Media/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-09-15 13:01:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

Append proper flag to buffer object

Browse Source 
Related-To: NEO-4338

Change-Id: I87604992fdfc20cd02773999f7c019344e8e3213
Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>
This commit is contained in:
Lukasz Jobczyk
2020-07-31 07:45:48 +02:00
committed by sys_ocldev
parent c08e3dde82
commit 70ac7ec80e
11 changed files with 49 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
opencl/source/os_interface/linux/drm_command_stream.h
View File

@ -51,7 +51,7 @@ class DrmCommandStreamReceiver : public DeviceCommandStreamReceiver<GfxFamily> {
protected:
MOCKABLE_VIRTUAL void flushInternal(const BatchBuffer &batchBuffer, const ResidencyContainer &allocationsForResidency);
MOCKABLE_VIRTUAL void exec(const BatchBuffer &batchBuffer, uint32_t drmContextId);
MOCKABLE_VIRTUAL void exec(const BatchBuffer &batchBuffer, uint32_t vmHandleId, uint32_t drmContextId);
std::vector<BufferObject *> residency;
std::vector<drm_i915_gem_exec_object2> execObjectsStorage;

3
opencl/source/os_interface/linux/drm_command_stream.inl
View File

@ -90,7 +90,7 @@ bool DrmCommandStreamReceiver<GfxFamily>::flush(BatchBuffer &batchBuffer, Reside
}
template <typename GfxFamily>
void DrmCommandStreamReceiver<GfxFamily>::exec(const BatchBuffer &batchBuffer, uint32_t drmContextId) {
void DrmCommandStreamReceiver<GfxFamily>::exec(const BatchBuffer &batchBuffer, uint32_t vmHandleId, uint32_t drmContextId) {
DrmAllocation *alloc = static_cast<DrmAllocation *>(batchBuffer.commandBufferAllocation);
DEBUG_BREAK_IF(!alloc);
BufferObject *bb = alloc->getBO();
@ -110,6 +110,7 @@ void DrmCommandStreamReceiver<GfxFamily>::exec(const BatchBuffer &batchBuffer, u
int err = bb->exec(static_cast<uint32_t>(alignUp(batchBuffer.usedSize - batchBuffer.startOffset, 8)),
batchBuffer.startOffset, execFlags,
batchBuffer.requiresCoherency,
vmHandleId,
drmContextId,
this->residency.data(), this->residency.size(),
this->execObjectsStorage.data());

2
opencl/source/os_interface/linux/drm_command_stream_bdw_plus.inl
View File

@ -14,7 +14,7 @@ namespace NEO {
template <typename GfxFamily>
void DrmCommandStreamReceiver<GfxFamily>::flushInternal(const BatchBuffer &batchBuffer, const ResidencyContainer &allocationsForResidency) {
this->processResidency(allocationsForResidency, 0u);
this->exec(batchBuffer, static_cast<const OsContextLinux *>(osContext)->getDrmContextIds()[0]);
this->exec(batchBuffer, 0u, static_cast<const OsContextLinux *>(osContext)->getDrmContextIds()[0]);
}
} // namespace NEO

20
opencl/test/unit_test/os_interface/linux/drm_buffer_object_tests.cpp
View File

@ -26,8 +26,8 @@ class TestedBufferObject : public BufferObject {
this->tiling_mode = mode;
}
void fillExecObject(drm_i915_gem_exec_object2 &execObject, uint32_t drmContextId) override {
BufferObject::fillExecObject(execObject, drmContextId);
void fillExecObject(drm_i915_gem_exec_object2 &execObject, uint32_t vmHandleId, uint32_t drmContextId) override {
BufferObject::fillExecObject(execObject, vmHandleId, drmContextId);
execObjectPointerFilled = &execObject;
}
@ -66,7 +66,7 @@ TEST_F(DrmBufferObjectTest, exec) {
mock->ioctl_res = 0;
drm_i915_gem_exec_object2 execObjectsStorage = {};
auto ret = bo->exec(0, 0, 0, false, 1, nullptr, 0u, &execObjectsStorage);
auto ret = bo->exec(0, 0, 0, false, 0, 1, nullptr, 0u, &execObjectsStorage);
EXPECT_EQ(mock->ioctl_res, ret);
EXPECT_EQ(0u, mock->execBuffer.flags);
}
@ -76,7 +76,7 @@ TEST_F(DrmBufferObjectTest, exec_ioctlFailed) {
mock->ioctl_res = -1;
mock->errnoValue = EFAULT;
drm_i915_gem_exec_object2 execObjectsStorage = {};
EXPECT_EQ(EFAULT, bo->exec(0, 0, 0, false, 1, nullptr, 0u, &execObjectsStorage));
EXPECT_EQ(EFAULT, bo->exec(0, 0, 0, false, 0, 1, nullptr, 0u, &execObjectsStorage));
}
TEST_F(DrmBufferObjectTest, setTiling_success) {
@ -105,7 +105,7 @@ TEST_F(DrmBufferObjectTest, givenAddressThatWhenSizeIsAddedCrosses32BitBoundaryW
memset(&execObject, 0, sizeof(execObject));
bo->setAddress(((uint64_t)1u << 32) - 0x1000u);
bo->setSize(0x1000);
bo->fillExecObject(execObject, 1);
bo->fillExecObject(execObject, 0, 1);
//base address + size > size of 32bit address space
EXPECT_TRUE(execObject.flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS);
}
@ -116,7 +116,7 @@ TEST_F(DrmBufferObjectTest, givenAddressThatWhenSizeIsAddedWithin32BitBoundaryWh
memset(&execObject, 0, sizeof(execObject));
bo->setAddress(((uint64_t)1u << 32) - 0x1000u);
bo->setSize(0xFFF);
bo->fillExecObject(execObject, 1);
bo->fillExecObject(execObject, 0, 1);
//base address + size < size of 32bit address space
EXPECT_TRUE(execObject.flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS);
}
@ -133,7 +133,7 @@ TEST_F(DrmBufferObjectTest, onPinIoctlFailed) {
bo->setAddress(reinterpret_cast<uint64_t>(buff.get()));
BufferObject *boArray[1] = {boToPin.get()};
auto ret = bo->pin(boArray, 1, 1);
auto ret = bo->pin(boArray, 1, 0, 1);
EXPECT_EQ(EINVAL, ret);
}
@ -144,7 +144,7 @@ TEST_F(DrmBufferObjectTest, whenPrintExecutionBufferIsSetToTrueThenMessageFoundI
drm_i915_gem_exec_object2 execObjectsStorage = {};
testing::internal::CaptureStdout();
auto ret = bo->exec(0, 0, 0, false, 1, nullptr, 0u, &execObjectsStorage);
auto ret = bo->exec(0, 0, 0, false, 0, 1, nullptr, 0u, &execObjectsStorage);
EXPECT_EQ(0, ret);
std::string output = testing::internal::GetCapturedStdout();
@ -170,7 +170,7 @@ TEST(DrmBufferObjectSimpleTest, givenInvalidBoWhenPinIsCalledThenErrorIsReturned
mock->errnoValue = EFAULT;
BufferObject *boArray[1] = {boToPin.get()};
auto ret = bo->pin(boArray, 1, 1);
auto ret = bo->pin(boArray, 1, 0, 1);
EXPECT_EQ(EFAULT, ret);
}
@ -200,7 +200,7 @@ TEST(DrmBufferObjectSimpleTest, givenArrayOfBosWhenPinnedThenAllBosArePinned) {
BufferObject *array[3] = {boToPin.get(), boToPin2.get(), boToPin3.get()};
bo->setAddress(reinterpret_cast<uint64_t>(buff.get()));
auto ret = bo->pin(array, 3, 1);
auto ret = bo->pin(array, 3, 0, 1);
EXPECT_EQ(mock->ioctl_res, ret);
EXPECT_LT(0u, mock->execBuffer.batch_len);

2
opencl/test/unit_test/os_interface/linux/drm_memory_manager_tests.cpp
View File

@ -2973,7 +2973,7 @@ TEST_F(DrmMemoryManagerWithExplicitExpectationsTest, givenDisabledForcePinAndEna
PinBufferObject(Drm *drm) : BufferObject(drm, 1, 0) {
}
int pin(BufferObject *const boToPin[], size_t numberOfBos, uint32_t drmContextId) override {
int pin(BufferObject *const boToPin[], size_t numberOfBos, uint32_t vmHandleId, uint32_t drmContextId) override {
for (size_t i = 0; i < numberOfBos; i++) {
pinnedBoArray[i] = boToPin[i];
}

5
shared/source/direct_submission/linux/drm_direct_submission.inl
View File

@ -54,12 +54,13 @@ bool DrmDirectSubmission<GfxFamily, Dispatcher>::submit(uint64_t gpuAddress, siz
drm_i915_gem_exec_object2 execObject{};
bool ret = false;
for (const auto &drmContextId : drmContextIds) {
for (uint32_t drmIterator = 0u; drmIterator < drmContextIds.size(); drmIterator++) {
ret |= bb->exec(static_cast<uint32_t>(size),
0,
execFlags,
false,
drmContextId,
drmIterator,
drmContextIds[drmIterator],
nullptr,
0,
&execObject);

1
shared/source/os_interface/linux/CMakeLists.txt
View File

@ -14,6 +14,7 @@ set(NEO_CORE_OS_INTERFACE_LINUX
${CMAKE_CURRENT_SOURCE_DIR}${BRANCH_DIR_SUFFIX}/drm_allocation_extended.cpp
${CMAKE_CURRENT_SOURCE_DIR}/drm_buffer_object.cpp
${CMAKE_CURRENT_SOURCE_DIR}/drm_buffer_object.h
${CMAKE_CURRENT_SOURCE_DIR}${BRANCH_DIR_SUFFIX}/drm_buffer_object_extended.cpp
${CMAKE_CURRENT_SOURCE_DIR}/drm_gem_close_worker.cpp
${CMAKE_CURRENT_SOURCE_DIR}/drm_gem_close_worker.h
${CMAKE_CURRENT_SOURCE_DIR}/drm_memory_manager.cpp

14
shared/source/os_interface/linux/drm_buffer_object.cpp
View File

@ -90,7 +90,7 @@ bool BufferObject::setTiling(uint32_t mode, uint32_t stride) {
return set_tiling.tiling_mode == mode;
}
void BufferObject::fillExecObject(drm_i915_gem_exec_object2 &execObject, uint32_t drmContextId) {
void BufferObject::fillExecObject(drm_i915_gem_exec_object2 &execObject, uint32_t vmHandleId, uint32_t drmContextId) {
execObject.handle = this->handle;
execObject.relocation_count = 0; //No relocations, we are SoftPinning
execObject.relocs_ptr = 0ul;
@ -99,13 +99,15 @@ void BufferObject::fillExecObject(drm_i915_gem_exec_object2 &execObject, uint32_
execObject.flags = EXEC_OBJECT_PINNED | EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
execObject.rsvd1 = drmContextId;
execObject.rsvd2 = 0;
this->fillExecObjectImpl(execObject, vmHandleId);
}
int BufferObject::exec(uint32_t used, size_t startOffset, unsigned int flags, bool requiresCoherency, uint32_t drmContextId, BufferObject *const residency[], size_t residencyCount, drm_i915_gem_exec_object2 *execObjectsStorage) {
int BufferObject::exec(uint32_t used, size_t startOffset, unsigned int flags, bool requiresCoherency, uint32_t vmHandleId, uint32_t drmContextId, BufferObject *const residency[], size_t residencyCount, drm_i915_gem_exec_object2 *execObjectsStorage) {
for (size_t i = 0; i < residencyCount; i++) {
residency[i]->fillExecObject(execObjectsStorage[i], drmContextId);
residency[i]->fillExecObject(execObjectsStorage[i], vmHandleId, drmContextId);
}
this->fillExecObject(execObjectsStorage[residencyCount], drmContextId);
this->fillExecObject(execObjectsStorage[residencyCount], vmHandleId, drmContextId);
drm_i915_gem_execbuffer2 execbuf{};
execbuf.buffers_ptr = reinterpret_cast<uintptr_t>(execObjectsStorage);
@ -176,9 +178,9 @@ void BufferObject::printExecutionBuffer(drm_i915_gem_execbuffer2 &execbuf, const
std::cout << logger << std::endl;
}
int BufferObject::pin(BufferObject *const boToPin[], size_t numberOfBos, uint32_t drmContextId) {
int BufferObject::pin(BufferObject *const boToPin[], size_t numberOfBos, uint32_t vmHandleId, uint32_t drmContextId) {
StackVec<drm_i915_gem_exec_object2, maxFragmentsCount + 1> execObject(numberOfBos + 1);
return this->exec(4u, 0u, 0u, false, drmContextId, boToPin, numberOfBos, &execObject[0]);
return this->exec(4u, 0u, 0u, false, vmHandleId, drmContextId, boToPin, numberOfBos, &execObject[0]);
}
} // namespace NEO

7
shared/source/os_interface/linux/drm_buffer_object.h
View File

@ -39,9 +39,9 @@ class BufferObject {
bool setTiling(uint32_t mode, uint32_t stride);
MOCKABLE_VIRTUAL int pin(BufferObject *const boToPin[], size_t numberOfBos, uint32_t drmContextId);
MOCKABLE_VIRTUAL int pin(BufferObject *const boToPin[], size_t numberOfBos, uint32_t vmHandleId, uint32_t drmContextId);
int exec(uint32_t used, size_t startOffset, unsigned int flags, bool requiresCoherency, uint32_t drmContextId, BufferObject *const residency[], size_t residencyCount, drm_i915_gem_exec_object2 *execObjectsStorage);
int exec(uint32_t used, size_t startOffset, unsigned int flags, bool requiresCoherency, uint32_t vmHandleId, uint32_t drmContextId, BufferObject *const residency[], size_t residencyCount, drm_i915_gem_exec_object2 *execObjectsStorage);
void bind(uint32_t vmHandleId);
void unbind(uint32_t vmHandleId);
@ -78,7 +78,8 @@ class BufferObject {
//Tiling
uint32_t tiling_mode;
MOCKABLE_VIRTUAL void fillExecObject(drm_i915_gem_exec_object2 &execObject, uint32_t drmContextId);
MOCKABLE_VIRTUAL void fillExecObject(drm_i915_gem_exec_object2 &execObject, uint32_t vmHandleId, uint32_t drmContextId);
void fillExecObjectImpl(drm_i915_gem_exec_object2 &execObject, uint32_t vmHandleId);
uint64_t gpuAddress = 0llu;

14
shared/source/os_interface/linux/drm_buffer_object_extended.cpp Normal file
View File

@ -0,0 +1,14 @@
/*
* Copyright (C) 2017-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "shared/source/os_interface/linux/drm_buffer_object.h"
namespace NEO {
void BufferObject::fillExecObjectImpl(drm_i915_gem_exec_object2 &execObject, uint32_t vmHandleId) {}
} // namespace NEO

8
shared/source/os_interface/linux/drm_memory_manager.cpp
View File

@ -181,7 +181,7 @@ NEO::BufferObject *DrmMemoryManager::allocUserptr(uintptr_t address, size_t size
void DrmMemoryManager::emitPinningRequest(BufferObject *bo, const AllocationData &allocationData) const {
if (forcePinEnabled && pinBBs.at(allocationData.rootDeviceIndex) != nullptr && allocationData.flags.forcePin && allocationData.size >= this->pinThreshold) {
pinBBs.at(allocationData.rootDeviceIndex)->pin(&bo, 1, getDefaultDrmContextId());
pinBBs.at(allocationData.rootDeviceIndex)->pin(&bo, 1, 0, getDefaultDrmContextId());
}
}
@ -276,7 +276,7 @@ GraphicsAllocation *DrmMemoryManager::allocateGraphicsMemoryWithGpuVa(const Allo
BufferObject *boPtr = bo.get();
if (forcePinEnabled && pinBBs.at(allocationData.rootDeviceIndex) != nullptr && alignedSize >= this->pinThreshold) {
pinBBs.at(allocationData.rootDeviceIndex)->pin(&boPtr, 1, osContextLinux->getContextId());
pinBBs.at(allocationData.rootDeviceIndex)->pin(&boPtr, 1, 0, osContextLinux->getDrmContextIds()[0]);
}
auto allocation = new DrmAllocation(allocationData.rootDeviceIndex, allocationData.type, bo.get(), res, bo->gpuAddress, alignedSize, MemoryPool::System4KBPages);
@ -310,7 +310,7 @@ DrmAllocation *DrmMemoryManager::allocateGraphicsMemoryForNonSvmHostPtr(const Al
if (validateHostPtrMemory) {
auto boPtr = bo.get();
int result = pinBBs.at(allocationData.rootDeviceIndex)->pin(&boPtr, 1, getDefaultDrmContextId());
int result = pinBBs.at(allocationData.rootDeviceIndex)->pin(&boPtr, 1, 0, getDefaultDrmContextId());
if (result != SUCCESS) {
unreference(bo.release(), true);
releaseGpuRange(reinterpret_cast<void *>(gpuVirtualAddress), alignedSize, allocationData.rootDeviceIndex);
@ -666,7 +666,7 @@ MemoryManager::AllocationStatus DrmMemoryManager::populateOsHandles(OsHandleStor
}
if (validateHostPtrMemory) {
int result = pinBBs.at(rootDeviceIndex)->pin(allocatedBos, numberOfBosAllocated, getDefaultDrmContextId());
int result = pinBBs.at(rootDeviceIndex)->pin(allocatedBos, numberOfBosAllocated, 0, getDefaultDrmContextId());
if (result == EFAULT) {
for (uint32_t i = 0; i < numberOfBosAllocated; i++) {