Feed MFC PAK with correct MV of VME output on haswell

This is helpful to improve the video compress rate of H264 encoding.

Signed-off-by: Zhao Yakui <yakui.zhao@intel.com>

diff --git a/src/gen75_mfc.c b/src/gen75_mfc.c
index 6660940..08fd59c 100644 (file)
--- a/src/gen75_mfc.c
+++ b/src/gen75_mfc.c
@@ -76,6 +76,8 @@ static struct i965_kernel gen75_mfc_kernels[] = {
 
 #define                INTER_MODE_MASK         0x03
 #define                INTER_8X8               0x03
+#define                INTER_16X8              0x01
+#define                INTER_8X16              0x02
 #define                SUBMB_SHAPE_MASK        0x00FF00
 
 #define                INTER_MV8               (4 << 20)
@@ -1053,6 +1055,41 @@ gen75_mfc_avc_pak_object_inter(VADriverContextP ctx, int x, int y, int end_mb, i
        unsigned int inter_msg = 0;
     if (batch == NULL)
         batch = encoder_context->base.batch;
+    {
+#define MSG_MV_OFFSET  4
+       unsigned int *mv_ptr;
+       mv_ptr = msg + MSG_MV_OFFSET;
+       /* MV of VME output is based on 16 sub-blocks. So it is necessary
+         * to convert them to be compatible with the format of AVC_PAK
+         * command.
+         */
+       if ((msg[0] & INTER_MODE_MASK) == INTER_8X16) {
+               /* MV[0] and MV[2] are replicated */
+               mv_ptr[4] = mv_ptr[0];
+               mv_ptr[5] = mv_ptr[1];
+               mv_ptr[2] = mv_ptr[8];
+               mv_ptr[3] = mv_ptr[9];
+               mv_ptr[6] = mv_ptr[8]; 
+               mv_ptr[7] = mv_ptr[9]; 
+       } else if ((msg[0] & INTER_MODE_MASK) == INTER_16X8) {
+               /* MV[0] and MV[1] are replicated */
+               mv_ptr[2] = mv_ptr[0];  
+               mv_ptr[3] = mv_ptr[1];
+               mv_ptr[4] = mv_ptr[16]; 
+               mv_ptr[5] = mv_ptr[17]; 
+               mv_ptr[6] = mv_ptr[24];
+               mv_ptr[7] = mv_ptr[25];
+       } else if (((msg[0] & INTER_MODE_MASK) == INTER_8X8) &&
+                       !(msg[1] & SUBMB_SHAPE_MASK)) {
+               /* Don't touch MV[0] or MV[1] */
+               mv_ptr[2] = mv_ptr[8];
+               mv_ptr[3] = mv_ptr[9];
+               mv_ptr[4] = mv_ptr[16];
+               mv_ptr[5] = mv_ptr[17];
+               mv_ptr[6] = mv_ptr[24];
+               mv_ptr[7] = mv_ptr[25];
+       }
+    }
 
     BEGIN_BCS_BATCH(batch, len_in_dwords);