AI/rocm-systems
2
0
Fork 0
程式碼 問題 合併請求 Actions 套件 專案 版本發布 Wiki 動態

P4 to Git Change 1338072 by wchau@wchau_WIN_OCL_HSA on 2016/11/08 10:02:39

瀏覽代碼 
SWDEV-95007 - OpenCL on PAL - DVR

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.hpp#14 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevicegl.cpp#3 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#18 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.hpp#9 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#37 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.hpp#16 edit


[ROCm/clr commit: 94bf561bd7]
此提交包含在：
foreman
2016-11-08 10:11:21 -05:00
父節點 64eb9eb9b9
當前提交 c59607b856
共有 6 個檔案被更改，包括 94 行新增 和 15 行删除
下載 Patch 檔 下載差異檔 展開所有檔案 折疊所有檔案
projects/clr/rocclr/runtime/device/pal/paldevice.hpp
+5 -2
查看文件
@@ -492,8 +492,11 @@ public:
//! Interop for GL device
bool initGLInteropPrivateExt(void* GLplatformContext, void* GLdeviceContext) const;
bool glCanInterop(void* GLplatformContext, void* GLdeviceContext) const;
bool resGLAssociate(void* GLContext, uint name, uint type,
void** handle, void** mbResHandle, size_t* offset) const;
bool resGLAssociate(void* GLContext, uint name, uint type, void** handle, void** mbResHandle, size_t* offset
#ifdef ATI_OS_WIN
, Pal::DoppDesktopInfo& doppDesktopInfo
#endif
) const;
bool resGLAcquire(void* GLplatformContext, void* mbResHandle, uint type) const;
bool resGLRelease(void* GLplatformContext, void* mbResHandle, uint type) const;
bool resGLFree(void* GLplatformContext, void* mbResHandle, uint type) const;
projects/clr/rocclr/runtime/device/pal/paldevicegl.cpp
+15 -1
查看文件
@@ -185,7 +185,11 @@ Device::resGLAssociate(
uint type,
void** handle,
void** mbResHandle,
size_t* offset) const
size_t* offset
#ifdef ATI_OS_WIN
, Pal::DoppDesktopInfo& doppDesktopInfo
#endif
) const
{
amd::ScopedLock lk(lockPAL());
@@ -218,6 +222,16 @@ Device::resGLAssociate(
*handle = reinterpret_cast<void*>(hData.handle);
*mbResHandle = reinterpret_cast<void*>(hData.mbResHandle);
*offset = static_cast<size_t>(hData.offset);
#ifdef ATI_OS_WIN
if (hData.isDoppDesktopTexture) {
doppDesktopInfo.gpuVirtAddr = hData.cardAddr;
doppDesktopInfo.vidPnSourceId = hData.vidpnSourceId;
}
else {
doppDesktopInfo.gpuVirtAddr = 0;
doppDesktopInfo.vidPnSourceId = 0;
}
#endif
return status;
}
projects/clr/rocclr/runtime/device/pal/palresource.cpp
+6 -4
查看文件
@@ -410,7 +410,7 @@ Resource::memTypeToHeap(Pal::GpuMemoryCreateInfo* createInfo)
break;
default:
createInfo->heaps[0] = Pal::GpuHeapLocal;
break;
break;
}
}
@@ -512,9 +512,11 @@ Resource::create(MemoryType memType, CreateParams* params)
mipLevel = oglRes->mipLevel_;
if (!dev().resGLAssociate(oglRes->glPlatformContext_, oglRes->handle_,
glType_, &openInfo.hExternalResource, &glInteropMbRes_, &offset_)) {
glType_, &openInfo.hExternalResource, &glInteropMbRes_, &offset_,
openInfo.doppDesktopInfo)) {
return false;
}
desc_.isDoppTexture_ = (openInfo.doppDesktopInfo.gpuVirtAddr != 0);
}
else {
D3DInteropParams* d3dRes = reinterpret_cast<D3DInteropParams*>(params);
@@ -849,7 +851,7 @@ Resource::create(MemoryType memType, CreateParams* params)
imgCreateInfo.fragments = 1;
Pal::ImageTiling tiling = Pal::ImageTiling::Optimal;
if (((memoryType() == Persistent) &&
if (((memoryType() == Persistent) &&
dev().settings().linearPersistentImage_) ||
(memoryType() == ImageBuffer)) {
tiling = Pal::ImageTiling::Linear;
@@ -1275,7 +1277,7 @@ Resource::partialMemCopyTo(
if (((copyRegion.gpuMemoryRowPitch % 4) != 0) ||
// another DRM restriciton... SI has 4 pixels
(copyRegion.gpuMemoryOffset % 4 != 0) ||
(dev().settings().sdamPageFaultWar_ &&
(dev().settings().sdamPageFaultWar_ &&
(copyRegion.imageOffset.x % dstResource.elementSize() != 0))) {
result = false;
}
projects/clr/rocclr/runtime/device/pal/palresource.hpp
+2 -1
查看文件
@@ -169,7 +169,7 @@ public:
size_t height_; //!< Resource height
size_t depth_; //!< Resource depth
uint baseLevel_; //!< The base level for the view
uint mipLevels_; //!< Number of mip levels
uint mipLevels_; //!< Number of mip levels
uint flags_; //!< Resource flags, used in creation
size_t pitch_; //!< Resource pitch, valid if locked
size_t slice_; //!< Resource slice, valid if locked
@@ -185,6 +185,7 @@ public:
uint SVMRes_ : 1; //!< SVM flag to the cal resource
uint scratch_ : 1; //!< Scratch buffer
uint isAllocExecute_ : 1; //!< SVM resource allocation attribute for shader\cmdbuf
uint isDoppTexture_ : 1; //!< PAL resource is for a DOPP desktop texture
};
uint state_;
};
projects/clr/rocclr/runtime/device/pal/palvirtual.cpp
+56 -7
查看文件
@@ -156,7 +156,7 @@ VirtualGPU::Queue::~Queue()
void
VirtualGPU::Queue::addCmdMemRef(Pal::IGpuMemory* iMem)
{
{
auto it = memReferences_.find(iMem);
if (it != memReferences_.end()) {
it->second = (it->second & FirstMemoryReference) | cmdBufIdSlot_;
@@ -174,6 +174,37 @@ VirtualGPU::Queue::removeCmdMemRef(Pal::IGpuMemory* iMem)
}
}
void
VirtualGPU::Queue::addCmdDoppRef(Pal::IGpuMemory* iMem, bool lastDoppCmd, bool pfpaDoppCmd)
{
for (int i = 0; i < palDoppRefs_.size(); i++) {
if (palDoppRefs_[i].pGpuMemory == iMem) {
// If both LAST_DOPP_SUBMISSION and PFPA_DOPP_SUBMISSION VCOPs are requested,
// the LAST_DOPP_SUBMISSION is send as requsted by KMD
//
if (palDoppRefs_[i].flags.lastPfpaCmd == 1) {
return; // no need to override the last submission command
}
if (lastDoppCmd) {
palDoppRefs_[i].flags.lastPfpaCmd = 1;
palDoppRefs_[i].flags.pfpa = 0;
}
else if (pfpaDoppCmd) {
palDoppRefs_[i].flags.pfpa = 1;
}
return;
}
}
// this is the first reference of the DOPP desktop texture, add it in the vector
Pal::DoppRef doppRef = {};
doppRef.flags.pfpa = pfpaDoppCmd ? 1 : 0;
doppRef.flags.lastPfpaCmd = lastDoppCmd ? 1 : 0;
doppRef.pGpuMemory = iMem;
palDoppRefs_.push_back(doppRef);
}
uint
VirtualGPU::Queue::submit(bool forceFlush)
{
@@ -196,6 +227,7 @@ VirtualGPU::Queue::flush()
LogError("PAL failed to finalize a command buffer!");
return false;
}
// Add memory references
for (auto it = memReferences_.begin(); it != memReferences_.end(); ++it) {
if (it->second & FirstMemoryReference) {
@@ -218,6 +250,8 @@ VirtualGPU::Queue::flush()
submitInfo.cmdBufferCount = 1;
submitInfo.ppCmdBuffers = &iCmdBuffs_[cmdBufIdSlot_];
submitInfo.pFence = iCmdFences_[cmdBufIdSlot_];
submitInfo.doppRefCount = palDoppRefs_.size();
submitInfo.pDoppRefs = palDoppRefs_.data();
// Submit command buffer to OS
if (Pal::Result::Success != iQueue_->Submit(submitInfo)) {
@@ -247,7 +281,7 @@ VirtualGPU::Queue::flush()
// Make sure the slot isn't busy
waifForFence(cmdBufIdSlot_);
// Progress retired TS
// Progress retired TS
if ((cmdBufIdCurrent_ > MaxCmdBuffers) &&
(cmbBufIdRetired_ < (cmdBufIdCurrent_ - MaxCmdBuffers))) {
cmbBufIdRetired_ = cmdBufIdCurrent_ - MaxCmdBuffers;
@@ -272,6 +306,9 @@ VirtualGPU::Queue::flush()
return false;
}
// Clear dopp references
palDoppRefs_.resize(0);
palMems_.resize(0);
// Remove old memory references
for (auto it = memReferences_.begin(); it != memReferences_.end();) {
@@ -299,7 +336,7 @@ VirtualGPU::Queue::waitForEvent(uint id)
uint slotId = id % MaxCmdBuffers;
bool result = waifForFence(slotId);
cmbBufIdRetired_ = id;
cmbBufIdRetired_ = id;
return result;
}
@@ -318,7 +355,7 @@ VirtualGPU::Queue::isDone(uint id)
if (Pal::Result::Success != iCmdFences_[id % MaxCmdBuffers]->GetStatus()) {
return false;
}
cmbBufIdRetired_ = id;
cmbBufIdRetired_ = id;
return true;
}
@@ -384,7 +421,7 @@ VirtualGPU::MemoryDependency::validate(
if (flushL1Cache) {
// Flush cache
gpu.flushCUCaches();
// Clear memory dependency state
const static bool All = true;
clear(!All);
@@ -2019,6 +2056,11 @@ VirtualGPU::submitKernelInternal(
// Add GSL handle to the memory list for VidMM
for (uint i = 0; i < dispMemList.size(); ++i) {
addVmMemory(dispMemList[i]);
if (dispMemList[i]->desc().isDoppTexture_) {
addDoppRef(dispMemList[i],
kernel.parameters().getExecNewVcop(),
kernel.parameters().getExecPfpaVcop());
}
}
// HW Debug for the kernel?
@@ -2312,9 +2354,9 @@ VirtualGPU::getGpuEvent(Pal::IGpuMemory* iMem)
return &gpuEvents_[iMem];
}
void
void
VirtualGPU::assignGpuEvent(Pal::IGpuMemory* iMem, GpuEvent gpuEvent)
{
{
auto it = gpuEvents_.find(iMem);
if (it != gpuEvents_.end()) {
@@ -3089,6 +3131,13 @@ VirtualGPU::addVmMemory(const Memory* memory)
return true;
}
bool
VirtualGPU::addDoppRef(const Memory* memory, bool lastDoppCmd, bool pfpaDoppCmd)
{
queues_[MainEngine]->addCmdDoppRef(memory->iMem(), lastDoppCmd, pfpaDoppCmd);
return true;
}
void
VirtualGPU::profileEvent(EngineType engine, bool type) const
{
projects/clr/rocclr/runtime/device/pal/palvirtual.hpp
+10
查看文件
@@ -76,6 +76,8 @@ public:
void addCmdMemRef(Pal::IGpuMemory* iMem);
void removeCmdMemRef(Pal::IGpuMemory* iMem);
void addCmdDoppRef(Pal::IGpuMemory* iMem, bool lastDoppCmd, bool pfpaDoppCmd);
void addMemRef(Pal::IGpuMemory* iMem) const
{
Pal::GpuMemoryRef memRef = {};
@@ -141,6 +143,7 @@ public:
Util::VirtualLinearAllocator vlAlloc_;
std::vector<Pal::GpuMemoryRef> palMemRefs_;
std::vector<Pal::IGpuMemory*> palMems_;
std::vector<Pal::DoppRef> palDoppRefs_;
};
struct CommandBatch : public amd::HeapObject
@@ -379,6 +382,13 @@ public:
const Memory* memory //!< GPU memory object
);
//! Adds a dopp desktop texture reference
bool addDoppRef(
const Memory* memory, //!< GPU memory object
bool lastDoopCmd, //!< is the last submission for the pre-present primary
bool pfpaDoppCmd //!< is a submission for the pre-present primary
);
//! Adds a stage write buffer into a list
void addXferWrite(Memory& memory);