P4 to Git Change 1379762 by rili@rili-opencl-pal-stg on 2017/03/01 14:16:43

SWDEV-95925 - Implement SVM fine grain system.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palprogram.cpp#34 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#25 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#19 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#44 edit


[ROCm/clr commit: 5c3609ac64]

projects/clr/rocclr/runtime/device/pal/palprogram.cpp

@@ -49,7 +49,7 @@ Segment::alloc(
 {
     align = amd::alignUp(align, sizeof(uint32_t));
     gpuAccess_ = new pal::Memory(prog.dev(), amd::alignUp(size, align));
-    if ((gpuAccess_ == nullptr) || !gpuAccess_->create(pal::Resource::Local)) {
+    if ((gpuAccess_ == nullptr) || !gpuAccess_->create(pal::Resource::Shader)) {
         delete gpuAccess_;
         gpuAccess_ = nullptr;
         return false;

projects/clr/rocclr/runtime/device/pal/palresource.cpp

@@ -437,6 +437,10 @@ Resource::create(MemoryType memType, CreateParams* params)
     amd::ScopedLock lk(dev().lockPAL());
 
     if (memType == Shader) {
+        if(dev().settings().svmFineGrainSystem_) {
+            desc_.isAllocExecute_ = true;
+            desc_.SVMRes_     = true;
+        }
         // force to use remote memory for HW DEBUG or use
         // local memory once we determine if FGS is supported
         // memType = (!dev().settings().enableHwDebug_) ? Local : RemoteUSWC;
@@ -573,6 +577,7 @@ Resource::create(MemoryType memType, CreateParams* params)
                 imgCreateInfo.samples   = 1;
                 imgCreateInfo.fragments = 1;
                 imgCreateInfo.tiling    = Pal::ImageTiling::Linear;
+                imgCreateInfo.depthPitch = desc().height_ * imgCreateInfo.rowPitch;
 
                 switch (misc) {
                 case 1:     // NV12 format
@@ -635,6 +640,9 @@ Resource::create(MemoryType memType, CreateParams* params)
                     }
                 }
                 result = image_->BindGpuMemory(iMem(), viewOffset);
+                if (result != Pal::Result::Success) {
+                    return false;
+                }
                 offset_ = static_cast<size_t>(viewOffset);
                 hwSrd_ = dev().srds().allocSrdSlot(reinterpret_cast<address*>(&hwState_));
                 if ((0 == hwSrd_) && (memoryType() != ImageView)) {
@@ -1010,6 +1018,10 @@ Resource::create(MemoryType memType, CreateParams* params)
             return false;
         }
 
+        if(dev().settings().svmFineGrainSystem_) {
+            desc_.SVMRes_ = true;
+        }
+
         // Ensure page alignment
         if ((uint64_t)(pinAddress) & (amd::Os::pageSize() - 1)) {
             return false;
@@ -1028,21 +1040,31 @@ Resource::create(MemoryType memType, CreateParams* params)
         return true;
     }
 
+    Pal::gpusize svmPtr = 0;
     if ((nullptr != params) &&
         (nullptr != params->owner_) &&
         (nullptr != params->owner_->getSvmPtr())) {
-        Pal::gpusize svmPtr = reinterpret_cast<Pal::gpusize>(params->owner_->getSvmPtr());
+        svmPtr = reinterpret_cast<Pal::gpusize>(params->owner_->getSvmPtr());
+        desc_.SVMRes_ = true;
         svmPtr = (svmPtr == 1) ? 0 : svmPtr;
+    }
+    if (desc_.SVMRes_) {
         // @todo 64K alignment is too big
         uint allocSize = amd::alignUp(desc().width_ * elementSize_, MaxGpuAlignment);
-        if (memoryType() == Remote) {
+        if ((memoryType() == RemoteUSWC) ||
+            (memoryType() == Remote))  {
             Pal::SvmGpuMemoryCreateInfo createInfo = {};
+            createInfo.isUsedForKernel = desc_.isAllocExecute_;
             createInfo.size = allocSize;
             createInfo.alignment = MaxGpuAlignment;
             if (svmPtr != 0) {
                 createInfo.flags.useReservedGpuVa = true;
                 createInfo.pReservedGpuVaOwner = params->svmBase_->iMem();
             }
+            else {
+                createInfo.flags.useReservedGpuVa = false;
+                createInfo.pReservedGpuVaOwner = nullptr;
+            }
             memRef_ = GpuMemoryReference::Create(dev(), createInfo);
         }
         else {
@@ -1063,8 +1085,11 @@ Resource::create(MemoryType memType, CreateParams* params)
             return false;
         }
         desc_.cardMemory_ = false;
-        desc_.SVMRes_ = true;
-        params->owner_->setSvmPtr(reinterpret_cast<void*>(memRef_->iMem()->Desc().gpuVirtAddr));
+        if ((nullptr != params) &&
+            (nullptr != params->owner_) &&
+            (nullptr != params->owner_->getSvmPtr())) {
+            params->owner_->setSvmPtr(reinterpret_cast<void*>(memRef_->iMem()->Desc().gpuVirtAddr));
+        }
         return true;
     }
 

projects/clr/rocclr/runtime/device/pal/palsettings.cpp

@@ -82,9 +82,6 @@ Settings::Settings()
     // GPU device by default
     apuSystem_  = false;
 
-    // Fine-Grained System is disabled by default
-    svmFineGrainSystem_  = false;
-
     // Disable 64 bit pointers support by default
     use64BitPtr_ = false;
 
@@ -361,12 +358,11 @@ Settings::create(
         // accessing persistent staged buffer may fail if LongIdleDetct is enabled.
         disablePersistent_ = true;
     }
-
-    svmFineGrainSystem_ = calAttr.isSVMFineGrainSystem;
-
-    svmAtomics_ = (calAttr.svmAtomics || calAttr.isSVMFineGrainSystem) ? true : false;
 */
 
+    svmFineGrainSystem_ = palProp.gpuMemoryProperties.flags.iommuv2Support;
+    svmAtomics_         = svmFineGrainSystem_;
+
     // SVM is not currently supported for DX Interop
 #if defined(_WIN32)
     enableExtension(ClKhrD3d9Sharing);

projects/clr/rocclr/runtime/device/pal/palvirtual.cpp

@@ -793,8 +793,7 @@ VirtualGPU::create(bool profiling, uint  deviceQueueSize, uint rtCUs,
         if (dev().numDMAEngines() != 0) {
             uint sdma;
             // If only 1 DMA engine is available then use that one
-            if ((dev().numDMAEngines() < 2) || (idx & 0x1) ||
-                (dev().settings().apuSystem_ && !dev().settings().svmFineGrainSystem_)) {
+            if ((dev().numDMAEngines() < 2) || ((idx & 0x1) && !dev().settings().svmFineGrainSystem_)) {
                 sdma = 0;
             }
             else {
@@ -2144,7 +2143,7 @@ VirtualGPU::submitKernelInternal(
         dispatchParam.pCpuAqlCode    = hsaKernel.cpuAqlCode();
         dispatchParam.hsaQueueVa     = hsaQueueMem_->vmAddress();
         dispatchParam.wavesPerSh     = hsaKernel.getWavesPerSH(this);
-
+        dispatchParam.useAtc         = dev().settings().svmFineGrainSystem_ ? true : false;
         // Run AQL dispatch in HW
         eventBegin(MainEngine);
         iCmd()->CmdDispatchAql(dispatchParam);