H.264: split luma dc idct out and implement MMX/SSE2 versions

About 2.5x the speed.

NOTE: the way that the asm code handles large qmuls is a bit suboptimal.
If x264-style dequant was used (separate shift and qmul values), it might
be possible to get some extra speed.

Originally committed as revision 26336 to svn://svn.ffmpeg.org/ffmpeg/trunk

1 files changed, 35 insertions, 0 deletions

diff --git a/libavcodec/h264idct.c b/libavcodec/h264idct.c
index 86c5ef2559..f5b05ac24f 100644
--- a/libavcodec/h264idct.c
+++ b/libavcodec/h264idct.c
@@ -216,3 +216,38 @@ void ff_h264_idct_add8_c(uint8_t **dest, const int *block_offset, DCTELEM *block
             ff_h264_idct_dc_add_c(dest[(i&4)>>2] + block_offset[i], block + i*16, stride);
     }
 }
+/**
+ * IDCT transforms the 16 dc values and dequantizes them.
+ * @param qp quantization parameter
+ */
+void ff_h264_luma_dc_dequant_idct_c(DCTELEM *output, DCTELEM *input, int qmul){
+#define stride 16
+    int i;
+    int temp[16];
+    static const uint8_t x_offset[4]={0, 2*stride, 8*stride, 10*stride};
+
+    for(i=0; i<4; i++){
+        const int z0= input[4*i+0] + input[4*i+1];
+        const int z1= input[4*i+0] - input[4*i+1];
+        const int z2= input[4*i+2] - input[4*i+3];
+        const int z3= input[4*i+2] + input[4*i+3];
+
+        temp[4*i+0]= z0+z3;
+        temp[4*i+1]= z0-z3;
+        temp[4*i+2]= z1-z2;
+        temp[4*i+3]= z1+z2;
+    }
+
+    for(i=0; i<4; i++){
+        const int offset= x_offset[i];
+        const int z0= temp[4*0+i] + temp[4*2+i];
+        const int z1= temp[4*0+i] - temp[4*2+i];
+        const int z2= temp[4*1+i] - temp[4*3+i];
+        const int z3= temp[4*1+i] + temp[4*3+i];
+
+        output[stride* 0+offset]= ((((z0 + z3)*qmul + 128 ) >> 8));
+        output[stride* 1+offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
+        output[stride* 4+offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
+        output[stride* 5+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
+    }
+}