AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)

Originally committed as revision 1464 to svn://svn.ffmpeg.org/ffmpeg/trunk

11 files changed, 466 insertions, 40 deletions

diff --git a/configure b/configure
index 8c414d6c4c..12390f82e2 100755
--- a/configure
+++ b/configure
@@ -644,6 +644,10 @@ fi
 if test "$altivec" = "yes" ; then
   echo "TARGET_ALTIVEC=yes" >> config.mak
   echo "#define HAVE_ALTIVEC 1" >> $TMPH
+  echo "// Enable the next line to get AltiVec performance report" >> $TMPH
+  echo "// #define ALTIVEC_TBL_PERFORMANCE_REPORT 1" >> $TMPH
+  echo "// Enable the next line to use the reference C code instead of AltiVec" >> $TMPH
+  echo "// #define ALTIVEC_USE_REFERENCE_C_CODE 1" >> $TMPH
 fi
 if test "$gprof" = "yes" ; then
   echo "TARGET_GPROF=yes" >> config.mak
diff --git a/ffmpeg.c b/ffmpeg.c
index 6a8534b1b6..0dabfd09c4 100644
--- a/ffmpeg.c
+++ b/ffmpeg.c
@@ -2829,5 +2829,12 @@ int main(int argc, char **argv)
         av_close_input_file(input_files[i]);
 
     av_free_static();
+
+    
+#ifdef ALTIVEC_TBL_PERFORMANCE_REPORT
+    extern void altivec_display_perf_report(void);
+    altivec_display_perf_report();
+#endif /* ALTIVEC_TBL_PERFORMANCE_REPORT */
+
     return 0;
 }
diff --git a/libavcodec/fft.c b/libavcodec/fft.c
index 079d2d6a6e..65eb575f36 100644
--- a/libavcodec/fft.c
+++ b/libavcodec/fft.c
@@ -53,12 +53,12 @@ int fft_init(FFTContext *s, int nbits, int inverse)
     /* compute constant table for HAVE_SSE version */
 #if (defined(HAVE_MMX) && defined(HAVE_BUILTIN_VECTOR)) || defined(HAVE_ALTIVEC)
     {
-        int has_vectors;
+        int has_vectors = 0;
 
 #if defined(HAVE_MMX)
         has_vectors = mm_support() & MM_SSE;
 #endif
-#if defined(HAVE_ALTIVEC)
+#if defined(HAVE_ALTIVEC) && !defined(ALTIVEC_USE_REFERENCE_C_CODE)
         has_vectors = mm_support() & MM_ALTIVEC;
 #endif
         if (has_vectors) {
diff --git a/libavcodec/ppc/dsputil_altivec.c b/libavcodec/ppc/dsputil_altivec.c
index 55ea7371d3..e7f34ee16c 100644
--- a/libavcodec/ppc/dsputil_altivec.c
+++ b/libavcodec/ppc/dsputil_altivec.c
@@ -24,6 +24,22 @@
 #include <sys/sysctl.h>
 #endif
 
+#ifdef ALTIVEC_TBL_PERFORMANCE_REPORT
+unsigned long long perfdata[altivec_perf_total][altivec_data_total];
+/* list below must match enum in dsputil_altivec.h */
+static unsigned char* perfname[] = {
+  "fft_calc",
+  "gmc1",
+  "dct_unquantize_h263",
+  "idct_add",
+  "idct_put",
+  "put_pixels_clamped",
+  "put_pixels16",
+  "avg_pixels16"
+};
+#include <stdio.h>
+#endif
+
 int pix_abs16x16_x2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 {
     int i;
@@ -594,7 +610,7 @@ int sad8x8_altivec(void *s, uint8_t *a, uint8_t *b, int stride) {
 }
 
 void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
-#if 0
+#ifdef ALTIVEC_USE_REFERENCE_C_CODE
     int i;
     for(i=0; i+7<w; i++){
         dst[i+0] += src[i+0];
@@ -608,38 +624,188 @@ void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
     }
     for(; i<w; i++)
         dst[i+0] += src[i+0];
-#else
+#else /* ALTIVEC_USE_REFERENCE_C_CODE */
     register int i;
-    register uint8_t *temp_src = src, *temp_dst = dst;
-    register vector unsigned char vdst, vsrc, temp1, temp2;
-    register vector unsigned char perm;
-    register int count = 0;
-
-    for (i = 0; (i < w) && ((unsigned long)temp_dst & 0x0000000F) ; i++)
-    {
-      dst[i] = src[i];
-      temp_src ++;
-      temp_dst ++;
-    }
-    /* temp_dst is a properly aligned pointer */
-    /* we still need to deal with ill-aligned src */
-    perm = vec_lvsl(0, temp_src);
-    temp1 = vec_ld(0, temp_src);
-    while ((i + 15) < w)
+    register vector unsigned char vdst, vsrc;
+    
+    /* dst and src are 16 bytes-aligned (guaranteed) */
+    for(i = 0 ; (i + 15) < w ; i++)
     {
-      temp2 = vec_ld(count + 16, temp_src);
-      vdst = vec_ld(count, temp_dst);
-      vsrc = vec_perm(temp1, temp2, perm);
-      temp1 = temp2;
+      vdst = vec_ld(i << 4, (unsigned char*)dst);
+      vsrc = vec_ld(i << 4, (unsigned char*)src);
       vdst = vec_add(vsrc, vdst);
-      vec_st(vdst, count, temp_dst);
-      count += 16;
+      vec_st(vdst, i << 4, (unsigned char*)dst);
     }
+    /* if w is not a multiple of 16 */
     for (; (i < w) ; i++)
     {
       dst[i] = src[i];
     }
-#endif
+#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
+}
+
+extern UINT8 cropTbl[];
+void put_pixels_clamped_altivec(const DCTELEM *block, UINT8 *restrict pixels,
+                                int line_size)
+{
+ALTIVEC_TBL_DECLARE(altivec_put_pixels_clamped_num, 1);
+#ifdef ALTIVEC_USE_REFERENCE_C_CODE
+    int i;
+    UINT8 *cm = cropTbl + MAX_NEG_CROP;
+
+ALTIVEC_TBL_START_COUNT(altivec_put_pixels_clamped_num, 1);
+    
+    /* read the pixels */
+    for(i=0;i<8;i++) {
+        pixels[0] = cm[block[0]];
+        pixels[1] = cm[block[1]];
+        pixels[2] = cm[block[2]];
+        pixels[3] = cm[block[3]];
+        pixels[4] = cm[block[4]];
+        pixels[5] = cm[block[5]];
+        pixels[6] = cm[block[6]];
+        pixels[7] = cm[block[7]];
+
+        pixels += line_size;
+        block += 8;
+    }
+
+ALTIVEC_TBL_STOP_COUNT(altivec_put_pixels_clamped_num, 1);
+
+#else /* ALTIVEC_USE_REFERENCE_C_CODE */
+    register const vector short vczero = (const vector short)(0);
+    register vector short
+      blockv0, blockv1, blockv2, blockv3,
+      blockv4, blockv5, blockv6, blockv7;
+    register vector unsigned char
+      pixelsv0, pixelsv1, pixelsv2, pixelsv3, pixelsv4,
+      pixelsv0old, pixelsv4old;
+
+ALTIVEC_TBL_START_COUNT(altivec_put_pixels_clamped_num, 1);
+
+    blockv0 = vec_ld(0, block);
+    blockv1 = vec_ld(16, block);
+    blockv2 = vec_ld(32, block);
+    blockv3 = vec_ld(48, block);
+    blockv4 = vec_ld(64, block);
+    blockv5 = vec_ld(80, block);
+    blockv6 = vec_ld(96, block);
+    blockv7 = vec_ld(112, block);
+    if (((unsigned long)pixels) & 0x0000000F)
+    {
+      pixelsv0old = vec_ld(-8, pixels);
+      pixelsv4old = vec_ld(56, pixels);
+      pixelsv0 = vec_packsu(vczero, blockv0);
+      pixelsv1 = vec_packsu(blockv1, blockv2);
+      pixelsv2 = vec_packsu(blockv3, blockv4);
+      pixelsv3 = vec_packsu(blockv5, blockv6);
+      pixelsv4 = vec_packsu(blockv5, vczero);
+      pixelsv0 = vec_perm(pixelsv0old, pixelsv0, vcprm(0, 1, s2, s3));
+      pixelsv4 = vec_perm(pixelsv4, pixelsv4old, vcprm(0, 1, s2, s3));
+      vec_st(pixelsv0, -8, pixels);
+      vec_st(pixelsv1, 8, pixels);
+      vec_st(pixelsv2, 24, pixels);
+      vec_st(pixelsv3, 40, pixels);
+      vec_st(pixelsv4, 56, pixels);
+    }
+    else
+    {
+      pixelsv0 = vec_packsu(blockv0, blockv1);
+      pixelsv1 = vec_packsu(blockv2, blockv3);
+      pixelsv2 = vec_packsu(blockv4, blockv5);
+      pixelsv3 = vec_packsu(blockv6, blockv7);
+      vec_st(pixelsv0, 0, pixels);
+      vec_st(pixelsv1, 16, pixels);
+      vec_st(pixelsv2, 32, pixels);
+      vec_st(pixelsv3, 48, pixels);
+    }
+
+ALTIVEC_TBL_STOP_COUNT(altivec_put_pixels_clamped_num, 1);
+#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
+}
+
+void put_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+ALTIVEC_TBL_DECLARE(altivec_put_pixels16_num, 1);
+#ifdef ALTIVEC_USE_REFERENCE_C_CODE
+    int i;
+
+ALTIVEC_TBL_START_COUNT(altivec_put_pixels16_num, 1);
+
+    for(i=0; i<h; i++) {
+      *((uint32_t*)(block )) = (((const struct unaligned_32 *) (pixels))->l);
+      *((uint32_t*)(block+4)) = (((const struct unaligned_32 *) (pixels+4))->l);
+      *((uint32_t*)(block+8)) = (((const struct unaligned_32 *) (pixels+8))->l);
+      *((uint32_t*)(block+12)) = (((const struct unaligned_32 *) (pixels+12))->l);
+      pixels+=line_size;
+      block +=line_size;
+    }
+
+ALTIVEC_TBL_STOP_COUNT(altivec_put_pixels16_num, 1);
+
+#else /* ALTIVEC_USE_REFERENCE_C_CODE */
+
+    register vector unsigned char perm = vec_lvsl(0, pixels); 
+    register vector unsigned char pixelsv1, pixelsv2;
+    int i;
+
+ALTIVEC_TBL_START_COUNT(altivec_put_pixels16_num, 1);
+
+    for(i=0; i<h; i++) {
+      pixelsv1 = vec_ld(0, (unsigned char*)pixels);
+      pixelsv2 = vec_ld(16, (unsigned char*)pixels);
+      vec_st(vec_perm(pixelsv1, pixelsv2, perm), 0, (unsigned char*)block);
+      pixels+=line_size;
+      block +=line_size;
+    }
+
+ALTIVEC_TBL_STOP_COUNT(altivec_put_pixels16_num, 1);
+
+#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
+}
+
+#define op_avg(a,b)  a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEUL)>>1) )
+void avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+ALTIVEC_TBL_DECLARE(altivec_avg_pixels16_num, 1);
+#ifdef ALTIVEC_USE_REFERENCE_C_CODE
+    int i;
+
+ALTIVEC_TBL_START_COUNT(altivec_avg_pixels16_num, 1);
+
+    for(i=0; i<h; i++) {
+      op_avg(*((uint32_t*)(block)),(((const struct unaligned_32 *)(pixels))->l));
+      op_avg(*((uint32_t*)(block+4)),(((const struct unaligned_32 *)(pixels+4))->l));
+      op_avg(*((uint32_t*)(block+8)),(((const struct unaligned_32 *)(pixels+8))->l));
+      op_avg(*((uint32_t*)(block+12)),(((const struct unaligned_32 *)(pixels+12))->l));
+      pixels+=line_size;
+      block +=line_size;
+    }
+
+ALTIVEC_TBL_STOP_COUNT(altivec_avg_pixels16_num, 1);
+
+#else /* ALTIVEC_USE_REFERENCE_C_CODE */
+
+    register vector unsigned char perm = vec_lvsl(0, pixels); 
+    register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
+    int i;
+
+ALTIVEC_TBL_START_COUNT(altivec_avg_pixels16_num, 1);
+
+    for(i=0; i<h; i++) {
+      pixelsv1 = vec_ld(0, (unsigned char*)pixels);
+      pixelsv2 = vec_ld(16, (unsigned char*)pixels);
+      blockv = vec_ld(0, block);
+      pixelsv = vec_perm(pixelsv1, pixelsv2, perm);
+      blockv = vec_avg(blockv,pixelsv);
+      vec_st(blockv, 0, (unsigned char*)block);
+      pixels+=line_size;
+      block +=line_size;
+    }
+
+ALTIVEC_TBL_STOP_COUNT(altivec_avg_pixels16_num, 1);
+
+#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
 }
 
 int has_altivec(void)
@@ -656,3 +822,22 @@ int has_altivec(void)
 #endif
     return 0;
 }
+
+#ifdef ALTIVEC_TBL_PERFORMANCE_REPORT
+void altivec_display_perf_report(void)
+{
+  int i;
+  fprintf(stderr, "AltiVec performance report\n Values are from the Time Base register, and represent 4 bus cycles.\n");
+  for(i = 0 ; i < altivec_perf_total ; i++)
+  {
+    if (perfdata[i][altivec_data_num] != (unsigned long long)0)
+      fprintf(stderr, " Function \"%s\":\n\tmin: %llu\n\tmax: %llu\n\tavg: %1.2lf (%llu)\n",
+              perfname[i],
+              perfdata[i][altivec_data_min],
+              perfdata[i][altivec_data_max],
+              (double)perfdata[i][altivec_data_sum] /
+              (double)perfdata[i][altivec_data_num],
+              perfdata[i][altivec_data_num]);
+  }
+}
+#endif /* ALTIVEC_TBL_PERFORMANCE_REPORT */
diff --git a/libavcodec/ppc/dsputil_altivec.h b/libavcodec/ppc/dsputil_altivec.h
index 4a6043fe2f..b7b4238db9 100644
--- a/libavcodec/ppc/dsputil_altivec.h
+++ b/libavcodec/ppc/dsputil_altivec.h
@@ -31,13 +31,16 @@ extern int pix_sum_altivec(UINT8 * pix, int line_size);
 extern void diff_pixels_altivec(DCTELEM* block, const UINT8* s1, const UINT8* s2, int stride);
 extern void get_pixels_altivec(DCTELEM* block, const UINT8 * pixels, int line_size);
 
-extern void gmc1_altivec(UINT8 *dst, UINT8 *src, int stride, int h, int x16, int y16, int rounder);
-
 extern void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w);
+extern void put_pixels_clamped_altivec(const DCTELEM *block, UINT8 *restrict pixels, int line_size);
+void put_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h);
+void avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h);
 
-extern int has_altivec(void);
+extern void gmc1_altivec(UINT8 *dst, UINT8 *src, int stride, int h, int x16, int y16, int rounder);
 
+extern int has_altivec(void);
 
+#ifdef HAVE_ALTIVEC
 
 // used to build registers permutation vectors (vcprm)
 // the 's' are for words in the _s_econd vector
@@ -63,3 +66,59 @@ extern int has_altivec(void);
 #define FLOAT_p 1.
 
 #define vcii(a,b,c,d) (const vector float)(FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d)
+
+#ifdef ALTIVEC_TBL_PERFORMANCE_REPORT
+void altivec_display_perf_report(void);
+/* if you add to the enum below, also add to the perfname array
+   in dsputil_altivec.c */
+enum altivec_perf_index {
+  altivec_fft_num = 0,
+  altivec_gmc1_num,
+  altivec_dct_unquantize_h263_num,
+  altivec_idct_add_num,
+  altivec_idct_put_num,
+  altivec_put_pixels_clamped_num,
+  altivec_put_pixels16_num,
+  altivec_avg_pixels16_num,
+  altivec_perf_total
+};
+enum altivec_data_index {
+  altivec_data_min = 0,
+  altivec_data_max,
+  altivec_data_sum,
+  altivec_data_num,
+  altivec_data_total
+};
+extern unsigned long long perfdata[altivec_perf_total][altivec_data_total];
+#define ALTIVEC_TBL_DECLARE(a, cond) register unsigned long tbl_start, tbl_stop
+#define ALTIVEC_TBL_START_COUNT(a, cond) do { asm("mftb %0" : "=r" (tbl_start)); } while (0)
+#define ALTIVEC_TBL_STOP_COUNT(a, cond) do {     \
+  asm volatile("mftb %0" : "=r" (tbl_stop));              \
+  if (tbl_stop > tbl_start)                      \
+  {                                              \
+    unsigned long diff =  tbl_stop - tbl_start;  \
+    if (cond)                                    \
+    {                                            \
+      if (diff < perfdata[a][altivec_data_min])  \
+        perfdata[a][altivec_data_min] = diff;    \
+      if (diff > perfdata[a][altivec_data_max])  \
+        perfdata[a][altivec_data_max] = diff;    \
+      perfdata[a][altivec_data_sum] += diff;     \
+      perfdata[a][altivec_data_num] ++;          \
+    }                                            \
+  }                                              \
+} while (0)
+#else /* ALTIVEC_TBL_PERFORMANCE_REPORT */
+#define ALTIVEC_TBL_DECLARE(a, cond)
+#define ALTIVEC_TBL_START_COUNT(a, cond)
+#define ALTIVEC_TBL_STOP_COUNT(a, cond)
+#endif /* ALTIVEC_TBL_PERFORMANCE_REPORT */
+
+#else /* HAVE_ALTIVEC */
+#ifdef ALTIVEC_USE_REFERENCE_C_CODE
+#error "I can't use ALTIVEC_USE_REFERENCE_C_CODE if I don't use HAVE_ALTIVEC"
+#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
+#ifdef ALTIVEC_TBL_PERFORMANCE_REPORT
+#error "I can't use ALTIVEC_TBL_PERFORMANCE_REPORT if I don't use HAVE_ALTIVEC"
+#endif /* ALTIVEC_TBL_PERFORMANCE_REPORT */
+#endif /* HAVE_ALTIVEC */
diff --git a/libavcodec/ppc/dsputil_ppc.c b/libavcodec/ppc/dsputil_ppc.c
index 055a8f09fc..ec2dda4690 100644
--- a/libavcodec/ppc/dsputil_ppc.c
+++ b/libavcodec/ppc/dsputil_ppc.c
@@ -60,11 +60,27 @@ void dsputil_init_ppc(DSPContext* c, unsigned mask)
         c->pix_sum = pix_sum_altivec;
         c->diff_pixels = diff_pixels_altivec;
         c->get_pixels = get_pixels_altivec;
-// next one disabled as it it untested.
+// next two disabled as they're untested.
 #if 0
         c->add_bytes= add_bytes_altivec;
+        c->put_pixels_clamped = put_pixels_clamped_altivec;
 #endif
+        c->put_pixels_tab[0][0] = put_pixels16_altivec;
+        c->avg_pixels_tab[0][0] = avg_pixels16_altivec;
 	c->gmc1 = gmc1_altivec;
+
+#ifdef ALTIVEC_TBL_PERFORMANCE_REPORT
+        {
+          int i;
+          for (i = 0 ; i < altivec_perf_total ; i++)
+          {
+            perfdata[i][altivec_data_min] = 0xFFFFFFFFFFFFFFFF;
+            perfdata[i][altivec_data_max] = 0x0000000000000000;
+            perfdata[i][altivec_data_sum] = 0x0000000000000000;
+            perfdata[i][altivec_data_num] = 0x0000000000000000;
+          }
+        }
+#endif
     } else
 #endif
     {
diff --git a/libavcodec/ppc/fft_altivec.c b/libavcodec/ppc/fft_altivec.c
index 4abdd4f740..55bfcdf4d2 100644
--- a/libavcodec/ppc/fft_altivec.c
+++ b/libavcodec/ppc/fft_altivec.c
@@ -22,6 +22,31 @@
 
 #include "dsputil_altivec.h"
 
+/*
+  those three macros are from libavcodec/fft.c
+  and are required for the reference C code
+*/
+/* butter fly op */
+#define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
+{\
+  FFTSample ax, ay, bx, by;\
+  bx=pre1;\
+  by=pim1;\
+  ax=qre1;\
+  ay=qim1;\
+  pre = (bx + ax);\
+  pim = (by + ay);\
+  qre = (bx - ax);\
+  qim = (by - ay);\
+}
+#define MUL16(a,b) ((a) * (b))
+#define CMUL(pre, pim, are, aim, bre, bim) \
+{\
+   pre = (MUL16(are, bre) - MUL16(aim, bim));\
+   pim = (MUL16(are, bim) + MUL16(bre, aim));\
+}
+
+
 /**
  * Do a complex FFT with the parameters defined in fft_init(). The
  * input data must be permuted before with s->revtab table. No
@@ -35,6 +60,84 @@
  */
 void fft_calc_altivec(FFTContext *s, FFTComplex *z)
 {
+ALTIVEC_TBL_DECLARE(altivec_fft_num, s->nbits >= 6);
+#ifdef ALTIVEC_USE_REFERENCE_C_CODE
+    int ln = s->nbits;
+    int	j, np, np2;
+    int	nblocks, nloops;
+    register FFTComplex *p, *q;
+    FFTComplex *exptab = s->exptab;
+    int l;
+    FFTSample tmp_re, tmp_im;
+    
+ALTIVEC_TBL_START_COUNT(altivec_fft_num, s->nbits >= 6);
+ 
+    np = 1 << ln;
+
+    /* pass 0 */
+
+    p=&z[0];
+    j=(np >> 1);
+    do {
+        BF(p[0].re, p[0].im, p[1].re, p[1].im, 
+           p[0].re, p[0].im, p[1].re, p[1].im);
+        p+=2;
+    } while (--j != 0);
+
+    /* pass 1 */
+
+    
+    p=&z[0];
+    j=np >> 2;
+    if (s->inverse) {
+        do {
+            BF(p[0].re, p[0].im, p[2].re, p[2].im, 
+               p[0].re, p[0].im, p[2].re, p[2].im);
+            BF(p[1].re, p[1].im, p[3].re, p[3].im, 
+               p[1].re, p[1].im, -p[3].im, p[3].re);
+            p+=4;
+        } while (--j != 0);
+    } else {
+        do {
+            BF(p[0].re, p[0].im, p[2].re, p[2].im, 
+               p[0].re, p[0].im, p[2].re, p[2].im);
+            BF(p[1].re, p[1].im, p[3].re, p[3].im, 
+               p[1].re, p[1].im, p[3].im, -p[3].re);
+            p+=4;
+        } while (--j != 0);
+    }
+    /* pass 2 .. ln-1 */
+
+    nblocks = np >> 3;
+    nloops = 1 << 2;
+    np2 = np >> 1;
+    do {
+        p = z;
+        q = z + nloops;
+        for (j = 0; j < nblocks; ++j) {
+            BF(p->re, p->im, q->re, q->im,
+               p->re, p->im, q->re, q->im);
+            
+            p++;
+            q++;
+            for(l = nblocks; l < np2; l += nblocks) {
+                CMUL(tmp_re, tmp_im, exptab[l].re, exptab[l].im, q->re, q->im);
+                BF(p->re, p->im, q->re, q->im,
+                   p->re, p->im, tmp_re, tmp_im);
+                p++;
+                q++;
+            }
+
+            p += nloops;
+            q += nloops;
+        }
+        nblocks = nblocks >> 1;
+        nloops = nloops << 1;
+    } while (nblocks != 0);
+
+ALTIVEC_TBL_STOP_COUNT(altivec_fft_num, s->nbits >= 6);
+
+#else /* ALTIVEC_USE_REFERENCE_C_CODE */
     register const vector float vczero = (const vector float)(0.);
     
     int ln = s->nbits;
@@ -44,6 +147,8 @@ void fft_calc_altivec(FFTContext *s, FFTComplex *z)
     FFTComplex *cptr, *cptr1;
     int k;
 
+ALTIVEC_TBL_START_COUNT(altivec_fft_num, s->nbits >= 6);
+
     np = 1 << ln;
 
     {
@@ -129,5 +234,8 @@ void fft_calc_altivec(FFTContext *s, FFTComplex *z)
         nblocks = nblocks >> 1;
         nloops = nloops << 1;
     } while (nblocks != 0);
-}
 
+ALTIVEC_TBL_STOP_COUNT(altivec_fft_num, s->nbits >= 6);
+
+#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
+}
diff --git a/libavcodec/ppc/gmc_altivec.c b/libavcodec/ppc/gmc_altivec.c
index 3bda9ac16f..c2d908d039 100644
--- a/libavcodec/ppc/gmc_altivec.c
+++ b/libavcodec/ppc/gmc_altivec.c
@@ -28,13 +28,15 @@
 */
 void gmc1_altivec(UINT8 *dst /* align 8 */, UINT8 *src /* align1 */, int stride, int h, int x16, int y16, int rounder)
 {
-#if 0
+ALTIVEC_TBL_DECLARE(altivec_gmc1_num, h == 8);
+#ifdef ALTIVEC_USE_REFERENCE_C_CODE
     const int A=(16-x16)*(16-y16);
     const int B=(   x16)*(16-y16);
     const int C=(16-x16)*(   y16);
     const int D=(   x16)*(   y16);
-    
     int i;
+
+ALTIVEC_TBL_START_COUNT(altivec_gmc1_num, h == 8);
     
     for(i=0; i<h; i++)
     {
@@ -49,7 +51,10 @@ void gmc1_altivec(UINT8 *dst /* align 8 */, UINT8 *src /* align1 */, int stride,
         dst+= stride;
         src+= stride;
     }
-#else
+
+ALTIVEC_TBL_STOP_COUNT(altivec_gmc1_num, h == 8);
+
+#else /* ALTIVEC_USE_REFERENCE_C_CODE */
     const unsigned short __attribute__ ((aligned(16))) rounder_a[8] =
       {rounder, rounder, rounder, rounder,
        rounder, rounder, rounder, rounder};
@@ -61,7 +66,6 @@ void gmc1_altivec(UINT8 *dst /* align 8 */, UINT8 *src /* align1 */, int stride,
         (   x16)*(   y16), /* D */
         0, 0, 0, 0         /* padding */
       };
-    
     register const vector unsigned char vczero = (const vector unsigned char)(0);
     register const vector unsigned short vcsr8 = (const vector unsigned short)(8);
     register vector unsigned char dstv, dstv2, src_0, src_1, srcvA, srcvB, srcvC, srcvD;
@@ -70,6 +74,9 @@ void gmc1_altivec(UINT8 *dst /* align 8 */, UINT8 *src /* align1 */, int stride,
     unsigned long dst_odd = (unsigned long)dst & 0x0000000F;
     unsigned long src_really_odd = (unsigned long)src & 0x0000000F;
 
+
+ALTIVEC_TBL_START_COUNT(altivec_gmc1_num, h == 8);
+
     tempA = vec_ld(0, (unsigned short*)ABCD);
     Av = vec_splat(tempA, 0);
     Bv = vec_splat(tempA, 1);
@@ -155,5 +162,8 @@ void gmc1_altivec(UINT8 *dst /* align 8 */, UINT8 *src /* align1 */, int stride,
       dst += stride;
       src += stride;
     }
-#endif
+
+ALTIVEC_TBL_STOP_COUNT(altivec_gmc1_num, h == 8);
+
+#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
 }
diff --git a/libavcodec/ppc/idct_altivec.c b/libavcodec/ppc/idct_altivec.c
index 8036d403fa..533ab3e4c8 100644
--- a/libavcodec/ppc/idct_altivec.c
+++ b/libavcodec/ppc/idct_altivec.c
@@ -38,6 +38,7 @@
 #include <stdlib.h>                                      /* malloc(), free() */
 #include <string.h>
 #include "../dsputil.h"
+#include "dsputil_altivec.h"
 
 #define vector_s16_t vector signed short
 #define vector_u16_t vector unsigned short
@@ -160,8 +161,17 @@ static const vector_s16_t constants[5] = {
 
 void idct_put_altivec(uint8_t* dest, int stride, vector_s16_t* block)
 {
+ALTIVEC_TBL_DECLARE(altivec_idct_put_num, 1);
+#ifdef ALTIVEC_USE_REFERENCE_C_CODE
+ALTIVEC_TBL_START_COUNT(altivec_idct_put_num, 1);
+    void simple_idct_put(UINT8 *dest, int line_size, INT16 *block);
+    simple_idct_put(dest, stride, (INT16*)block);
+ALTIVEC_TBL_STOP_COUNT(altivec_idct_put_num, 1);
+#else /* ALTIVEC_USE_REFERENCE_C_CODE */
     vector_u8_t tmp;
 
+ALTIVEC_TBL_START_COUNT(altivec_idct_put_num, 1);
+
     IDCT
 
 #define COPY(dest,src)						\
@@ -177,16 +187,28 @@ void idct_put_altivec(uint8_t* dest, int stride, vector_s16_t* block)
     COPY (dest, vx5)	dest += stride;
     COPY (dest, vx6)	dest += stride;
     COPY (dest, vx7)
+
+ALTIVEC_TBL_STOP_COUNT(altivec_idct_put_num, 1);
+#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
 }
 
 void idct_add_altivec(uint8_t* dest, int stride, vector_s16_t* block)
 {
+ALTIVEC_TBL_DECLARE(altivec_idct_add_num, 1);
+#ifdef ALTIVEC_USE_REFERENCE_C_CODE
+ALTIVEC_TBL_START_COUNT(altivec_idct_add_num, 1);
+    void simple_idct_add(UINT8 *dest, int line_size, INT16 *block);
+    simple_idct_add(dest, stride, (INT16*)block);
+ALTIVEC_TBL_STOP_COUNT(altivec_idct_add_num, 1);
+#else /* ALTIVEC_USE_REFERENCE_C_CODE */
     vector_u8_t tmp;
     vector_s16_t tmp2, tmp3;
     vector_u8_t perm0;
     vector_u8_t perm1;
     vector_u8_t p0, p1, p;
 
+ALTIVEC_TBL_START_COUNT(altivec_idct_add_num, 1);
+
     IDCT
 
     p0 = vec_lvsl (0, dest);
@@ -212,5 +234,8 @@ void idct_add_altivec(uint8_t* dest, int stride, vector_s16_t* block)
     ADD (dest, vx5, perm1)	dest += stride;
     ADD (dest, vx6, perm0)	dest += stride;
     ADD (dest, vx7, perm1)
+
+ALTIVEC_TBL_STOP_COUNT(altivec_idct_add_num, 1);
+#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
 }
 
diff --git a/libavcodec/ppc/mpegvideo_altivec.c b/libavcodec/ppc/mpegvideo_altivec.c
index a63df307c0..dfa4c0229a 100644
--- a/libavcodec/ppc/mpegvideo_altivec.c
+++ b/libavcodec/ppc/mpegvideo_altivec.c
@@ -20,6 +20,7 @@
 #include <stdio.h>
 #include "../dsputil.h"
 #include "../mpegvideo.h"
+#include "dsputil_altivec.h"
 
 // Swaps two variables (used for altivec registers)
 #define SWAP(a,b) \
@@ -510,10 +511,13 @@ int dct_quantize_altivec(MpegEncContext* s,
 void dct_unquantize_h263_altivec(MpegEncContext *s, 
                                  DCTELEM *block, int n, int qscale)
 {
+ALTIVEC_TBL_DECLARE(altivec_dct_unquantize_h263_num, 1);
     int i, level, qmul, qadd;
     int nCoeffs;
     
     assert(s->block_last_index[n]>=0);
+
+ALTIVEC_TBL_START_COUNT(altivec_dct_unquantize_h263_num, 1);
     
     qadd = (qscale - 1) | 1;
     qmul = qscale << 1;
@@ -533,7 +537,7 @@ void dct_unquantize_h263_altivec(MpegEncContext *s,
         nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
     }
 
-#if 0
+#ifdef ALTIVEC_USE_REFERENCE_C_CODE
     for(;i<=nCoeffs;i++) {
         level = block[i];
         if (level) {
@@ -545,7 +549,7 @@ void dct_unquantize_h263_altivec(MpegEncContext *s,
             block[i] = level;
         }
     }
-#else
+#else /* ALTIVEC_USE_REFERENCE_C_CODE */
     {
       register const vector short vczero = (const vector short)(0);
       short __attribute__ ((aligned(16))) qmul8[] =
@@ -572,6 +576,7 @@ void dct_unquantize_h263_altivec(MpegEncContext *s,
       qaddv = vec_ld(0, qadd8);
       nqaddv = vec_ld(0, nqadd8);
 
+#if 0 // block *is* 16 bytes-aligned, it seems.
       // first make sure block[j] is 16 bytes-aligned
       for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
         level = block[j];
@@ -584,6 +589,7 @@ void dct_unquantize_h263_altivec(MpegEncContext *s,
             block[j] = level;
         }
       }
+#endif
       
       // vectorize all the 16 bytes-aligned blocks
       // of 8 elements
@@ -622,5 +628,7 @@ void dct_unquantize_h263_altivec(MpegEncContext *s,
         block[0] = backup_0;
       }
     }
-#endif
+#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
+
+ALTIVEC_TBL_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);
 }
diff --git a/libavcodec/ppc/mpegvideo_ppc.c b/libavcodec/ppc/mpegvideo_ppc.c
index b1c834f828..78105526e1 100644
--- a/libavcodec/ppc/mpegvideo_ppc.c
+++ b/libavcodec/ppc/mpegvideo_ppc.c
@@ -44,7 +44,11 @@ void MPV_common_init_ppc(MpegEncContext *s)
         {
             s->idct_put = idct_put_altivec;
             s->idct_add = idct_add_altivec;
+#ifndef ALTIVEC_USE_REFERENCE_C_CODE
             s->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
+#else /* ALTIVEC_USE_REFERENCE_C_CODE */
+            s->idct_permutation_type = FF_NO_IDCT_PERM;
+#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
         }
 
         // Test to make sure that the dct required alignments are met.