commit 3ea5b9cc71960a95a34adfde600c8fb87d4f053a
author Jason Sams <jsams@google.com> Wed Nov 14 18:08:34 2012 -0800
committer Jason Sams <jsams@google.com> Wed Nov 14 18:08:34 2012 -0800
tree fb0188736a6259a28e00825328e0a94ae12ca194
parent c5b34c6ff0f1c2a25c45af932e1d7a7b494a34b0

Fix performance when reading or writing large buffers.

Blur intrinsic which uses ~25mb of data would spill the
L2 cache when a smarter walking pattern could reduce this
hit.  We now vary the chunk size on both the processor
count and data size.

N7 execution time drops 1959ms to 930ms
Mako 470ms to 385ms
Manta, no change.

Change-Id: I57e36e2fec3bb51033a9bd9f3040963f5b071eb0

driver/rsdBcc.cpp

diff --git a/driver/rsdBcc.cpp b/driver/rsdBcc.cpp
index ce43331..460967e 100644
--- a/driver/rsdBcc.cpp
+++ b/driver/rsdBcc.cpp
@@ -342,16 +342,39 @@
     RsdHal * dc = (RsdHal *)mtls->rsc->mHal.drv;
 
     if ((dc->mWorkers.mCount >= 1) && s->mHal.info.isThreadable && !dc->mInForEach) {
+        const size_t targetByteChunk = 16 * 1024;
         dc->mInForEach = true;
         if (mtls->fep.dimY > 1) {
-            mtls->mSliceSize = mtls->fep.dimY / (dc->mWorkers.mCount * 4);
+            uint32_t s1 = mtls->fep.dimY / ((dc->mWorkers.mCount + 1) * 4);
+            uint32_t s2 = 0;
+
+            // This chooses our slice size to rate limit atomic ops to
+            // one per 16k bytes of reads/writes.
+            if (mtls->fep.yStrideOut) {
+                s2 = targetByteChunk / mtls->fep.yStrideOut;
+            } else {
+                s2 = targetByteChunk / mtls->fep.yStrideIn;
+            }
+            mtls->mSliceSize = rsMin(s1, s2);
+
             if(mtls->mSliceSize < 1) {
                 mtls->mSliceSize = 1;
             }
 
             rsdLaunchThreads(mrsc, wc_xy, mtls);
         } else {
-            mtls->mSliceSize = mtls->fep.dimX / (dc->mWorkers.mCount * 4);
+            uint32_t s1 = mtls->fep.dimX / ((dc->mWorkers.mCount + 1) * 4);
+            uint32_t s2 = 0;
+
+            // This chooses our slice size to rate limit atomic ops to
+            // one per 16k bytes of reads/writes.
+            if (mtls->fep.eStrideOut) {
+                s2 = targetByteChunk / mtls->fep.eStrideOut;
+            } else {
+                s2 = targetByteChunk / mtls->fep.eStrideIn;
+            }
+            mtls->mSliceSize = rsMin(s1, s2);
+
             if(mtls->mSliceSize < 1) {
                 mtls->mSliceSize = 1;
             }