P4 to Git Change 1458879 by gandryey@gera-w8 on 2017/09/14 11:41:44

SWDEV-129129 - [[CQE OCL][Vega vs Fiji] Upto 12% Performance drop observed on VEGA10 compared to FIJI while running BlackMagic Davinci Resolve
	More benchmark tuning:
	- Keep system memory locked in the resource cache. That removes huge amount of lock/unlock calls to OS due to the resource creation and destruciton
	- Reduce the command buffer size to 256 commands and incrrease the amount of CBs to 16
	- Increase the amount of resident resources to 2048

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#574 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palmemory.cpp#17 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#37 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#58 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.hpp#31 edit


[ROCm/clr commit: 4066449a8b]

projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp

@@ -1337,6 +1337,12 @@ gpu::Memory* Device::createBuffer(amd::Memory& owner, bool directAccess) const {
       (owner.forceSysMemAlloc() || (owner.getMemFlags() & CL_MEM_SVM_FINE_GRAIN_BUFFER))
       ? Resource::Remote
       : Resource::Local;
+  
+  // Check if runtime can force a tiny buffer into USWC memory
+  if ((size <= (GPU_MAX_REMOTE_MEM_SIZE * Ki)) && (type == Resource::Local) &&
+    (owner.getMemFlags() & CL_MEM_READ_ONLY)) {
+    type = Resource::RemoteUSWC;
+  }
 
   if (owner.getMemFlags() & CL_MEM_BUS_ADDRESSABLE_AMD) {
     type = Resource::BusAddressable;

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

@@ -416,7 +416,8 @@ Memory::~Memory() {
   if ((owner() != nullptr) && isHostMemDirectAccess() && !(flags_ & SubMemoryObject) &&
       (memoryType() != Resource::ExternalPhysical)) {
     // Unmap memory if direct access was requested
-    unmap(nullptr);
+    // Note: runtime will perform unmap on the actual resource destruction
+    //unmap(nullptr);
   }
 }
 

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

@@ -744,6 +744,12 @@ bool Resource::create(MemoryType memType, CreateParams* params) {
           }
         }
       }
+      // Check if memory is locked already and restore CPU pointer
+      if (memRef_->cpuAddress_ != nullptr) {
+        address_ = memRef_->cpuAddress_;
+        memRef_->cpuAddress_ = nullptr;
+        mapCount_++;
+      }
       Pal::BufferViewInfo viewInfo = {};
       viewInfo.gpuAddr = memRef_->iMem()->Desc().gpuVirtAddr + offset();
       viewInfo.range = memRef_->iMem()->Desc().size;
@@ -890,6 +896,12 @@ bool Resource::create(MemoryType memType, CreateParams* params) {
         LogWarning("Image is bigger than the original mem object!");
       }
     }
+    // Check if memory is locked already and restore CPU pointer
+    if (memRef_->cpuAddress_ != nullptr) {
+      address_ = memRef_->cpuAddress_;
+      memRef_->cpuAddress_ = nullptr;
+      mapCount_++;
+    }
 
     result = image_->BindGpuMemory(memRef_->gpuMem_, offset_);
     if (result != Pal::Result::Success) {
@@ -1070,6 +1082,12 @@ bool Resource::create(MemoryType memType, CreateParams* params) {
       return false;
     }
   }
+  // Check if memory is locked already and restore CPU pointer
+  if (memRef_->cpuAddress_ != nullptr) {
+    address_ = memRef_->cpuAddress_;
+    memRef_->cpuAddress_ = nullptr;
+    mapCount_++;
+  }
   return true;
 }
 
@@ -1079,7 +1097,7 @@ void Resource::free() {
   }
 
   // Sanity check for the map calls
-  if (mapCount_ != 0) {
+  if ((mapCount_ != 0) && (memoryType() != Remote) && (memoryType() != RemoteUSWC)) {
     LogWarning("Resource wasn't unlocked, but destroyed!");
   }
   const bool wait =
@@ -1105,10 +1123,16 @@ void Resource::free() {
   if (renames_.size() == 0) {
     // Destroy GSL resource
     if (iMem() != 0) {
-      //! @note: This is a workaround for bad applications that
-      //! don't unmap memory
       if (mapCount_ != 0) {
-        unmap(nullptr);
+        if ((memoryType() != Remote) && (memoryType() != RemoteUSWC)) {
+          //! @note: This is a workaround for bad applications that
+          //! don't unmap memory
+          unmap(nullptr);
+        } else {
+          // Delay CPU address unmap until memRef_ destruction
+          assert(memRef_->cpuAddress_ == nullptr && "Memref shouldn't have a valid CPU address");
+          memRef_->cpuAddress_ = address_;
+        }
       }
 
       // Add resource to the cache if it's not assigned to a specific queue

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

@@ -298,7 +298,7 @@ bool VirtualGPU::Queue::flush() {
   palSdiRefs_.resize(0);
 
   // Remove old memory references
-  if ((memReferences_.size() > 1024) || (residency_size_ > residency_limit_)) {
+  if ((memReferences_.size() > 2048) || (residency_size_ > residency_limit_)) {
     for (auto it = memReferences_.begin(); it != memReferences_.end();) {
       if (it->second == cmdBufIdSlot_) {
         palMems_.push_back(it->first->iMem());

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

@@ -41,8 +41,8 @@ class VirtualGPU : public device::VirtualDevice {
  public:
   class Queue : public amd::HeapObject {
    public:
-    static const uint MaxCmdBuffers = 8;
-    static const uint MaxCommands = 512;
+    static const uint MaxCmdBuffers = 16;
+    static const uint MaxCommands = 256;
     static const uint StartCmdBufIdx = 1;
     static const uint FirstMemoryReference = 0x80000000;
     static const uint64_t WaitTimeoutInNsec = 6000000000;