fixed vob video issue

git-svn-id: svn://svn.handbrake.fr/HandBrake/branches/opencl@5378 b64f7644-9d1e-0410-96f1-a4d463321fa5

3 files changed, 73 insertions, 49 deletions

diff --git a/libhb/decavcodec.c b/libhb/decavcodec.c
index 3df00006a..9c1bc1ef9 100644
--- a/libhb/decavcodec.c
+++ b/libhb/decavcodec.c
@@ -622,57 +622,55 @@ static hb_buffer_t *copy_frame( hb_work_private_t *pv, AVFrame *frame )
         return buf;
     }
     else
-    {
 #endif
-    hb_buffer_t *buf = hb_video_buffer_init( w, h );
-    uint8_t *dst = buf->data;
-
-    if (context->pix_fmt != AV_PIX_FMT_YUV420P || w != context->width ||
-        h != context->height)
     {
-        // have to convert to our internal color space and/or rescale
-        AVPicture dstpic;
-        hb_avpicture_fill(&dstpic, buf);
+        hb_buffer_t *buf = hb_video_buffer_init( w, h );
+        uint8_t *dst = buf->data;
 
-        if (pv->sws_context == NULL            ||
-            pv->sws_width   != context->width  ||
-            pv->sws_height  != context->height ||
-            pv->sws_pix_fmt != context->pix_fmt)
+        if (context->pix_fmt != AV_PIX_FMT_YUV420P || w != context->width ||
+            h != context->height)
         {
-            if (pv->sws_context != NULL)
-                sws_freeContext(pv->sws_context);
-            pv->sws_context = hb_sws_get_context(context->width,
-                                                 context->height,
-                                                 context->pix_fmt,
-                                                 w, h, AV_PIX_FMT_YUV420P,
-                                                  SWS_LANCZOS|SWS_ACCURATE_RND);
-            pv->sws_width   = context->width;
-            pv->sws_height  = context->height;
-            pv->sws_pix_fmt = context->pix_fmt;
+            // have to convert to our internal color space and/or rescale
+            AVPicture dstpic;
+            hb_avpicture_fill(&dstpic, buf);
+
+            if (pv->sws_context == NULL            ||
+                pv->sws_width   != context->width  ||
+                pv->sws_height  != context->height ||
+                pv->sws_pix_fmt != context->pix_fmt)
+            {
+                if (pv->sws_context != NULL)
+                    sws_freeContext(pv->sws_context);
+                pv->sws_context = hb_sws_get_context(context->width,
+                                                     context->height,
+                                                     context->pix_fmt,
+                                                     w, h, AV_PIX_FMT_YUV420P,
+                                                     SWS_LANCZOS|SWS_ACCURATE_RND);
+                pv->sws_width   = context->width;
+                pv->sws_height  = context->height;
+                pv->sws_pix_fmt = context->pix_fmt;
+            }
+            sws_scale(pv->sws_context,
+                      (const uint8_t* const *)frame->data, frame->linesize,
+                      0, context->height, dstpic.data, dstpic.linesize);
         }
-        sws_scale(pv->sws_context,
-                  (const uint8_t* const *)frame->data, frame->linesize,
-                  0, context->height, dstpic.data, dstpic.linesize);
-    }
-    else
-    {
-        w = buf->plane[0].stride;
-        h = buf->plane[0].height;
-        dst = buf->plane[0].data;
-        copy_plane( dst, frame->data[0], w, frame->linesize[0], h );
-        w = buf->plane[1].stride;
-        h = buf->plane[1].height;
-        dst = buf->plane[1].data;
-        copy_plane( dst, frame->data[1], w, frame->linesize[1], h );
-        w = buf->plane[2].stride;
-        h = buf->plane[2].height;
-        dst = buf->plane[2].data;
-        copy_plane( dst, frame->data[2], w, frame->linesize[2], h );
+        else
+        {
+            w = buf->plane[0].stride;
+            h = buf->plane[0].height;
+            dst = buf->plane[0].data;
+            copy_plane( dst, frame->data[0], w, frame->linesize[0], h );
+            w = buf->plane[1].stride;
+            h = buf->plane[1].height;
+            dst = buf->plane[1].data;
+            copy_plane( dst, frame->data[1], w, frame->linesize[1], h );
+            w = buf->plane[2].stride;
+            h = buf->plane[2].height;
+            dst = buf->plane[2].data;
+            copy_plane( dst, frame->data[2], w, frame->linesize[2], h );
+        }
+        return buf;
     }
-    return buf;
-#ifdef USE_HWD
-}
-#endif
 }
 
 #ifdef USE_HWD
diff --git a/libhb/oclnv12toyuv.c b/libhb/oclnv12toyuv.c
index 26b725bfa..a540e49c1 100644
--- a/libhb/oclnv12toyuv.c
+++ b/libhb/oclnv12toyuv.c
@@ -100,6 +100,24 @@ static int hb_init_nv12toyuv_ocl( KernelEnv *kenv, int w, int h, hb_va_dxva2_t *
     return 0;
 }
 
+static uint8_t *copy_plane( uint8_t *dst, uint8_t* src, int dstride, int sstride,
+                            int h )
+{
+    if ( dstride == sstride )
+    {
+        memcpy( dst, src, dstride * h );
+        return dst + dstride * h;
+    }
+    int lbytes = dstride <= sstride? dstride : sstride;
+    while ( --h >= 0 )
+    {
+        memcpy( dst, src, lbytes );
+        src += sstride;
+        dst += dstride;
+    }
+    return dst;
+}
+
 /**
  * Run nv12 to yuv kernel.
  */
@@ -158,9 +176,17 @@ static int hb_nv12toyuv( void **userdata, KernelEnv *kenv )
         AVPicture           pic_crop;
         clEnqueueReadBuffer( kenv->command_queue, dxva2->cl_mem_yuv, CL_TRUE, 0, in_bytes, dxva2->nv12toyuv_tmp_out, 0, NULL, NULL );
         hb_buffer_t *in = hb_video_buffer_init( w, h );
-        memcpy( in->plane[0].data, dxva2->nv12toyuv_tmp_out, w * h );
-        memcpy( in->plane[1].data, dxva2->nv12toyuv_tmp_out + w * h, ( w * h )>>2 );
-        memcpy( in->plane[2].data, dxva2->nv12toyuv_tmp_out + w * h + ( ( w * h )>>2 ), ( w * h )>>2 );
+		
+        int wmp = in->plane[0].stride;
+        int hmp = in->plane[0].height;
+        copy_plane( in->plane[0].data, dxva2->nv12toyuv_tmp_out, wmp, w, hmp );
+        wmp = in->plane[1].stride;
+        hmp = in->plane[1].height;
+        copy_plane( in->plane[1].data, dxva2->nv12toyuv_tmp_out + w * h, wmp, w>>1, hmp );
+        wmp = in->plane[2].stride;
+        hmp = in->plane[2].height;
+        copy_plane( in->plane[2].data, dxva2->nv12toyuv_tmp_out + w * h +( ( w * h )>>2 ), wmp, w>>1, hmp );
+
         hb_avpicture_fill( &pic_in, in );
         av_picture_crop( &pic_crop, &pic_in, in->f.fmt, crop[0], crop[2] );
         int i, ww = w - ( crop[2] + crop[3] ), hh = h - ( crop[0] + crop[1] );
diff --git a/libhb/vadxva2.c b/libhb/vadxva2.c
index 369f51b48..9cd780b3c 100644
--- a/libhb/vadxva2.c
+++ b/libhb/vadxva2.c
@@ -592,7 +592,7 @@ void hb_init_filter( cl_mem src, int srcwidth, int srcheight, uint8_t* dst, int
 
     cl_mem cl_outbuf;
 
-    if( !hb_create_buffer( &(cl_outbuf), CL_MEM_WRITE_ONLY, STEP ) )
+    if( !hb_create_buffer( &(cl_outbuf), CL_MEM_WRITE_ONLY, OUTSTEP ) )
     {   
         hb_log("av_create_buffer cl_outbuf Error");
         return;