Media
/
compute-runtime
mirror of https://github.com/intel/compute-runtime.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Reuse SVM storage on writeBuffer call. 

- Instead of creating new allocation, look if it already exists if so
re-use it.

Change-Id: I23bc4ac8b8e59e96fce7d48546b76289bedc157f
Signed-off-by: Michal Mrozek <michal.mrozek@intel.com>
This commit is contained in:
Michal Mrozek 2020-01-22 14:50:16 +01:00
parent 0d7cc3a8c0
commit 35b59b7cbe
3 changed files with 126 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Jenkinsfile vendored
View File

@ -1,5 +1,5 @@
#!groovy
dependenciesRevision='31422303aeb91920dbad182d97ef610bf75b5599-1360'
strategy='EQUAL'
allowedCD=259
allowedCD=257
allowedF=11

11
runtime/command_queue/enqueue_write_buffer.h
View File

@ -1,5 +1,5 @@
/*
* Copyright (C) 2017-2019 Intel Corporation
* Copyright (C) 2017-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@ -7,6 +7,7 @@
#pragma once
#include "core/helpers/string.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_stream/command_stream_receiver.h"
@ -65,6 +66,14 @@ cl_int CommandQueueHw<GfxFamily>::enqueueWriteBuffer(
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) {
mapAllocation = svmEntry->cpuAllocation ? svmEntry->cpuAllocation : svmEntry->gpuAllocation;
}
}
if (mapAllocation) {
surfaces[1] = &mapSurface;
mapSurface.setGraphicsAllocation(mapAllocation);

115
unit_tests/memory_manager/unified_memory_manager_tests.cpp
View File

@ -5,12 +5,15 @@
*
*/
#include "core/memory_manager/allocations_list.h"
#include "core/unit_tests/helpers/debug_manager_state_restore.h"
#include "core/unit_tests/page_fault_manager/mock_cpu_page_fault_manager.h"
#include "runtime/api/api.h"
#include "runtime/command_stream/command_stream_receiver.h"
#include "runtime/mem_obj/mem_obj_helper.h"
#include "test.h"
#include "unit_tests/mocks/mock_command_queue.h"
#include "unit_tests/mocks/mock_context.h"
#include "unit_tests/mocks/mock_execution_environment.h"
#include "unit_tests/mocks/mock_memory_manager.h"
#include "unit_tests/mocks/mock_svm_manager.h"
@ -576,3 +579,115 @@ TEST_F(ShareableUnifiedMemoryManagerPropertiesTest, givenShareableUnifiedPropert
EXPECT_TRUE(memoryManager->shareablePassed);
svmManager->freeSVMAlloc(ptr);
}
TEST(UnfiedSharedMemoryTransferCalls, givenHostUSMllocationWhenPointerIsUsedAsWriteBufferSourceThenUSMAllocationIsReused) {
MockContext mockContext;
cl_context clContext = &mockContext;
auto status = CL_SUCCESS;
auto hostMemory = clHostMemAllocINTEL(clContext, nullptr, 4096u, 0u, &status);
auto svmAllocation = mockContext.getSVMAllocsManager()->getSVMAlloc(hostMemory);
ASSERT_EQ(CL_SUCCESS, status);
auto buffer = clCreateBuffer(clContext, CL_MEM_READ_WRITE, 4096u, nullptr, &status);
ASSERT_EQ(CL_SUCCESS, status);
cl_device_id clDevice = mockContext.getDevice(0u);
auto commandQueue = clCreateCommandQueue(clContext, clDevice, 0u, &status);
ASSERT_EQ(CL_SUCCESS, status);
status = clEnqueueWriteBuffer(commandQueue, buffer, false, 0u, 4096u, hostMemory, 0u, nullptr, nullptr);
ASSERT_EQ(CL_SUCCESS, status);
auto neoQueue = castToObject<CommandQueue>(commandQueue);
auto &temporaryAllocations = neoQueue->getGpgpuCommandStreamReceiver().getTemporaryAllocations();
EXPECT_TRUE(temporaryAllocations.peekIsEmpty());
auto osContextId = neoQueue->getGpgpuCommandStreamReceiver().getOsContext().getContextId();
EXPECT_EQ(1u, 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, givenSharedUSMllocationWithoutLocalMemoryWhenPointerIsUsedAsWriteBufferSourceThenUSMAllocationIsReused) {
DebugManagerStateRestore restore;
DebugManager.flags.EnableLocalMemory.set(0);
MockContext mockContext;
cl_context clContext = &mockContext;
cl_device_id clDevice = mockContext.getDevice(0u);
auto status = CL_SUCCESS;
auto sharedMemory = clSharedMemAllocINTEL(clContext, clDevice, nullptr, 4096u, 0u, &status);
auto svmAllocation = mockContext.getSVMAllocsManager()->getSVMAlloc(sharedMemory);
ASSERT_EQ(CL_SUCCESS, status);
auto buffer = clCreateBuffer(clContext, CL_MEM_READ_WRITE, 4096u, nullptr, &status);
ASSERT_EQ(CL_SUCCESS, status);
auto commandQueue = clCreateCommandQueue(clContext, clDevice, 0u, &status);
ASSERT_EQ(CL_SUCCESS, status);
status = clEnqueueWriteBuffer(commandQueue, buffer, false, 0u, 4096u, sharedMemory, 0u, nullptr, nullptr);
ASSERT_EQ(CL_SUCCESS, status);
auto neoQueue = castToObject<CommandQueue>(commandQueue);
auto &temporaryAllocations = neoQueue->getGpgpuCommandStreamReceiver().getTemporaryAllocations();
EXPECT_TRUE(temporaryAllocations.peekIsEmpty());
auto osContextId = neoQueue->getGpgpuCommandStreamReceiver().getOsContext().getContextId();
EXPECT_EQ(1u, 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) {
DebugManagerStateRestore restore;
DebugManager.flags.EnableLocalMemory.set(1);
MockContext mockContext;
cl_context clContext = &mockContext;
cl_device_id clDevice = mockContext.getDevice(0u);
auto status = CL_SUCCESS;
auto sharedMemory = clSharedMemAllocINTEL(clContext, clDevice, nullptr, 4096u, 0u, &status);
auto svmAllocation = mockContext.getSVMAllocsManager()->getSVMAlloc(sharedMemory);
ASSERT_EQ(CL_SUCCESS, status);
auto buffer = clCreateBuffer(clContext, CL_MEM_READ_WRITE, 4096u, nullptr, &status);
ASSERT_EQ(CL_SUCCESS, status);
auto commandQueue = clCreateCommandQueue(clContext, clDevice, 0u, &status);
ASSERT_EQ(CL_SUCCESS, status);
auto neoQueue = castToObject<CommandQueue>(commandQueue);
auto osContextId = neoQueue->getGpgpuCommandStreamReceiver().getOsContext().getContextId();
EXPECT_EQ(1u, svmAllocation->cpuAllocation->getTaskCount(osContextId));
status = clEnqueueWriteBuffer(commandQueue, buffer, false, 0u, 4096u, sharedMemory, 0u, nullptr, nullptr);
ASSERT_EQ(CL_SUCCESS, status);
auto &temporaryAllocations = neoQueue->getGpgpuCommandStreamReceiver().getTemporaryAllocations();
EXPECT_TRUE(temporaryAllocations.peekIsEmpty());
EXPECT_EQ(2u, svmAllocation->cpuAllocation->getTaskCount(osContextId));
status = clReleaseMemObject(buffer);
ASSERT_EQ(CL_SUCCESS, status);
status = clMemFreeINTEL(clContext, sharedMemory);
ASSERT_EQ(CL_SUCCESS, status);
clReleaseCommandQueue(commandQueue);
}