P4 to Git Change 1999659 by vsytchen@vsytchen-remote-ocl-win10 on 2019/09/17 14:52:23

SWDEV-193973 - Add MALL support for the PAL backend

	ReviewBoardURL = http://ocltc.amd.com/reviews/r/17995/diff/

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#78 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#99 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.hpp#26 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#318 edit


[ROCm/clr commit: 46f37631fb]

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

@@ -419,6 +419,19 @@ void Resource::memTypeToHeap(Pal::GpuMemoryCreateInfo* createInfo) {
       createInfo->heaps[0] = Pal::GpuHeapLocal;
       break;
   }
+
+#if !IS_MAINLINE
+  // Pick the appropriate mall policy based on the mem type
+  switch (memoryType()) {
+    case Local:
+    case Scratch:
+      createInfo->mallPolicy = static_cast<Pal::GpuMemMallPolicy>(dev().settings().mallPolicy_);
+      break;
+    default:
+      createInfo->mallPolicy = Pal::GpuMemMallPolicy::Never;
+      break;
+  }
+#endif
 }
 
 // ================================================================================================
@@ -1873,7 +1886,7 @@ bool MemorySubAllocator::CreateChunk(const Pal::IGpuMemory* reserved_va) {
   createInfo.priority = Pal::GpuMemPriority::Normal;
   createInfo.heapCount = 1;
   createInfo.heaps[0] = Pal::GpuHeapInvisible;
-  createInfo.flags.peerWritable = device_->P2PAccessAllowed();
+  createInfo.flags.peerWritable = device_->P2PAccessAllowed();.mallPolicy_);)
   GpuMemoryReference* mem_ref = GpuMemoryReference::Create(*device_, createInfo);
   if (mem_ref != nullptr) {
     return InitAllocator(mem_ref);
@@ -1892,7 +1905,7 @@ bool CoarseMemorySubAllocator::CreateChunk(const Pal::IGpuMemory* reserved_va) {
   createInfo.pReservedGpuVaOwner = reserved_va;
   createInfo.heapCount = 2;
   createInfo.heaps[0] = Pal::GpuHeapInvisible;
-  createInfo.heaps[1] = Pal::GpuHeapLocal;
+  createInfo.heaps[1] = Pal::GpuHeapLocal;.mallPolicy_);)
   GpuMemoryReference* mem_ref = GpuMemoryReference::Create(*device_, createInfo);
   if (mem_ref != nullptr) {
     return InitAllocator(mem_ref);

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

@@ -151,6 +151,7 @@ Settings::Settings() {
   enableHwP2P_ = false;
   imageBufferWar_ = false;
   disableSdma_ = PAL_DISABLE_SDMA;
+  mallPolicy_ = 0;
 }
 
 bool Settings::create(const Pal::DeviceProperties& palProp,
@@ -551,6 +552,10 @@ void Settings::override() {
     enableCoopGroups_ = GPU_ENABLE_COOP_GROUPS;
     enableCoopMultiDeviceGroups_ = GPU_ENABLE_COOP_GROUPS;
   }
+
+  if (!flagIsDefault(PAL_MALL_POLICY)) {
+    mallPolicy_ = PAL_MALL_POLICY;
+  }
 }
 
 }  // namespace pal

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

@@ -94,6 +94,7 @@ class Settings : public device::Settings {
   uint64_t maxAllocSize_;        //!< Maximum single allocation size
   uint rgpSqttDispCount_;        //!< The number of dispatches captured in SQTT
   uint maxCmdBuffers_;           //!< Maximum number of command buffers allocated per queue
+  uint mallPolicy_;              //!< 0 - default, 1 - always bypass, 2 - always put
 
   uint64_t subAllocationMinSize_;    //!< Minimum size allowed for suballocations
   uint64_t subAllocationMaxSize_;    //!< Maximum size allowed with suballocations

projects/clr/rocclr/runtime/utils/flags.hpp

@@ -167,6 +167,11 @@ release_on_stg(bool, PAL_DISABLE_SDMA, false,                                 \
         "1 = Disable SDMA for PAL")                                           \
 release(uint, PAL_RGP_DISP_COUNT, 50,                                         \
         "The number of dispatches for RGP capture with SQTT")                 \
+release(uint, PAL_MALL_POLICY, 0,                                             \
+        "Controls the behaviour of allocations with respect to the MALL"      \
+        "0 = MALL policy is decided by KMD"                                   \
+        "1 = Allocations are never put through the MALL"                      \
+        "2 = Allocations will always be put through the MALL")                \
 release(bool, GPU_ENABLE_WAVE32_MODE, true,                                   \
         "Enables Wave32 compilation in HW if available")                      \
 release(bool, GPU_ENABLE_LC, true,                                            \