Fix cursor garbare after suspend/hibernate/resume on VX855/VX900 chipsets #405

2 files changed, 77 insertions, 30 deletions

diff --git a/src/via_swov.c b/src/via_swov.c
index 1f9aff9..c2c7d6f 100644
--- a/src/via_swov.c
+++ b/src/via_swov.c
@@ -392,6 +392,7 @@ viaOverlayGetV1V3Format(VIAPtr pVia, int vport, /* 1 or 3, as in V1 or V3 */
     if (videoFlag & VIDEO_HQV_INUSE) {
         switch (pVia->swov.SrcFourCC) {
             case FOURCC_YV12:
+            case FOURCC_I420:
             case FOURCC_XVMC:
                 *pHQVCtl |= HQV_YUV420;
                 break;
@@ -421,6 +422,7 @@ viaOverlayGetV1V3Format(VIAPtr pVia, int vport, /* 1 or 3, as in V1 or V3 */
     } else {
         switch (pVia->swov.SrcFourCC) {
             case FOURCC_YV12:
+            case FOURCC_I420:
             case FOURCC_XVMC:
                 if (vport == 1) {
                     *pVidCtl |= V1_YCbCr420;
@@ -503,6 +505,7 @@ viaOverlayGetSrcStartAddress(VIAPtr pVia,
                 break;
 
             case FOURCC_YV12:
+            case FOURCC_I420:
             case FOURCC_XVMC:
 
                 if (videoFlag & VIDEO_HQV_INUSE)
@@ -763,6 +766,7 @@ viaOverlayGetFetch(VIAPtr pVia, unsigned long videoFlag,
 
     switch (pVia->swov.SrcFourCC) {
         case FOURCC_YV12:
+        case FOURCC_I420:
         case FOURCC_XVMC:
             n = 0; /* 2^n = 1 byte per pixel (Y channel in planar YUV) */
             break;
@@ -1146,7 +1150,7 @@ AddHQVSurface(ScrnInfoPtr pScrn, unsigned int numbuf, CARD32 fourcc)
         !(pVia->swov.gdwVideoFlagSW & VIDEO_1_INUSE))
         proReg = PRO_HQV1_OFFSET;
 
-    isplanar = ((fourcc == FOURCC_YV12) || (fourcc == FOURCC_XVMC));
+    isplanar = ((fourcc == FOURCC_YV12) || (fourcc == FOURCC_I420) || (fourcc == FOURCC_XVMC));
 
     width = pVia->swov.SWDevice.gdwSWSrcWidth;
     height = pVia->swov.SWDevice.gdwSWSrcHeight;
@@ -1189,6 +1193,7 @@ CreateSurface(ScrnInfoPtr pScrn, CARD32 FourCC, CARD16 Width,
     isplanar = FALSE;
     switch (FourCC) {
         case FOURCC_YV12:
+        case FOURCC_I420:
         case FOURCC_XVMC:
             isplanar = TRUE;
             pitch = ALIGN_TO(Width, 32);
@@ -1281,9 +1286,10 @@ ViaSwovSurfaceCreate(ScrnInfoPtr pScrn, viaPortPrivPtr pPriv,
             break;
 
         case FOURCC_YV12:
+        case FOURCC_I420:
             retCode = CreateSurface(pScrn, FourCC, Width, Height, TRUE);
             if (retCode == Success)
-                retCode = AddHQVSurface(pScrn, numbuf, FOURCC_YV12);
+                retCode = AddHQVSurface(pScrn, numbuf, FourCC);
             break;
 
         case FOURCC_XVMC:
@@ -1345,6 +1351,7 @@ ViaSwovSurfaceDestroy(ScrnInfoPtr pScrn, viaPortPrivPtr pPriv)
                 break;
 
             case FOURCC_YV12:
+            case FOURCC_I420:
                 VIAFreeLinear(&pVia->swov.SWfbMem);
             case FOURCC_XVMC:
                 pVia->swov.SrcFourCC = 0;
@@ -1509,6 +1516,7 @@ SetupFIFOs(VIAPtr pVia, unsigned long videoFlag,
 {
     if (miniCtl & V1_Y_INTERPOLY) {
         if (pVia->swov.SrcFourCC == FOURCC_YV12
+            || pVia->swov.SrcFourCC == FOURCC_I420
             || pVia->swov.SrcFourCC == FOURCC_XVMC) {
             if (videoFlag & VIDEO_HQV_INUSE) {
                 if (videoFlag & VIDEO_1_INUSE)
@@ -1542,6 +1550,7 @@ SetupFIFOs(VIAPtr pVia, unsigned long videoFlag,
         }
     } else {
         if (pVia->swov.SrcFourCC == FOURCC_YV12
+            || pVia->swov.SrcFourCC == FOURCC_I420
             || pVia->swov.SrcFourCC == FOURCC_XVMC) {
             if (videoFlag & VIDEO_HQV_INUSE) {
                 if (videoFlag & VIDEO_1_INUSE)
@@ -1891,6 +1900,7 @@ Upd_Video(ScrnInfoPtr pScrn, unsigned long videoFlag,
     pVia->swov.overlayRecordV1.dwOffset = dwOffset;
 
     if (pVia->swov.SrcFourCC == FOURCC_YV12
+        || pVia->swov.SrcFourCC == FOURCC_I420
         || pVia->swov.SrcFourCC == FOURCC_XVMC) {
 
         YCBCRREC YCbCr;
@@ -1983,6 +1993,7 @@ Upd_Video(ScrnInfoPtr pScrn, unsigned long videoFlag,
             SetHQVFetch(pVia, hqvSrcFetch, oriSrcHeight);
 
         if (pVia->swov.SrcFourCC == FOURCC_YV12
+            || pVia->swov.SrcFourCC == FOURCC_I420
             || pVia->swov.SrcFourCC == FOURCC_XVMC) {
             if (videoFlag & VIDEO_1_INUSE)
                 SaveVideoRegister(pVia, V1_STRIDE, srcPitch << 1);
@@ -2292,6 +2303,7 @@ VIAVidUpdateOverlay(ScrnInfoPtr pScrn, LPDDUPDATEOVERLAY pUpdate)
         (pVia->swov.SrcFourCC == FOURCC_RV16) ||
         (pVia->swov.SrcFourCC == FOURCC_RV32) ||
         (pVia->swov.SrcFourCC == FOURCC_YV12) ||
+        (pVia->swov.SrcFourCC == FOURCC_I420) ||
         (pVia->swov.SrcFourCC == FOURCC_XVMC)) {
         videoFlag = pVia->swov.gdwVideoFlagSW;
     }
@@ -2366,6 +2378,7 @@ VIAVidUpdateOverlay(ScrnInfoPtr pScrn, LPDDUPDATEOVERLAY pUpdate)
         (pVia->swov.SrcFourCC == FOURCC_RV16) ||
         (pVia->swov.SrcFourCC == FOURCC_RV32) ||
         (pVia->swov.SrcFourCC == FOURCC_YV12) ||
+        (pVia->swov.SrcFourCC == FOURCC_I420) ||
         (pVia->swov.SrcFourCC == FOURCC_XVMC)) {
         pVia->swov.SWDevice.gdwSWDstLeft = pUpdate->DstLeft + panDX;
         pVia->swov.SWDevice.gdwSWDstTop = pUpdate->DstTop + panDY;
@@ -2425,6 +2438,7 @@ ViaOverlayHide(ScrnInfoPtr pScrn)
         (pVia->swov.SrcFourCC == FOURCC_RV16) ||
         (pVia->swov.SrcFourCC == FOURCC_RV32) ||
         (pVia->swov.SrcFourCC == FOURCC_YV12) ||
+        (pVia->swov.SrcFourCC == FOURCC_I420) ||
         (pVia->swov.SrcFourCC == FOURCC_XVMC))
         videoFlag = pVia->swov.gdwVideoFlagSW;
 
diff --git a/src/via_video.c b/src/via_video.c
index 58538d8..1797cfd 100644
--- a/src/via_video.c
+++ b/src/via_video.c
@@ -117,10 +117,14 @@ static int viaSetPortAttribute(ScrnInfoPtr, Atom, INT32, pointer);
 static int viaPutImage(ScrnInfoPtr, short, short, short, short, short, short,
     short, short, int, unsigned char *, short, short, Bool,
     RegionPtr, pointer, DrawablePtr);
-static void nv12Blit(unsigned char *nv12Chroma,
-    const unsigned char *uBuffer,
-    const unsigned char *vBuffer,
-    unsigned width, unsigned srcPitch, unsigned dstPitch, unsigned lines);
+static void nv12Blit(unsigned char *dst,
+    const unsigned char *src1,
+    const unsigned char *src2,
+    unsigned srcPitch, unsigned dstPitch, unsigned width, unsigned height);
+static void UVBlit(unsigned char *dst,
+    const unsigned char *src1,
+    const unsigned char *src2,
+    unsigned srcPitch, unsigned dstPitch, unsigned width, unsigned height);
 
 static Atom xvBrightness, xvContrast, xvColorKey, xvHue, xvSaturation,
     xvAutoPaint;
@@ -158,11 +162,12 @@ static XF86AttributeRec AttributesG[NUM_ATTRIBUTES_G] = {
     {XvSettable | XvGettable, 0, 1, "XV_AUTOPAINT_COLORKEY"}
 };
 
-#define NUM_IMAGES_G 6
+#define NUM_IMAGES_G 7
 
 static XF86ImageRec ImagesG[NUM_IMAGES_G] = {
     XVIMAGE_YUY2,
     XVIMAGE_YV12,
+    XVIMAGE_I420,
     {
         /*
          * Below, a dummy picture type that is used in XvPutImage only to do
@@ -963,6 +968,7 @@ Flip(VIAPtr pVia, viaPortPrivPtr pPriv, int fourcc,
                 proReg) & ~HQV_FLIP_ODD) | HQV_SW_FLIP | HQV_FLIP_STATUS);
             break;
         case FOURCC_YV12:
+        case FOURCC_I420:
         default:
             while ((VIDInD(HQV_CONTROL + proReg) & HQV_SW_FLIP)
                     && --count);
@@ -984,22 +990,7 @@ Flip(VIAPtr pVia, viaPortPrivPtr pPriv, int fourcc,
     }
 }
 
-/*
- * Slow and dirty. NV12 blit.
- */
 
-static void
-nv12cp(unsigned char *dst,
-    const unsigned char *src, int dstPitch, int w, int h, int yuv422)
-{
-    /* 
-     * Blit luma component as a fake YUY2 assembler blit. 
-     */
-
-    (*viaFastVidCpy) (dst, src, dstPitch, w >> 1, h, TRUE);
-    nv12Blit(dst + dstPitch * h, src + w * h + (w >> 1) * (h >> 1),
-            src + w * h, w >> 1, w >> 1, dstPitch, h >> 1);
-}
 
 #ifdef XF86DRI
 
@@ -1023,7 +1014,7 @@ viaDmaBlitImage(VIAPtr pVia,
 
     bounceBuffer = ((unsigned long)src & 15);
     nv12Conversion = (pVia->VideoEngine == VIDEO_ENGINE_CME && 
-        id == FOURCC_YV12);
+        (id == FOURCC_YV12 || id == FOURCC_I420));
 
     switch (id) {
         case FOURCC_YUY2:
@@ -1038,6 +1029,7 @@ viaDmaBlitImage(VIAPtr pVia,
             break;
 
         case FOURCC_YV12:
+        case FOURCC_I420:
         default:
             bounceStride = ALIGN_TO(width, 16);
             bounceLines = height;
@@ -1053,8 +1045,9 @@ viaDmaBlitImage(VIAPtr pVia,
                 pPort->dmaBounceBuffer = 0;
             }
             size = bounceStride * bounceLines + 16;
-            if (FOURCC_YV12 == id)
+            if (id == FOURCC_YV12 || id == FOURCC_I420) {
                 size += ALIGN_TO(bounceStride >> 1, 16) * bounceLines;
+            }
             pPort->dmaBounceBuffer = (unsigned char *)malloc(size);
             pPort->dmaBounceLines = bounceLines;
             pPort->dmaBounceStride = bounceStride;
@@ -1093,14 +1086,22 @@ viaDmaBlitImage(VIAPtr pVia,
 
     lumaSync = blit.sync;
 
-    if (id == FOURCC_YV12) {
+    if (id == FOURCC_YV12 || id == FOURCC_I420) {
         unsigned tmp = ALIGN_TO(width >> 1, 16);
 
         if (nv12Conversion) {
-            nv12Blit(bounceBase + bounceStride * height,
-                src + bounceStride * height + tmp * (height >> 1),
-                src + bounceStride * height, width >> 1, tmp,
-                bounceStride, height >> 1);
+            if (id == FOURCC_YV12) {
+                nv12Blit(bounceBase + bounceStride * height,
+                    src + bounceStride * height + tmp * (height >> 1),
+                    src + bounceStride * height, 
+                    tmp, bounceStride, width >> 1, height >> 1);
+            } else if (id == FOURCC_I420) {
+                /* TODO Check if this code will work properly with FOURCC_I420 */
+                nv12Blit(bounceBase + bounceStride * height,
+                    src + bounceStride * height, 
+                    src + bounceStride * height + tmp * (height >> 1),
+                    tmp, bounceStride, width >> 1, height >> 1);
+            }
         } else if (bounceBuffer) {
             (*viaFastVidCpy) (base + bounceStride * height,
                     src + bounceStride * height, tmp, tmp >> 1, height, 1);
@@ -1227,13 +1228,45 @@ viaPutImage(ScrnInfoPtr pScrn,
                     switch (id) {
                         case FOURCC_YV12:
                             if (pVia->VideoEngine == VIDEO_ENGINE_CME) {
-                                nv12cp(pVia->swov.SWDevice.
+                                int srcYSize  = width * height;
+                                /* Copy Y component */
+                                (*viaFastVidCpy) (pVia->swov.SWDevice.
+                                    lpSWOverlaySurface[pVia->dwFrameNum & 1], 
+                                    buf, dstPitch, width >> 1, height, TRUE);
+
+                                /* Copy U and V component */
+                                nv12Blit(pVia->swov.SWDevice.lpSWOverlaySurface[pVia->dwFrameNum & 1] + dstPitch * height, 
+                                    buf + srcYSize + (srcYSize >> 2), /* Size of UV component is 1/4 of Y */
+                                    buf + srcYSize, 
+                                    width >> 1, dstPitch, width >> 1, height >> 1);
+                            } else {
+                                (*viaFastVidCpy)(pVia->swov.SWDevice.
                                     lpSWOverlaySurface[pVia->dwFrameNum & 1],
                                     buf, dstPitch, width, height, 0);
+                            }
+                            break;
+                        case FOURCC_I420:
+                            if (pVia->VideoEngine == VIDEO_ENGINE_CME) {
+                                int srcYSize  = width * height;
+
+                                (*viaFastVidCpy) (pVia->swov.SWDevice.
+                                   lpSWOverlaySurface[pVia->dwFrameNum & 1], 
+                                   buf, dstPitch, width >> 1, height, TRUE);
+
+                                nv12Blit(pVia->swov.SWDevice.lpSWOverlaySurface[pVia->dwFrameNum & 1] + dstPitch*height, 
+                                   buf + srcYSize, 
+                                   buf + srcYSize + (srcYSize >> 2), /* UV component is 1/4 of Y */
+                                   width >> 1, dstPitch, width >> 1, height >> 1);
                             } else {
+                                int srcYSize  = width * height;
                                 (*viaFastVidCpy)(pVia->swov.SWDevice.
                                     lpSWOverlaySurface[pVia->dwFrameNum & 1],
                                     buf, dstPitch, width, height, 0);
+
+                                nv12Blit(pVia->swov.SWDevice.lpSWOverlaySurface[pVia->dwFrameNum & 1] + dstPitch * height, 
+                                    buf + srcYSize + (srcYSize >> 2), 
+                                    buf + srcYSize, 
+                                    width >> 1, dstPitch, width, height >> 1);
                             }
                             break;
                         case FOURCC_RV32: