From 0e85ccf084f48381b8e227c0052260b49094d1c0 Mon Sep 17 00:00:00 2001
From: Michal Mrozek <michal.mrozek@intel.com>
Date: Mon, 24 Feb 2020 13:56:41 +0100
Subject: [PATCH] Choose cpu copy for 32 bit application using local memory.

Change-Id: I74aed9475185b09d4569fafb3427052fff73fd89
Signed-off-by: Michal Mrozek <michal.mrozek@intel.com>
---
 opencl/source/mem_obj/buffer.cpp              | 19 +++++++++++--------
 .../read_write_buffer_cpu_copy.cpp            | 18 ++++++++++++++++++
 2 files changed, 29 insertions(+), 8 deletions(-)

diff --git a/opencl/source/mem_obj/buffer.cpp b/opencl/source/mem_obj/buffer.cpp
index 4f2b2db1a1..c782cb36be 100644
--- a/opencl/source/mem_obj/buffer.cpp
+++ b/opencl/source/mem_obj/buffer.cpp
@@ -518,19 +518,22 @@ bool Buffer::isReadWriteOnCpuAllowed() {
 }
 
 bool Buffer::isReadWriteOnCpuPreffered(void *ptr, size_t size) {
-    //if buffer is not zero copy and pointer is aligned it will be more beneficial to do the transfer on GPU
-    if (!isMemObjZeroCopy() && (reinterpret_cast<uintptr_t>(ptr) & (MemoryConstants::cacheLineSize - 1)) == 0) {
-        return false;
-    }
+    if (MemoryPool::isSystemMemoryPool(graphicsAllocation->getMemoryPool())) {
+        //if buffer is not zero copy and pointer is aligned it will be more beneficial to do the transfer on GPU
+        if (!isMemObjZeroCopy() && (reinterpret_cast<uintptr_t>(ptr) & (MemoryConstants::cacheLineSize - 1)) == 0) {
+            return false;
+        }
 
-    //on low power devices larger transfers are better on the GPU
-    if (context->getDevice(0)->getDeviceInfo().platformLP && size > maxBufferSizeForReadWriteOnCpu) {
-        return false;
+        //on low power devices larger transfers are better on the GPU
+        if (context->getDevice(0)->getDeviceInfo().platformLP && size > maxBufferSizeForReadWriteOnCpu) {
+            return false;
+        }
+        return true;
     }
 
     //if we are not in System Memory Pool, it is more beneficial to do the transfer on GPU
     //for 32 bit applications, utilize CPU transfers here.
-    if (!MemoryPool::isSystemMemoryPool(graphicsAllocation->getMemoryPool()) && is64bit) {
+    if (is64bit) {
         return false;
     }
 
diff --git a/opencl/test/unit_test/command_queue/read_write_buffer_cpu_copy.cpp b/opencl/test/unit_test/command_queue/read_write_buffer_cpu_copy.cpp
index 346699b6c1..00b604e9ee 100644
--- a/opencl/test/unit_test/command_queue/read_write_buffer_cpu_copy.cpp
+++ b/opencl/test/unit_test/command_queue/read_write_buffer_cpu_copy.cpp
@@ -7,6 +7,7 @@
 
 #include "shared/source/gmm_helper/gmm.h"
 #include "shared/source/helpers/basic_math.h"
+#include "shared/test/unit_test/helpers/debug_manager_state_restore.h"
 
 #include "opencl/test/unit_test/command_queue/enqueue_read_buffer_fixture.h"
 #include "opencl/test/unit_test/mocks/mock_command_queue.h"
@@ -251,3 +252,20 @@ TEST(ReadWriteBufferOnCpu, given32BitApplicationWhenLocalMemoryPoolAllocationIsA
     EXPECT_TRUE(buffer->isReadWriteOnCpuAllowed());
     EXPECT_TRUE(buffer->isReadWriteOnCpuPreffered(reinterpret_cast<void *>(0x1000), MemoryConstants::pageSize));
 }
+
+TEST(ReadWriteBufferOnCpu, given32BitAppAndLocalMemoryBufferWhenAskedForCpuTransferPreferenceThenTrueIsReturned) {
+    if (is64bit) {
+        GTEST_SKIP();
+    }
+    DebugManagerStateRestore restorer;
+    DebugManager.flags.EnableLocalMemory.set(true);
+    auto device = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(nullptr));
+    MockContext ctx(device.get());
+
+    cl_int retVal = 0;
+    std::unique_ptr<Buffer> buffer(Buffer::create(&ctx, CL_MEM_READ_WRITE, MemoryConstants::pageSize, nullptr, retVal));
+    ASSERT_NE(nullptr, buffer.get());
+
+    EXPECT_TRUE(buffer->isReadWriteOnCpuAllowed());
+    EXPECT_TRUE(buffer->isReadWriteOnCpuPreffered(reinterpret_cast<void *>(0x1000), MemoryConstants::pageSize));
+}
\ No newline at end of file