/*** 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 #include #include "cpu-x86.h" #include "sample-util.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 | */ \ " 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" /* .. | 0 | vh | */ \ " 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 */ \ " cmovae %4, %3 \n\t" /* if (tmp >= 0) 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 int channel_overread_table[8] = {8,8,8,12,8,10,12,14}; static void pa_volume_s16ne_sse2(int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) { pa_reg_x86 channel, temp; /* Channels must be at least 8 and always a multiple of the original number. * This is also the max amount we overread the volume array, which should * have enough padding. */ if (channels < 8) channels = channel_overread_table[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) #if defined (__i386__) : "m" ((pa_reg_x86)channels) #else : "r" ((pa_reg_x86)channels) #endif : "cc" ); } static void pa_volume_s16re_sse2(int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) { pa_reg_x86 channel, temp; /* Channels must be at least 8 and always a multiple of the original number. * This is also the max amount we overread the volume array, which should * have enough padding. */ if (channels < 8) channels = channel_overread_table[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) #if defined (__i386__) : "m" ((pa_reg_x86)channels) #else : "r" ((pa_reg_x86)channels) #endif : "cc" ); } #undef RUN_TEST #ifdef RUN_TEST #define CHANNELS 2 #define SAMPLES 1022 #define TIMES 1000 #define TIMES2 100 #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; pa_usec_t start, stop; int k; pa_usec_t min = INT_MAX, max = 0; double s1 = 0, s2 = 0; func = pa_get_volume_func(PA_SAMPLE_S16NE); printf("checking SSE2 %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] = PA_CLAMP_VOLUME(rand() >> 1); for (padding = 0; padding < PADDING; padding++, i++) volumes[i] = volumes[padding]; func(samples_ref, volumes, CHANNELS, sizeof(samples)); pa_volume_s16ne_sse2(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]); } } for (k = 0; k < TIMES2; k++) { start = pa_rtclock_now(); for (j = 0; j < TIMES; j++) { memcpy(samples, samples_orig, sizeof(samples)); pa_volume_s16ne_sse2(samples, volumes, CHANNELS, sizeof(samples)); } stop = pa_rtclock_now(); if (min > (stop - start)) min = stop - start; if (max < (stop - start)) max = stop - start; s1 += stop - start; s2 += (stop - start) * (stop - start); } pa_log_info("SSE: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1, (long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2); min = INT_MAX; max = 0; s1 = s2 = 0; for (k = 0; k < TIMES2; k++) { start = pa_rtclock_now(); for (j = 0; j < TIMES; j++) { memcpy(samples_ref, samples_orig, sizeof(samples)); func(samples_ref, volumes, CHANNELS, sizeof(samples)); } stop = pa_rtclock_now(); if (min > (stop - start)) min = stop - start; if (max < (stop - start)) max = stop - start; s1 += stop - start; s2 += (stop - start) * (stop - start); } pa_log_info("ref: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1, (long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2); pa_assert_se(memcmp(samples_ref, samples, sizeof(samples)) == 0); } #endif #endif /* defined (__i386__) || defined (__amd64__) */ void pa_volume_func_init_sse(pa_cpu_x86_flag_t flags) { #if defined (__i386__) || defined (__amd64__) #ifdef RUN_TEST run_test(); #endif if (flags & PA_CPU_X86_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__) */ }