P4 to Git Change 1333973 by rili@rili-opencl-pal-stg on 2016/10/31 10:41:30

SWDEV-95903 -  Implement SVM on PAL feature.
	1. Implement SVM Coarse Grain
	2. Implement SVM Fine Grain Buffer

	Review:http://ocltc.amd.com/reviews/r/11706/

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#30 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#17 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.hpp#8 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#8 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#36 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#259 edit

rocclr/runtime/device/pal/paldevice.cpp

@@ -1063,8 +1063,12 @@ Device::init()
     info.flags.disableGpuTimeout = true;
 #ifdef ATI_BITS_32
     info.flags.force32BitVaSpace = true;
+    info.flags.enableSvmMode     = false;
+#else
+    info.flags.enableSvmMode     = true;
 #endif
     info.pSettingsPath = "OCL";
+    info.maxSvmSize    = static_cast<Pal::gpusize>(OCL_SET_SVM_SIZE * Mi);
 
     // PAL init
     if (Pal::Result::Success !=

rocclr/runtime/device/pal/palresource.cpp

@@ -80,6 +80,32 @@ GpuMemoryReference::Create(
     return memRef;
 }
 
+GpuMemoryReference*
+GpuMemoryReference::Create(
+    const Device&   dev,
+    const Pal::SvmGpuMemoryCreateInfo& createInfo)
+{
+    Pal::Result result;
+    size_t gpuMemSize = dev.iDev()->GetSvmGpuMemorySize(createInfo, &result);
+    if (result != Pal::Result::Success) {
+        return nullptr;
+    }
+
+    GpuMemoryReference*  memRef = new (gpuMemSize) GpuMemoryReference();
+    if (memRef != nullptr) {
+        result = dev.iDev()->CreateSvmGpuMemory(createInfo,
+            &memRef[1], &memRef->gpuMem_);
+        if (result != Pal::Result::Success) {
+            memRef->release();
+            return nullptr;
+        }
+    }
+    // Update free memory size counters
+    const_cast<Device&>(dev).updateFreeMemory(
+        Pal::GpuHeap::GpuHeapGartCacheable, createInfo.size, false);
+    return memRef;
+}
+
 GpuMemoryReference*
 GpuMemoryReference::Create(
     const Device&   dev,
@@ -997,6 +1023,36 @@ Resource::create(MemoryType memType, CreateParams* params)
         return true;
     }
 
+    if ((nullptr != params) &&
+        (nullptr != params->owner_) &&
+        (nullptr != params->owner_->getSvmPtr())) {
+        // @todo 64K alignment is too big
+        uint allocSize = amd::alignUp(desc().width_ * elementSize_, MaxGpuAlignment);
+        if (memoryType() == Remote) {
+            Pal::SvmGpuMemoryCreateInfo createInfo = {};
+            createInfo.size = allocSize;
+            createInfo.alignment = MaxGpuAlignment;
+            memRef_ = GpuMemoryReference::Create(dev(), createInfo);
+        }
+        else {
+            Pal::GpuMemoryCreateInfo createInfo = {};
+            createInfo.size = allocSize;
+            createInfo.alignment = MaxGpuAlignment;
+            createInfo.vaRange = Pal::VaRange::Svm;
+            createInfo.priority = Pal::GpuMemPriority::Normal;
+            memTypeToHeap(&createInfo);
+            memRef_ = GpuMemoryReference::Create(dev(), createInfo);
+        }
+        if (nullptr == memRef_) {
+            LogError("Failed PAL memory allocation!");
+            return false;
+        }
+        desc_.cardMemory_ = false;
+        desc_.SVMRes_ = true;
+        params->owner_->setSvmPtr(reinterpret_cast<void*>(memRef_->iMem()->Desc().gpuVirtAddr));
+        return true;
+    }
+
     Pal::GpuMemoryCreateInfo createInfo = {};
     createInfo.size = desc().width_ * elementSize_;
     // @todo 64K alignment is too big

rocclr/runtime/device/pal/palresource.hpp

@@ -28,6 +28,10 @@ public:
         const Device&   dev,
         const Pal::PinnedGpuMemoryCreateInfo& createInfo);
 
+    static GpuMemoryReference* Create(
+        const Device&   dev,
+        const Pal::SvmGpuMemoryCreateInfo& createInfo);
+
     static GpuMemoryReference* Create(
         const Device&   dev,
         const Pal::ExternalResourceOpenInfo& openInfo);

rocclr/runtime/device/pal/palsettings.cpp

@@ -74,6 +74,9 @@ 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;
 

rocclr/runtime/device/pal/palvirtual.cpp

@@ -3192,8 +3192,7 @@ VirtualGPU::processMemObjectsHSA(
                 svmMem = amd::SvmManager::FindSvmBuffer(
                     *reinterpret_cast<void* const*>(params + desc.offset_));
                 if (!svmMem) {
-                    Unimplemented();
-                    //flushCUCaches();
+                    flushCUCaches();
                     // Clear memory dependency state
                     const static bool All = true;
                     memoryDependency().clear(!All);

rocclr/runtime/utils/flags.hpp

@@ -171,7 +171,7 @@ release(int, AMD_GPU_FORCE_SINGLE_FP_DENORM, -1,                              \
         "Force denorm for single precision: -1 - don't force, 0 - disable, 1 - enable") \
 debug(bool, OCL_FORCE_CPU_SVM, false,                                         \
         "force svm support for CPU")                                          \
-release(uint, OCL_SET_SVM_SIZE, 4096,                                           \
+release(uint, OCL_SET_SVM_SIZE, 16384,                                        \
         "set SVM space size for discrete GPU")                                \
 debug(uint, OCL_SYSMEM_REQUIREMENT, 2,                                        \
         "Use flag to change the minimum requirement of system memory not to downgrade")        \