PPC perf, PPC clear_block, AltiVec put_pixels8_xy2 patch by (Romain Dolbeau <dolbeau at irisa dot fr>)

Originally committed as revision 1476 to svn://svn.ffmpeg.org/ffmpeg/trunk

1 files changed, 159 insertions, 8 deletions

diff --git a/libavcodec/ppc/dsputil_ppc.c b/libavcodec/ppc/dsputil_ppc.c
index ec2dda4690..7dcca85a7c 100644
--- a/libavcodec/ppc/dsputil_ppc.c
+++ b/libavcodec/ppc/dsputil_ppc.c
@@ -19,6 +19,8 @@
 
 #include "../dsputil.h"
 
+#include "dsputil_ppc.h"
+
 #ifdef HAVE_ALTIVEC
 #include "dsputil_altivec.h"
 #endif
@@ -36,12 +38,149 @@ int mm_support(void)
     return result;
 }
 
+#ifdef POWERPC_TBL_PERFORMANCE_REPORT
+unsigned long long perfdata[powerpc_perf_total][powerpc_data_total];
+/* list below must match enum in dsputil_altivec.h */
+static unsigned char* perfname[] = {
+  "fft_calc_altivec",
+  "gmc1_altivec",
+  "dct_unquantize_h263_altivec",
+  "idct_add_altivec",
+  "idct_put_altivec",
+  "put_pixels_clamped_altivec",
+  "put_pixels16_altivec",
+  "avg_pixels16_altivec",
+  "avg_pixels8_altivec",
+  "put_pixels8_xy2_altivec",
+  "clear_blocks_dcbz32_ppc"
+};
+#ifdef POWERPC_PERF_USE_PMC
+unsigned long long perfdata_miss[powerpc_perf_total][powerpc_data_total];
+#endif
+#include <stdio.h>
+#endif
+
+#ifdef POWERPC_TBL_PERFORMANCE_REPORT
+void powerpc_display_perf_report(void)
+{
+  int i;
+#ifndef POWERPC_PERF_USE_PMC
+  fprintf(stderr, "AltiVec performance report\n Values are from the Time Base register, and represent 4 bus cycles.\n");
+#else /* POWERPC_PERF_USE_PMC */
+  fprintf(stderr, "AltiVec performance report\n Values are from the PMC registers, and represent whatever the registers are set to record.\n");
+#endif /* POWERPC_PERF_USE_PMC */
+  for(i = 0 ; i < powerpc_perf_total ; i++)
+  {
+    if (perfdata[i][powerpc_data_num] != (unsigned long long)0)
+      fprintf(stderr, " Function \"%s\" (pmc1):\n\tmin: %llu\n\tmax: %llu\n\tavg: %1.2lf (%llu)\n",
+              perfname[i],
+              perfdata[i][powerpc_data_min],
+              perfdata[i][powerpc_data_max],
+              (double)perfdata[i][powerpc_data_sum] /
+              (double)perfdata[i][powerpc_data_num],
+              perfdata[i][powerpc_data_num]);
+#ifdef POWERPC_PERF_USE_PMC
+    if (perfdata_miss[i][powerpc_data_num] != (unsigned long long)0)
+      fprintf(stderr, " Function \"%s\" (pmc2):\n\tmin: %llu\n\tmax: %llu\n\tavg: %1.2lf (%llu)\n",
+              perfname[i],
+              perfdata_miss[i][powerpc_data_min],
+              perfdata_miss[i][powerpc_data_max],
+              (double)perfdata_miss[i][powerpc_data_sum] /
+              (double)perfdata_miss[i][powerpc_data_num],
+              perfdata_miss[i][powerpc_data_num]);
+#endif
+  }
+}
+#endif /* POWERPC_TBL_PERFORMANCE_REPORT */
+
+/* ***** WARNING ***** WARNING ***** WARNING ***** */
+/*
+  clear_blocks_dcbz32_ppc will not work properly
+  on PowerPC processors with a cache line size
+  not equal to 32 bytes.
+  Fortunately all processor used by Apple up to
+  at least the 7450 (aka second generation G4)
+  use 32 bytes cache line.
+  This is due to the use of the 'dcbz' instruction.
+  It simply clear to zero a single cache line,
+  so you need to know the cache line size to use it !
+  It's absurd, but it's fast...
+*/
+void clear_blocks_dcbz32_ppc(DCTELEM *blocks)
+{
+POWERPC_TBL_DECLARE(powerpc_clear_blocks_dcbz32, 1);
+    register int misal = ((unsigned long)blocks & 0x00000010);
+    register int i = 0;
+POWERPC_TBL_START_COUNT(powerpc_clear_blocks_dcbz32, 1);
+#if 1
+    if (misal) {
+      ((unsigned long*)blocks)[0] = 0L;
+      ((unsigned long*)blocks)[1] = 0L;
+      ((unsigned long*)blocks)[2] = 0L;
+      ((unsigned long*)blocks)[3] = 0L;
+      vec_st((vector short)(0), 0, blocks);
+      i += 16;
+    }
+    for ( ; i < sizeof(DCTELEM)*6*64 ; i += 32) {
+      asm volatile("dcbz %0,%1" : : "r" (blocks), "r" (i) : "memory");
+    }
+    if (misal) {
+      ((unsigned long*)blocks)[188] = 0L;
+      ((unsigned long*)blocks)[189] = 0L;
+      ((unsigned long*)blocks)[190] = 0L;
+      ((unsigned long*)blocks)[191] = 0L;
+      i += 16;
+    }
+#else
+    memset(blocks, 0, sizeof(DCTELEM)*6*64);
+#endif
+POWERPC_TBL_STOP_COUNT(powerpc_clear_blocks_dcbz32, 1);
+}
+
+/* check dcbz report how many bytes are set to 0 by dcbz */
+long check_dcbz_effect(void)
+{
+  register char *fakedata = (char*)malloc(1024);
+  register char *fakedata_middle;
+  register long zero = 0;
+  register long i = 0;
+  long count = 0;
+
+  if (fakedata == NULL)
+  {
+    return 0L;
+  }
+
+
+  fakedata_middle = (fakedata + 512);
+
+  memset(fakedata, 0xFF, 1024);
+
+  asm volatile("dcbz %0, %1" : : "r" (fakedata_middle), "r" (zero));
+
+  for (i = 0; i < 1024 ; i ++)
+  {
+    if (fakedata[i] == (char)0)
+      count++;
+  }
+
+  free(fakedata);
+  
+  return count;
+}
+
 void dsputil_init_ppc(DSPContext* c, unsigned mask)
 {
     // Common optimisations whether Altivec or not
 
-    // ... pending ...
-
+  switch (check_dcbz_effect()) {
+  case 32:
+    c->clear_blocks = clear_blocks_dcbz32_ppc;
+    break;
+  default:
+    break;
+  }
+  
 #if HAVE_ALTIVEC
     if (has_altivec()) {
         mm_flags |= MM_ALTIVEC;
@@ -67,17 +206,29 @@ void dsputil_init_ppc(DSPContext* c, unsigned mask)
 #endif
         c->put_pixels_tab[0][0] = put_pixels16_altivec;
         c->avg_pixels_tab[0][0] = avg_pixels16_altivec;
+// next one disabled as it's untested.
+#if 0
+        c->avg_pixels_tab[1][0] = avg_pixels8_altivec;
+#endif
+        c->put_pixels_tab[1][3] = put_pixels8_xy2_altivec;
+        
 	c->gmc1 = gmc1_altivec;
 
-#ifdef ALTIVEC_TBL_PERFORMANCE_REPORT
+#ifdef POWERPC_TBL_PERFORMANCE_REPORT
         {
           int i;
-          for (i = 0 ; i < altivec_perf_total ; i++)
+          for (i = 0 ; i < powerpc_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;
+            perfdata[i][powerpc_data_min] = 0xFFFFFFFFFFFFFFFF;
+            perfdata[i][powerpc_data_max] = 0x0000000000000000;
+            perfdata[i][powerpc_data_sum] = 0x0000000000000000;
+            perfdata[i][powerpc_data_num] = 0x0000000000000000;
+#ifdef POWERPC_PERF_USE_PMC
+            perfdata_miss[i][powerpc_data_min] = 0xFFFFFFFFFFFFFFFF;
+            perfdata_miss[i][powerpc_data_max] = 0x0000000000000000;
+            perfdata_miss[i][powerpc_data_sum] = 0x0000000000000000;
+            perfdata_miss[i][powerpc_data_num] = 0x0000000000000000;
+#endif
           }
         }
 #endif