Mailinglist Archive: radeonhd (307 mails)

[Yang Zhao <yang@xxxxxxxxxx>]

I poked around in R600OverlapCopy() and noticed that in the case where
the offset between src and dst is horizontal, the simple all-at-once
copy seems to work correctly.

For the case where the offset is only vertical, I've modified the
line-by-line copy to instead copy in chunks equal to height of the
non-overlapping area.

The patch has only been tested on a RV770, so I will claim nothing
regarding other GPUs. At least for this particular machine, EXA is now
snappy and corruption free.

-- 
Yang Zhao
http://yangman.ca
From 32be7c13b5b1da4c08070751dd9ef8d218c8a832 Mon Sep 17 00:00:00 2001
From: Yang Zhao <yang@xxxxxxxxxx>
Date: Fri, 6 Feb 2009 00:31:39 -0800
Subject: [PATCH] r6xx/r7xx EXA: Optimize overlapping copy

When source and destination blocks are only offset horizontally, it
appears to be unnecessary to perform careful, segment-by-segment copy.
The code path that does this is taken out completely.

For the case where offset is only vertical, copying is now done by
height of the non-overlapping area each time, instead of always
line-by-line.
---
 src/r600_exa.c |   79 ++++++++++++++++++++-----------------------------------
 1 files changed, 29 insertions(+), 50 deletions(-)

diff --git a/src/r600_exa.c b/src/r600_exa.c
index 7d7c80d..df17704 100644
--- a/src/r600_exa.c
+++ b/src/r600_exa.c
@@ -719,58 +719,37 @@ R600OverlapCopy(PixmapPtr pDst,
     struct r6xx_accel_state *accel_state = rhdPtr->TwoDPrivate;
     uint32_t dst_pitch = exaGetPixmapPitch(pDst) / 
(pDst->drawable.bitsPerPixel / 8);
     uint32_t dst_offset = exaGetPixmapOffset(pDst) + rhdPtr->FbIntAddress + 
rhdPtr->FbScanoutStart;
-    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
-           if (srcX < dstX) { // right
-               // copy right to left
-               for (i = w; 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 + i - 1, srcY, dstX + i - 
1, dstY, 1, h);
-                   R600DoCopy(pScrn);
-               }
-           } else { //left
-               // copy left to right
-               for (i = 0; i < w; 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 + i, srcY, dstX + i, dstY, 
1, 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);
-               }
-           } 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);
-               }
-           }
-       }
+                  dstX, dstX + w, dstY, dstY + h) && (srcY != dstY)) {
+        if (srcY > dstY) { // up
+            // copy top to bottom
+            chunk = srcY - dstY;
+            for (i = 0; i < h; i += chunk) {
+                if (chunk > h - i) chunk = 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, 
chunk);
+                R600DoCopy(pScrn);
+            }
+        } else { // down
+            // copy bottom to top
+            chunk = dstY - srcY;
+            for (i = h; i > 0; i -= chunk) {
+                if (chunk > i) chunk = 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 - chunk, dstX, dstY 
+ i - chunk, w, chunk);
+                R600DoCopy(pScrn);
+            }
+        }
     } else {
        R600DoPrepareCopy(pScrn,
                          dst_pitch, pDst->drawable.width, 
pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-- 
1.6.0.6