P4 to Git Change 1597023 by vsytchen@vsytchen-win10 on 2018/08/22 15:12:23

SWDEV-159881 - [OCL][ROCm] Add SVM coarse-grain buffer support with device memory

	1. Use the system memory pool for coarse grain allocations on APUs

	ReviewBoardURL = http://ocltc.amd.com/reviews/r/15671/diff/

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocmemory.cpp#39 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#64 edit


[ROCm/clr commit: 1bf70582f0]

projects/clr/rocclr/runtime/device/rocm/rocmemory.cpp

@@ -589,7 +589,8 @@ void Buffer::destroy() {
   cl_mem_flags memFlags = owner()->getMemFlags();
 
   if (owner()->getSvmPtr() != nullptr) {
-    if (dev().forceFineGrain(owner())) {
+    if (dev().forceFineGrain(owner()) ||
+        dev().isFineGrainedSystem(true)) {
       memFlags |= CL_MEM_SVM_FINE_GRAIN_BUFFER;
     }
     const bool isFineGrain = memFlags & CL_MEM_SVM_FINE_GRAIN_BUFFER;
@@ -657,7 +658,8 @@ bool Buffer::create() {
   cl_mem_flags memFlags = owner()->getMemFlags();
 
   if (owner()->getSvmPtr() != nullptr) {
-    if (dev().forceFineGrain(owner())) {
+    if (dev().forceFineGrain(owner()) ||
+        dev().isFineGrainedSystem(true)) {
       memFlags |= CL_MEM_SVM_FINE_GRAIN_BUFFER;
       flags_ |= HostMemoryDirectAccess;
     }

projects/clr/rocclr/runtime/device/rocm/rocvirtual.cpp

@@ -1215,7 +1215,7 @@ void VirtualGPU::submitSvmCopyMemory(amd::SvmCopyMemoryCommand& cmd) {
 
   profilingBegin(cmd);
   // no op for FGS supported device
-  if (!dev().isFineGrainedSystem()) {
+  if (!dev().isFineGrainedSystem(true)) {
     amd::Coord3D srcOrigin(0, 0, 0);
     amd::Coord3D dstOrigin(0, 0, 0);
     amd::Coord3D size(cmd.srcSize(), 1, 1);
@@ -1374,7 +1374,7 @@ void VirtualGPU::submitSvmMapMemory(amd::SvmMapMemoryCommand& cmd) {
   profilingBegin(cmd);
 
   // no op for FGS supported device
-  if (!dev().isFineGrainedSystem() &&
+  if (!dev().isFineGrainedSystem(true) &&
       !dev().forceFineGrain(cmd.getSvmMem())) {
     // Make sure we have memory for the command execution
     Memory* memory = dev().getRocMemory(cmd.getSvmMem());
@@ -1410,7 +1410,7 @@ void VirtualGPU::submitSvmUnmapMemory(amd::SvmUnmapMemoryCommand& cmd) {
   profilingBegin(cmd);
 
   // no op for FGS supported device
-  if (!dev().isFineGrainedSystem() &&
+  if (!dev().isFineGrainedSystem(true) &&
       !dev().forceFineGrain(cmd.getSvmMem())) {
     Memory* memory = dev().getRocMemory(cmd.getSvmMem());
     const device::Memory::WriteMapInfo* writeMapInfo = memory->writeMapInfo(cmd.svmPtr());
@@ -1700,7 +1700,7 @@ void VirtualGPU::submitSvmFillMemory(amd::SvmFillMemoryCommand& cmd) {
 
   amd::Memory* dstMemory = amd::MemObjMap::FindMemObj(cmd.dst());
 
-  if (!dev().isFineGrainedSystem() ||
+  if (!dev().isFineGrainedSystem(true) ||
       ((dstMemory != nullptr) &&
        !dev().forceFineGrain(dstMemory))) {
     size_t patternSize = cmd.patternSize();