From 5994573b3fde49b7d64bb00e323a92104f5769f5 Mon Sep 17 00:00:00 2001
From: foreman <dl.constructicon@amd.com>
Date: Tue, 19 Dec 2017 13:29:53 -0500
Subject: [PATCH] P4 to Git Change 1496286 by gandryey@gera-w8 on 2017/12/19
 13:24:13

	SWDEV-140401 - [CQE OCL][PAL][Vega10] Observed ~ 80 % performance drop while running GE_HealthCare - PenalityTest
	- Port the logic to consider cache line size for the local workgroup detection.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#41 edit
---
 rocclr/runtime/device/pal/palkernel.cpp | 37 +++++++++++++++++++------
 1 file changed, 28 insertions(+), 9 deletions(-)

diff --git a/rocclr/runtime/device/pal/palkernel.cpp b/rocclr/runtime/device/pal/palkernel.cpp
index 3fbb6c0b9b..cc081d7db2 100644
--- a/rocclr/runtime/device/pal/palkernel.cpp
+++ b/rocclr/runtime/device/pal/palkernel.cpp
@@ -813,10 +813,15 @@ void HSAILKernel::findLocalWorkSize(size_t workDim, const amd::NDRange& gblWorkS
             tmp /= div;
           }
 
+          // Assuming DWORD access
+          const uint cacheLineMatch = dev().settings().cacheLineSize_ >> 2;
+
           // Check if partial dispatch is enabled and
           if (dev().settings().partialDispatch_ &&
-              // we couldn't find optimal workload
-              (lclWorkSize.product() % workGroupInfo()->wavefrontSize_) != 0) {
+            // we couldn't find optimal workload
+            (((lclWorkSize.product() % workGroupInfo()->wavefrontSize_) != 0) ||
+                // or size is too small for the cache line
+            (lclWorkSize[0] < cacheLineMatch))) {
             size_t maxSize = 0;
             size_t maxDim = 0;
             for (uint d = 0; d < workDim; ++d) {
@@ -825,14 +830,28 @@ void HSAILKernel::findLocalWorkSize(size_t workDim, const amd::NDRange& gblWorkS
                 maxDim = d;
               }
             }
-            // Check if a local workgroup has the most optimal size
-            if (thrPerGrp > maxSize) {
-              thrPerGrp = maxSize;
+            // Use X dimension as high priority. Runtime will assume that
+            // X dimension is more important for the address calculation
+            if ((maxDim != 0) && (gblWorkSize[0] >= (cacheLineMatch / 2))) {
+              lclWorkSize[0] = cacheLineMatch;
+              thrPerGrp /= cacheLineMatch;
+              lclWorkSize[maxDim] = thrPerGrp;
+              for (uint d = 1; d < workDim; ++d) {
+                if (d != maxDim) {
+                  lclWorkSize[d] = 1;
+                }
+              }
             }
-            lclWorkSize[maxDim] = thrPerGrp;
-            for (uint d = 0; d < workDim; ++d) {
-              if (d != maxDim) {
-                lclWorkSize[d] = 1;
+            else {
+              // Check if a local workgroup has the most optimal size
+              if (thrPerGrp > maxSize) {
+                thrPerGrp = maxSize;
+              }
+              lclWorkSize[maxDim] = thrPerGrp;
+              for (uint d = 0; d < workDim; ++d) {
+                if (d != maxDim) {
+                  lclWorkSize[d] = 1;
+                }
               }
             }
           }