[FFmpeg-devel] [PATCH 1/2] PPC: simplify loading some values into altivec registers

[Mans Rullgard]

Instead of filling a local array with the desired value and loading it,
load a single element and vec_splat() it to fill the vector.
---
 libavcodec/ppc/gmc_altivec.c       |    7 +++----
 libavcodec/ppc/mpegvideo_altivec.c |   25 +++++++------------------
 2 files changed, 10 insertions(+), 22 deletions(-)

diff --git a/libavcodec/ppc/gmc_altivec.c b/libavcodec/ppc/gmc_altivec.c
index c77c716..70c0cf9 100644
--- a/libavcodec/ppc/gmc_altivec.c
+++ b/libavcodec/ppc/gmc_altivec.c
@@ -23,6 +23,7 @@
 #include "libavcodec/dsputil.h"
 #include "dsputil_ppc.h"
 #include "util_altivec.h"
+#include "types_altivec.h"
 
 /*
   altivec-enhanced gmc1. ATM this code assume stride is a multiple of 8,
@@ -32,9 +33,7 @@
 void gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, int stride, int h, int x16, int y16, int rounder)
 {
 POWERPC_PERF_DECLARE(altivec_gmc1_num, GMC1_PERF_COND);
-    const DECLARE_ALIGNED_16(unsigned short, rounder_a[8]) =
-        {rounder, rounder, rounder, rounder,
-         rounder, rounder, rounder, rounder};
+    const DECLARE_ALIGNED_16(unsigned short, rounder_a) = rounder;
     const DECLARE_ALIGNED_16(unsigned short, ABCD[8]) =
         {
             (16-x16)*(16-y16), /* A */
@@ -60,7 +59,7 @@ POWERPC_PERF_START_COUNT(altivec_gmc1_num, GMC1_PERF_COND);
     Cv = vec_splat(tempA, 2);
     Dv = vec_splat(tempA, 3);
 
-    rounderV = vec_ld(0, (unsigned short*)rounder_a);
+    rounderV = vec_splat((vec_u16)vec_lde(0, &rounder_a), 0);
 
     // we'll be able to pick-up our 9 char elements
     // at src from those 32 bytes
diff --git a/libavcodec/ppc/mpegvideo_altivec.c b/libavcodec/ppc/mpegvideo_altivec.c
index 74775f0..8348e68 100644
--- a/libavcodec/ppc/mpegvideo_altivec.c
+++ b/libavcodec/ppc/mpegvideo_altivec.c
@@ -28,6 +28,8 @@
 
 #include "dsputil_ppc.h"
 #include "util_altivec.h"
+#include "types_altivec.h"
+
 // Swaps two variables (used for altivec registers)
 #define SWAP(a,b) \
 do { \
@@ -504,29 +506,16 @@ POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1);
 
     {
         register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
-        DECLARE_ALIGNED_16(short, qmul8[]) =
-            {
-              qmul, qmul, qmul, qmul,
-              qmul, qmul, qmul, qmul
-            };
-        DECLARE_ALIGNED_16(short, qadd8[]) =
-            {
-              qadd, qadd, qadd, qadd,
-              qadd, qadd, qadd, qadd
-            };
-        DECLARE_ALIGNED_16(short, nqadd8[]) =
-            {
-              -qadd, -qadd, -qadd, -qadd,
-              -qadd, -qadd, -qadd, -qadd
-            };
+        DECLARE_ALIGNED_16(short, qmul8) = qmul;
+        DECLARE_ALIGNED_16(short, qadd8) = qadd;
         register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
         register vector bool short blockv_null, blockv_neg;
         register short backup_0 = block[0];
         register int j = 0;
 
-        qmulv = vec_ld(0, qmul8);
-        qaddv = vec_ld(0, qadd8);
-        nqaddv = vec_ld(0, nqadd8);
+        qmulv = vec_splat((vec_s16)vec_lde(0, &qmul8), 0);
+        qaddv = vec_splat((vec_s16)vec_lde(0, &qadd8), 0);
+        nqaddv = vec_sub(vczero, qaddv);
 
 #if 0   // block *is* 16 bytes-aligned, it seems.
         // first make sure block[j] is 16 bytes-aligned
-- 
1.6.4