3dnow2 implementation of imdct.

6% faster vorbis and wma.

Originally committed as revision 5954 to svn://svn.ffmpeg.org/ffmpeg/trunk

5 files changed, 93 insertions, 7 deletions

diff --git a/libavcodec/dsputil.h b/libavcodec/dsputil.h
index a608350294..1083c39c71 100644
--- a/libavcodec/dsputil.h
+++ b/libavcodec/dsputil.h
@@ -594,6 +594,8 @@ void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
    FFTSample type */
 typedef float FFTSample;
 
+struct MDCTContext;
+
 typedef struct FFTComplex {
     FFTSample re, im;
 } FFTComplex;
@@ -605,6 +607,8 @@ typedef struct FFTContext {
     FFTComplex *exptab;
     FFTComplex *exptab1; /* only used by SSE code */
     void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
+    void (*imdct_calc)(struct MDCTContext *s, FFTSample *output,
+                       const FFTSample *input, FFTSample *tmp);
 } FFTContext;
 
 int ff_fft_init(FFTContext *s, int nbits, int inverse);
@@ -635,6 +639,8 @@ typedef struct MDCTContext {
 int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
 void ff_imdct_calc(MDCTContext *s, FFTSample *output,
                 const FFTSample *input, FFTSample *tmp);
+void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output,
+                        const FFTSample *input, FFTSample *tmp);
 void ff_mdct_calc(MDCTContext *s, FFTSample *out,
                const FFTSample *input, FFTSample *tmp);
 void ff_mdct_end(MDCTContext *s);
diff --git a/libavcodec/fft.c b/libavcodec/fft.c
index 1c63f6889f..3d5c221ebf 100644
--- a/libavcodec/fft.c
+++ b/libavcodec/fft.c
@@ -54,6 +54,7 @@ int ff_fft_init(FFTContext *s, int nbits, int inverse)
         s->exptab[i].im = s1;
     }
     s->fft_calc = ff_fft_calc_c;
+    s->imdct_calc = ff_imdct_calc;
     s->exptab1 = NULL;
 
     /* compute constant table for HAVE_SSE version */
@@ -62,11 +63,7 @@ int ff_fft_init(FFTContext *s, int nbits, int inverse)
         int has_vectors = 0;
 
 #if defined(HAVE_MMX)
-#ifdef HAVE_MM3DNOW
         has_vectors = mm_support() & (MM_3DNOW | MM_3DNOWEXT | MM_SSE | MM_SSE2);
-#else
-        has_vectors = mm_support() & (MM_SSE | MM_SSE2);
-#endif
 #endif
 #if defined(HAVE_ALTIVEC) && !defined(ALTIVEC_USE_REFERENCE_C_CODE)
         has_vectors = mm_support() & MM_ALTIVEC;
@@ -98,6 +95,8 @@ int ff_fft_init(FFTContext *s, int nbits, int inverse)
             } while (nblocks != 0);
             av_freep(&s->exptab);
 #if defined(HAVE_MMX)
+            if (has_vectors & MM_3DNOWEXT)
+                s->imdct_calc = ff_imdct_calc_3dn2;
 #ifdef HAVE_MM3DNOW
             if (has_vectors & MM_3DNOWEXT)
                 /* 3DNowEx for Athlon(XP) */
diff --git a/libavcodec/i386/fft_3dn2.c b/libavcodec/i386/fft_3dn2.c
index aa8f0aee2e..40ec9d8eb1 100644
--- a/libavcodec/i386/fft_3dn2.c
+++ b/libavcodec/i386/fft_3dn2.c
@@ -1,6 +1,6 @@
 /*
  * FFT/MDCT transform with Extended 3DNow! optimizations
- * Copyright (c) 2006 Zuxy MENG Jie.
+ * Copyright (c) 2006 Zuxy MENG Jie, Loren Merritt
  * Based on fft_sse.c copyright (c) 2002 Fabrice Bellard.
  *
  * This library is free software; you can redistribute it and/or
@@ -134,3 +134,84 @@ void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z)
 }
 
 #endif
+
+void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output,
+                        const FFTSample *input, FFTSample *tmp)
+{
+    int k, n8, n4, n2, n;
+    const uint16_t *revtab = s->fft.revtab;
+    const FFTSample *tcos = s->tcos;
+    const FFTSample *tsin = s->tsin;
+    const FFTSample *in1, *in2;
+    FFTComplex *z = (FFTComplex *)tmp;
+
+    n = 1 << s->nbits;
+    n2 = n >> 1;
+    n4 = n >> 2;
+    n8 = n >> 3;
+
+    /* pre rotation */
+    in1 = input;
+    in2 = input + n2 - 1;
+    for(k = 0; k < n4; k++) {
+        asm volatile(
+            "movd       %1, %%mm0 \n\t"
+            "movd       %3, %%mm1 \n\t"
+            "punpckldq  %2, %%mm0 \n\t"
+            "punpckldq  %4, %%mm1 \n\t"
+            "movq    %%mm0, %%mm2 \n\t"
+            "pfmul   %%mm1, %%mm0 \n\t"
+            "pswapd  %%mm1, %%mm1 \n\t"
+            "pfmul   %%mm1, %%mm2 \n\t"
+            "pfpnacc %%mm2, %%mm0 \n\t"
+            "movq    %%mm0, %0    \n\t"
+            :"=m"(z[revtab[k]])
+            :"m"(in2[-2*k]), "m"(in1[2*k]),
+             "m"(tcos[k]), "m"(tsin[k])
+        );
+    }
+
+    ff_fft_calc(&s->fft, z);
+
+    /* post rotation + reordering */
+    for(k = 0; k < n4; k++) {
+        asm volatile(
+            "movq       %0, %%mm0 \n\t"
+            "movd       %1, %%mm1 \n\t"
+            "punpckldq  %2, %%mm1 \n\t"
+            "movq    %%mm0, %%mm2 \n\t"
+            "pfmul   %%mm1, %%mm0 \n\t"
+            "pswapd  %%mm1, %%mm1 \n\t"
+            "pfmul   %%mm1, %%mm2 \n\t"
+            "pfpnacc %%mm2, %%mm0 \n\t"
+            "movq    %%mm0, %0    \n\t"
+            :"+m"(z[k])
+            :"m"(tcos[k]), "m"(tsin[k])
+        );
+    }
+
+    asm volatile("movd %0, %%mm7" ::"r"(1<<31));
+    for(k = 0; k < n8; k++) {
+        asm volatile(
+            "movq         %4, %%mm0 \n\t"
+            "pswapd       %5, %%mm1 \n\t"
+            "movq      %%mm0, %%mm2 \n\t"
+            "pxor      %%mm7, %%mm2 \n\t"
+            "punpckldq %%mm1, %%mm2 \n\t"
+            "pswapd    %%mm2, %%mm3 \n\t"
+            "punpckhdq %%mm1, %%mm0 \n\t"
+            "pswapd    %%mm0, %%mm4 \n\t"
+            "pxor      %%mm7, %%mm0 \n\t"
+            "pxor      %%mm7, %%mm4 \n\t"
+            "movq      %%mm0, %0    \n\t" // { -z[n8+k].im, z[n8-1-k].re }
+            "movq      %%mm4, %1    \n\t" // { -z[n8-1-k].re, z[n8+k].im }
+            "movq      %%mm2, %2    \n\t" // { -z[n8+k].re, z[n8-1-k].im }
+            "movq      %%mm3, %3    \n\t" // { z[n8-1-k].im, -z[n8+k].re }
+            :"=m"(output[2*k]), "=m"(output[n2-2-2*k]),
+             "=m"(output[n2+2*k]), "=m"(output[n-2-2*k])
+            :"m"(z[n8+k]), "m"(z[n8-1-k])
+            :"memory"
+        );
+    }
+    asm volatile("emms");
+}
diff --git a/libavcodec/vorbis.c b/libavcodec/vorbis.c
index 0bd318d6b1..bf2cb358bf 100644
--- a/libavcodec/vorbis.c
+++ b/libavcodec/vorbis.c
@@ -1598,7 +1598,7 @@ static int vorbis_parse_audio_packet(vorbis_context *vc) {
 
         saved_start=vc->saved_start;
 
-        ff_imdct_calc(vc->modes[mode_number].blockflag ? &vc->mdct1 : &vc->mdct0, buf, ch_floor_ptr, buf_tmp);
+        vc->mdct0.fft.imdct_calc(vc->modes[mode_number].blockflag ? &vc->mdct1 : &vc->mdct0, buf, ch_floor_ptr, buf_tmp);
 
         if (vc->modes[mode_number].blockflag) {
             // -- overlap/add
diff --git a/libavcodec/wmadec.c b/libavcodec/wmadec.c
index 227c9695ba..6f33d2a8f8 100644
--- a/libavcodec/wmadec.c
+++ b/libavcodec/wmadec.c
@@ -1113,7 +1113,7 @@ static int wma_decode_block(WMADecodeContext *s)
 
             n = s->block_len;
             n4 = s->block_len / 2;
-            ff_imdct_calc(&s->mdct_ctx[bsize],
+            s->mdct_ctx[bsize].fft.imdct_calc(&s->mdct_ctx[bsize],
                           output, s->coefs[ch], s->mdct_tmp);
 
             /* XXX: optimize all that by build the window and