Make LOAD4/STORE4 macros more generic.

Patch by Victor Pollex victor pollex web de
Original thread: [PATCH] mmx implementation of vc-1 inverse transformations
Date: 06/21/2008 03:37 PM

Originally committed as revision 14108 to svn://svn.ffmpeg.org/ffmpeg/trunk

2 files changed, 18 insertions, 18 deletions

diff --git a/libavcodec/i386/dsputil_mmx.h b/libavcodec/i386/dsputil_mmx.h
index 142854684b..d944dbd43c 100644
--- a/libavcodec/i386/dsputil_mmx.h
+++ b/libavcodec/i386/dsputil_mmx.h
@@ -57,6 +57,18 @@ extern const uint64_t ff_pb_FC;
 extern const double ff_pd_1[2];
 extern const double ff_pd_2[2];
 
+#define LOAD4(stride,in,a,b,c,d)\
+    "movq 0*"#stride"+"#in", "#a"\n\t"\
+    "movq 1*"#stride"+"#in", "#b"\n\t"\
+    "movq 2*"#stride"+"#in", "#c"\n\t"\
+    "movq 3*"#stride"+"#in", "#d"\n\t"
+
+#define STORE4(stride,out,a,b,c,d)\
+    "movq "#a", 0*"#stride"+"#out"\n\t"\
+    "movq "#b", 1*"#stride"+"#out"\n\t"\
+    "movq "#c", 2*"#stride"+"#out"\n\t"\
+    "movq "#d", 3*"#stride"+"#out"\n\t"
+
 /* in/out: mma=mma+mmb, mmb=mmb-mma */
 #define SUMSUB_BA( a, b ) \
     "paddw "#b", "#a" \n\t"\
diff --git a/libavcodec/i386/dsputilenc_mmx.c b/libavcodec/i386/dsputilenc_mmx.c
index e7893de487..be423f8f24 100644
--- a/libavcodec/i386/dsputilenc_mmx.c
+++ b/libavcodec/i386/dsputilenc_mmx.c
@@ -998,18 +998,6 @@ static void sub_hfyu_median_prediction_mmx2(uint8_t *dst, uint8_t *src1, uint8_t
     "paddusw %%xmm1, %%xmm0           \n\t"
 #endif
 
-#define LOAD4(o, a, b, c, d)\
-    "movq "#o"(%1),    "#a"           \n\t"\
-    "movq "#o"+8(%1),  "#b"           \n\t"\
-    "movq "#o"+16(%1), "#c"           \n\t"\
-    "movq "#o"+24(%1), "#d"           \n\t"\
-
-#define STORE4(o, a, b, c, d)\
-    "movq "#a", "#o"(%1)              \n\t"\
-    "movq "#b", "#o"+8(%1)            \n\t"\
-    "movq "#c", "#o"+16(%1)           \n\t"\
-    "movq "#d", "#o"+24(%1)           \n\t"\
-
 /* FIXME: HSUM_* saturates at 64k, while an 8x8 hadamard or dct block can get up to
  * about 100k on extreme inputs. But that's very unlikely to occur in natural video,
  * and it's even more unlikely to not have any alternative mvs/modes with lower cost. */
@@ -1053,11 +1041,11 @@ static int hadamard8_diff_##cpu(void *s, uint8_t *src1, uint8_t *src2, int strid
         "movq %%mm7, 96(%1)             \n\t"\
 \
         TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)\
-        STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2)\
+        STORE4(8,  0(%1), %%mm0, %%mm3, %%mm7, %%mm2)\
 \
         "movq 96(%1), %%mm7             \n\t"\
         TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)\
-        STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6)\
+        STORE4(8, 64(%1), %%mm4, %%mm7, %%mm0, %%mm6)\
 \
         : "=r" (sum)\
         : "r"(temp)\
@@ -1071,7 +1059,7 @@ static int hadamard8_diff_##cpu(void *s, uint8_t *src1, uint8_t *src2, int strid
         "movq %%mm7, 96(%1)             \n\t"\
 \
         TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)\
-        STORE4(32, %%mm0, %%mm3, %%mm7, %%mm2)\
+        STORE4(8, 32(%1), %%mm0, %%mm3, %%mm7, %%mm2)\
 \
         "movq 96(%1), %%mm7             \n\t"\
         TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)\
@@ -1079,7 +1067,7 @@ static int hadamard8_diff_##cpu(void *s, uint8_t *src1, uint8_t *src2, int strid
         "movq %%mm6, %%mm7              \n\t"\
         "movq %%mm0, %%mm6              \n\t"\
 \
-        LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3)\
+        LOAD4(8, 64(%1), %%mm0, %%mm1, %%mm2, %%mm3)\
 \
         HADAMARD48\
         "movq %%mm7, 64(%1)             \n\t"\
@@ -1095,8 +1083,8 @@ static int hadamard8_diff_##cpu(void *s, uint8_t *src1, uint8_t *src2, int strid
         "paddusw %%mm1, %%mm0           \n\t"\
         "movq %%mm0, 64(%1)             \n\t"\
 \
-        LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)\
-        LOAD4(32, %%mm4, %%mm5, %%mm6, %%mm7)\
+        LOAD4(8,  0(%1), %%mm0, %%mm1, %%mm2, %%mm3)\
+        LOAD4(8, 32(%1), %%mm4, %%mm5, %%mm6, %%mm7)\
 \
         HADAMARD48\
         "movq %%mm7, (%1)               \n\t"\