Support HQV command submission fully

6 files changed, 244 insertions, 118 deletions

diff --git a/src/ochr_ioctl.c b/src/ochr_ioctl.c
index 697cfe8..ce1324f 100644
--- a/src/ochr_ioctl.c
+++ b/src/ochr_ioctl.c
@@ -185,6 +185,92 @@ struct via_validate_buffer
 };
 
 static int
+ochr_apply_yuv_reloc(uint32_t *cmdbuf,
+		    uint32_t num_buffers,
+		    struct via_validate_buffer *buffers,
+		    const struct via_yuv_reloc *reloc)
+{
+	uint32_t *buf = cmdbuf + reloc->base.offset;
+	const struct via_reloc_bufaddr *baddr = &reloc->addr;
+	const struct via_validate_buffer *val_buf;
+	uint32_t val;
+	int i;
+
+	if (reloc->planes > 3)
+		return -EINVAL;
+
+	if (baddr->index > num_buffers)
+		return -EINVAL;
+
+	val_buf = &buffers[baddr->index];
+	if (val_buf->po_correct)
+		return 0;
+
+	val = val_buf->offset + baddr->delta;
+
+	for(i=0; i<reloc->planes; ++i) {
+		*buf++ = val + reloc->plane_offs[i];
+		++buf;
+	}
+	
+	return 0;
+}
+
+int
+ochr_yuv_relocation(struct _ViaCommandBuffer *cBuf,
+		   struct _DriBufferObject *buffer,
+		   uint32_t delta, 
+		   int planes, uint32_t plane_0, uint32_t plane_1,
+		    uint32_t plane_2,
+		   uint64_t flags, uint64_t mask)
+{
+    struct via_yuv_reloc reloc;
+    struct via_validate_buffer fake;
+    int itemLoc;
+    struct _ValidateNode *node;
+    struct via_validate_req *val_req;
+    int ret;
+    uint32_t tmp;
+    uint32_t *cmdbuf = (uint32_t *) cBuf->buf + (cBuf->pos - planes * 2);
+
+    ret = driBOAddListItem(cBuf->validate_list, buffer,
+			   flags, mask, &itemLoc, &node);
+    if (ret)
+	return ret;
+
+    val_req = ochrValReq(node);
+
+    if (!(val_req->presumed_flags & VIA_USE_PRESUMED)) {
+	val_req->presumed_gpu_offset = (uint64_t) driBOOffset(buffer) - 
+	    driBOPoolOffset(buffer);
+	val_req->presumed_flags |= VIA_USE_PRESUMED;
+    }
+
+    fake.po_correct = 0;
+    fake.offset = val_req->presumed_gpu_offset;
+    reloc.base.type = VIA_RELOC_YUV;
+    reloc.base.offset = 1;
+    reloc.addr.index = 0;
+    reloc.addr.delta = delta + driBOPoolOffset(buffer);
+    reloc.planes = planes;
+    reloc.plane_offs[0] = plane_0;
+    reloc.plane_offs[1] = plane_1;
+    reloc.plane_offs[2] = plane_2;
+
+    tmp = cmdbuf - (uint32_t *) cBuf->buf;
+
+    ret = ochr_apply_yuv_reloc(cmdbuf, 1, &fake, &reloc);
+
+    reloc.addr.index = itemLoc;
+    reloc.base.offset = tmp + 1;
+
+    assert(ret == 0);
+
+    return ochr_add_reloc(cBuf->reloc_info, &reloc, sizeof(reloc));
+}
+
+
+static int
 ochr_apply_2d_reloc(uint32_t * cmdbuf,
 		    uint32_t num_buffers,
 		    const struct via_validate_buffer *buffers,
diff --git a/src/ochr_ioctl.h b/src/ochr_ioctl.h
index b2505be..666bf63 100644
--- a/src/ochr_ioctl.h
+++ b/src/ochr_ioctl.h
@@ -37,6 +37,14 @@ extern int
 ochr_dest_relocation(struct _ViaCommandBuffer *cBuf,
 		     struct _DriBufferObject *dstBuffer,
 		     uint32_t delta, uint64_t flags, uint64_t mask);
+extern int
+ochr_yuv_relocation(struct _ViaCommandBuffer *cBuf,
+		    struct _DriBufferObject *buffer,
+		    uint32_t delta, int planes, 
+		    uint32_t plane_0, uint32_t plane_1,
+		    uint32_t plane_2,
+		    uint64_t flags, uint64_t mask);
+
 
 extern int ochr_execbuf(int fd, struct _ViaCommandBuffer *cBuf);
 #endif
diff --git a/src/via_dmabuffer.h b/src/via_dmabuffer.h
index bc2535b..478b259 100644
--- a/src/via_dmabuffer.h
+++ b/src/via_dmabuffer.h
@@ -139,6 +139,25 @@ typedef struct _ViaCommandBuffer
 #define OUT_RING_SubA(val1, val2) \
   OUT_RING(((val1) << HC_SubA_SHIFT) | ((val2) & HC_Para_MASK))
 
+#define BEGIN_RING_H6(size) { \
+	    BEGIN_RING(((size) + 4 + 3) & ~3);	\
+	    cb->mode = 6;			\
+	    cb->header_start = cb->pos;		\
+	    cb->pos += 4;			\
+	}
+#define ADVANCE_RING_H6() {					\
+	    CARD32 *hs = cb->buf + cb->header_start;		\
+	    *hs++ = VIA_VIDEO_HEADER6;				\
+	    *hs++ = (cb->pos - cb->header_start - 4) >> 1;	\
+	    *hs++ = 0x00F60000;					\
+	    *hs++ = 0x00000000;					\
+	    while (cb->pos & 3) 				\
+		cb->buf[cb->pos++] = 0x00000000;		\
+	    ADVANCE_RING_VARIABLE;				\
+	}
+	
+	
+
 extern int viaSetupCBuffer(ScrnInfoPtr pScrn, ViaCommandBuffer * buf,
     unsigned size);
 extern void viaTearDownCBuffer(ViaCommandBuffer * buf);
diff --git a/src/via_driver.h b/src/via_driver.h
index 7399909..4016f4f 100644
--- a/src/via_driver.h
+++ b/src/via_driver.h
@@ -223,7 +223,6 @@ enum _ViaScanoutTypes {
     VIA_SCANOUT_DISPLAY,
     VIA_SCANOUT_CURSOR,
     VIA_SCANOUT_OVERLAY,
-    VIA_SCANOUT_HQV,
     VIA_SCANOUT_NUM
 };
 
@@ -476,6 +475,8 @@ void viaRestoreVideo(ScrnInfoPtr pScrn);
 void viaSetColorSpace(VIAPtr pVia, int hue, int saturation, int brightness, int contrast,
 		      Bool reset);
 void VIAVidAdjustFrame(ScrnInfoPtr pScrn, int x, int y);
+void viaVideoFlip(VIAPtr pVia, int fourcc, unsigned long DisplayBufferIndex);
+
 
 /* In via_xwmc.c */
 
diff --git a/src/via_swov.c b/src/via_swov.c
index 929f547..e35c273 100644
--- a/src/via_swov.c
+++ b/src/via_swov.c
@@ -122,22 +122,6 @@ viaWaitVBI(VIAPtr pVia)
     while (IN_VIDEO_DISPLAY) ;
 }
 
-static void
-viaWaitHQVDone(VIAPtr pVia)
-{
-    CARD32 volatile *pdwState;
-    unsigned long proReg = 0;
-
-    if (pVia->ChipId == PCI_CHIP_VT3259
-        && !(pVia->swov.gdwVideoFlagSW & VIDEO_1_INUSE))
-        proReg = PRO_HQV1_OFFSET;
-
-    pdwState = (CARD32 volatile *)(pVia->VidMapBase + (HQV_CONTROL + proReg));
-    if (pVia->swov.MPEG_ON) {
-        while ((*pdwState & HQV_SW_FLIP)) ;
-    }
-}
-
 /*
  * Send all data in VidRegBuffer to the hardware.
  */
@@ -1047,7 +1031,6 @@ CreateSurface(ScrnInfoPtr pScrn, CARD32 FourCC, CARD16 Width,
 					&hqvBuf->buf, 0, 
 					DRM_BO_FLAG_MEM_VRAM |
 					DRM_BO_FLAG_READ |
-					DRM_BO_FLAG_NO_EVICT |
 					hqvFlag, 0);
 		if (retCode) {
 		    hqvBuf->buf = NULL;
@@ -1580,12 +1563,49 @@ SetVideoWindow(ScrnInfoPtr pScrn, unsigned long videoFlag,
     }
 }
 
+static void viaSetHqvSrc(ScrnInfoPtr pScrn, int fourcc, 
+			 struct _HQVBuffer *hqvBuf)
+{
+    VIAPtr pVia = VIAPTR(pScrn);
+    int proReg = 0;
+
+    if (pVia->ChipId == PCI_CHIP_VT3259
+        && !(pVia->swov.gdwVideoFlagSW & VIDEO_1_INUSE)) {
+        proReg += PRO_HQV1_OFFSET;
+    }
+
+    switch (fourcc) {
+    case FOURCC_UYVY:
+    case FOURCC_YUY2:
+    case FOURCC_RV15:
+    case FOURCC_RV16:
+    case FOURCC_RV32:
+	SaveVideoRegister(pVia, HQV_SRC_STARTADDR_Y + proReg,
+			  hqvBuf->pinnedOffset + hqvBuf->deltaY);
+	break;
+    case FOURCC_YV12:
+    default:
+	SaveVideoRegister(pVia, HQV_SRC_STARTADDR_Y + proReg, 
+			  hqvBuf->pinnedOffset + hqvBuf->deltaY);
+	if (pVia->VideoEngine == VIDEO_ENGINE_CME) {
+	    SaveVideoRegister(pVia, HQV_SRC_STARTADDR_U + proReg,
+			      hqvBuf->pinnedOffset + hqvBuf->deltaU); 
+	} else {
+	    SaveVideoRegister(pVia, HQV_SRC_STARTADDR_U + proReg, 
+			      hqvBuf->pinnedOffset + hqvBuf->deltaV); 
+	    SaveVideoRegister(pVia, HQV_SRC_STARTADDR_V + proReg, 
+			      hqvBuf->pinnedOffset + hqvBuf->deltaU);
+	}
+	break;
+    }
+}
+
 /*
  * Upd_Video()
  */
 static Bool
 Upd_Video(ScrnInfoPtr pScrn, unsigned long videoFlag,
-          unsigned long startAddr, LPDDUPDATEOVERLAY pUpdate,
+          LPDDUPDATEOVERLAY pUpdate,
           unsigned long srcPitch,
           unsigned long oriSrcWidth, unsigned long oriSrcHeight,
           unsigned long deinterlaceMode,
@@ -1608,6 +1628,7 @@ Upd_Video(ScrnInfoPtr pScrn, unsigned long videoFlag,
     unsigned long hqvSrcFetch = 0, hqvOffset = 0;
     unsigned long dwOffset = 0, fetch = 0, tmp = 0;
     unsigned long proReg = 0;
+    int ret;
 
     DBG_DD(ErrorF("videoflag=%p\n", videoFlag));
 
@@ -1674,44 +1695,16 @@ Upd_Video(ScrnInfoPtr pScrn, unsigned long videoFlag,
 
     pVia->swov.overlayRecordV1.dwOffset = dwOffset;
 
-    /*
-     * FIXME: Note that the HQV source is initialized to its own destination
-     * surfaces here. Should be investigated in case of strange
-     * behaviour during video startup.
-     */
-
     if (pVia->swov.SrcFourCC == FOURCC_YV12
         || pVia->swov.SrcFourCC == FOURCC_XVMC) {
 
         if (videoFlag & VIDEO_HQV_INUSE) {
-	    int bufNum = 1 - (pVia->dwFrameNum & 1);
-	    struct _HQVBuffer *hqvBuf = &pVia->swov.SWDevice.hqvBuf[bufNum];
-
             SetVideoStart(pVia, videoFlag, hwDiff->dwThreeHQVBuffer ? 3 : 2,
                           pVia->swov.overlayRecordV1.dwHQVAddr[0] + dwOffset,
                           pVia->swov.overlayRecordV1.dwHQVAddr[1] + dwOffset,
                           pVia->swov.overlayRecordV1.dwHQVAddr[2] + dwOffset);
-
-            if (pVia->swov.SrcFourCC != FOURCC_XVMC) {
-                if (pVia->VideoEngine == VIDEO_ENGINE_CME) {
-                    SaveVideoRegister(pVia, HQV_SRC_STARTADDR_Y + proReg,
-                                      hqvBuf->pinnedOffset + hqvBuf->deltaY);
-                    SaveVideoRegister(pVia, HQV_SRC_STARTADDR_U + proReg,
-                                      hqvBuf->pinnedOffset + hqvBuf->deltaU);
-                } else {
-                    SaveVideoRegister(pVia, HQV_SRC_STARTADDR_Y,
-				      hqvBuf->pinnedOffset + hqvBuf->deltaY);
-                    SaveVideoRegister(pVia, HQV_SRC_STARTADDR_U,
-				      hqvBuf->pinnedOffset + hqvBuf->deltaV);
-                    SaveVideoRegister(pVia, HQV_SRC_STARTADDR_V, 
-				      hqvBuf->pinnedOffset + hqvBuf->deltaU);
-                }
-            }
         } 
     } else {
-	int bufNum = 1 - (pVia->dwFrameNum & 1);
-	struct _HQVBuffer *hqvBuf = &pVia->swov.SWDevice.hqvBuf[bufNum];
-
         if (videoFlag & VIDEO_HQV_INUSE) {
             hqvSrcWidth = (unsigned long)pUpdate->SrcRight - pUpdate->SrcLeft;
             hqvDstWidth = (unsigned long)pUpdate->DstRight - pUpdate->DstLeft;
@@ -1726,10 +1719,6 @@ Upd_Video(ScrnInfoPtr pScrn, unsigned long videoFlag,
     
             if (pVia->VideoEngine == VIDEO_ENGINE_CME)
                 SaveVideoRegister(pVia, 0x1cc + proReg, dwOffset);
-
-            SaveVideoRegister(pVia, HQV_SRC_STARTADDR_Y + proReg, 
-			      hqvBuf->pinnedOffset + hqvBuf->deltaY);
-
         } 
     }
 
@@ -1737,18 +1726,6 @@ Upd_Video(ScrnInfoPtr pScrn, unsigned long videoFlag,
                                srcWidth, dstWidth, oriSrcWidth, &hqvSrcFetch);
     DBG_DD(ErrorF("===fetch= 0x%lx\n", fetch));
 
-#if 0
-    /* For DCT450 test-BOB INTERLEAVE */
-    if ((deinterlaceMode & DDOVER_INTERLEAVED)
-        && (deinterlaceMode & DDOVER_BOB)) {
-        if (videoFlag & VIDEO_HQV_INUSE)
-            hqvCtl |= HQV_FIELD_2_FRAME | HQV_FRAME_2_FIELD | HQV_DEINTERLACE;
-    } else if (deinterlaceMode & DDOVER_BOB) {
-        if (videoFlag & VIDEO_HQV_INUSE)
-            /* The HQV source data line count should be two times of the original line count */
-            hqvCtl |= HQV_FIELD_2_FRAME | HQV_DEINTERLACE;
-    }
-#endif
 
     if (videoFlag & VIDEO_HQV_INUSE) {
         if (!(deinterlaceMode & DDOVER_INTERLEAVED)
@@ -1887,12 +1864,25 @@ Upd_Video(ScrnInfoPtr pScrn, unsigned long videoFlag,
 
     /* Set up video control */
     if (videoFlag & VIDEO_HQV_INUSE) {
-
         if (!pVia->swov.SWVideo_ON) {
+	    unsigned bufNum = 1 - (pVia->dwFrameNum & 1);
+	    struct _HQVBuffer *hqvBuf = &pVia->swov.SWDevice.hqvBuf[bufNum];
+
             DBG_DD(ErrorF("    First HQV\n"));
 
-            FlushVidRegBuffer(pVia);
+	    /*
+	     * Since HQV initialization seems to require a tight 
+	     * interaction with the hardware, we pin the HQV source
+	     * buffer here, and unpin it when the HQV is idle and
+	     * initialization done.
+	     */
 
+	    ret = driBOSetStatus(hqvBuf->buf, DRM_BO_FLAG_NO_EVICT, 0);
+	    hqvBuf->pinnedOffset = driBOOffset(hqvBuf->buf);
+
+	    viaSetHqvSrc(pScrn, pVia->swov.SrcFourCC, hqvBuf);	    
+            FlushVidRegBuffer(pVia);
+	    
             DBG_DD(ErrorF(" Wait flips"));
 
             if (hwDiff->dwHQVInitPatch) {
@@ -1940,6 +1930,8 @@ Upd_Video(ScrnInfoPtr pScrn, unsigned long videoFlag,
                 DBG_DD(ErrorF(" Wait flips6"));
             }
 
+	    ret = driBOSetStatus(hqvBuf->buf, 0, DRM_BO_FLAG_NO_EVICT);
+
             if (videoFlag & VIDEO_1_INUSE) {
                 VIDOutD(V1_CONTROL, vidCtl);
                 VIDOutD(V_COMPOSE_MODE, compose | V1_COMMAND_FIRE);
@@ -1961,16 +1953,17 @@ Upd_Video(ScrnInfoPtr pScrn, unsigned long videoFlag,
             DBG_DD(ErrorF(" Done flips"));
         } else {
             DBG_DD(ErrorF("    Normal called\n"));
-            SaveVideoRegister(pVia, HQV_CONTROL + proReg,
-                              hqvCtl | HQV_FLIP_STATUS);
             SetVideoControl(pVia, videoFlag, vidCtl);
             FireVideoCommand(pVia, videoFlag, compose);
-            viaWaitHQVDone(pVia);
             FlushVidRegBuffer(pVia);
+
+	    viaVideoFlip(pVia, pVia->swov.SrcFourCC, 
+			 1 - (pVia->dwFrameNum & 1));	    
         }
     }
     pVia->swov.SWVideo_ON = TRUE;
 
+    pVia->swov.hqvCtl = hqvCtl;
     DBG_DD(ErrorF(" Done Upd_Video"));
 
     return TRUE;
@@ -1991,7 +1984,6 @@ VIAVidUpdateOverlay(ScrnInfoPtr pScrn, LPDDUPDATEOVERLAY pUpdate)
 
     unsigned long flags = pUpdate->dwFlags;
     unsigned long videoFlag = 0;
-    unsigned long startAddr = 0;
     unsigned long deinterlaceMode = 0;
 
     unsigned long haveColorKey = 0, haveChromaKey = 0;
@@ -2042,7 +2034,6 @@ VIAVidUpdateOverlay(ScrnInfoPtr pScrn, LPDDUPDATEOVERLAY pUpdate)
     ResetVidRegBuffer(pVia);
 
     /* For SW decode HW overlay use */
-    startAddr = VIDInD(HQV_SRC_STARTADDR_Y + proReg);
 
     if (flags & DDOVER_KEYDEST) {
         haveColorKey = 1;
@@ -2130,14 +2121,14 @@ VIAVidUpdateOverlay(ScrnInfoPtr pScrn, LPDDUPDATEOVERLAY pUpdate)
 
     /* Update the overlay */
 
-    if (!Upd_Video(pScrn, videoFlag, startAddr, pUpdate,
+    if (!Upd_Video(pScrn, videoFlag, pUpdate,
                    pVia->swov.SWDevice.dwPitch, ovlV1->dwV1OriWidth,
                    ovlV1->dwV1OriHeight, deinterlaceMode, haveColorKey,
                    haveChromaKey, colorKeyLow, colorKeyHigh, chromaKeyLow,
-                   chromaKeyHigh))
+                   chromaKeyHigh)) {
+	pVia->swov.SWVideo_ON = FALSE;
         return FALSE;
-
-    pVia->swov.SWVideo_ON = FALSE;
+    }
 
     return TRUE;
 
@@ -2194,3 +2185,4 @@ ViaOverlayHide(ScrnInfoPtr pScrn)
     pVia->swov.SWVideo_ON = FALSE;
     pVia->VideoStatus &= ~VIDEO_SWOV_ON;
 }
+
diff --git a/src/via_video.c b/src/via_video.c
index 0d2e8fd..723283c 100644
--- a/src/via_video.c
+++ b/src/via_video.c
@@ -1056,19 +1056,23 @@ viaQueryBestSize(ScrnInfoPtr pScrn,
         *p_w = 2048;
 }
 
-/*
- *  To do SW Flip
- */
-static void
-Flip(VIAPtr pVia, viaPortPrivPtr pPriv, int fourcc,
-        unsigned long DisplayBufferIndex)
+void
+viaVideoFlip(VIAPtr pVia, int fourcc,
+	     unsigned long DisplayBufferIndex)
 {
-    unsigned long proReg = 0;
+    unsigned long proReg = 0x200;
+    uint64_t hqvFlag = VIA_BO_FLAG_HQV0;
     struct _HQVBuffer *hqvBuf = &pVia->swov.SWDevice.hqvBuf[DisplayBufferIndex];
+   
+    int ret;
+    RING_VARS;
+    BEGIN_RING_H6(8);
 
     if (pVia->ChipId == PCI_CHIP_VT3259
-        && !(pVia->swov.gdwVideoFlagSW & VIDEO_1_INUSE))
-        proReg = PRO_HQV1_OFFSET;
+        && !(pVia->swov.gdwVideoFlagSW & VIDEO_1_INUSE)) {
+        proReg += PRO_HQV1_OFFSET;
+	hqvFlag = VIA_BO_FLAG_HQV1;
+    }
 
     switch (fourcc) {
         case FOURCC_UYVY:
@@ -1076,32 +1080,46 @@ Flip(VIAPtr pVia, viaPortPrivPtr pPriv, int fourcc,
         case FOURCC_RV15:
         case FOURCC_RV16:
         case FOURCC_RV32:
-            while ((VIDInD(HQV_CONTROL + proReg) & HQV_SW_FLIP));
-            VIDOutD(HQV_SRC_STARTADDR_Y + proReg,
-		    hqvBuf->pinnedOffset + hqvBuf->deltaY);
-            VIDOutD(HQV_CONTROL + proReg,
-                (VIDInD(HQV_CONTROL +
-                proReg) & ~HQV_FLIP_ODD) | HQV_SW_FLIP | HQV_FLIP_STATUS);
+            OUT_RING_QW(HQV_SRC_STARTADDR_Y + proReg, 0);
+	    ret = ochr_yuv_relocation(cb, hqvBuf->buf, 0, 1,
+				      hqvBuf->deltaY, 0, 0, 
+				      DRM_BO_FLAG_MEM_VRAM |
+				      hqvFlag,
+				      DRM_BO_MASK_MEM |
+				      hqvFlag);
             break;
         case FOURCC_YV12:
         default:
-            while ((VIDInD(HQV_CONTROL + proReg) & HQV_SW_FLIP));
-            VIDOutD(HQV_SRC_STARTADDR_Y + proReg,
-		    hqvBuf->pinnedOffset + hqvBuf->deltaY);
+	    OUT_RING_QW(HQV_SRC_STARTADDR_Y + proReg, 0);
             if (pVia->VideoEngine == VIDEO_ENGINE_CME) {
-                VIDOutD(HQV_SRC_STARTADDR_U + proReg,
-			hqvBuf->pinnedOffset + hqvBuf->deltaU);
+		OUT_RING_QW(HQV_SRC_STARTADDR_U + proReg, 0);
+		ret = ochr_yuv_relocation(cb, hqvBuf->buf, 0, 2,
+					  hqvBuf->deltaY, hqvBuf->deltaU, 0, 
+					  DRM_BO_FLAG_MEM_VRAM |
+					  hqvFlag,
+					  DRM_BO_MASK_MEM |
+					  hqvFlag);
             } else {
-                VIDOutD(HQV_SRC_STARTADDR_U,
-			hqvBuf->pinnedOffset + hqvBuf->deltaV);
-                VIDOutD(HQV_SRC_STARTADDR_V,
-			hqvBuf->pinnedOffset + hqvBuf->deltaU);
+		OUT_RING_QW(HQV_SRC_STARTADDR_U + proReg, 0);
+		OUT_RING_QW(HQV_SRC_STARTADDR_V + proReg, 0);
+		ret = ochr_yuv_relocation(cb, hqvBuf->buf, 0, 3,
+					  hqvBuf->deltaY, hqvBuf->deltaV, 
+					  hqvBuf->deltaU,
+					  DRM_BO_FLAG_MEM_VRAM |
+					  hqvFlag,
+					  DRM_BO_MASK_MEM |
+					  hqvFlag);
             }
-            VIDOutD(HQV_CONTROL + proReg,
-            (VIDInD(HQV_CONTROL +
-                proReg) & ~HQV_FLIP_ODD) | HQV_SW_FLIP | HQV_FLIP_STATUS);
+
         break;
     }
+
+    OUT_RING_QW(HQV_CONTROL + proReg, 
+		(pVia->swov.hqvCtl & ~HQV_FLIP_ODD) | 
+		HQV_SW_FLIP | HQV_FLIP_STATUS);
+
+    ADVANCE_RING_H6();
+    FLUSH_RING;
 }
 
 /*
@@ -1302,11 +1320,6 @@ viaPutImage(ScrnInfoPtr pScrn,
 	    void *virtual;
 
 
-	    if (virtual == NULL) {
-		viaXvError(pScrn, pPriv, xve_mem);
-		return BadAlloc;
-	    }
-
             DBG_DD(ErrorF(" via_video.c :              : S/W Overlay! \n"));
             /*  Allocate video memory(CreateSurface),
              *  add codes to judge if need to re-create surface
@@ -1326,6 +1339,12 @@ viaPutImage(ScrnInfoPtr pScrn,
 	    hqvBuf = &pVia->swov.SWDevice.hqvBuf[pVia->dwFrameNum & 1];
 	    virtual = driBOMap(hqvBuf->buf, WS_DRI_MAP_WRITE);
 
+	    if (virtual == NULL) {
+		viaXvError(pScrn, pPriv, xve_mem);
+		return BadAlloc;
+	    }
+
+
             /*  Copy image data from system memory to video memory
              *  TODO: use DRM's DMA feature to accelerate data copy
              */
@@ -1410,17 +1429,6 @@ viaPutImage(ScrnInfoPtr pScrn,
                 dwUseExtendedFIFO = 1;
             }
 
-            if (FOURCC_XVMC != id) {
-
-                /*
-                 * XvMC flipping is done in the client lib.
-                 */
-
-                DBG_DD(ErrorF("             : Flip\n"));
-                Flip(pVia, pPriv, id, pVia->dwFrameNum & 1);
-            }
-
-            pVia->dwFrameNum++;
 
             /* If the dest rec. & extendFIFO doesn't change, don't do UpdateOverlay 
              * unless the surface clipping has changed */
@@ -1431,10 +1439,22 @@ viaPutImage(ScrnInfoPtr pScrn,
                     && (pVia->old_dwUseExtendedFIFO == dwUseExtendedFIFO)
                     && (pVia->VideoStatus & VIDEO_SWOV_ON) &&
                     RegionsEqual(&pPriv->clip, clipBoxes)) {
+		if (FOURCC_XVMC != id) {
+		
+		    /*
+		     * XvMC flipping is done in the client lib.
+		     */
+
+		    DBG_DD(ErrorF("             : Flip\n"));
+		    viaVideoFlip(pVia, id, pVia->dwFrameNum & 1);
+		    pVia->dwFrameNum++;
+		}
+
                 viaXvError(pScrn, pPriv, xve_none);
                 return Success;
             }
 
+	    pVia->dwFrameNum++;
             pPriv->old_src_x = src_x;
             pPriv->old_src_y = src_y;
             pPriv->old_src_w = src_w;
@@ -1472,7 +1492,7 @@ viaPutImage(ScrnInfoPtr pScrn,
                         (" via_video.c : call v4l updateoverlay fail. \n"));
             } else {
                 DBG_DD(ErrorF(" via_video.c : PutImage done OK\n"));
-                viaXvError(pScrn, pPriv, xve_none);
+		viaXvError(pScrn, pPriv, xve_none);
                 return Success;
             }
             break;