Add support for SVM pointer reuse to enqueueReadBuffer.

Change-Id: I7a6718b2ebe48912a19af3da5e233acd84bdd3ef
Signed-off-by: Michal Mrozek <michal.mrozek@intel.com>

Author: MichalMrozek

runtime/command_queue/enqueue_read_buffer.h

@@ -1,12 +1,13 @@
 /*
- * Copyright (C) 2017-2019 Intel Corporation
+ * Copyright (C) 2017-2020 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
  */
 
 #pragma once
 #include "core/helpers/cache_policy.h"
+#include "core/memory_manager/unified_memory_manager.h"
 #include "runtime/built_ins/built_ins.h"
 #include "runtime/command_queue/command_queue_hw.h"
 #include "runtime/command_queue/enqueue_common.h"
@@ -69,6 +70,21 @@ cl_int CommandQueueHw<GfxFamily>::enqueueReadBuffer(
     GeneralSurface mapSurface;
     Surface *surfaces[] = {&bufferSurf, nullptr};
 
+    //check if we are dealing with SVM pointer here for which we already have an allocation
+    if (!mapAllocation && this->getContext().getSVMAllocsManager()) {
+        auto svmEntry = this->getContext().getSVMAllocsManager()->getSVMAlloc(ptr);
+        if (svmEntry) {
+            if (svmEntry->memoryType == DEVICE_UNIFIED_MEMORY) {
+                return CL_INVALID_OPERATION;
+            }
+            if ((svmEntry->gpuAllocation->getGpuAddress() + svmEntry->size) < (castToUint64(ptr) + size)) {
+                return CL_INVALID_OPERATION;
+            }
+
+            mapAllocation = svmEntry->cpuAllocation ? svmEntry->cpuAllocation : svmEntry->gpuAllocation;
+        }
+    }
+
     if (mapAllocation) {
         surfaces[1] = &mapSurface;
         mapSurface.setGraphicsAllocation(mapAllocation);

unit_tests/memory_manager/unified_memory_manager_tests.cpp

@@ -580,7 +580,7 @@ TEST_F(ShareableUnifiedMemoryManagerPropertiesTest, givenShareableUnifiedPropert
     svmManager->freeSVMAlloc(ptr);
 }
 
-TEST(UnfiedSharedMemoryTransferCalls, givenHostUSMllocationWhenPointerIsUsedAsWriteBufferSourceThenUSMAllocationIsReused) {
+TEST(UnfiedSharedMemoryTransferCalls, givenHostUSMllocationWhenPointerIsUsedForTransferCallsThenUSMAllocationIsReused) {
     MockContext mockContext;
     cl_context clContext = &mockContext;
 
@@ -608,13 +608,18 @@ TEST(UnfiedSharedMemoryTransferCalls, givenHostUSMllocationWhenPointerIsUsedAsWr
 
     EXPECT_EQ(1u, svmAllocation->gpuAllocation->getTaskCount(osContextId));
 
+    status = clEnqueueReadBuffer(commandQueue, buffer, false, 0u, 4096u, hostMemory, 0u, nullptr, nullptr);
+    ASSERT_EQ(CL_SUCCESS, status);
+    EXPECT_TRUE(temporaryAllocations.peekIsEmpty());
+    EXPECT_EQ(2u, svmAllocation->gpuAllocation->getTaskCount(osContextId));
+
     status = clReleaseMemObject(buffer);
     ASSERT_EQ(CL_SUCCESS, status);
     status = clMemFreeINTEL(clContext, hostMemory);
     ASSERT_EQ(CL_SUCCESS, status);
     clReleaseCommandQueue(commandQueue);
 }
-TEST(UnfiedSharedMemoryTransferCalls, givenDeviceUsmAllocationWhenItIsPassedToWriteBufferAsSourceThenErrorIsReturned) {
+TEST(UnfiedSharedMemoryTransferCalls, givenDeviceUsmAllocationWhenPtrIsUsedForTransferCallsThenErrorIsReturned) {
     MockContext mockContext;
     cl_context clContext = &mockContext;
 
@@ -633,14 +638,17 @@ TEST(UnfiedSharedMemoryTransferCalls, givenDeviceUsmAllocationWhenItIsPassedToWr
     status = clEnqueueWriteBuffer(commandQueue, buffer, false, 0u, 4096u, deviceMemory, 0u, nullptr, nullptr);
     EXPECT_EQ(CL_INVALID_OPERATION, status);
 
+    status = clEnqueueReadBuffer(commandQueue, buffer, false, 0u, 4096u, deviceMemory, 0u, nullptr, nullptr);
+    ASSERT_EQ(CL_INVALID_OPERATION, status);
+
     status = clReleaseMemObject(buffer);
     ASSERT_EQ(CL_SUCCESS, status);
     status = clMemFreeINTEL(clContext, deviceMemory);
     ASSERT_EQ(CL_SUCCESS, status);
     clReleaseCommandQueue(commandQueue);
 }
 
-TEST(UnfiedSharedMemoryTransferCalls, givenHostAllocationThatIsSmallerThenWriteBufferTranfserSizeWhenTransferCallIsEmittedThenErrorIsReturned) {
+TEST(UnfiedSharedMemoryTransferCalls, givenHostAllocationThatIsSmallerThenTransferRequirementsThenErrorIsReturned) {
     MockContext mockContext;
     cl_context clContext = &mockContext;
 
@@ -660,14 +668,17 @@ TEST(UnfiedSharedMemoryTransferCalls, givenHostAllocationThatIsSmallerThenWriteB
     status = clEnqueueWriteBuffer(commandQueue, buffer, false, 0u, 4096u, hostMemory, 0u, nullptr, nullptr);
     EXPECT_EQ(CL_INVALID_OPERATION, status);
 
+    status = clEnqueueReadBuffer(commandQueue, buffer, false, 0u, 4096u, hostMemory, 0u, nullptr, nullptr);
+    ASSERT_EQ(CL_INVALID_OPERATION, status);
+
     status = clReleaseMemObject(buffer);
     ASSERT_EQ(CL_SUCCESS, status);
     status = clMemFreeINTEL(clContext, hostMemory);
     ASSERT_EQ(CL_SUCCESS, status);
     clReleaseCommandQueue(commandQueue);
 }
 
-TEST(UnfiedSharedMemoryTransferCalls, givenSharedUSMllocationWithoutLocalMemoryWhenPointerIsUsedAsWriteBufferSourceThenUSMAllocationIsReused) {
+TEST(UnfiedSharedMemoryTransferCalls, givenSharedUSMllocationWithoutLocalMemoryWhenPointerIsUsedAsTranfserParameterThenUSMAllocationIsReused) {
     DebugManagerStateRestore restore;
     DebugManager.flags.EnableLocalMemory.set(0);
 
@@ -697,14 +708,19 @@ TEST(UnfiedSharedMemoryTransferCalls, givenSharedUSMllocationWithoutLocalMemoryW
 
     EXPECT_EQ(1u, svmAllocation->gpuAllocation->getTaskCount(osContextId));
 
+    status = clEnqueueReadBuffer(commandQueue, buffer, false, 0u, 4096u, sharedMemory, 0u, nullptr, nullptr);
+    ASSERT_EQ(CL_SUCCESS, status);
+    EXPECT_TRUE(temporaryAllocations.peekIsEmpty());
+    EXPECT_EQ(2u, svmAllocation->gpuAllocation->getTaskCount(osContextId));
+
     status = clReleaseMemObject(buffer);
     ASSERT_EQ(CL_SUCCESS, status);
     status = clMemFreeINTEL(clContext, sharedMemory);
     ASSERT_EQ(CL_SUCCESS, status);
     clReleaseCommandQueue(commandQueue);
 }
 
-TEST(UnfiedSharedMemoryTransferCalls, givenSharedUSMllocationWithLocalMemoryWhenPointerIsUsedAsWriteBufferSourceThenUSMAllocationIsReused) {
+TEST(UnfiedSharedMemoryTransferCalls, givenSharedUSMllocationWithLocalMemoryWhenPointerIsUsedAsTransferParameterThenUSMAllocationIsReused) {
     DebugManagerStateRestore restore;
     DebugManager.flags.EnableLocalMemory.set(1);
 
@@ -737,6 +753,11 @@ TEST(UnfiedSharedMemoryTransferCalls, givenSharedUSMllocationWithLocalMemoryWhen
 
     EXPECT_EQ(2u, svmAllocation->cpuAllocation->getTaskCount(osContextId));
 
+    status = clEnqueueReadBuffer(commandQueue, buffer, false, 0u, 4096u, sharedMemory, 0u, nullptr, nullptr);
+    ASSERT_EQ(CL_SUCCESS, status);
+    EXPECT_TRUE(temporaryAllocations.peekIsEmpty());
+    EXPECT_EQ(3u, svmAllocation->cpuAllocation->getTaskCount(osContextId));
+
     status = clReleaseMemObject(buffer);
     ASSERT_EQ(CL_SUCCESS, status);
     status = clMemFreeINTEL(clContext, sharedMemory);