P4 to Git Change 1789549 by vsytchen@vsytchen-remote-ocl-win10 on 2019/05/30 10:15:29

SWDEV-188631 - Allocating large buffers produce wrong kernel result on Windows

	1. Set a limit for USWC allocations to 2GB on Windows.
	2. Allocations larger than the specified limit will get placed into pinned memory instead.

	ReviewBoardURL = http://ocltc.amd.com/reviews/r/17407/diff/

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#138 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#310 edit


[ROCm/clr commit: 66cda9d97a]

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

@@ -1511,7 +1511,8 @@ pal::Memory* Device::createBuffer(amd::Memory& owner, bool directAccess) const {
         (settings().remoteAlloc_)) {
       // Allocate remote memory if AHP allocation and context has just 1 device
       if ((owner.getMemFlags() & CL_MEM_ALLOC_HOST_PTR) &&
-          (owner.getContext().devices().size() == 1)) {
+          (owner.getContext().devices().size() == 1) &&
+          (owner.getSize() < static_cast<size_t>(GPU_MAX_USWC_ALLOC_SIZE) * Mi)) {
         if (owner.getMemFlags() &
             (CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_HOST_NO_ACCESS)) {
           // GPU will be reading from this host memory buffer,

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

@@ -183,6 +183,9 @@ release(bool, GPU_ENABLE_WGP_MODE, true,                                      \
         "Enables WGP Mode in HW if available")                                \
 release(bool, OCL_DUMP_CODE_OBJECT, false,                                    \
         "Enable dump code object")                                    \
+release(uint, GPU_MAX_USWC_ALLOC_SIZE, 2048,                                  \
+        "Set a limit in Mb on the maximum USWC allocation size"               \
+        "-1 = No limit")
 
 namespace amd {