vp8/armv6: mc: avoid boolean expression in calculation

[ffmpeg] / libavcodec / simple_idct_template.c

diff --git a/libavcodec/simple_idct_template.c b/libavcodec/simple_idct_template.c
index 6d3f6f764de52d840d54a3d161337b98de9c244e..171cdf65ff1147f6cb3f3268acf7cbb6b64b5d24 100644 (file)
--- a/libavcodec/simple_idct_template.c
+++ b/libavcodec/simple_idct_template.c
@@ -25,10 +25,10 @@
  * simpleidct in C.
  */
 
-/*
-  based upon some outcommented c code from mpeg2dec (idct_mmx.c
-  written by Aaron Holtzman <aholtzma@ess.engr.uvic.ca>)
- */
+/* Based upon some commented-out C code from mpeg2dec (idct_mmx.c
+ * written by Aaron Holtzman <aholtzma@ess.engr.uvic.ca>). */
+
+#include "simple_idct.h"
 
 #include "bit_depth_template.c"
 
@@ -85,29 +85,41 @@
 
 #endif
 
-static inline void FUNC(idctRowCondDC)(DCTELEM *row)
+static inline void FUNC(idctRowCondDC)(int16_t *row, int extra_shift)
 {
     int a0, a1, a2, a3, b0, b1, b2, b3;
 
 #if HAVE_FAST_64BIT
 #define ROW0_MASK (0xffffLL << 48 * HAVE_BIGENDIAN)
-    if (((((uint64_t *)row)[0] & ~ROW0_MASK) | ((uint64_t *)row)[1]) == 0) {
-        uint64_t temp = (row[0] << DC_SHIFT) & 0xffff;
-        temp += temp << 16;
-        temp += temp << 32;
-        ((uint64_t *)row)[0] = temp;
-        ((uint64_t *)row)[1] = temp;
+    if (((AV_RN64A(row) & ~ROW0_MASK) | AV_RN64A(row+4)) == 0) {
+        uint64_t temp;
+        if (DC_SHIFT - extra_shift > 0) {
+            temp = (row[0] * (1 << (DC_SHIFT - extra_shift))) & 0xffff;
+        } else {
+            temp = (row[0] >> (extra_shift - DC_SHIFT)) & 0xffff;
+        }
+        temp += temp * (1 << 16);
+        temp += temp * ((uint64_t) 1 << 32);
+        AV_WN64A(row, temp);
+        AV_WN64A(row + 4, temp);
         return;
     }
 #else
-    if (!(((uint32_t*)row)[1] |
-          ((uint32_t*)row)[2] |
-          ((uint32_t*)row)[3] |
+    if (!(AV_RN32A(row+2) |
+          AV_RN32A(row+4) |
+          AV_RN32A(row+6) |
           row[1])) {
-        uint32_t temp = (row[0] << DC_SHIFT) & 0xffff;
-        temp += temp << 16;
-        ((uint32_t*)row)[0]=((uint32_t*)row)[1] =
-            ((uint32_t*)row)[2]=((uint32_t*)row)[3] = temp;
+        uint32_t temp;
+        if (DC_SHIFT - extra_shift > 0) {
+            temp = (row[0] * (1 << (DC_SHIFT - extra_shift))) & 0xffff;
+        } else {
+            temp = (row[0] >> (extra_shift - DC_SHIFT)) & 0xffff;
+        }
+        temp += temp * (1 << 16);
+        AV_WN32A(row, temp);
+        AV_WN32A(row+2, temp);
+        AV_WN32A(row+4, temp);
+        AV_WN32A(row+6, temp);
         return;
     }
 #endif
@@ -150,14 +162,14 @@ static inline void FUNC(idctRowCondDC)(DCTELEM *row)
         MAC(b3, -W1, row[7]);
     }
 
-    row[0] = (a0 + b0) >> ROW_SHIFT;
-    row[7] = (a0 - b0) >> ROW_SHIFT;
-    row[1] = (a1 + b1) >> ROW_SHIFT;
-    row[6] = (a1 - b1) >> ROW_SHIFT;
-    row[2] = (a2 + b2) >> ROW_SHIFT;
-    row[5] = (a2 - b2) >> ROW_SHIFT;
-    row[3] = (a3 + b3) >> ROW_SHIFT;
-    row[4] = (a3 - b3) >> ROW_SHIFT;
+    row[0] = (a0 + b0) >> (ROW_SHIFT + extra_shift);
+    row[7] = (a0 - b0) >> (ROW_SHIFT + extra_shift);
+    row[1] = (a1 + b1) >> (ROW_SHIFT + extra_shift);
+    row[6] = (a1 - b1) >> (ROW_SHIFT + extra_shift);
+    row[2] = (a2 + b2) >> (ROW_SHIFT + extra_shift);
+    row[5] = (a2 - b2) >> (ROW_SHIFT + extra_shift);
+    row[3] = (a3 + b3) >> (ROW_SHIFT + extra_shift);
+    row[4] = (a3 - b3) >> (ROW_SHIFT + extra_shift);
 }
 
 #define IDCT_COLS do {                                  \
@@ -211,56 +223,54 @@ static inline void FUNC(idctRowCondDC)(DCTELEM *row)
     } while (0)
 
 static inline void FUNC(idctSparseColPut)(pixel *dest, int line_size,
-                                          DCTELEM *col)
+                                          int16_t *col)
 {
     int a0, a1, a2, a3, b0, b1, b2, b3;
-    INIT_CLIP;
 
     IDCT_COLS;
 
-    dest[0] = CLIP((a0 + b0) >> COL_SHIFT);
+    dest[0] = av_clip_pixel((a0 + b0) >> COL_SHIFT);
     dest += line_size;
-    dest[0] = CLIP((a1 + b1) >> COL_SHIFT);
+    dest[0] = av_clip_pixel((a1 + b1) >> COL_SHIFT);
     dest += line_size;
-    dest[0] = CLIP((a2 + b2) >> COL_SHIFT);
+    dest[0] = av_clip_pixel((a2 + b2) >> COL_SHIFT);
     dest += line_size;
-    dest[0] = CLIP((a3 + b3) >> COL_SHIFT);
+    dest[0] = av_clip_pixel((a3 + b3) >> COL_SHIFT);
     dest += line_size;
-    dest[0] = CLIP((a3 - b3) >> COL_SHIFT);
+    dest[0] = av_clip_pixel((a3 - b3) >> COL_SHIFT);
     dest += line_size;
-    dest[0] = CLIP((a2 - b2) >> COL_SHIFT);
+    dest[0] = av_clip_pixel((a2 - b2) >> COL_SHIFT);
     dest += line_size;
-    dest[0] = CLIP((a1 - b1) >> COL_SHIFT);
+    dest[0] = av_clip_pixel((a1 - b1) >> COL_SHIFT);
     dest += line_size;
-    dest[0] = CLIP((a0 - b0) >> COL_SHIFT);
+    dest[0] = av_clip_pixel((a0 - b0) >> COL_SHIFT);
 }
 
 static inline void FUNC(idctSparseColAdd)(pixel *dest, int line_size,
-                                          DCTELEM *col)
+                                          int16_t *col)
 {
     int a0, a1, a2, a3, b0, b1, b2, b3;
-    INIT_CLIP;
 
     IDCT_COLS;
 
-    dest[0] = CLIP(dest[0] + ((a0 + b0) >> COL_SHIFT));
+    dest[0] = av_clip_pixel(dest[0] + ((a0 + b0) >> COL_SHIFT));
     dest += line_size;
-    dest[0] = CLIP(dest[0] + ((a1 + b1) >> COL_SHIFT));
+    dest[0] = av_clip_pixel(dest[0] + ((a1 + b1) >> COL_SHIFT));
     dest += line_size;
-    dest[0] = CLIP(dest[0] + ((a2 + b2) >> COL_SHIFT));
+    dest[0] = av_clip_pixel(dest[0] + ((a2 + b2) >> COL_SHIFT));
     dest += line_size;
-    dest[0] = CLIP(dest[0] + ((a3 + b3) >> COL_SHIFT));
+    dest[0] = av_clip_pixel(dest[0] + ((a3 + b3) >> COL_SHIFT));
     dest += line_size;
-    dest[0] = CLIP(dest[0] + ((a3 - b3) >> COL_SHIFT));
+    dest[0] = av_clip_pixel(dest[0] + ((a3 - b3) >> COL_SHIFT));
     dest += line_size;
-    dest[0] = CLIP(dest[0] + ((a2 - b2) >> COL_SHIFT));
+    dest[0] = av_clip_pixel(dest[0] + ((a2 - b2) >> COL_SHIFT));
     dest += line_size;
-    dest[0] = CLIP(dest[0] + ((a1 - b1) >> COL_SHIFT));
+    dest[0] = av_clip_pixel(dest[0] + ((a1 - b1) >> COL_SHIFT));
     dest += line_size;
-    dest[0] = CLIP(dest[0] + ((a0 - b0) >> COL_SHIFT));
+    dest[0] = av_clip_pixel(dest[0] + ((a0 - b0) >> COL_SHIFT));
 }
 
-static inline void FUNC(idctSparseCol)(DCTELEM *col)
+static inline void FUNC(idctSparseCol)(int16_t *col)
 {
     int a0, a1, a2, a3, b0, b1, b2, b3;
 
@@ -276,7 +286,7 @@ static inline void FUNC(idctSparseCol)(DCTELEM *col)
     col[56] = ((a0 - b0) >> COL_SHIFT);
 }
 
-void FUNC(ff_simple_idct_put)(uint8_t *dest_, int line_size, DCTELEM *block)
+void FUNC(ff_simple_idct_put)(uint8_t *dest_, int line_size, int16_t *block)
 {
     pixel *dest = (pixel *)dest_;
     int i;
@@ -284,13 +294,13 @@ void FUNC(ff_simple_idct_put)(uint8_t *dest_, int line_size, DCTELEM *block)
     line_size /= sizeof(pixel);
 
     for (i = 0; i < 8; i++)
-        FUNC(idctRowCondDC)(block + i*8);
+        FUNC(idctRowCondDC)(block + i*8, 0);
 
     for (i = 0; i < 8; i++)
         FUNC(idctSparseColPut)(dest + i, line_size, block + i);
 }
 
-void FUNC(ff_simple_idct_add)(uint8_t *dest_, int line_size, DCTELEM *block)
+void FUNC(ff_simple_idct_add)(uint8_t *dest_, int line_size, int16_t *block)
 {
     pixel *dest = (pixel *)dest_;
     int i;
@@ -298,18 +308,18 @@ void FUNC(ff_simple_idct_add)(uint8_t *dest_, int line_size, DCTELEM *block)
     line_size /= sizeof(pixel);
 
     for (i = 0; i < 8; i++)
-        FUNC(idctRowCondDC)(block + i*8);
+        FUNC(idctRowCondDC)(block + i*8, 0);
 
     for (i = 0; i < 8; i++)
         FUNC(idctSparseColAdd)(dest + i, line_size, block + i);
 }
 
-void FUNC(ff_simple_idct)(DCTELEM *block)
+void FUNC(ff_simple_idct)(int16_t *block)
 {
     int i;
 
     for (i = 0; i < 8; i++)
-        FUNC(idctRowCondDC)(block + i*8);
+        FUNC(idctRowCondDC)(block + i*8, 0);
 
     for (i = 0; i < 8; i++)
         FUNC(idctSparseCol)(block + i);