AI/rocm-systems
2
0
Garpu 0
Kode Masalah Tarik Permintaan Actions Packages Projects Rilis Wiki Kegiatan

P4 to Git Change 2005779 by vsytchen@vsytchen-hip-win10 on 2019/09/27 18:12:14

Telusuri Sumber 
SWDEV-192384 - [HIP CQE][HIPonPAL][19.40] hipBindTexRef1DFetch, hipTextureRef2D are failed on all ASICs for both Win/Lnx

	The runtime cannot trivially determine all the resources that will be used by a kernel, thus it can fail to make all of them resident.

	1. Add new runtime flag PAL_ALWAYS_RESIDENT. Enabling this setting will cause resources to become resident at allocation time.
	2. Set the default value of the above flag to true for HIP and false for OCL.

	ReviewBoardURL = http://ocltc.amd.com/reviews/r/18054/diff/

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#79 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.hpp#30 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#100 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.hpp#27 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#153 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#319 edit
This commit is contained in:
foreman
2019-09-27 18:23:31 -04:00
orang tua bf668d922c
melakukan 2404ade2ef
6 mengubah file dengan 40 tambahan dan 2 penghapusan
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/runtime/device/pal/palresource.cpp
+29
Melihat File
@@ -28,6 +28,17 @@
namespace pal {
// ================================================================================================
Pal::Result GpuMemoryReference::MakeResident() const {
Pal::Result result = Pal::Result::Success;
if (device_.settings().alwaysResident_) {
Pal::GpuMemoryRef memRef = {};
memRef.pGpuMemory = gpuMem_;
result = device_.iDev()->AddGpuMemoryReferences(1, &memRef, nullptr, Pal::GpuMemoryRefCantTrim);
}
return result;
}
// ================================================================================================
GpuMemoryReference* GpuMemoryReference::Create(const Device& dev,
const Pal::GpuMemoryCreateInfo& createInfo) {
@@ -46,6 +57,9 @@ GpuMemoryReference* GpuMemoryReference::Create(const Device& dev,
// If cache was freed, then try to allocate again
result = dev.iDev()->CreateGpuMemory(createInfo, &memRef[1], &memRef->gpuMem_);
}
if (result == Pal::Result::Success) {
result = memRef->MakeResident();
}
if (result != Pal::Result::Success) {
memRef->release();
return nullptr;
@@ -71,6 +85,9 @@ GpuMemoryReference* GpuMemoryReference::Create(const Device& dev,
Pal::VaRange vaRange = Pal::VaRange::Default;
if (memRef != nullptr) {
result = dev.iDev()->CreatePinnedGpuMemory(createInfo, &memRef[1], &memRef->gpuMem_);
if (result == Pal::Result::Success) {
result = memRef->MakeResident();
}
if (result != Pal::Result::Success) {
memRef->release();
return nullptr;
@@ -93,6 +110,9 @@ GpuMemoryReference* GpuMemoryReference::Create(const Device& dev,
GpuMemoryReference* memRef = new (gpuMemSize) GpuMemoryReference(dev);
if (memRef != nullptr) {
result = dev.iDev()->CreateSvmGpuMemory(createInfo, &memRef[1], &memRef->gpuMem_);
if (result == Pal::Result::Success) {
result = memRef->MakeResident();
}
if (result != Pal::Result::Success) {
memRef->release();
return nullptr;
@@ -117,6 +137,9 @@ GpuMemoryReference* GpuMemoryReference::Create(const Device& dev,
if (memRef != nullptr) {
result = dev.iDev()->OpenExternalSharedGpuMemory(openInfo, &memRef[1], &createInfo,
&memRef->gpuMem_);
if (result == Pal::Result::Success) {
result = memRef->MakeResident();
}
if (result != Pal::Result::Success) {
memRef->release();
return nullptr;
@@ -144,6 +167,9 @@ GpuMemoryReference* GpuMemoryReference::Create(const Device& dev,
if (memRef != nullptr) {
result = dev.iDev()->OpenExternalSharedImage(openInfo, imgMem, &memRef[1], &createInfo, image,
&memRef->gpuMem_);
if (result == Pal::Result::Success) {
result = memRef->MakeResident();
}
if (result != Pal::Result::Success) {
memRef->release();
return nullptr;
@@ -164,6 +190,9 @@ GpuMemoryReference* GpuMemoryReference::Create(const Device& dev,
GpuMemoryReference* memRef = new (gpuMemSize) GpuMemoryReference(dev);
if (memRef != nullptr) {
result = dev.iDev()->OpenPeerGpuMemory(openInfo, &memRef[1], &memRef->gpuMem_);
if (result == Pal::Result::Success) {
result = memRef->MakeResident();
}
if (result != Pal::Result::Success) {
memRef->release();
return nullptr;
rocclr/runtime/device/pal/palresource.hpp
+2
Melihat File
@@ -43,6 +43,8 @@ class GpuMemoryReference : public amd::ReferenceCountedObject {
//! Get PAL memory object
Pal::IGpuMemory* iMem() const { return gpuMem_; }
Pal::Result MakeResident() const;
Pal::IGpuMemory* gpuMem_; //!< PAL GPU memory object
void* cpuAddress_; //!< CPU address of this memory
const Device& device_; //!< GPU device
rocclr/runtime/device/pal/palsettings.cpp
+1
Melihat File
@@ -152,6 +152,7 @@ Settings::Settings() {
imageBufferWar_ = false;
disableSdma_ = PAL_DISABLE_SDMA;
mallPolicy_ = 0;
alwaysResident_ = amd::IS_HIP ? true : ((!flagIsDefault(PAL_ALWAYS_RESIDENT)) ? PAL_ALWAYS_RESIDENT : false);
}
bool Settings::create(const Pal::DeviceProperties& palProp,
rocclr/runtime/device/pal/palsettings.hpp
+2 -1
Melihat File
@@ -64,7 +64,8 @@ class Settings : public device::Settings {
uint enableHwP2P_ : 1; //!< Forces HW P2P path for testing
uint imageBufferWar_ : 1; //!< Image buffer workaround for Gfx10
uint disableSdma_ : 1; //!< Disable SDMA support
uint reserved_ : 8;
uint alwaysResident_ : 1; //!< Make resources resident at allocation time
uint reserved_ : 7;
};
uint value_;
};
rocclr/runtime/device/pal/palvirtual.cpp
+3
Melihat File
@@ -254,6 +254,9 @@ Pal::Result VirtualGPU::Queue::UpdateAppPowerProfile() {
}
void VirtualGPU::Queue::addCmdMemRef(GpuMemoryReference* mem) {
if (gpu_.dev().settings().alwaysResident_) {
return;
}
Pal::IGpuMemory* iMem = mem->iMem();
auto it = memReferences_.find(mem);
if (it != memReferences_.end()) {
rocclr/runtime/utils/flags.hpp
+3 -1
Melihat File
@@ -202,7 +202,9 @@ release(uint, AMD_SERIALIZE_KERNEL, 0, \
"0x2 = Wait for completion after enqueue 0x3 = both") \
release(uint, AMD_SERIALIZE_COPY, 0, \
"Serialize copies, 0x1 = Wait for completion before enqueue" \
"0x2 = Wait for completion after enqueue 0x3 = both")
"0x2 = Wait for completion after enqueue 0x3 = both") \
release(bool, PAL_ALWAYS_RESIDENT, false, \
"Force memory resources to become resident at allocation time")
namespace amd {