From a456182888fa049d73f0deb5b37fac082b758fbd Mon Sep 17 00:00:00 2001
From: jatang <jatang@amd.com>
Date: Mon, 5 Jun 2023 18:23:42 -0400
Subject: [PATCH] SWDEV-373396 - OCLCreateBuffer: make the CPU checkResult() 4
 times faster

Change-Id: If20cd6b509896a748f74ab5296cb85f2c4a9f04c
---
 opencl/amdocl/cl_memobj.cpp                   |  4 +--
 .../ocltst/module/runtime/OCLCreateBuffer.cpp | 31 ++++++++++++-------
 .../ocltst/module/runtime/OCLCreateBuffer.h   |  8 +++--
 3 files changed, 26 insertions(+), 17 deletions(-)

diff --git a/opencl/amdocl/cl_memobj.cpp b/opencl/amdocl/cl_memobj.cpp
index 7a4851f110..468d79db6c 100644
--- a/opencl/amdocl/cl_memobj.cpp
+++ b/opencl/amdocl/cl_memobj.cpp
@@ -4240,8 +4240,8 @@ RUNTIME_ENTRY(cl_int, clEnqueueFillBuffer,
     return CL_INVALID_VALUE;
   }
 
-  // Offset must be a multiple of pattern_size
-  if ((offset % pattern_size) != 0) {
+  // Offset and size must be multiple of pattern_size
+  if (!(amd::isMultipleOf(offset, pattern_size) && amd::isMultipleOf(size, pattern_size))) {
     return CL_INVALID_VALUE;
   }
 
diff --git a/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.cpp b/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.cpp
index 6120d5b12f..39b034123e 100644
--- a/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.cpp
+++ b/opencl/tests/ocltst/module/runtime/OCLCreateBuffer.cpp
@@ -62,6 +62,10 @@ void OCLCreateBuffer::open(unsigned int test, char *units, double &conversion,
   maxSize_ = 1000;
 #endif // EMU_ENV
   cl_mem buf = NULL;
+
+  // Make sure to use a size that's multiple of 8 (64bit).
+  maxSize_ &= 0xFFFFFFFFFFFFFFF8;
+
   buf = _wrapper->clCreateBuffer(context_, CL_MEM_READ_WRITE, maxSize_, NULL,
                                  &error_);
   CHECK_RESULT((error_ != CL_SUCCESS), "clCreateBuffer() failed");
@@ -72,7 +76,7 @@ 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(
@@ -90,6 +94,7 @@ void OCLCreateBuffer::run(void) {
   }
 #endif
   void *resultBuf = NULL;
+
   // Reduce the buffer for the step transfers ahead of the allocation,
   // since huge buffers may cause paging and very low performance
   maxSteps /= 16;
@@ -98,16 +103,14 @@ void OCLCreateBuffer::run(void) {
     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);
 
@@ -129,7 +132,7 @@ 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;
@@ -142,14 +145,18 @@ 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;
+    size_t chk_cnt = readSize / sizeof(cl_ulong);
+
+    cl_ulong *cc = reinterpret_cast<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: