This file is part of PulseAudio.
Copyright 2004-2006 Lennart Poettering
Copyright 2009 Wim Taymans <wim.taymans@collabora.co.uk>
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, see <http://www.gnu.org/licenses/>.
***/
#ifdef HAVE_CONFIG_H
#include <config.h>
#endif
#include <pulse/rtclock.h>
#include <pulsecore/random.h>
#include <pulsecore/macro.h>
#include <pulsecore/endianmacros.h>
#include "cpu-x86.h"
#include "sample-util.h"
#if (!defined(__APPLE__) && !defined(__FreeBSD__) && !defined(__FreeBSD_kernel__) && defined (__i386__)) || defined (__amd64__)
* in v: 2 int32_t volumes, fixed point 16:16
* out s: contains scaled and clamped int16_t samples.
*
* We calculate the high 32 bits of a 32x16 multiply which we then
* clamp to 16 bits. The calculation is:
*
* vl = (v & 0xffff)
* vh = (v >> 16)
* s = ((s * vl) >> 16) + (s * vh);
*
* For the first multiply we have to do a sign correction as we need to
* multiply a signed int with an unsigned int. Hacker's delight 8-3 gives a
* simple formula to correct the sign of the high word after the signed
* multiply.
*/
#define VOLUME_32x16(s,v) \
" pxor %%mm4, %%mm4 \n\t" \
" punpcklwd %%mm4, "#s" \n\t" \
" pcmpgtw "#v", %%mm4 \n\t" \
" pand "#s", %%mm4 \n\t" \
" movq "#s", %%mm5 \n\t" \
" pmulhw "#v", "#s" \n\t" \
" paddw %%mm4, "#s" \n\t" \
" pslld $16, "#s" \n\t" \
" psrld $16, "#v" \n\t" \
" psrad $16, "#s" \n\t" \
" pmaddwd %%mm5, "#v" \n\t" \
" paddd "#s", "#v" \n\t" \
" packssdw "#v", "#v" \n\t"
* a <= b. */
#define MOD_ADD(a,b) \
" add "#a", %3 \n\t" \
" mov %3, %4 \n\t" \
" sub "#b", %4 \n\t" \
" cmovae %4, %3 \n\t"
#define SWAP_16(s) \
" movq "#s", %%mm4 \n\t" \
" psrlw $8, %%mm4 \n\t" \
" psllw $8, "#s" \n\t" \
" por %%mm4, "#s" \n\t"
#define SWAP_16_2(s1,s2) \
" movq "#s1", %%mm4 \n\t" \
" movq "#s2", %%mm5 \n\t" \
" psrlw $8, %%mm4 \n\t" \
" psrlw $8, %%mm5 \n\t" \
" psllw $8, "#s1" \n\t" \
" psllw $8, "#s2" \n\t" \
" por %%mm4, "#s1" \n\t" \
" por %%mm5, "#s2" \n\t"
static void pa_volume_s16ne_mmx(int16_t *samples, const int32_t *volumes, unsigned channels, unsigned length) {
pa_reg_x86 channel, temp;
* This is also the max amount we overread the volume array, which should
* have enough padding. */
channels = channels == 3 ? 6 : PA_MAX (4U, channels);
__asm__ __volatile__ (
" xor %3, %3 \n\t"
" sar $1, %2 \n\t"
" test $1, %2 \n\t"
" je 2f \n\t"
" movd (%q1, %3, 4), %%mm0 \n\t"
" movw (%0), %w4 \n\t"
" movd %4, %%mm1 \n\t"
VOLUME_32x16 (%%mm1, %%mm0)
" movd %%mm0, %4 \n\t"
" movw %w4, (%0) \n\t"
" add $2, %0 \n\t"
MOD_ADD ($1, %5)
"2: \n\t"
" sar $1, %2 \n\t"
" test $1, %2 \n\t"
" je 4f \n\t"
"3: \n\t"
" movq (%q1, %3, 4), %%mm0 \n\t"
" movd (%0), %%mm1 \n\t"
VOLUME_32x16 (%%mm1, %%mm0)
" movd %%mm0, (%0) \n\t"
" add $4, %0 \n\t"
MOD_ADD ($2, %5)
"4: \n\t"
" sar $1, %2 \n\t"
" cmp $0, %2 \n\t"
" je 6f \n\t"
"5: \n\t"
" movq (%q1, %3, 4), %%mm0 \n\t"
" movq 8(%q1, %3, 4), %%mm2 \n\t"
" movd (%0), %%mm1 \n\t"
" movd 4(%0), %%mm3 \n\t"
VOLUME_32x16 (%%mm1, %%mm0)
VOLUME_32x16 (%%mm3, %%mm2)
" movd %%mm0, (%0) \n\t"
" movd %%mm2, 4(%0) \n\t"
" 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" (channel), "=&r" (temp)
#if defined (__i386__)
: "m" (channels)
#else
: "r" ((pa_reg_x86)channels)
#endif
: "cc"
);
}
static void pa_volume_s16re_mmx(int16_t *samples, const int32_t *volumes, unsigned channels, unsigned length) {
pa_reg_x86 channel, temp;
* This is also the max amount we overread the volume array, which should
* have enough padding. */
channels = channels == 3 ? 6 : PA_MAX (4U, channels);
__asm__ __volatile__ (
" xor %3, %3 \n\t"
" sar $1, %2 \n\t"
" pcmpeqw %%mm6, %%mm6 \n\t"
" pcmpeqw %%mm7, %%mm7 \n\t"
" pslld $16, %%mm6 \n\t"
" psrld $31, %%mm7 \n\t"
" test $1, %2 \n\t"
" je 2f \n\t"
" movd (%q1, %3, 4), %%mm0 \n\t"
" movw (%0), %w4 \n\t"
" rorw $8, %w4 \n\t"
" movd %4, %%mm1 \n\t"
VOLUME_32x16 (%%mm1, %%mm0)
" movd %%mm0, %4 \n\t"
" 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"
" test $1, %2 \n\t"
" je 4f \n\t"
"3: \n\t"
" movq (%q1, %3, 4), %%mm0 \n\t"
" movd (%0), %%mm1 \n\t"
SWAP_16 (%%mm1)
VOLUME_32x16 (%%mm1, %%mm0)
SWAP_16 (%%mm0)
" movd %%mm0, (%0) \n\t"
" add $4, %0 \n\t"
MOD_ADD ($2, %5)
"4: \n\t"
" sar $1, %2 \n\t"
" cmp $0, %2 \n\t"
" je 6f \n\t"
"5: \n\t"
" movq (%q1, %3, 4), %%mm0 \n\t"
" movq 8(%q1, %3, 4), %%mm2 \n\t"
" movd (%0), %%mm1 \n\t"
" movd 4(%0), %%mm3 \n\t"
SWAP_16_2 (%%mm1, %%mm3)
VOLUME_32x16 (%%mm1, %%mm0)
VOLUME_32x16 (%%mm3, %%mm2)
SWAP_16_2 (%%mm0, %%mm2)
" movd %%mm0, (%0) \n\t"
" movd %%mm2, 4(%0) \n\t"
" 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" (channel), "=&r" (temp)
#if defined (__i386__)
: "m" (channels)
#else
: "r" ((pa_reg_x86)channels)
#endif
: "cc"
);
}
#endif
void pa_volume_func_init_mmx(pa_cpu_x86_flag_t flags) {
#if (!defined(__APPLE__) && !defined(__FreeBSD__) && !defined(__FreeBSD_kernel__) && defined (__i386__)) || defined (__amd64__)
if ((flags & PA_CPU_X86_MMX) && (flags & PA_CPU_X86_CMOV)) {
pa_log_info("Initialising MMX optimized volume functions.");
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
}