R6xx/R7xx EXA: Optimize overlapping copy

Overlapping copy is now done in chunks proportional to the
non-overlapping area.

Diagonal overlaps are also handled properly.

1 files changed, 58 insertions, 28 deletions

diff --git a/src/r600_exa.c b/src/r600_exa.c
index a4e2a4dc..fa99a6ee 100644
--- a/src/r600_exa.c
+++ b/src/r600_exa.c
@@ -675,8 +675,6 @@ R600PrepareCopy(PixmapPtr pSrc,   PixmapPtr pDst,
 	accel_state->rop = rop;
 	accel_state->planemask = planemask;
 
-	return FALSE;
-
 #ifdef SHOW_VERTEXES
 	ErrorF("same surface!\n");
 #endif
@@ -719,58 +717,90 @@ R600OverlapCopy(PixmapPtr pDst,
     struct radeon_accel_state *accel_state = info->accel_state;
     uint32_t dst_pitch = exaGetPixmapPitch(pDst) / (pDst->drawable.bitsPerPixel / 8);
     uint32_t dst_offset = exaGetPixmapOffset(pDst) + info->fbLocation + pScrn->fbOffset;
-    int i;
+    int i, chunk;
 
     if (is_overlap(srcX, srcX + w, srcY, srcY + h,
 		   dstX, dstX + w, dstY, dstY + h)) {
-	if (srcY == dstY) { // left/right
+        /* Diagonally offset overlap is reduced to a horizontal-only offset by first
+         * copying the vertically non-overlapping portion, then adjusting coordinates
+         */
+	if (srcX != dstX) { // left/right or diagonal
+            if (srcY > dstY ) { // diagonal up
+                chunk = srcY - dstY;
+                R600DoPrepareCopy(pScrn,
+                                  dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
+                                  dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
+                                  accel_state->rop, accel_state->planemask);
+                R600AppendCopyVertex(pScrn, srcX, srcY, dstX, dstY, w, chunk);
+                R600DoCopy(pScrn);
+
+                h = h - chunk;
+                srcY = srcY + chunk;
+                dstY = dstY + chunk;
+            } else if (srcY < dstY) { // diagonal down
+                chunk = dstY - srcY;
+                R600DoPrepareCopy(pScrn,
+                                  dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
+                                  dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
+                                  accel_state->rop, accel_state->planemask);
+                R600AppendCopyVertex(pScrn, srcX, srcY + h - chunk, dstX, dstY + h - chunk, w, chunk);
+                R600DoCopy(pScrn);
+
+                h = h - chunk;
+            }
+
 	    if (srcX < dstX) { // right
 		// copy right to left
-		for (i = w; i > 0; i--) {
+                chunk = dstX - srcX;
+		for (i = w; i > 0; i -= chunk) {
 		    R600DoPrepareCopy(pScrn,
 				      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
 				      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
 				      accel_state->rop, accel_state->planemask);
-
-		    R600AppendCopyVertex(pScrn, srcX + i - 1, srcY, dstX + i - 1, dstY, 1, h);
+		    R600AppendCopyVertex(pScrn, srcX + i - chunk, srcY, dstX + i - chunk, dstY, chunk, h);
 		    R600DoCopy(pScrn);
 		}
 	    } else { //left
 		// copy left to right
-		for (i = 0; i < w; i++) {
+                chunk = srcX - dstX;
+		for (i = 0; i < w; i += chunk) {
 		    R600DoPrepareCopy(pScrn,
 				      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
 				      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
 				      accel_state->rop, accel_state->planemask);
 
-		    R600AppendCopyVertex(pScrn, srcX + i, srcY, dstX + i, dstY, 1, h);
+		    R600AppendCopyVertex(pScrn, srcX + i, srcY, dstX + i, dstY, chunk, h);
 		    R600DoCopy(pScrn);
 		}
 	    }
 	} else { //up/down
 	    if (srcY > dstY) { // up
 		// copy top to bottom
-		for (i = 0; i < h; i++) {
-		    R600DoPrepareCopy(pScrn,
-				      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-				      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-				      accel_state->rop, accel_state->planemask);
-
-		    R600AppendCopyVertex(pScrn, srcX, srcY + i, dstX, dstY + i, w, 1);
-		    R600DoCopy(pScrn);
-		}
+                for (i = 0; i < h; i += chunk) {
+                chunk = srcY - dstY;
+                    R600DoPrepareCopy(pScrn,
+                                      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
+                                      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
+                                      accel_state->rop, accel_state->planemask);
+
+                    if (chunk > h - i) chunk = h - i;
+                    R600AppendCopyVertex(pScrn, srcX, srcY + i, dstX, dstY + i, w, chunk);
+                    R600DoCopy(pScrn);
+                }
 	    } else { // down
 		// copy bottom to top
-		for (i = h; i > 0; i--) {
-		    R600DoPrepareCopy(pScrn,
-				      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-				      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-				      accel_state->rop, accel_state->planemask);
-
-		    R600AppendCopyVertex(pScrn, srcX, srcY + i - 1, dstX, dstY + i - 1, w, 1);
-		    R600DoCopy(pScrn);
-		}
-	    }
+		chunk = dstY - srcY;
+                for (i = h; i > 0; i -= chunk) {
+                    R600DoPrepareCopy(pScrn,
+                                      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
+                                      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
+                                      accel_state->rop, accel_state->planemask);
+
+                    if (chunk > i) chunk = i;
+                    R600AppendCopyVertex(pScrn, srcX, srcY + i - chunk, dstX, dstY + i - chunk, w, chunk);
+                    R600DoCopy(pScrn);
+                }
+            }
 	}
     } else {
 	R600DoPrepareCopy(pScrn,