SSE-optimized MP3 floating point windowing functions

Originally committed as revision 23750 to svn://svn.ffmpeg.org/ffmpeg/trunk

diff --git a/libavcodec/mpegaudio.h b/libavcodec/mpegaudio.h
index 52d8320..da81ac5 100644 (file)
--- a/libavcodec/mpegaudio.h
+++ b/libavcodec/mpegaudio.h
@@ -183,6 +183,8 @@ void ff_mpa_synth_filter_float(MPADecodeContext *s,
                          OUT_INT *samples, int incr,
                          INTFLOAT sb_samples[SBLIMIT]);
 
+void ff_mpegaudiodec_init_mmx(MPADecodeContext *s);
+
 /* fast header check for resync */
 static inline int ff_mpa_check_header(uint32_t header){
     /* header */

diff --git a/libavcodec/mpegaudiodec.c b/libavcodec/mpegaudiodec.c
index bdf01ea..9ee5b60 100644 (file)
--- a/libavcodec/mpegaudiodec.c
+++ b/libavcodec/mpegaudiodec.c
@@ -112,7 +112,7 @@ static const int32_t scale_factor_mult2[3][3] = {
     SCALE_GEN(4.0 / 9.0), /* 9 steps */
 };
 
-DECLARE_ALIGNED(16, MPA_INT, RENAME(ff_mpa_synth_window))[512];
+DECLARE_ALIGNED(16, MPA_INT, RENAME(ff_mpa_synth_window))[512+256];
 
 /**
  * Convert region offsets to region sizes and truncate
@@ -308,7 +308,9 @@ static av_cold int decode_init(AVCodecContext * avctx)
 
     s->avctx = avctx;
     s->apply_window_mp3 = apply_window_mp3_c;
-
+#if HAVE_MMX
+    ff_mpegaudiodec_init_mmx(s);
+#endif
     avctx->sample_fmt= OUT_FMT;
     s->error_recognition= avctx->error_recognition;
 
@@ -820,7 +822,7 @@ static inline int round_sample(int64_t *sum)
 
 void av_cold RENAME(ff_mpa_synth_init)(MPA_INT *window)
 {
-    int i;
+    int i, j;
 
     /* max = 18760, max sum over all 16 coefs : 44736 */
     for(i=0;i<257;i++) {
@@ -837,6 +839,15 @@ void av_cold RENAME(ff_mpa_synth_init)(MPA_INT *window)
         if (i != 0)
             window[512 - i] = v;
     }
+
+    // Needed for avoiding shuffles in ASM implementations
+    for(i=0; i < 8; i++)
+        for(j=0; j < 16; j++)
+            window[512+16*i+j] = window[64*i+32-j];
+
+    for(i=0; i < 8; i++)
+        for(j=0; j < 16; j++)
+            window[512+128+16*i+j] = window[64*i+48-j];
 }
 
 static void apply_window_mp3_c(MPA_INT *synth_buf, MPA_INT *window,

diff --git a/libavcodec/x86/Makefile b/libavcodec/x86/Makefile
index f4d0c84..5e8b67f 100644 (file)
--- a/libavcodec/x86/Makefile
+++ b/libavcodec/x86/Makefile
@@ -10,6 +10,11 @@ YASM-OBJS-$(CONFIG_GPL)                += x86/h264_deblock_sse2.o       \
                                           x86/h264_idct_sse2.o          \
 
 MMX-OBJS-$(CONFIG_CAVS_DECODER)        += x86/cavsdsp_mmx.o
+MMX-OBJS-$(CONFIG_MP1FLOAT_DECODER)    += x86/mpegaudiodec_mmx.o
+MMX-OBJS-$(CONFIG_MP2FLOAT_DECODER)    += x86/mpegaudiodec_mmx.o
+MMX-OBJS-$(CONFIG_MP3FLOAT_DECODER)    += x86/mpegaudiodec_mmx.o
+MMX-OBJS-$(CONFIG_MP3ON4FLOAT_DECODER) += x86/mpegaudiodec_mmx.o
+MMX-OBJS-$(CONFIG_MP3ADUFLOAT_DECODER) += x86/mpegaudiodec_mmx.o
 MMX-OBJS-$(CONFIG_ENCODERS)            += x86/dsputilenc_mmx.o
 MMX-OBJS-$(CONFIG_GPL)                 += x86/idct_mmx.o
 MMX-OBJS-$(CONFIG_LPC)                 += x86/lpc_mmx.o

diff --git a/libavcodec/x86/mpegaudiodec_mmx.c b/libavcodec/x86/mpegaudiodec_mmx.c
new file mode 100644 (file)
index 0000000..39122b8
--- /dev/null
+++ b/libavcodec/x86/mpegaudiodec_mmx.c
@@ -0,0 +1,156 @@
+/*
+ * MMX optimized MP3 decoding functions
+ * Copyright (c) 2010 Vitor Sessak
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg 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.
+ *
+ * FFmpeg 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/x86_cpu.h"
+
+#define CONFIG_FLOAT 1
+#include "libavcodec/mpegaudio.h"
+
+#define MACS(rt, ra, rb) rt+=(ra)*(rb)
+#define MLSS(rt, ra, rb) rt-=(ra)*(rb)
+
+#define SUM8(op, sum, w, p)               \
+{                                         \
+    op(sum, (w)[0 * 64], (p)[0 * 64]);    \
+    op(sum, (w)[1 * 64], (p)[1 * 64]);    \
+    op(sum, (w)[2 * 64], (p)[2 * 64]);    \
+    op(sum, (w)[3 * 64], (p)[3 * 64]);    \
+    op(sum, (w)[4 * 64], (p)[4 * 64]);    \
+    op(sum, (w)[5 * 64], (p)[5 * 64]);    \
+    op(sum, (w)[6 * 64], (p)[6 * 64]);    \
+    op(sum, (w)[7 * 64], (p)[7 * 64]);    \
+}
+
+static void apply_window(const float *buf, const float *win1,
+                         const float *win2, float *sum1, float *sum2, int len)
+{
+    x86_reg count = - 4*len;
+    const float *win1a = win1+len;
+    const float *win2a = win2+len;
+    const float *bufa  = buf+len;
+    float *sum1a = sum1+len;
+    float *sum2a = sum2+len;
+
+
+#define MULT(a, b)                                 \
+    "movaps " #a "(%0,%5), %%xmm1           \n\t"  \
+    "movaps " #a "(%2,%5), %%xmm2           \n\t"  \
+    "mulps         %%xmm2, %%xmm1           \n\t"  \
+    "subps         %%xmm1, %%xmm0           \n\t"  \
+    "mulps  " #b "(%1,%5), %%xmm2           \n\t"  \
+    "subps         %%xmm2, %%xmm4           \n\t"  \
+
+    __asm__ volatile(
+            "1:                                   \n\t"
+            "xorps       %%xmm0, %%xmm0           \n\t"
+            "xorps       %%xmm4, %%xmm4           \n\t"
+
+            MULT(   0,   0)
+            MULT( 256,  64)
+            MULT( 512, 128)
+            MULT( 768, 192)
+            MULT(1024, 256)
+            MULT(1280, 320)
+            MULT(1536, 384)
+            MULT(1792, 448)
+
+            "movaps      %%xmm0, (%3,%5)          \n\t"
+            "movaps      %%xmm4, (%4,%5)          \n\t"
+            "addl           $16,  %5              \n\t"
+            "jl              1b                   \n\t"
+            :"+&r"(win1a), "+&r"(win2a), "+&r"(bufa), "+&r"(sum1a), "+&r"(sum2a), "+&r"(count)
+            );
+
+#undef MULT
+}
+
+static void apply_window_mp3(float *in, float *win, int *unused, float *out,
+                             int incr)
+{
+    LOCAL_ALIGNED_16(float, suma, [17]);
+    LOCAL_ALIGNED_16(float, sumb, [17]);
+    LOCAL_ALIGNED_16(float, sumc, [17]);
+    LOCAL_ALIGNED_16(float, sumd, [17]);
+
+    float sum;
+
+    /* copy to avoid wrap */
+    memcpy(in + 512, in, 32 * sizeof(*in));
+
+    apply_window(in + 16, win     , win + 512, suma, sumc, 16);
+    apply_window(in + 32, win + 48, win + 640, sumb, sumd, 16);
+
+    SUM8(MACS, suma[0], win + 32, in + 48);
+
+    sumc[ 0] = 0;
+    sumb[16] = 0;
+    sumd[16] = 0;
+
+#define SUMS(suma, sumb, sumc, sumd, out1, out2)               \
+            "movups " #sumd "(%4),       %%xmm0          \n\t" \
+            "shufps         $0x1b,       %%xmm0, %%xmm0  \n\t" \
+            "subps  " #suma "(%1),       %%xmm0          \n\t" \
+            "movaps        %%xmm0," #out1 "(%0)          \n\t" \
+\
+            "movups " #sumc "(%3),       %%xmm0          \n\t" \
+            "shufps         $0x1b,       %%xmm0, %%xmm0  \n\t" \
+            "addps  " #sumb "(%2),       %%xmm0          \n\t" \
+            "movaps        %%xmm0," #out2 "(%0)          \n\t"
+
+    if (incr == 1) {
+        __asm__ volatile(
+            SUMS( 0, 48,  4, 52,  0, 112)
+            SUMS(16, 32, 20, 36, 16,  96)
+            SUMS(32, 16, 36, 20, 32,  80)
+            SUMS(48,  0, 52,  4, 48,  64)
+
+            :"+&r"(out)
+            :"r"(&suma[0]), "r"(&sumb[0]), "r"(&sumc[0]), "r"(&sumd[0])
+            :"memory"
+            );
+        out += 16*incr;
+    } else {
+        int j;
+        float *out2 = out + 32 * incr;
+        out[0  ]  = -suma[   0];
+        out += incr;
+        out2 -= incr;
+        for(j=1;j<16;j++) {
+            *out  = -suma[   j] + sumd[16-j];
+            *out2 =  sumb[16-j] + sumc[   j];
+            out  += incr;
+            out2 -= incr;
+        }
+    }
+
+    sum = 0;
+    SUM8(MLSS, sum, win + 16 + 32, in + 32);
+    *out = sum;
+}
+
+void ff_mpegaudiodec_init_mmx(MPADecodeContext *s)
+{
+    mm_flags = mm_support();
+
+    if (mm_flags & FF_MM_SSE2) {
+        s->apply_window_mp3 = apply_window_mp3;
+    }
+}