floatdsp: move vector_fmul_reverse from dsputil to avfloatdsp.

Now, nellymoserenc and aacenc no longer depends on dsputil. Independent
of this patch, wmaprodec also does not depend on dsputil, so I removed
it from there also.

32 files changed, 231 insertions, 310 deletions

diff --git a/libavcodec/aacdec.c b/libavcodec/aacdec.c
index d59dea4f0e..0c4e356d8f 100644
--- a/libavcodec/aacdec.c
+++ b/libavcodec/aacdec.c
@@ -2067,9 +2067,9 @@ static void windowing_and_mdct_ltp(AACContext *ac, float *out,
         ac->fdsp.vector_fmul(in + 448, in + 448, swindow_prev, 128);
     }
     if (ics->window_sequence[0] != LONG_START_SEQUENCE) {
-        ac->dsp.vector_fmul_reverse(in + 1024, in + 1024, lwindow, 1024);
+        ac->fdsp.vector_fmul_reverse(in + 1024, in + 1024, lwindow, 1024);
     } else {
-        ac->dsp.vector_fmul_reverse(in + 1024 + 448, in + 1024 + 448, swindow, 128);
+        ac->fdsp.vector_fmul_reverse(in + 1024 + 448, in + 1024 + 448, swindow, 128);
         memset(in + 1024 + 576, 0, 448 * sizeof(float));
     }
     ac->mdct_ltp.mdct_calc(&ac->mdct_ltp, out, in);
@@ -2122,17 +2122,17 @@ static void update_ltp(AACContext *ac, SingleChannelElement *sce)
     if (ics->window_sequence[0] == EIGHT_SHORT_SEQUENCE) {
         memcpy(saved_ltp,       saved, 512 * sizeof(float));
         memset(saved_ltp + 576, 0,     448 * sizeof(float));
-        ac->dsp.vector_fmul_reverse(saved_ltp + 448, ac->buf_mdct + 960,     &swindow[64],      64);
+        ac->fdsp.vector_fmul_reverse(saved_ltp + 448, ac->buf_mdct + 960,     &swindow[64],      64);
         for (i = 0; i < 64; i++)
             saved_ltp[i + 512] = ac->buf_mdct[1023 - i] * swindow[63 - i];
     } else if (ics->window_sequence[0] == LONG_START_SEQUENCE) {
         memcpy(saved_ltp,       ac->buf_mdct + 512, 448 * sizeof(float));
         memset(saved_ltp + 576, 0,                  448 * sizeof(float));
-        ac->dsp.vector_fmul_reverse(saved_ltp + 448, ac->buf_mdct + 960,     &swindow[64],      64);
+        ac->fdsp.vector_fmul_reverse(saved_ltp + 448, ac->buf_mdct + 960,     &swindow[64],      64);
         for (i = 0; i < 64; i++)
             saved_ltp[i + 512] = ac->buf_mdct[1023 - i] * swindow[63 - i];
     } else { // LONG_STOP or ONLY_LONG
-        ac->dsp.vector_fmul_reverse(saved_ltp,       ac->buf_mdct + 512,     &lwindow[512],     512);
+        ac->fdsp.vector_fmul_reverse(saved_ltp,       ac->buf_mdct + 512,     &lwindow[512],     512);
         for (i = 0; i < 512; i++)
             saved_ltp[i + 512] = ac->buf_mdct[1023 - i] * lwindow[511 - i];
     }
diff --git a/libavcodec/aacenc.c b/libavcodec/aacenc.c
index 6f582ca8aa..00a6d0397d 100644
--- a/libavcodec/aacenc.c
+++ b/libavcodec/aacenc.c
@@ -183,7 +183,7 @@ static void put_audio_specific_config(AVCodecContext *avctx)
 }
 
 #define WINDOW_FUNC(type) \
-static void apply_ ##type ##_window(DSPContext *dsp, AVFloatDSPContext *fdsp, \
+static void apply_ ##type ##_window(AVFloatDSPContext *fdsp, \
                                     SingleChannelElement *sce, \
                                     const float *audio)
 
@@ -193,8 +193,8 @@ WINDOW_FUNC(only_long)
     const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
     float *out = sce->ret_buf;
 
-    fdsp->vector_fmul       (out,        audio,        lwindow, 1024);
-    dsp->vector_fmul_reverse(out + 1024, audio + 1024, pwindow, 1024);
+    fdsp->vector_fmul        (out,        audio,        lwindow, 1024);
+    fdsp->vector_fmul_reverse(out + 1024, audio + 1024, pwindow, 1024);
 }
 
 WINDOW_FUNC(long_start)
@@ -205,7 +205,7 @@ WINDOW_FUNC(long_start)
 
     fdsp->vector_fmul(out, audio, lwindow, 1024);
     memcpy(out + 1024, audio + 1024, sizeof(out[0]) * 448);
-    dsp->vector_fmul_reverse(out + 1024 + 448, audio + 1024 + 448, swindow, 128);
+    fdsp->vector_fmul_reverse(out + 1024 + 448, audio + 1024 + 448, swindow, 128);
     memset(out + 1024 + 576, 0, sizeof(out[0]) * 448);
 }
 
@@ -218,7 +218,7 @@ WINDOW_FUNC(long_stop)
     memset(out, 0, sizeof(out[0]) * 448);
     fdsp->vector_fmul(out + 448, audio + 448, swindow, 128);
     memcpy(out + 576, audio + 576, sizeof(out[0]) * 448);
-    dsp->vector_fmul_reverse(out + 1024, audio + 1024, lwindow, 1024);
+    fdsp->vector_fmul_reverse(out + 1024, audio + 1024, lwindow, 1024);
 }
 
 WINDOW_FUNC(eight_short)
@@ -230,15 +230,15 @@ WINDOW_FUNC(eight_short)
     int w;
 
     for (w = 0; w < 8; w++) {
-        fdsp->vector_fmul       (out, in, w ? pwindow : swindow, 128);
+        fdsp->vector_fmul        (out, in, w ? pwindow : swindow, 128);
         out += 128;
         in  += 128;
-        dsp->vector_fmul_reverse(out, in, swindow, 128);
+        fdsp->vector_fmul_reverse(out, in, swindow, 128);
         out += 128;
     }
 }
 
-static void (*const apply_window[4])(DSPContext *dsp, AVFloatDSPContext *fdsp,
+static void (*const apply_window[4])(AVFloatDSPContext *fdsp,
                                      SingleChannelElement *sce,
                                      const float *audio) = {
     [ONLY_LONG_SEQUENCE]   = apply_only_long_window,
@@ -253,7 +253,7 @@ static void apply_window_and_mdct(AACEncContext *s, SingleChannelElement *sce,
     int i;
     float *output = sce->ret_buf;
 
-    apply_window[sce->ics.window_sequence[0]](&s->dsp, &s->fdsp, sce, audio);
+    apply_window[sce->ics.window_sequence[0]](&s->fdsp, sce, audio);
 
     if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE)
         s->mdct1024.mdct_calc(&s->mdct1024, sce->coeffs, output);
@@ -694,7 +694,6 @@ static av_cold int dsp_init(AVCodecContext *avctx, AACEncContext *s)
 {
     int ret = 0;
 
-    ff_dsputil_init(&s->dsp, avctx);
     avpriv_float_dsp_init(&s->fdsp, avctx->flags & CODEC_FLAG_BITEXACT);
 
     // window init
diff --git a/libavcodec/aacenc.h b/libavcodec/aacenc.h
index 81ffb973d6..cc2eb7717d 100644
--- a/libavcodec/aacenc.h
+++ b/libavcodec/aacenc.h
@@ -58,7 +58,6 @@ typedef struct AACEncContext {
     PutBitContext pb;
     FFTContext mdct1024;                         ///< long (1024 samples) frame transform context
     FFTContext mdct128;                          ///< short (128 samples) frame transform context
-    DSPContext  dsp;
     AVFloatDSPContext fdsp;
     float *planar_samples[6];                    ///< saved preprocessed input
 
diff --git a/libavcodec/aacsbr.c b/libavcodec/aacsbr.c
index 0b96abb95b..8b290340d4 100644
--- a/libavcodec/aacsbr.c
+++ b/libavcodec/aacsbr.c
@@ -1150,7 +1150,7 @@ static void sbr_dequant(SpectralBandReplication *sbr, int id_aac)
  * @param   x       pointer to the beginning of the first sample window
  * @param   W       array of complex-valued samples split into subbands
  */
-static void sbr_qmf_analysis(DSPContext *dsp, FFTContext *mdct,
+static void sbr_qmf_analysis(AVFloatDSPContext *dsp, FFTContext *mdct,
                              SBRDSPContext *sbrdsp, const float *in, float *x,
                              float z[320], float W[2][32][32][2], int buf_idx)
 {
@@ -1663,7 +1663,7 @@ void ff_sbr_apply(AACContext *ac, SpectralBandReplication *sbr, int id_aac,
     }
     for (ch = 0; ch < nch; ch++) {
         /* decode channel */
-        sbr_qmf_analysis(&ac->dsp, &sbr->mdct_ana, &sbr->dsp, ch ? R : L, sbr->data[ch].analysis_filterbank_samples,
+        sbr_qmf_analysis(&ac->fdsp, &sbr->mdct_ana, &sbr->dsp, ch ? R : L, sbr->data[ch].analysis_filterbank_samples,
                          (float*)sbr->qmf_filter_scratch,
                          sbr->data[ch].W, sbr->data[ch].Ypos);
         sbr_lf_gen(ac, sbr, sbr->X_low, sbr->data[ch].W, sbr->data[ch].Ypos);
diff --git a/libavcodec/arm/Makefile b/libavcodec/arm/Makefile
index 71048f9c4c..5ebda62cf6 100644
--- a/libavcodec/arm/Makefile
+++ b/libavcodec/arm/Makefile
@@ -56,9 +56,6 @@ ARMV6-OBJS                             += arm/dsputil_init_armv6.o      \
 
 VFP-OBJS-$(HAVE_ARMV6)                 += arm/fmtconvert_vfp.o
 
-VFP-OBJS                               += arm/dsputil_vfp.o             \
-                                          arm/dsputil_init_vfp.o        \
-
 NEON-OBJS-$(CONFIG_FFT)                += arm/fft_neon.o                \
                                           arm/fft_fixed_neon.o          \
 
diff --git a/libavcodec/arm/dsputil_init_arm.c b/libavcodec/arm/dsputil_init_arm.c
index 0c1563df93..9feda140ff 100644
--- a/libavcodec/arm/dsputil_init_arm.c
+++ b/libavcodec/arm/dsputil_init_arm.c
@@ -121,6 +121,5 @@ void ff_dsputil_init_arm(DSPContext* c, AVCodecContext *avctx)
 
     if (have_armv5te(cpu_flags)) ff_dsputil_init_armv5te(c, avctx);
     if (have_armv6(cpu_flags))   ff_dsputil_init_armv6(c, avctx);
-    if (have_vfp(cpu_flags))     ff_dsputil_init_vfp(c, avctx);
     if (have_neon(cpu_flags))    ff_dsputil_init_neon(c, avctx);
 }
diff --git a/libavcodec/arm/dsputil_init_neon.c b/libavcodec/arm/dsputil_init_neon.c
index 0d23b26826..bb0c4af69a 100644
--- a/libavcodec/arm/dsputil_init_neon.c
+++ b/libavcodec/arm/dsputil_init_neon.c
@@ -144,8 +144,6 @@ void ff_avg_h264_chroma_mc2_neon(uint8_t *, uint8_t *, int, int, int, int);
 
 void ff_butterflies_float_neon(float *v1, float *v2, int len);
 float ff_scalarproduct_float_neon(const float *v1, const float *v2, int len);
-void ff_vector_fmul_reverse_neon(float *dst, const float *src0,
-                                 const float *src1, int len);
 
 void ff_vector_clipf_neon(float *dst, const float *src, float min, float max,
                           int len);
@@ -298,7 +296,6 @@ void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx)
 
     c->butterflies_float          = ff_butterflies_float_neon;
     c->scalarproduct_float        = ff_scalarproduct_float_neon;
-    c->vector_fmul_reverse        = ff_vector_fmul_reverse_neon;
     c->vector_clipf               = ff_vector_clipf_neon;
     c->vector_clip_int32          = ff_vector_clip_int32_neon;
 
diff --git a/libavcodec/arm/dsputil_init_vfp.c b/libavcodec/arm/dsputil_init_vfp.c
deleted file mode 100644
index d77d686578..0000000000
--- a/libavcodec/arm/dsputil_init_vfp.c
+++ /dev/null
@@ -1,30 +0,0 @@
-/*
- * Copyright (c) 2008 Siarhei Siamashka <ssvb@users.sourceforge.net>
- *
- * This file is part of Libav.
- *
- * Libav 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.
- *
- * Libav 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 Libav; if not, write to the Free Software
- * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
- */
-
-#include "libavcodec/dsputil.h"
-#include "dsputil_arm.h"
-
-void ff_vector_fmul_reverse_vfp(float *dst, const float *src0,
-                                const float *src1, int len);
-
-void ff_dsputil_init_vfp(DSPContext* c, AVCodecContext *avctx)
-{
-    c->vector_fmul_reverse = ff_vector_fmul_reverse_vfp;
-}
diff --git a/libavcodec/arm/dsputil_neon.S b/libavcodec/arm/dsputil_neon.S
index 5e512a7a21..4ceecbcb4e 100644
--- a/libavcodec/arm/dsputil_neon.S
+++ b/libavcodec/arm/dsputil_neon.S
@@ -556,30 +556,6 @@ NOVFP   vmov.32         r0,  d0[0]
         bx              lr
 endfunc
 
-function ff_vector_fmul_reverse_neon, export=1
-        add             r2,  r2,  r3,  lsl #2
-        sub             r2,  r2,  #32
-        mov             r12, #-32
-        vld1.32         {q0-q1},  [r1,:128]!
-        vld1.32         {q2-q3},  [r2,:128], r12
-1:      pld             [r1, #32]
-        vrev64.32       q3,  q3
-        vmul.f32        d16, d0,  d7
-        vmul.f32        d17, d1,  d6
-        pld             [r2, #-32]
-        vrev64.32       q2,  q2
-        vmul.f32        d18, d2,  d5
-        vmul.f32        d19, d3,  d4
-        subs            r3,  r3,  #8
-        beq             2f
-        vld1.32         {q0-q1},  [r1,:128]!
-        vld1.32         {q2-q3},  [r2,:128], r12
-        vst1.32         {q8-q9},  [r0,:128]!
-        b               1b
-2:      vst1.32         {q8-q9},  [r0,:128]!
-        bx              lr
-endfunc
-
 function ff_vector_clipf_neon, export=1
 VFP     vdup.32         q1,  d0[1]
 VFP     vdup.32         q0,  d0[0]
diff --git a/libavcodec/arm/dsputil_vfp.S b/libavcodec/arm/dsputil_vfp.S
deleted file mode 100644
index 9df955dbf9..0000000000
--- a/libavcodec/arm/dsputil_vfp.S
+++ /dev/null
@@ -1,106 +0,0 @@
-/*
- * Copyright (c) 2008 Siarhei Siamashka <ssvb@users.sourceforge.net>
- *
- * This file is part of Libav.
- *
- * Libav 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.
- *
- * Libav 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 Libav; if not, write to the Free Software
- * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
- */
-
-#include "config.h"
-#include "libavutil/arm/asm.S"
-
-/*
- * VFP is a floating point coprocessor used in some ARM cores. VFP11 has 1 cycle
- * throughput for almost all the instructions (except for double precision
- * arithmetics), but rather high latency. Latency is 4 cycles for loads and 8 cycles
- * for arithmetic operations. Scheduling code to avoid pipeline stalls is very
- * important for performance. One more interesting feature is that VFP has
- * independent load/store and arithmetics pipelines, so it is possible to make
- * them work simultaneously and get more than 1 operation per cycle. Load/store
- * pipeline can process 2 single precision floating point values per cycle and
- * supports bulk loads and stores for large sets of registers. Arithmetic operations
- * can be done on vectors, which allows to keep the arithmetics pipeline busy,
- * while the processor may issue and execute other instructions. Detailed
- * optimization manuals can be found at http://www.arm.com
- */
-
-/**
- * ARM VFP optimized implementation of 'vector_fmul_reverse_c' function.
- * Assume that len is a positive number and is multiple of 8
- */
-@ void ff_vector_fmul_reverse_vfp(float *dst, const float *src0,
-@                                 const float *src1, int len)
-function ff_vector_fmul_reverse_vfp, export=1
-        vpush           {d8-d15}
-        add             r2,  r2,  r3, lsl #2
-        vldmdb          r2!, {s0-s3}
-        vldmia          r1!, {s8-s11}
-        vldmdb          r2!, {s4-s7}
-        vldmia          r1!, {s12-s15}
-        vmul.f32        s8,  s3,  s8
-        vmul.f32        s9,  s2,  s9
-        vmul.f32        s10, s1,  s10
-        vmul.f32        s11, s0,  s11
-1:
-        subs            r3,  r3,  #16
-        it              ge
-        vldmdbge        r2!, {s16-s19}
-        vmul.f32        s12, s7,  s12
-        it              ge
-        vldmiage        r1!, {s24-s27}
-        vmul.f32        s13, s6,  s13
-        it              ge
-        vldmdbge        r2!, {s20-s23}
-        vmul.f32        s14, s5,  s14
-        it              ge
-        vldmiage        r1!, {s28-s31}
-        vmul.f32        s15, s4,  s15
-        it              ge
-        vmulge.f32      s24, s19, s24
-        it              gt
-        vldmdbgt        r2!, {s0-s3}
-        it              ge
-        vmulge.f32      s25, s18, s25
-        vstmia          r0!, {s8-s13}
-        it              ge
-        vmulge.f32      s26, s17, s26
-        it              gt
-        vldmiagt        r1!, {s8-s11}
-        itt             ge
-        vmulge.f32      s27, s16, s27
-        vmulge.f32      s28, s23, s28
-        it              gt
-        vldmdbgt        r2!, {s4-s7}
-        it              ge
-        vmulge.f32      s29, s22, s29
-        vstmia          r0!, {s14-s15}
-        ittt            ge
-        vmulge.f32      s30, s21, s30
-        vmulge.f32      s31, s20, s31
-        vmulge.f32      s8,  s3,  s8
-        it              gt
-        vldmiagt        r1!, {s12-s15}
-        itttt           ge
-        vmulge.f32      s9,  s2,  s9
-        vmulge.f32      s10, s1,  s10
-        vstmiage        r0!, {s24-s27}
-        vmulge.f32      s11, s0,  s11
-        it              ge
-        vstmiage        r0!, {s28-s31}
-        bgt             1b
-
-        vpop            {d8-d15}
-        bx              lr
-endfunc
diff --git a/libavcodec/dsputil.c b/libavcodec/dsputil.c
index 4cbb48b0cd..0590e0c525 100644
--- a/libavcodec/dsputil.c
+++ b/libavcodec/dsputil.c
@@ -2353,13 +2353,6 @@ WRAPPER8_16_SQ(quant_psnr8x8_c, quant_psnr16_c)
 WRAPPER8_16_SQ(rd8x8_c, rd16_c)
 WRAPPER8_16_SQ(bit8x8_c, bit16_c)
 
-static void vector_fmul_reverse_c(float *dst, const float *src0, const float *src1, int len){
-    int i;
-    src1 += len-1;
-    for(i=0; i<len; i++)
-        dst[i] = src0[i] * src1[-i];
-}
-
 static void butterflies_float_c(float *restrict v1, float *restrict v2,
                                 int len)
 {
@@ -2707,7 +2700,6 @@ av_cold void ff_dsputil_init(DSPContext* c, AVCodecContext *avctx)
     c->try_8x8basis= try_8x8basis_c;
     c->add_8x8basis= add_8x8basis_c;
 
-    c->vector_fmul_reverse = vector_fmul_reverse_c;
     c->vector_clipf = vector_clipf_c;
     c->scalarproduct_int16 = scalarproduct_int16_c;
     c->scalarproduct_and_madd_int16 = scalarproduct_and_madd_int16_c;
diff --git a/libavcodec/dsputil.h b/libavcodec/dsputil.h
index b49ea1e81d..c197041071 100644
--- a/libavcodec/dsputil.h
+++ b/libavcodec/dsputil.h
@@ -340,8 +340,6 @@ typedef struct DSPContext {
 
     void (*h261_loop_filter)(uint8_t *src, int stride);
 
-    /* assume len is a multiple of 16, and arrays are 32-byte aligned */
-    void (*vector_fmul_reverse)(float *dst, const float *src0, const float *src1, int len);
     /* assume len is a multiple of 8, and arrays are 16-byte aligned */
     void (*vector_clipf)(float *dst /* align 16 */, const float *src /* align 16 */, float min, float max, int len /* align 16 */);
     /**
diff --git a/libavcodec/nellymoserenc.c b/libavcodec/nellymoserenc.c
index 122282623c..34a15e5a50 100644
--- a/libavcodec/nellymoserenc.c
+++ b/libavcodec/nellymoserenc.c
@@ -55,7 +55,6 @@
 typedef struct NellyMoserEncodeContext {
     AVCodecContext  *avctx;
     int             last_frame;
-    DSPContext      dsp;
     AVFloatDSPContext fdsp;
     FFTContext      mdct_ctx;
     AudioFrameQueue afq;
@@ -122,12 +121,12 @@ static void apply_mdct(NellyMoserEncodeContext *s)
     float *in1 = s->buf + NELLY_BUF_LEN;
     float *in2 = s->buf + 2 * NELLY_BUF_LEN;
 
-    s->fdsp.vector_fmul       (s->in_buff,                 in0, ff_sine_128, NELLY_BUF_LEN);
-    s->dsp.vector_fmul_reverse(s->in_buff + NELLY_BUF_LEN, in1, ff_sine_128, NELLY_BUF_LEN);
+    s->fdsp.vector_fmul        (s->in_buff,                 in0, ff_sine_128, NELLY_BUF_LEN);
+    s->fdsp.vector_fmul_reverse(s->in_buff + NELLY_BUF_LEN, in1, ff_sine_128, NELLY_BUF_LEN);
     s->mdct_ctx.mdct_calc(&s->mdct_ctx, s->mdct_out, s->in_buff);
 
-    s->fdsp.vector_fmul       (s->in_buff,                 in1, ff_sine_128, NELLY_BUF_LEN);
-    s->dsp.vector_fmul_reverse(s->in_buff + NELLY_BUF_LEN, in2, ff_sine_128, NELLY_BUF_LEN);
+    s->fdsp.vector_fmul        (s->in_buff,                 in1, ff_sine_128, NELLY_BUF_LEN);
+    s->fdsp.vector_fmul_reverse(s->in_buff + NELLY_BUF_LEN, in2, ff_sine_128, NELLY_BUF_LEN);
     s->mdct_ctx.mdct_calc(&s->mdct_ctx, s->mdct_out + NELLY_BUF_LEN, s->in_buff);
 }
 
@@ -173,7 +172,6 @@ static av_cold int encode_init(AVCodecContext *avctx)
     s->avctx = avctx;
     if ((ret = ff_mdct_init(&s->mdct_ctx, 8, 0, 32768.0)) < 0)
         goto error;
-    ff_dsputil_init(&s->dsp, avctx);
     avpriv_float_dsp_init(&s->fdsp, avctx->flags & CODEC_FLAG_BITEXACT);
 
     /* Generate overlap window */
diff --git a/libavcodec/ppc/Makefile b/libavcodec/ppc/Makefile
index e1ebf26d0f..cdc6435be1 100644
--- a/libavcodec/ppc/Makefile
+++ b/libavcodec/ppc/Makefile
@@ -15,7 +15,6 @@ ALTIVEC-OBJS-$(CONFIG_VP8_DECODER)     += ppc/vp8dsp_altivec.o
 
 ALTIVEC-OBJS                           += ppc/dsputil_altivec.o         \
                                           ppc/fdct_altivec.o            \
-                                          ppc/float_altivec.o           \
                                           ppc/fmtconvert_altivec.o      \
                                           ppc/gmc_altivec.o             \
                                           ppc/idct_altivec.o            \
diff --git a/libavcodec/ppc/dsputil_ppc.c b/libavcodec/ppc/dsputil_ppc.c
index a3069402da..24dbb8c10e 100644
--- a/libavcodec/ppc/dsputil_ppc.c
+++ b/libavcodec/ppc/dsputil_ppc.c
@@ -160,7 +160,6 @@ void ff_dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx)
 
     if (av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC) {
         ff_dsputil_init_altivec(c, avctx);
-        ff_float_init_altivec(c, avctx);
         ff_int_init_altivec(c, avctx);
         c->gmc1 = ff_gmc1_altivec;
 
diff --git a/libavcodec/ppc/float_altivec.c b/libavcodec/ppc/float_altivec.c
deleted file mode 100644
index e336cbed5f..0000000000
--- a/libavcodec/ppc/float_altivec.c
+++ /dev/null
@@ -1,57 +0,0 @@
-/*
- * Copyright (c) 2006 Luca Barbato <lu_zero@gentoo.org>
- *
- * This file is part of Libav.
- *
- * Libav 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.
- *
- * Libav 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 Libav; if not, write to the Free Software
- * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
- */
-
-#include "libavutil/ppc/util_altivec.h"
-#include "libavcodec/dsputil.h"
-
-#include "dsputil_altivec.h"
-
-static void vector_fmul_reverse_altivec(float *dst, const float *src0,
-                                        const float *src1, int len)
-{
-    int i;
-    vector float d, s0, s1, h0, l0,
-                 s2, s3, zero = (vector float)vec_splat_u32(0);
-    src1 += len-4;
-    for(i=0; i<len-7; i+=8) {
-        s1 = vec_ld(0, src1-i);              // [a,b,c,d]
-        s0 = vec_ld(0, src0+i);
-        l0 = vec_mergel(s1, s1);             // [c,c,d,d]
-        s3 = vec_ld(-16, src1-i);
-        h0 = vec_mergeh(s1, s1);             // [a,a,b,b]
-        s2 = vec_ld(16, src0+i);
-        s1 = vec_mergeh(vec_mergel(l0,h0),   // [d,b,d,b]
-                        vec_mergeh(l0,h0));  // [c,a,c,a]
-                                             // [d,c,b,a]
-        l0 = vec_mergel(s3, s3);
-        d = vec_madd(s0, s1, zero);
-        h0 = vec_mergeh(s3, s3);
-        vec_st(d, 0, dst+i);
-        s3 = vec_mergeh(vec_mergel(l0,h0),
-                        vec_mergeh(l0,h0));
-        d = vec_madd(s2, s3, zero);
-        vec_st(d, 16, dst+i);
-    }
-}
-
-void ff_float_init_altivec(DSPContext* c, AVCodecContext *avctx)
-{
-    c->vector_fmul_reverse = vector_fmul_reverse_altivec;
-}
diff --git a/libavcodec/wmadec.c b/libavcodec/wmadec.c
index 3da09c59b1..7d5ca1b606 100644
--- a/libavcodec/wmadec.c
+++ b/libavcodec/wmadec.c
@@ -401,7 +401,7 @@ static void wma_window(WMACodecContext *s, float *out)
         block_len = s->block_len;
         bsize = s->frame_len_bits - s->block_len_bits;
 
-        s->dsp.vector_fmul_reverse(out, in, s->windows[bsize], block_len);
+        s->fdsp.vector_fmul_reverse(out, in, s->windows[bsize], block_len);
 
     } else {
         block_len = 1 << s->next_block_len_bits;
@@ -410,7 +410,7 @@ static void wma_window(WMACodecContext *s, float *out)
 
         memcpy(out, in, n*sizeof(float));
 
-        s->dsp.vector_fmul_reverse(out+n, in+n, s->windows[bsize], block_len);
+        s->fdsp.vector_fmul_reverse(out+n, in+n, s->windows[bsize], block_len);
 
         memset(out+n+block_len, 0, n*sizeof(float));
     }
diff --git a/libavcodec/wmaenc.c b/libavcodec/wmaenc.c
index 044114b516..bf8c2674b9 100644
--- a/libavcodec/wmaenc.c
+++ b/libavcodec/wmaenc.c
@@ -112,7 +112,7 @@ static void apply_window_and_mdct(AVCodecContext * avctx, const AVFrame *frame)
     for (ch = 0; ch < avctx->channels; ch++) {
         memcpy(s->output, s->frame_out[ch], window_len * sizeof(*s->output));
         s->fdsp.vector_fmul_scalar(s->frame_out[ch], audio[ch], n, len);
-        s->dsp.vector_fmul_reverse(&s->output[window_len], s->frame_out[ch], win, len);
+        s->fdsp.vector_fmul_reverse(&s->output[window_len], s->frame_out[ch], win, len);
         s->fdsp.vector_fmul(s->frame_out[ch], s->frame_out[ch], win, len);
         mdct->mdct_calc(mdct, s->coefs[ch], s->output);
     }
diff --git a/libavcodec/wmaprodec.c b/libavcodec/wmaprodec.c
index f04b43fd61..b54484ea9d 100644
--- a/libavcodec/wmaprodec.c
+++ b/libavcodec/wmaprodec.c
@@ -170,7 +170,6 @@ typedef struct WMAProDecodeCtx {
     /* generic decoder variables */
     AVCodecContext*  avctx;                         ///< codec context for av_log
     AVFrame          frame;                         ///< AVFrame for decoded output
-    DSPContext       dsp;                           ///< accelerated DSP functions
     AVFloatDSPContext fdsp;
     uint8_t          frame_data[MAX_FRAMESIZE +
                       FF_INPUT_BUFFER_PADDING_SIZE];///< compressed frame data
@@ -281,7 +280,6 @@ static av_cold int decode_init(AVCodecContext *avctx)
     int num_possible_block_sizes;
 
     s->avctx = avctx;
-    ff_dsputil_init(&s->dsp, avctx);
     avpriv_float_dsp_init(&s->fdsp, avctx->flags & CODEC_FLAG_BITEXACT);
 
     init_put_bits(&s->pb, s->frame_data, MAX_FRAMESIZE);
diff --git a/libavcodec/x86/dsputil.asm b/libavcodec/x86/dsputil.asm
index 26b9dee5db..27e77d565d 100644
--- a/libavcodec/x86/dsputil.asm
+++ b/libavcodec/x86/dsputil.asm
@@ -567,43 +567,6 @@ VECTOR_CLIP_INT32 11, 1, 1, 0
 VECTOR_CLIP_INT32 6, 1, 0, 0
 %endif
 
-;-----------------------------------------------------------------------------
-; void vector_fmul_reverse(float *dst, const float *src0, const float *src1,
-;                          int len)
-;-----------------------------------------------------------------------------
-%macro VECTOR_FMUL_REVERSE 0
-cglobal vector_fmul_reverse, 4,4,2, dst, src0, src1, len
-    lea       lenq, [lend*4 - 2*mmsize]
-ALIGN 16
-.loop:
-%if cpuflag(avx)
-    vmovaps     xmm0, [src1q + 16]
-    vinsertf128 m0, m0, [src1q], 1
-    vshufps     m0, m0, m0, q0123
-    vmovaps     xmm1, [src1q + mmsize + 16]
-    vinsertf128 m1, m1, [src1q + mmsize], 1
-    vshufps     m1, m1, m1, q0123
-%else
-    mova    m0, [src1q]
-    mova    m1, [src1q + mmsize]
-    shufps  m0, m0, q0123
-    shufps  m1, m1, q0123
-%endif
-    mulps   m0, m0, [src0q + lenq + mmsize]
-    mulps   m1, m1, [src0q + lenq]
-    mova    [dstq + lenq + mmsize], m0
-    mova    [dstq + lenq], m1
-    add     src1q, 2*mmsize
-    sub     lenq,  2*mmsize
-    jge     .loop
-    REP_RET
-%endmacro
-
-INIT_XMM sse
-VECTOR_FMUL_REVERSE
-INIT_YMM avx
-VECTOR_FMUL_REVERSE
-
 ; %1 = aligned/unaligned
 %macro BSWAP_LOOPS  1
     mov      r3, r2
diff --git a/libavcodec/x86/dsputil_mmx.c b/libavcodec/x86/dsputil_mmx.c
index 4ed48d64a4..503764817a 100644
--- a/libavcodec/x86/dsputil_mmx.c
+++ b/libavcodec/x86/dsputil_mmx.c
@@ -1848,11 +1848,6 @@ int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
 
 float ff_scalarproduct_float_sse(const float *v1, const float *v2, int order);
 
-void ff_vector_fmul_reverse_sse(float *dst, const float *src0,
-                                const float *src1, int len);
-void ff_vector_fmul_reverse_avx(float *dst, const float *src0,
-                                const float *src1, int len);
-
 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
                                    int32_t min, int32_t max, unsigned int len);
 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
@@ -2135,8 +2130,6 @@ static void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx, int mm_flags)
 #endif /* HAVE_INLINE_ASM */
 
 #if HAVE_YASM
-    c->vector_fmul_reverse = ff_vector_fmul_reverse_sse;
-
     c->scalarproduct_float          = ff_scalarproduct_float_sse;
 #endif /* HAVE_YASM */
 }
@@ -2288,7 +2281,6 @@ static void dsputil_init_avx(DSPContext *c, AVCodecContext *avctx, int mm_flags)
             c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_10_avx;
         }
     }
-    c->vector_fmul_reverse = ff_vector_fmul_reverse_avx;
 #endif /* HAVE_AVX_EXTERNAL */
 }
 
diff --git a/libavutil/arm/float_dsp_init_neon.c b/libavutil/arm/float_dsp_init_neon.c
index 41e513fcdc..c6f02bd2c5 100644
--- a/libavutil/arm/float_dsp_init_neon.c
+++ b/libavutil/arm/float_dsp_init_neon.c
@@ -38,6 +38,9 @@ void ff_vector_fmul_window_neon(float *dst, const float *src0,
 void ff_vector_fmul_add_neon(float *dst, const float *src0, const float *src1,
                              const float *src2, int len);
 
+void ff_vector_fmul_reverse_neon(float *dst, const float *src0,
+                                 const float *src1, int len);
+
 void ff_float_dsp_init_neon(AVFloatDSPContext *fdsp)
 {
     fdsp->vector_fmul = ff_vector_fmul_neon;
@@ -45,4 +48,5 @@ void ff_float_dsp_init_neon(AVFloatDSPContext *fdsp)
     fdsp->vector_fmul_scalar = ff_vector_fmul_scalar_neon;
     fdsp->vector_fmul_window = ff_vector_fmul_window_neon;
     fdsp->vector_fmul_add    = ff_vector_fmul_add_neon;
+    fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_neon;
 }
diff --git a/libavutil/arm/float_dsp_init_vfp.c b/libavutil/arm/float_dsp_init_vfp.c
index ef808d8a5b..7247d56762 100644
--- a/libavutil/arm/float_dsp_init_vfp.c
+++ b/libavutil/arm/float_dsp_init_vfp.c
@@ -25,10 +25,14 @@
 void ff_vector_fmul_vfp(float *dst, const float *src0, const float *src1,
                         int len);
 
+void ff_vector_fmul_reverse_vfp(float *dst, const float *src0,
+                                const float *src1, int len);
+
 void ff_float_dsp_init_vfp(AVFloatDSPContext *fdsp)
 {
     int cpu_flags = av_get_cpu_flags();
 
     if (!have_vfpv3(cpu_flags))
         fdsp->vector_fmul = ff_vector_fmul_vfp;
+    fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_vfp;
 }
diff --git a/libavutil/arm/float_dsp_neon.S b/libavutil/arm/float_dsp_neon.S
index 100eb02455..d00e59de8f 100644
--- a/libavutil/arm/float_dsp_neon.S
+++ b/libavutil/arm/float_dsp_neon.S
@@ -220,3 +220,27 @@ function ff_vector_fmul_add_neon, export=1
 2:      vst1.32         {q12-q13},[r0,:128]!
         bx              lr
 endfunc
+
+function ff_vector_fmul_reverse_neon, export=1
+        add             r2,  r2,  r3,  lsl #2
+        sub             r2,  r2,  #32
+        mov             r12, #-32
+        vld1.32         {q0-q1},  [r1,:128]!
+        vld1.32         {q2-q3},  [r2,:128], r12
+1:      pld             [r1, #32]
+        vrev64.32       q3,  q3
+        vmul.f32        d16, d0,  d7
+        vmul.f32        d17, d1,  d6
+        pld             [r2, #-32]
+        vrev64.32       q2,  q2
+        vmul.f32        d18, d2,  d5
+        vmul.f32        d19, d3,  d4
+        subs            r3,  r3,  #8
+        beq             2f
+        vld1.32         {q0-q1},  [r1,:128]!
+        vld1.32         {q2-q3},  [r2,:128], r12
+        vst1.32         {q8-q9},  [r0,:128]!
+        b               1b
+2:      vst1.32         {q8-q9},  [r0,:128]!
+        bx              lr
+endfunc
diff --git a/libavutil/arm/float_dsp_vfp.S b/libavutil/arm/float_dsp_vfp.S
index 3931828381..82952807de 100644
--- a/libavutil/arm/float_dsp_vfp.S
+++ b/libavutil/arm/float_dsp_vfp.S
@@ -66,3 +66,72 @@ function ff_vector_fmul_vfp, export=1
         vpop            {d8-d15}
         bx              lr
 endfunc
+
+/**
+ * ARM VFP optimized implementation of 'vector_fmul_reverse_c' function.
+ * Assume that len is a positive number and is multiple of 8
+ */
+@ void ff_vector_fmul_reverse_vfp(float *dst, const float *src0,
+@                                 const float *src1, int len)
+function ff_vector_fmul_reverse_vfp, export=1
+        vpush           {d8-d15}
+        add             r2,  r2,  r3, lsl #2
+        vldmdb          r2!, {s0-s3}
+        vldmia          r1!, {s8-s11}
+        vldmdb          r2!, {s4-s7}
+        vldmia          r1!, {s12-s15}
+        vmul.f32        s8,  s3,  s8
+        vmul.f32        s9,  s2,  s9
+        vmul.f32        s10, s1,  s10
+        vmul.f32        s11, s0,  s11
+1:
+        subs            r3,  r3,  #16
+        it              ge
+        vldmdbge        r2!, {s16-s19}
+        vmul.f32        s12, s7,  s12
+        it              ge
+        vldmiage        r1!, {s24-s27}
+        vmul.f32        s13, s6,  s13
+        it              ge
+        vldmdbge        r2!, {s20-s23}
+        vmul.f32        s14, s5,  s14
+        it              ge
+        vldmiage        r1!, {s28-s31}
+        vmul.f32        s15, s4,  s15
+        it              ge
+        vmulge.f32      s24, s19, s24
+        it              gt
+        vldmdbgt        r2!, {s0-s3}
+        it              ge
+        vmulge.f32      s25, s18, s25
+        vstmia          r0!, {s8-s13}
+        it              ge
+        vmulge.f32      s26, s17, s26
+        it              gt
+        vldmiagt        r1!, {s8-s11}
+        itt             ge
+        vmulge.f32      s27, s16, s27
+        vmulge.f32      s28, s23, s28
+        it              gt
+        vldmdbgt        r2!, {s4-s7}
+        it              ge
+        vmulge.f32      s29, s22, s29
+        vstmia          r0!, {s14-s15}
+        ittt            ge
+        vmulge.f32      s30, s21, s30
+        vmulge.f32      s31, s20, s31
+        vmulge.f32      s8,  s3,  s8
+        it              gt
+        vldmiagt        r1!, {s12-s15}
+        itttt           ge
+        vmulge.f32      s9,  s2,  s9
+        vmulge.f32      s10, s1,  s10
+        vstmiage        r0!, {s24-s27}
+        vmulge.f32      s11, s0,  s11
+        it              ge
+        vstmiage        r0!, {s28-s31}
+        bgt             1b
+
+        vpop            {d8-d15}
+        bx              lr
+endfunc
diff --git a/libavutil/float_dsp.c b/libavutil/float_dsp.c
index 456f83b656..aa5293710b 100644
--- a/libavutil/float_dsp.c
+++ b/libavutil/float_dsp.c
@@ -79,6 +79,16 @@ static void vector_fmul_add_c(float *dst, const float *src0, const float *src1,
         dst[i] = src0[i] * src1[i] + src2[i];
 }
 
+static void vector_fmul_reverse_c(float *dst, const float *src0,
+                                  const float *src1, int len)
+{
+    int i;
+
+    src1 += len-1;
+    for (i = 0; i < len; i++)
+        dst[i] = src0[i] * src1[-i];
+}
+
 void avpriv_float_dsp_init(AVFloatDSPContext *fdsp, int bit_exact)
 {
     fdsp->vector_fmul = vector_fmul_c;
@@ -87,6 +97,7 @@ void avpriv_float_dsp_init(AVFloatDSPContext *fdsp, int bit_exact)
     fdsp->vector_dmul_scalar = vector_dmul_scalar_c;
     fdsp->vector_fmul_window = vector_fmul_window_c;
     fdsp->vector_fmul_add = vector_fmul_add_c;
+    fdsp->vector_fmul_reverse = vector_fmul_reverse_c;
 
 #if ARCH_ARM
     ff_float_dsp_init_arm(fdsp);
diff --git a/libavutil/float_dsp.h b/libavutil/float_dsp.h
index b45c4560ff..96fcdab1e6 100644
--- a/libavutil/float_dsp.h
+++ b/libavutil/float_dsp.h
@@ -118,6 +118,25 @@ typedef struct AVFloatDSPContext {
      */
     void (*vector_fmul_add)(float *dst, const float *src0, const float *src1,
                             const float *src2, int len);
+
+    /**
+     * Calculate the product of two vectors of floats, and store the result
+     * in a vector of floats. The second vector of floats is iterated over
+     * in reverse order.
+     *
+     * @param dst  output vector
+     *             constraints: 32-byte aligned
+     * @param src0 first input vector
+     *             constraints: 32-byte aligned
+     * @param src1 second input vector
+     *             constraints: 32-byte aligned
+     * @param src1 third input vector
+     *             constraints: 32-byte aligned
+     * @param len  number of elements in the input
+     *             constraints: multiple of 16
+     */
+    void (*vector_fmul_reverse)(float *dst, const float *src0,
+                                const float *src1, int len);
 } AVFloatDSPContext;
 
 /**
diff --git a/libavutil/ppc/float_dsp_altivec.c b/libavutil/ppc/float_dsp_altivec.c
index 87379e92f8..fee4e7c7db 100644
--- a/libavutil/ppc/float_dsp_altivec.c
+++ b/libavutil/ppc/float_dsp_altivec.c
@@ -93,3 +93,32 @@ void ff_vector_fmul_add_altivec(float *dst, const float *src0,
         vec_st(t0, 0, dst + i);
     }
 }
+
+void ff_vector_fmul_reverse_altivec(float *dst, const float *src0,
+                                    const float *src1, int len)
+{
+    int i;
+    vector float d, s0, s1, h0, l0,
+                s2, s3, zero = (vector float) vec_splat_u32(0);
+
+    src1 += len-4;
+    for(i = 0; i < len - 7; i += 8) {
+        s1 = vec_ld(0, src1 - i);              // [a,b,c,d]
+        s0 = vec_ld(0, src0 + i);
+        l0 = vec_mergel(s1, s1);               // [c,c,d,d]
+        s3 = vec_ld(-16, src1 - i);
+        h0 = vec_mergeh(s1, s1);               // [a,a,b,b]
+        s2 = vec_ld(16, src0 + i);
+        s1 = vec_mergeh(vec_mergel(l0, h0),    // [d,b,d,b]
+                        vec_mergeh(l0, h0));   // [c,a,c,a]
+        // [d,c,b,a]
+        l0 = vec_mergel(s3, s3);
+        d = vec_madd(s0, s1, zero);
+        h0 = vec_mergeh(s3, s3);
+        vec_st(d, 0, dst + i);
+        s3 = vec_mergeh(vec_mergel(l0, h0),
+                        vec_mergeh(l0, h0));
+        d = vec_madd(s2, s3, zero);
+        vec_st(d, 16, dst + i);
+    }
+}
diff --git a/libavutil/ppc/float_dsp_altivec.h b/libavutil/ppc/float_dsp_altivec.h
index 38a9e15434..b484a630e8 100644
--- a/libavutil/ppc/float_dsp_altivec.h
+++ b/libavutil/ppc/float_dsp_altivec.h
@@ -32,4 +32,7 @@ extern void ff_vector_fmul_add_altivec(float *dst, const float *src0,
                                        const float *src1, const float *src2,
                                        int len);
 
+extern void ff_vector_fmul_reverse_altivec(float *dst, const float *src0,
+                                           const float *src1, int len);
+
 #endif /* AVUTIL_PPC_FLOAT_DSP_ALTIVEC_H */
diff --git a/libavutil/ppc/float_dsp_init.c b/libavutil/ppc/float_dsp_init.c
index f3d8a42f53..bd682f5a30 100644
--- a/libavutil/ppc/float_dsp_init.c
+++ b/libavutil/ppc/float_dsp_init.c
@@ -33,6 +33,7 @@ void ff_float_dsp_init_ppc(AVFloatDSPContext *fdsp, int bit_exact)
 
     fdsp->vector_fmul = ff_vector_fmul_altivec;
     fdsp->vector_fmul_add = ff_vector_fmul_add_altivec;
+    fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_altivec;
 
     if (!bit_exact) {
         fdsp->vector_fmul_window = ff_vector_fmul_window_altivec;
diff --git a/libavutil/x86/float_dsp.asm b/libavutil/x86/float_dsp.asm
index 70fc1d0310..126f3495c4 100644
--- a/libavutil/x86/float_dsp.asm
+++ b/libavutil/x86/float_dsp.asm
@@ -190,3 +190,40 @@ INIT_XMM sse
 VECTOR_FMUL_ADD
 INIT_YMM avx
 VECTOR_FMUL_ADD
+
+;-----------------------------------------------------------------------------
+; void vector_fmul_reverse(float *dst, const float *src0, const float *src1,
+;                          int len)
+;-----------------------------------------------------------------------------
+%macro VECTOR_FMUL_REVERSE 0
+cglobal vector_fmul_reverse, 4,4,2, dst, src0, src1, len
+    lea       lenq, [lend*4 - 2*mmsize]
+ALIGN 16
+.loop:
+%if cpuflag(avx)
+    vmovaps     xmm0, [src1q + 16]
+    vinsertf128 m0, m0, [src1q], 1
+    vshufps     m0, m0, m0, q0123
+    vmovaps     xmm1, [src1q + mmsize + 16]
+    vinsertf128 m1, m1, [src1q + mmsize], 1
+    vshufps     m1, m1, m1, q0123
+%else
+    mova    m0, [src1q]
+    mova    m1, [src1q + mmsize]
+    shufps  m0, m0, q0123
+    shufps  m1, m1, q0123
+%endif
+    mulps   m0, m0, [src0q + lenq + mmsize]
+    mulps   m1, m1, [src0q + lenq]
+    mova    [dstq + lenq + mmsize], m0
+    mova    [dstq + lenq], m1
+    add     src1q, 2*mmsize
+    sub     lenq,  2*mmsize
+    jge     .loop
+    REP_RET
+%endmacro
+
+INIT_XMM sse
+VECTOR_FMUL_REVERSE
+INIT_YMM avx
+VECTOR_FMUL_REVERSE
diff --git a/libavutil/x86/float_dsp_init.c b/libavutil/x86/float_dsp_init.c
index 0a2ad5ba0e..9f63b4c057 100644
--- a/libavutil/x86/float_dsp_init.c
+++ b/libavutil/x86/float_dsp_init.c
@@ -46,6 +46,11 @@ void ff_vector_fmul_add_sse(float *dst, const float *src0, const float *src1,
 void ff_vector_fmul_add_avx(float *dst, const float *src0, const float *src1,
                             const float *src2, int len);
 
+void ff_vector_fmul_reverse_sse(float *dst, const float *src0,
+                                const float *src1, int len);
+void ff_vector_fmul_reverse_avx(float *dst, const float *src0,
+                                const float *src1, int len);
+
 #if HAVE_6REGS && HAVE_INLINE_ASM
 static void vector_fmul_window_3dnowext(float *dst, const float *src0,
                                         const float *src1, const float *win,
@@ -129,6 +134,7 @@ void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
         fdsp->vector_fmac_scalar = ff_vector_fmac_scalar_sse;
         fdsp->vector_fmul_scalar = ff_vector_fmul_scalar_sse;
         fdsp->vector_fmul_add    = ff_vector_fmul_add_sse;
+        fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_sse;
     }
     if (EXTERNAL_SSE2(mm_flags)) {
         fdsp->vector_dmul_scalar = ff_vector_dmul_scalar_sse2;
@@ -138,5 +144,6 @@ void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
         fdsp->vector_fmac_scalar = ff_vector_fmac_scalar_avx;
         fdsp->vector_dmul_scalar = ff_vector_dmul_scalar_avx;
         fdsp->vector_fmul_add    = ff_vector_fmul_add_avx;
+        fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_avx;
     }
 }