1. The pitches for y/u/v surfaces are 1024bytes aligned.

2. clean code

6 files changed, 73 insertions, 183 deletions

diff --git a/src/i830.h b/src/i830.h
index cde15ad6..de377499 100644
--- a/src/i830.h
+++ b/src/i830.h
@@ -313,7 +313,8 @@ typedef struct _I830Rec {
    /* For XvMC */
    void *xvmc;
    Bool XvMCEnabled;
-
+   Bool IsXvMCSurface;
+ 
    XF86ModReqInfo shadowReq; /* to test for later libshadow */
    Rotation rotation;
    void (*PointerMoved)(int, int, int);
diff --git a/src/i830_video.c b/src/i830_video.c
index f7f2f740..015b4268 100644
--- a/src/i830_video.c
+++ b/src/i830_video.c
@@ -2382,8 +2382,14 @@ I830PutImage(ScrnInfoPtr pScrn,
     switch (id) {
     case FOURCC_YV12:
     case FOURCC_I420:
-	srcPitch = (width + 3) & ~3;
-	srcPitch2 = ((width >> 1) + 3) & ~3;
+        if (pI830->IsXvMCSurface) {
+            srcPitch = (width + 0x400) & ~0x3ff;
+            srcPitch2 = ((width >> 1) + 0x400) & ~0x3ff;
+        } else {
+            srcPitch = (width + 0x3) & ~0x3;
+            srcPitch2 = ((width >> 1) + 0x3) & ~0x3;
+        }
+
 	if (pPriv->textured && IS_I965G(pI830))
 	    destId = FOURCC_YUY2;
 	break;
diff --git a/src/i915_hwmc.c b/src/i915_hwmc.c
index 0f57a5fb..b5bd13e1 100644
--- a/src/i915_hwmc.c
+++ b/src/i915_hwmc.c
@@ -191,41 +191,6 @@ static XF86MCAdaptorPtr ppAdapt[1] =
     (XF86MCAdaptorPtr)&pAdapt
 };
 
-static unsigned int stride(int w)
-{
-    return (w + 3) & ~3;
-}
-
-static unsigned long size_y420(int w, int h)
-{
-   unsigned cpp = 1;
-   unsigned yPitch = stride(w) * cpp;
-   
-   return h * yPitch;
-}
-
-static unsigned long size_uv420(int w, int h)
-{
-   unsigned cpp = 1;
-   unsigned uvPitch = stride(w >> 1) * cpp;
-
-   return h / 2 * uvPitch;
-}
-
-static unsigned long size_yuv420(int w, int h)
-{
-    unsigned cpp = 1;
-    unsigned yPitch = stride(w) * cpp;
-    unsigned uvPitch = stride(w >> 1) * cpp;
-
-    return h * (yPitch + uvPitch);
-}
-
-static unsigned long size_xx44(int w, int h)
-{
-    return h * stride(w);
-}
-
 /*
  * Init and clean up the screen private parts of XvMC.
  */
@@ -641,7 +606,7 @@ static int I915XvMCCreateSurface(ScrnInfoPtr pScrn, XvMCSurfacePtr pSurf,
     }
 
     ctx = pSurf->context;
-    bufsize = size_yuv420(ctx->width, ctx->height);
+    bufsize = SIZE_YUV420(ctx->width, ctx->height);
 
     if (!i830_allocate_xvmc_buffer(pScrn, "XvMC surface", 
                                    &(sfpriv->surface), bufsize,
@@ -727,7 +692,7 @@ static int I915XvMCCreateSubpicture (ScrnInfoPtr pScrn, XvMCSubpicturePtr pSubp,
     }
 
     ctx = pSubp->context;
-    bufsize = size_xx44(ctx->width, ctx->height);
+    bufsize = SIZE_XX44(ctx->width, ctx->height);
 
     if (!i830_allocate_xvmc_buffer(pScrn, "XvMC surface", 
                                    &(sfpriv->surface), bufsize,
@@ -900,12 +865,13 @@ static int I915XvMCInterceptPutImage(ScrnInfoPtr pScrn, short src_x, short src_y
                                      short height, Bool sync, RegionPtr clipBoxes, pointer data,
                                      DrawablePtr pDraw)
 {
+    I830Ptr pI830 = I830PTR(pScrn);
     I830PortPrivPtr pPriv = (I830PortPrivPtr)data;
     I915XvMCXVPriv *vx = (I915XvMCXVPriv *)pPriv->xvmc_priv;
+    int ret;
 
     if (I830PTR(pScrn)->XvMCEnabled) {
         if (FOURCC_XVMC == id) {
-            I830Ptr pI830 = I830PTR(pScrn);
             I915XvMCPtr pXvMC = pI830->xvmc;
             I915XvMCCommandBuffer *i915XvMCData = (I915XvMCCommandBuffer *)buf;
             int i;
@@ -933,6 +899,7 @@ static int I915XvMCInterceptPutImage(ScrnInfoPtr pScrn, short src_x, short src_y
                    mem = pXvMC->sfprivs[i915XvMCData->srfNo]->surface;
                    buf = pI830->FbBase + mem->offset;
                    id = i915XvMCData->real_id;
+                   pI830->IsXvMCSurface = 1;
                    break;
                 }
 
@@ -944,8 +911,10 @@ static int I915XvMCInterceptPutImage(ScrnInfoPtr pScrn, short src_x, short src_y
         }
     }
 
-    return vx->PutImage(pScrn, src_x, src_y, drw_x, drw_y, src_w, src_h,
+    ret = vx->PutImage(pScrn, src_x, src_y, drw_x, drw_y, src_w, src_h,
                         drw_w, drw_h, id, buf, width, height, sync, clipBoxes, data, pDraw);
+    pI830->IsXvMCSurface = 0;
+    return ret;
 }
 
 /*********************************** Public Function **************************************/
diff --git a/src/i915_hwmc.h b/src/i915_hwmc.h
index a4b4220a..f2570ceb 100644
--- a/src/i915_hwmc.h
+++ b/src/i915_hwmc.h
@@ -1,6 +1,12 @@
 #ifndef _I915_HWMC_H
 #define _I915_HWMC_H
 
+#define STRIDE(w)               (((w) + 0x3ff) & ~0x3ff)
+#define SIZE_Y420(w, h)         (h * STRIDE(w))
+#define SIZE_UV420(w, h)        ((h >> 1) * STRIDE(w >> 1))
+#define SIZE_YUV420(w, h)       (h * (STRIDE(w) + STRIDE(w >> 1)))
+#define SIZE_XX44(w, h)         (h * STRIDE(w))
+
 #define FOURCC_XVMC     (('C' << 24) + ('M' << 16) + ('V' << 8) + 'X')
 
 #define I915_NUM_XVMC_ATTRIBUTES       0x02
diff --git a/src/xvmc/I915XvMC.c b/src/xvmc/I915XvMC.c
index b4d1f574..3b086c2b 100644
--- a/src/xvmc/I915XvMC.c
+++ b/src/xvmc/I915XvMC.c
@@ -34,10 +34,10 @@
 
 #define YOFFSET(surface)        (surface->srf.offset)
 #define UOFFSET(surface)        (surface->srf.offset + \
-                                 size_y(surface->width, surface->height) + \
-                                 size_uv(surface->width, surface->height))
+                                 SIZE_Y420(surface->width, surface->height) + \
+                                 SIZE_UV420(surface->width, surface->height))
 #define VOFFSET(surface)        (surface->srf.offset + \
-                                 size_y(surface->width, surface->height))
+                                 SIZE_Y420(surface->width, surface->height))
 
 /* Lookup tables to speed common calculations */
 _STATIC_ unsigned mb_bytes[] = {
@@ -173,67 +173,6 @@ _STATIC_ void UNLOCK_HARDWARE(i915XvMCContext *pI915XvMC)
     pthread_mutex_unlock(&pI915XvMC->ctxmutex);
 }
 
-_STATIC_ unsigned stride(int w)
-{
-    return (w + 3) & ~3;
-}
-
-_STATIC_ unsigned long size_y(int w, int h)
-{
-    unsigned cpp = 1;
-    unsigned yPitch = stride(w) * cpp;
-
-    return h * yPitch;
-}
-
-_STATIC_ unsigned long size_uv(int w, int h)
-{
-    unsigned cpp = 1;
-    unsigned uvPitch = stride(w >> 1) * cpp;
-    
-    return h / 2 * uvPitch;
-}
-
-_STATIC_ unsigned long size_yuv420(int w, int h)
-{
-    unsigned cpp = 1;
-    unsigned yPitch = stride(w) * cpp;
-    unsigned uvPitch = stride(w >> 1) * cpp;
-
-    return h * (yPitch + uvPitch);
-}
-
-_STATIC_ void i915_clear_surface(i915XvMCContext *pI915XvMC, i915XvMCSurface *surface)
-{
-    struct i915_color_blt color_blt;
-
-    if (surface->processing)
-        return;
-
-    memset(&color_blt, 0, sizeof(color_blt));
-    color_blt.dw0.type = CMD_2D;
-    color_blt.dw0.opcode = OPC_COLOR_BLT;
-    color_blt.dw0.bpp_mask = 0x1;
-    color_blt.dw0.length = 3;
-    color_blt.dw1.color_depth = 0;      /* 8bit */
-    color_blt.dw1.rop = 0xf0;
-    color_blt.dw1.pitch = surface->yStride;
-    color_blt.dw2.height = surface->height;
-    color_blt.dw2.width = surface->width;
-    color_blt.dw3.address = YOFFSET(surface);
-    color_blt.dw4.pattern = 0x7f;
-    intelBatchbufferData(pI915XvMC, &color_blt, sizeof(color_blt), 0);
-
-    color_blt.dw1.pitch = surface->uvStride;
-    color_blt.dw2.height = (surface->height >> 1);
-    color_blt.dw2.width = (surface->width >> 1);
-    color_blt.dw3.address = VOFFSET(surface);
-    intelBatchbufferData(pI915XvMC, &color_blt, sizeof(color_blt), 0);
-
-    color_blt.dw3.address = UOFFSET(surface);
-    intelBatchbufferData(pI915XvMC, &color_blt, sizeof(color_blt), 0);
-}
-
 _STATIC_ void i915_flush(i915XvMCContext *pI915XvMC, int map, int render)
 {
     struct i915_mi_flush mi_flush;
@@ -247,11 +186,6 @@ _STATIC_ void i915_flush(i915XvMCContext *pI915XvMC, int map, int render)
     intelBatchbufferData(pI915XvMC, &mi_flush, sizeof(mi_flush), 0);
 }
 
-_STATIC_ void i915_flush_with_flush_bit_clear(i915XvMCContext *pI915XvMC)
-{
-    i915_flush(pI915XvMC, 1, 0);
-}
-
 /* for MC picture rendering */
 _STATIC_ void i915_mc_static_indirect_state_buffer(XvMCContext *context, 
                                                    XvMCSurface *surface,
@@ -535,8 +469,6 @@ _STATIC_ void i915_mc_load_sis_msb_buffers(XvMCContext *context)
     unsigned size;
     int mem_select = 1;
 
-    i915_flush_with_flush_bit_clear(pI915XvMC);
-
     /* 3DSTATE_LOAD_INDIRECT */
     size = sizeof(*load_indirect) + sizeof(*sis) + sizeof(*msb);
     base = calloc(1, size);
@@ -668,10 +600,10 @@ _STATIC_ void i915_mc_mpeg_macroblock_1fbmv(XvMCContext *context, XvMCMacroBlock
     macroblock_1fbmv.header.dw1.coded_block_pattern = mb->coded_block_pattern; 
     macroblock_1fbmv.header.dw1.skipped_macroblocks = 0;      
 
-    fmv.s[0] = mb->PMV[0][0][1];
-    fmv.s[1] = mb->PMV[0][0][0];
-    bmv.s[0] = mb->PMV[0][1][1];
-    bmv.s[1] = mb->PMV[0][1][0];
+    fmv.s[0] = mb->PMV[0][0][0];
+    fmv.s[1] = mb->PMV[0][0][1];
+    bmv.s[0] = mb->PMV[0][1][0];
+    bmv.s[1] = mb->PMV[0][1][1];
 
     macroblock_1fbmv.dw2 = fmv.u[0];
     macroblock_1fbmv.dw3 = bmv.u[0];
@@ -707,34 +639,34 @@ _STATIC_ void i915_mc_mpeg_macroblock_2fbmv(XvMCContext *context, XvMCMacroBlock
     macroblock_2fbmv.header.dw1.coded_block_pattern = mb->coded_block_pattern;
     macroblock_2fbmv.header.dw1.skipped_macroblocks = 0;
 
-    fmv.s[0] = mb->PMV[0][0][1];
-    fmv.s[1] = mb->PMV[0][0][0];
-    fmv.s[2] = mb->PMV[1][0][1];
-    fmv.s[3] = mb->PMV[1][0][0];
-    bmv.s[0] = mb->PMV[0][1][1];
-    bmv.s[1] = mb->PMV[0][1][0];
-    bmv.s[2] = mb->PMV[1][1][1];
-    bmv.s[3] = mb->PMV[1][1][0];
+    fmv.s[0] = mb->PMV[0][0][0];
+    fmv.s[1] = mb->PMV[0][0][1];
+    fmv.s[2] = mb->PMV[1][0][0];
+    fmv.s[3] = mb->PMV[1][0][1];
+    bmv.s[0] = mb->PMV[0][1][0];
+    bmv.s[1] = mb->PMV[0][1][1];
+    bmv.s[2] = mb->PMV[1][1][0];
+    bmv.s[3] = mb->PMV[1][1][1];
 
     if ((ps & XVMC_FRAME_PICTURE) == XVMC_FRAME_PICTURE) {
         if ((mb->motion_type & 3) == XVMC_PREDICTION_FIELD) {
-            fmv.s[0] = mb->PMV[0][0][1] >> 1;
-            fmv.s[1] = mb->PMV[0][0][0];
-            fmv.s[2] = mb->PMV[1][0][1] >> 1;
-            fmv.s[3] = mb->PMV[1][0][0];
-            bmv.s[0] = mb->PMV[0][1][1] >> 1;
-            bmv.s[1] = mb->PMV[0][1][0];
-            bmv.s[2] = mb->PMV[1][1][1] >> 1;
-            bmv.s[3] = mb->PMV[1][1][0];
+            fmv.s[0] = mb->PMV[0][0][0];
+            fmv.s[1] = mb->PMV[0][0][1] >> 1;
+            fmv.s[2] = mb->PMV[1][0][0];
+            fmv.s[3] = mb->PMV[1][0][1] >> 1;
+            bmv.s[0] = mb->PMV[0][1][0];
+            bmv.s[1] = mb->PMV[0][1][1] >> 1;
+            bmv.s[2] = mb->PMV[1][1][0];
+            bmv.s[3] = mb->PMV[1][1][1] >> 1;
         } else if ((mb->motion_type & 3) == XVMC_PREDICTION_DUAL_PRIME) {
-            fmv.s[0] = mb->PMV[0][0][1] >> 1;
-            fmv.s[1] = mb->PMV[0][0][0];
-            fmv.s[2] = mb->PMV[0][0][1] >> 1;  // MPEG2 MV[0][1] isn't used
-            fmv.s[3] = mb->PMV[0][0][0];
-            bmv.s[0] = mb->PMV[1][0][1] >> 1;
-            bmv.s[1] = mb->PMV[1][0][0];
-            bmv.s[2] = mb->PMV[1][1][1] >> 1;
-            bmv.s[3] = mb->PMV[1][1][0];
+            fmv.s[0] = mb->PMV[0][0][0];
+            fmv.s[1] = mb->PMV[0][0][1] >> 1;
+            fmv.s[2] = mb->PMV[0][0][0];
+            fmv.s[3] = mb->PMV[0][0][1] >> 1;  // MPEG2 MV[0][1] isn't used
+            bmv.s[0] = mb->PMV[1][0][0];
+            bmv.s[1] = mb->PMV[1][0][1] >> 1;
+            bmv.s[2] = mb->PMV[1][1][0];
+            bmv.s[3] = mb->PMV[1][1][1] >> 1;
         }
     }
 
@@ -1072,9 +1004,6 @@ _STATIC_ void i915_mc_one_time_state_initialization(XvMCContext *context)
     intelBatchbufferData(pI915XvMC, base, size, 0);
     free(base);
 
-    /* flush */
-    i915_flush_with_flush_bit_clear(pI915XvMC);
-
     /* 3DSTATE_LOAD_INDIRECT */
     size = sizeof(*load_indirect) + sizeof(*dis) + sizeof(*ssb) + sizeof(*psp) + sizeof(*psc);
     base = calloc(1, size);
@@ -1147,9 +1076,6 @@ _STATIC_ void i915_mc_invalidate_subcontext_buffers(XvMCContext *context, unsign
     unsigned size;
     void *base = NULL, *ptr = NULL;
 
-    /* flush */
-    i915_flush_with_flush_bit_clear(pI915XvMC);
-
     size = sizeof(*load_indirect);
     if (mask & BLOCK_SIS)
         size += sizeof(*sis);
@@ -1364,8 +1290,7 @@ _STATIC_ void i915_yuv2rgb_map_state_buffer(XvMCSurface *target_surface)
     memset(tm, 0, sizeof(*tm));
     tm->tm0.v_ls_offset = 0;
     tm->tm0.v_ls = 0;
-    tm->tm0.base_address = privTarget->srf.offset +
-        size_y(w, h);
+    tm->tm0.base_address = VOFFSET(privTarget);
     tm->tm1.tile_walk = TILEWALK_XMAJOR;
     tm->tm1.tiled_surface = 0;
     tm->tm1.utilize_fence_regs = 1;
@@ -1383,7 +1308,7 @@ _STATIC_ void i915_yuv2rgb_map_state_buffer(XvMCSurface *target_surface)
     memset(tm, 0, sizeof(*tm));
     tm->tm0.v_ls_offset = 0;
     tm->tm0.v_ls = 0;
-    tm->tm0.base_address = privTarget->srf.offset;
+    tm->tm0.base_address = YOFFSET(privTarget);
     tm->tm1.tile_walk = TILEWALK_XMAJOR;
     tm->tm1.tiled_surface = 0;
     tm->tm1.utilize_fence_regs = 1;
@@ -1401,8 +1326,7 @@ _STATIC_ void i915_yuv2rgb_map_state_buffer(XvMCSurface *target_surface)
     memset(tm, 0, sizeof(*tm));
     tm->tm0.v_ls_offset = 0;
     tm->tm0.v_ls = 0;
-    tm->tm0.base_address = privTarget->srf.offset +
-        size_y(w, h) + size_uv(w, h);
+    tm->tm0.base_address = UOFFSET(privTarget);
     tm->tm1.tile_walk = TILEWALK_XMAJOR;
     tm->tm1.tiled_surface = 0;
     tm->tm1.utilize_fence_regs = 1;
@@ -1675,9 +1599,6 @@ _STATIC_ void i915_yuv2rgb_proc(XvMCSurface *surface)
     scissor_rectangle.dw2.max_y = 2047;
     intelBatchbufferData(pI915XvMC, &scissor_rectangle, sizeof(scissor_rectangle), 0);
 
-    /* flush */
-    i915_flush_with_flush_bit_clear(pI915XvMC);
-
     /* 3DSTATE_LOAD_INDIRECT */
     size = sizeof(*load_indirect) + sizeof(*sis) + sizeof(*ssb) + sizeof(*msb) + sizeof(*psp);
     base = calloc(1, size);
@@ -2071,8 +1992,8 @@ Status XvMCCreateContext(Display *display, XvPortID port,
 
     /* Initialize private context values */
     setupAttribDesc(display, port, &pI915XvMC->attrib, pI915XvMC->attribDesc);
-    pI915XvMC->yStride = stride(width) * 1;     /* cpp = 1 */
-    pI915XvMC->uvStride = stride(width >> 1) * 1;
+    pI915XvMC->yStride = STRIDE(width);
+    pI915XvMC->uvStride = STRIDE(width >> 1);
     pI915XvMC->haveXv = 0;
     pI915XvMC->attribChanged = 1;
     pI915XvMC->dual_prime = 0;
@@ -2082,7 +2003,6 @@ Status XvMCCreateContext(Display *display, XvPortID port,
     pthread_mutex_init(&pI915XvMC->ctxmutex, NULL);
     intelInitBatchBuffer(pI915XvMC);
     pI915XvMC->ref = 1;
-    pI915XvMC->inited_mc = 0;
     return Success;
 }
 
@@ -2144,7 +2064,6 @@ Status XvMCCreateSurface(Display *display, XvMCContext *context, XvMCSurface *su
     pI915Surface->uvStride = pI915XvMC->uvStride;
     pI915Surface->width = context->width;
     pI915Surface->height = context->height;
-    pI915Surface->processing = 0;
     pI915Surface->privContext = pI915XvMC;
     pI915Surface->privSubPic = NULL;
     pI915Surface->srf.map = NULL;
@@ -2415,27 +2334,9 @@ Status XvMCRenderSurface(Display *display, XvMCContext *context,
     }
 
     LOCK_HARDWARE(pI915XvMC);
-    // if (!pI915XvMC->inited_mc) {
-    i915_clear_surface(pI915XvMC, privTarget);
-    i915_mc_invalidate_subcontext_buffers(context, BLOCK_SIS | BLOCK_DIS | BLOCK_SSB 
-                                          | BLOCK_MSB | BLOCK_PSP | BLOCK_PSC);
-    i915_mc_sampler_state_buffer(context);
-    i915_mc_pixel_shader_program_buffer(context);
-    i915_mc_pixel_shader_constants_buffer(context);
-    i915_mc_one_time_state_initialization(context);
-    pI915XvMC->inited_mc = 1;
-    //}
-    intelFlushBatch(pI915XvMC, TRUE);
-    UNLOCK_HARDWARE(pI915XvMC);
-
     corrdata_ptr = pI915XvMC->corrdata.map;
     corrdata_size = 0;
 
-    if (macroblock_array->macro_blocks[first_macroblock].y == (target_surface->height >> 4) - 1)
-        privTarget->processing = 0;
-    else
-        privTarget->processing = 1;
-
     for (i = first_macroblock; i < (num_macroblocks + first_macroblock); i++) {
         int bspm = 0;
         mb = &macroblock_array->macro_blocks[i];
@@ -2461,19 +2362,29 @@ Status XvMCRenderSurface(Display *display, XvMCContext *context,
         }
         
         bspm = mb_bytes[mb->coded_block_pattern];
+
+        if (!bspm)
+            continue;
+
         corrdata_size += bspm;
 
         if (corrdata_size > pI915XvMC->corrdata.size) {
             printf("Error, correction data buffer overflow\n");
             break;
         }
-
         memcpy(corrdata_ptr, block_ptr, bspm);
         corrdata_ptr += bspm;
     } 
 
-    /* Lock */
-    LOCK_HARDWARE(pI915XvMC);
+    i915_flush(pI915XvMC, 1, 0);
+    i915_mc_invalidate_subcontext_buffers(context, BLOCK_SIS | BLOCK_DIS | BLOCK_SSB 
+                                          | BLOCK_MSB | BLOCK_PSP | BLOCK_PSC);
+
+    i915_mc_sampler_state_buffer(context);
+    i915_mc_pixel_shader_program_buffer(context);
+    i915_mc_pixel_shader_constants_buffer(context);
+    i915_mc_one_time_state_initialization(context);
+
     i915_mc_static_indirect_state_buffer(context, target_surface, 
                                          picture_structure, flags,
                                          picture_coding_type);
@@ -2524,7 +2435,6 @@ Status XvMCRenderSurface(Display *display, XvMCContext *context,
                 break;
             }
         }       /* Field Picture */
-        i915_flush(pI915XvMC, 0, 1);
     }
 
     intelFlushBatch(pI915XvMC, TRUE);
diff --git a/src/xvmc/I915XvMC.h b/src/xvmc/I915XvMC.h
index 9dd73e7c..7f47f433 100644
--- a/src/xvmc/I915XvMC.h
+++ b/src/xvmc/I915XvMC.h
@@ -77,7 +77,6 @@ typedef struct _i915XvMCContext {
     int lock;   /* Lightweight lock to avoid locking twice */
     int locked;
     volatile drmI830Sarea *sarea;
-    int inited_mc;
     
     drmLock *driHwLock;
     drm_context_t hHWContext; /* drmcontext; */
@@ -159,7 +158,6 @@ typedef struct _i915XvMCSurface {
     unsigned uvStride;
     unsigned width;                    /* Dimensions */
     unsigned height;
-    unsigned processing;
     i915XvMCDrmMap srf;
     i915XvMCContext *privContext;
     i915XvMCSubpicture *privSubPic;     /* Subpicture to be blended when