_32x16 (%%xmm1, %%xmm0) - " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */ - " movw %w4, (%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" - - /* FIXME, we can do aligned access of the volume values if we can guarantee - * that the array is 16 bytes aligned, we probably have to do the odd values - * after this then. */ - "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" ((pa_reg_x86)channels) - : "cc" - ); + pa_reg_x86 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 (8, channels); + + __asm__ __volatile__ ( + " xor %3, %3 \n\t" + " sar $1, %2 \n\t" /* length /= sizeof (int16_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), %w4 \n\t" /* .. | p0 | */ + " movd %4, %%xmm1 \n\t" + VOLUME_32x16 (%%xmm1, %%xmm0) + " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */ + " movw %w4, (%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" + + /* FIXME, we can do aligned access of the volume values if we can guarantee + * that the array is 16 bytes aligned, we probably have to do the odd values + * after this then. */ + "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" ((pa_reg_x86)channels) + : "cc" + ); } static void pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) { - pa_reg_x86 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 (8, channels); - - __asm__ __volatile__ ( - " xor %3, %3 \n\t" - " sar $1, %2 \n\t" /* length /= sizeof (int16_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), %w4 \n\t" /* .. | p0 | */ - " rorw $8, %w4 \n\t" - " movd %4, %%xmm1 \n\t" - VOLUME_32x16 (%%xmm1, %%xmm0) - " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */ - " rorw $8, %w4 \n\t" - " movw %w4, (%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 | */ - SWAP_16 (%%xmm1) - VOLUME_32x16 (%%xmm1, %%xmm0) - SWAP_16 (%%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" - - /* FIXME, we can do aligned access of the volume values if we can guarantee - * that the array is 16 bytes aligned, we probably have to do the odd values - * after this then. */ - "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 | */ - SWAP_16 (%%xmm1) - VOLUME_32x16 (%%xmm1, %%xmm0) - SWAP_16 (%%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 | */ - SWAP_16_2 (%%xmm1, %%xmm3) - VOLUME_32x16 (%%xmm1, %%xmm0) - VOLUME_32x16 (%%xmm3, %%xmm2) - SWAP_16_2 (%%xmm0, %%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" ((pa_reg_x86)channels) - : "cc" - ); + pa_reg_x86 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 (8, channels); + + __asm__ __volatile__ ( + " xor %3, %3 \n\t" + " sar $1, %2 \n\t" /* length /= sizeof (int16_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), %w4 \n\t" /* .. | p0 | */ + " rorw $8, %w4 \n\t" + " movd %4, %%xmm1 \n\t" + VOLUME_32x16 (%%xmm1, %%xmm0) + " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */ + " rorw $8, %w4 \n\t" + " movw %w4, (%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 | */ + SWAP_16 (%%xmm1) + VOLUME_32x16 (%%xmm1, %%xmm0) + SWAP_16 (%%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" + + /* FIXME, we can do aligned access of the volume values if we can guarantee + * that the array is 16 bytes aligned, we probably have to do the odd values + * after this then. */ + "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 | */ + SWAP_16 (%%xmm1) + VOLUME_32x16 (%%xmm1, %%xmm0) + SWAP_16 (%%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 | */ + SWAP_16_2 (%%xmm1, %%xmm3) + VOLUME_32x16 (%%xmm1, %%xmm0) + VOLUME_32x16 (%%xmm3, %%xmm2) + SWAP_16_2 (%%xmm0, %%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" ((pa_reg_x86)channels) + : "cc" + ); } #undef RUN_TEST @@ -251,64 +251,64 @@ pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi #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 + PADDING]; - int i, j, padding; - pa_do_volume_func_t func; - struct timeval start, stop; - - func = pa_get_volume_func (PA_SAMPLE_S16NE); - - printf ("checking SSE %zd\n", sizeof (samples)); - - 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() >> 1; - for (padding = 0; padding < PADDING; padding++, i++) - volumes[i] = volumes[padding]; - - func (samples_ref, volumes, CHANNELS, sizeof (samples)); - pa_volume_s16ne_sse (samples, volumes, CHANNELS, sizeof (samples)); - for (i = 0; i < SAMPLES; i++) { - if (samples[i] != samples_ref[i]) { - printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i], - samples_orig[i], volumes[i % CHANNELS]); - } - } + int16_t samples[SAMPLES]; + int16_t samples_ref[SAMPLES]; + int16_t samples_orig[SAMPLES]; + int32_t volumes[CHANNELS + PADDING]; + int i, j, padding; + pa_do_volume_func_t func; + struct timeval start, stop; - pa_gettimeofday(&start); - for (j = 0; j < TIMES; j++) { - memcpy (samples, samples_orig, sizeof (samples)); - pa_volume_s16ne_sse (samples, volumes, CHANNELS, sizeof (samples)); - } - pa_gettimeofday(&stop); - pa_log_info("SSE: %llu usec.", (long long unsigned int)pa_timeval_diff (&stop, &start)); + func = pa_get_volume_func (PA_SAMPLE_S16NE); + + printf ("checking SSE %zd\n", sizeof (samples)); + + 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() >> 1; + for (padding = 0; padding < PADDING; padding++, i++) + volumes[i] = volumes[padding]; - pa_gettimeofday(&start); - for (j = 0; j < TIMES; j++) { - memcpy (samples_ref, samples_orig, sizeof (samples)); func (samples_ref, volumes, CHANNELS, sizeof (samples)); - } - pa_gettimeofday(&stop); - pa_log_info("ref: %llu usec.", (long long unsigned int)pa_timeval_diff (&stop, &start)); + pa_volume_s16ne_sse (samples, volumes, CHANNELS, sizeof (samples)); + for (i = 0; i < SAMPLES; i++) { + if (samples[i] != samples_ref[i]) { + printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i], + samples_orig[i], volumes[i % CHANNELS]); + } + } + + pa_gettimeofday(&start); + for (j = 0; j < TIMES; j++) { + memcpy (samples, samples_orig, sizeof (samples)); + pa_volume_s16ne_sse (samples, volumes, CHANNELS, sizeof (samples)); + } + pa_gettimeofday(&stop); + pa_log_info("SSE: %llu usec.", (long long unsigned int)pa_timeval_diff (&stop, &start)); + + pa_gettimeofday(&start); + for (j = 0; j < TIMES; j++) { + memcpy (samples_ref, samples_orig, sizeof (samples)); + func (samples_ref, volumes, CHANNELS, sizeof (samples)); + } + pa_gettimeofday(&stop); + pa_log_info("ref: %llu usec.", (long long unsigned int)pa_timeval_diff (&stop, &start)); } #endif #endif /* defined (__i386__) || defined (__amd64__) */ void pa_volume_func_init_sse (pa_cpu_x86_flag_t flags) { #if defined (__i386__) || defined (__amd64__) - pa_log_info("Initialising SSE optimized functions."); + pa_log_info("Initialising SSE optimized functions."); #ifdef RUN_TEST - run_test (); + run_test (); #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); + 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); #endif /* defined (__i386__) || defined (__amd64__) */ } -- cgit From f8ffe0dabcedf56437c00feb895d7d7229971ba0 Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Thu, 20 Aug 2009 12:30:48 +0200 Subject: svolume: cleanups Use PA_MAX Use pa_rtclock_now() for benchmarks --- src/pulsecore/svolume_sse.c | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index b5e3687f..54af4a57 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -81,7 +81,7 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi /* 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); + channels = PA_MAX (8U, channels); __asm__ __volatile__ ( " xor %3, %3 \n\t" @@ -161,7 +161,7 @@ pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi /* 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); + channels = PA_MAX (8U, channels); __asm__ __volatile__ ( " xor %3, %3 \n\t" @@ -257,7 +257,7 @@ static void run_test (void) { int32_t volumes[CHANNELS + PADDING]; int i, j, padding; pa_do_volume_func_t func; - struct timeval start, stop; + pa_usec_t start, stop; func = pa_get_volume_func (PA_SAMPLE_S16NE); @@ -281,21 +281,21 @@ static void run_test (void) { } } - pa_gettimeofday(&start); + start = pa_rtclock_now(); for (j = 0; j < TIMES; j++) { memcpy (samples, samples_orig, sizeof (samples)); pa_volume_s16ne_sse (samples, volumes, CHANNELS, sizeof (samples)); } - pa_gettimeofday(&stop); - pa_log_info("SSE: %llu usec.", (long long unsigned int)pa_timeval_diff (&stop, &start)); + stop = pa_rtclock_now(); + pa_log_info("SSE: %llu usec.", (long long unsigned int)(stop - start)); - pa_gettimeofday(&start); + start = pa_rtclock_now(); for (j = 0; j < TIMES; j++) { memcpy (samples_ref, samples_orig, sizeof (samples)); func (samples_ref, volumes, CHANNELS, sizeof (samples)); } - pa_gettimeofday(&stop); - pa_log_info("ref: %llu usec.", (long long unsigned int)pa_timeval_diff (&stop, &start)); + stop = pa_rtclock_now(); + pa_log_info("ref: %llu usec.", (long long unsigned int)(stop - start)); } #endif #endif /* defined (__i386__) || defined (__amd64__) */ -- cgit From b5ac3839e18524524fa3e0da7ec68dbce16e8203 Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Mon, 7 Sep 2009 17:21:21 +0200 Subject: x86: only install some functions when SSE2 Remap and volume functions use SSE2 instructions so only install them when SSE2 is present. --- src/pulsecore/svolume_sse.c | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index 54af4a57..d9dcf476 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -75,7 +75,7 @@ " por %%xmm5, "#s2" \n\t" static void -pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) +pa_volume_s16ne_sse2 (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) { pa_reg_x86 channel, temp; @@ -155,7 +155,7 @@ pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsi } static void -pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) +pa_volume_s16re_sse2 (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) { pa_reg_x86 channel, temp; @@ -308,7 +308,9 @@ void pa_volume_func_init_sse (pa_cpu_x86_flag_t flags) { run_test (); #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); + if (flags & PA_CPU_X86_SSE2) { + pa_set_volume_func (PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_sse2); + pa_set_volume_func (PA_SAMPLE_S16RE, (pa_do_volume_func_t) pa_volume_s16re_sse2); + } #endif /* defined (__i386__) || defined (__amd64__) */ } -- cgit From 71e066c873e5bd31bd446ac0f8d0e97cc0b12ace Mon Sep 17 00:00:00 2001 From: Lennart Poettering Date: Wed, 9 Sep 2009 04:28:22 +0200 Subject: simd: be more precise which SIMD optimizations we activate --- src/pulsecore/svolume_sse.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) (limited to 'src/pulsecore/svolume_sse.c') diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c index d9dcf476..bbd73a9b 100644 --- a/src/pulsecore/svolume_sse.c +++ b/src/pulsecore/svolume_sse.c @@ -302,15 +302,16 @@ static void run_test (void) { void pa_volume_func_init_sse (pa_cpu_x86_flag_t flags) { #if defined (__i386__) || defined (__amd64__) - pa_log_info("Initialising SSE optimized functions."); #ifdef RUN_TEST run_test (); #endif if (flags & PA_CPU_X86_SSE2) { - pa_set_volume_func (PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_sse2); - pa_set_volume_func (PA_SAMPLE_S16RE, (pa_do_volume_func_t) pa_volume_s16re_sse2); + pa_log_info("Initialising SSE2 optimized functions."); + + pa_set_volume_func (PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_sse2); + pa_set_volume_func (PA_SAMPLE_S16RE, (pa_do_volume_func_t) pa_volume_s16re_sse2); } #endif /* defined (__i386__) || defined (__amd64__) */ } -- cgit