[PATCH] Add memcpy-based UTS/DFS.

EXA hits more optimized paths when it does not have to fallback because of
missing UTS/DFS.

1 files changed, 68 insertions, 6 deletions

diff --git a/src/atimach64exa.c b/src/atimach64exa.c
index 8ba3e09..2adf2ca 100644
--- a/src/atimach64exa.c
+++ b/src/atimach64exa.c
@@ -56,6 +56,8 @@
 #include "config.h"
 #endif
 
+#include <string.h>
+
 #include "ati.h"
 #include "atichip.h"
 #include "atidri.h"
@@ -299,7 +301,8 @@ Mach64Copy
      * for the fact that the problem occurs less often (but still occurs) when
      * copying larger rectangles.
      */
-    if ((pATI->Chip >= ATI_CHIP_264VTB) && !pATI->OptionDevel) {
+    if ((pATI->Chip >= ATI_CHIP_264VTB) && !pATI->OptionDevel)
+    {
         exaMarkSync(pScreenInfo->pScreen); /* Force sync. */
         exaWaitSync(pScreenInfo->pScreen); /* Sync and notify EXA. */
     }
@@ -384,6 +387,60 @@ Mach64Solid
 
 static void Mach64DoneSolid(PixmapPtr pPixmap) { }
 
+/*
+ * Memcpy-based UTS.
+ */
+static Bool
+Mach64UploadToScreen(PixmapPtr pDst, int x, int y, int w, int h,
+    char *src, int src_pitch)
+{
+    char  *dst        = pDst->devPrivate.ptr;
+    int    dst_pitch  = exaGetPixmapPitch(pDst);
+
+    int bpp    = pDst->drawable.bitsPerPixel;
+    int cpp    = (bpp + 7) / 8;
+    int wBytes = w * cpp;
+
+    exaWaitSync(pDst->drawable.pScreen);
+
+    dst += (x * cpp) + (y * dst_pitch);
+
+    while (h--) {
+        memcpy(dst, src, wBytes);
+        src += src_pitch;
+        dst += dst_pitch;
+    }
+
+    return TRUE;
+}
+
+/*
+ * Memcpy-based DFS.
+ */
+static Bool
+Mach64DownloadFromScreen(PixmapPtr pSrc, int x, int y, int w, int h,
+    char *dst, int dst_pitch)
+{
+    char  *src        = pSrc->devPrivate.ptr;
+    int    src_pitch  = exaGetPixmapPitch(pSrc);
+
+    int bpp    = pSrc->drawable.bitsPerPixel;
+    int cpp    = (bpp + 7) / 8;
+    int wBytes = w * cpp;
+
+    exaWaitSync(pSrc->drawable.pScreen);
+
+    src += (x * cpp) + (y * src_pitch);
+
+    while (h--) {
+        memcpy(dst, src, wBytes);
+        src += src_pitch;
+        dst += dst_pitch;
+    }
+
+    return TRUE;
+}
+
 /* Compute log base 2 of val. */
 static __inline__ int Mach64Log2(int val)
 {
@@ -562,7 +619,8 @@ Mach64SetupMemEXA(ScreenPtr pScreen)
                    "Will use depth buffer at offset 0x%x\n",
                    pATIDRIServer->depthOffset);
 
-        if (pATIDRIServer->textureSize > 0) {
+        if (pATIDRIServer->textureSize > 0)
+        {
             xf86DrvMsg(pScreen->myNum, X_INFO,
                    "Will use %d kB for local textures at offset 0x%x\n",
                    pATIDRIServer->textureSize/1024,
@@ -571,10 +629,8 @@ Mach64SetupMemEXA(ScreenPtr pScreen)
     }
 #endif /* XF86DRI_DEVEL */
 
-    /* FIXME: set to 64 if that helps for hostdata blits or textures */
-    /* FIXME: must be multiple of 8 pixels; 32 is ok for 16bpp, 32bpp */
-    pExa->pixmapOffsetAlign = 32;
-    pExa->pixmapPitchAlign = 32;
+    pExa->pixmapOffsetAlign = 64;
+    pExa->pixmapPitchAlign = 64;
 
     pExa->flags = EXA_OFFSCREEN_PIXMAPS;
 
@@ -618,6 +674,12 @@ Bool ATIMach64ExaInit(ScreenPtr pScreen)
     pExa->Copy = Mach64Copy;
     pExa->DoneCopy = Mach64DoneCopy;
 
+    /* EXA hits more optimized paths when it does not have to fallback because
+     * of missing UTS/DFS, hook memcpy-based UTS/DFS.
+     */
+    pExa->UploadToScreen = Mach64UploadToScreen;
+    pExa->DownloadFromScreen = Mach64DownloadFromScreen;
+
     if (!exaDriverInit(pScreen, pATI->pExa)) {
 	xfree(pATI->pExa);
 	pATI->pExa = NULL;