From 6bf7514adb8c8a14bd9dd6865803743c73e5e824 Mon Sep 17 00:00:00 2001
From: jatang <jatang@amd.com>
Date: Fri, 24 Mar 2023 11:58:29 -0400
Subject: [PATCH] SWDEV-373396 - OCLCreateBuffer: make the CPU checkResult() 4
 times faster

Change-Id: Ia6e614d76ea0532e7e6424bd33a5e81fbd49db9e
---
 .../ocltst/module/runtime/OCLCreateBuffer.cpp | 26 +++++++++----------
 .../ocltst/module/runtime/OCLCreateBuffer.h   |  8 +++---
 2 files changed, 18 insertions(+), 16 deletions(-)

diff --git a/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.cpp b/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.cpp
index cd6ea46eb5..a9d96bcc20 100644
--- a/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.cpp
+++ b/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.cpp
@@ -72,7 +72,7 @@ void OCLCreateBuffer::open(unsigned int test, char *units, double &conversion,
 void OCLCreateBuffer::run(void) {
   CPerfCounter timer;
 
-  cl_uchar pattern = PATTERN;
+  cl_ulong pattern = PATTERN_20_64BIT;
   timer.Reset();
   timer.Start();
   error_ = /*_wrapper->*/ clEnqueueFillBuffer(
@@ -90,22 +90,20 @@ void OCLCreateBuffer::run(void) {
   }
 #endif
   void *resultBuf = NULL;
-  ;
+
   while ((resultBuf = malloc(maxSteps)) == NULL) {
     maxSteps /= 2;
     continue;
   }
 
-  checkResult(maxSteps, resultBuf, pattern);
+  checkResult(maxSteps, resultBuf, PATTERN_20_64BIT);
 
-  pattern += 1;
-
-  memset(resultBuf, pattern, maxSteps);
+  memset(resultBuf, PATTERN_2A_08BIT, maxSteps);
 
   writeBuffer(maxSteps, resultBuf);
 
   memset(resultBuf, 0x00, maxSteps);
-  checkResult(maxSteps, resultBuf, pattern);
+  checkResult(maxSteps, resultBuf, PATTERN_2A_64BIT);
 
   free(resultBuf);
 
@@ -127,7 +125,7 @@ void OCLCreateBuffer::run(void) {
 }
 
 void OCLCreateBuffer::checkResult(size_t maxSteps, void *resultBuf,
-                                  cl_uchar pattern) {
+                                  cl_ulong pattern) {
   size_t startPoint = 0;
   while ((startPoint) < maxSize_) {
     cl_event ee;
@@ -140,14 +138,16 @@ void OCLCreateBuffer::checkResult(size_t maxSteps, void *resultBuf,
         resultBuf, 0, NULL, &ee);
     CHECK_RESULT((error_ != CL_SUCCESS), "clEnqueueReadBuffer() failed");
     _wrapper->clFinish(cmdQueues_[_deviceId]);
-    size_t cnt = 0;
-    cl_uchar *cc = (cl_uchar *)resultBuf;
-    for (size_t i = 0; i < readSize; i++) {
+
+    size_t err_cnt = 0, chk_cnt = readSize / sizeof(cl_ulong);
+    cl_ulong *cc = (cl_ulong *)resultBuf;
+
+    for (size_t i = 0; i < chk_cnt; i++) {
       if (cc[i] != pattern) {
-        cnt++;
+        err_cnt++;
       }
     }
-    if (cnt != 0) {
+    if (err_cnt != 0) {
       error_ = -1;
       CHECK_RESULT((error_ != CL_SUCCESS), "checkResult() failed");
       break;
diff --git a/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.h b/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.h
index d02117c5b8..2c63984096 100644
--- a/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.h
+++ b/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.h
@@ -22,7 +22,10 @@
 #define _OCL_CREATE_BUFFER_H_
 
 #include "OCLTestImp.h"
-#define PATTERN 0x20
+#define PATTERN_20_08BIT 0x20
+#define PATTERN_20_64BIT 0x2020202020202020
+#define PATTERN_2A_08BIT 0x2a
+#define PATTERN_2A_64BIT 0x2a2a2a2a2a2a2a2a
 
 class OCLCreateBuffer : public OCLTestImp {
  public:
@@ -34,8 +37,7 @@ class OCLCreateBuffer : public OCLTestImp {
                     unsigned int deviceID);
   virtual void run(void);
   virtual void writeBuffer(size_t tmpMaxSize, void* dataBuf);
-  virtual void checkResult(size_t tmpMaxSize, void* resultBuf,
-                           cl_uchar pattern);
+  virtual void checkResult(size_t tmpMaxSize, void* resultBuf, cl_ulong pattern);
   virtual unsigned int close(void);
 
  private: