From dc82afd21c68341a8a782a3a7fbae79e3c2d0e02 Mon Sep 17 00:00:00 2001 From: foreman Date: Fri, 13 Nov 2015 19:47:56 -0500 Subject: [PATCH] P4 to Git Change 1211287 by xcui@merged_opencl_jxcwin on 2015/11/13 19:19:52 SWDEV-77172 - temporary disable the FGS optimization for a failure on stg. code review: http://ocltc.amd.com/reviews/r/8992/ precheckin: http://ocltc.amd.com:8111/viewModification.html?modId=62238&personal=true&buildTypeId=&tab=vcsModificationBuilds&show_all_builds=true Affected files ... ... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#263 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#537 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.hpp#156 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.cpp#310 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#391 edit ... //depot/stg/opencl/drivers/opencl/runtime/platform/context.cpp#37 edit [ROCm/clr commit: 768a8e7e78db54bad1732fa93217c9654fd94789] --- projects/clr/rocclr/runtime/device/device.hpp | 4 ++-- projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp | 4 +--- projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp | 1 - projects/clr/rocclr/runtime/device/gpu/gpukernel.cpp | 2 +- projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp | 2 +- projects/clr/rocclr/runtime/platform/context.cpp | 4 ++-- 6 files changed, 7 insertions(+), 10 deletions(-) diff --git a/projects/clr/rocclr/runtime/device/device.hpp b/projects/clr/rocclr/runtime/device/device.hpp index db4fdd2229..8eb1469a3b 100644 --- a/projects/clr/rocclr/runtime/device/device.hpp +++ b/projects/clr/rocclr/runtime/device/device.hpp @@ -1572,8 +1572,8 @@ public: } //! check svm FGS support capability. - inline bool isFineGrainedSystem() const { - return (info().svmCapabilities_ & CL_DEVICE_SVM_FINE_GRAIN_SYSTEM) != 0 ? true : false; + inline bool isFineGrainedSystem(bool FGSOPT = false) const { + return FGSOPT && (info().svmCapabilities_ & CL_DEVICE_SVM_FINE_GRAIN_SYSTEM) != 0 ? true : false; } //! Return this device's type. diff --git a/projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp b/projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp index 0e76778d7b..1caa33381c 100644 --- a/projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp +++ b/projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp @@ -2254,10 +2254,8 @@ Device::svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_me size = amd::alignUp(size, alignment); amd::Memory* mem = NULL; - freeCPUMem_ = false; if (NULL == svmPtr) { if (isFineGrainedSystem()) { - freeCPUMem_ = true; return amd::Os::alignedMalloc(size, alignment); } @@ -2301,7 +2299,7 @@ Device::svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_me void Device::svmFree(void *ptr) const { - if (freeCPUMem_) { + if (isFineGrainedSystem()) { amd::Os::alignedFree(ptr); } else { diff --git a/projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp b/projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp index 5653973eb4..987b9f0279 100644 --- a/projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp +++ b/projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp @@ -630,7 +630,6 @@ private: SrdManager* srdManager_; //!< SRD manager object static AppProfile appProfile_; //!< application profile - mutable bool freeCPUMem_; //!< flag to mark GPU free SVM CPU mem }; /*@}*/} // namespace gpu diff --git a/projects/clr/rocclr/runtime/device/gpu/gpukernel.cpp b/projects/clr/rocclr/runtime/device/gpu/gpukernel.cpp index 20d4a66ab4..c0e2f9fc2e 100644 --- a/projects/clr/rocclr/runtime/device/gpu/gpukernel.cpp +++ b/projects/clr/rocclr/runtime/device/gpu/gpukernel.cpp @@ -3803,7 +3803,7 @@ HSAILKernel::loadArguments( } // If finegrainsystem is present then the pointer can be malloced by the app and // passed to kernel directly. If so copy the pointer location to aqlArgBuf - else if (dev().isFineGrainedSystem()) { + else if (dev().isFineGrainedSystem(true)) { return NULL; } break; diff --git a/projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp b/projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp index 680eee3404..cbae8f0814 100644 --- a/projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp +++ b/projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp @@ -3195,7 +3195,7 @@ VirtualGPU::processMemObjectsHSA( // so we can avoid checks of the aliased objects memoryDependency().newKernel(); - bool deviceSupportFGS = 0 != dev().isFineGrainedSystem(); + bool deviceSupportFGS = 0 != dev().isFineGrainedSystem(true); bool supportFineGrainedSystem = deviceSupportFGS; FGSStatus status = kernelParams.getSvmSystemPointersSupport(); switch (status) { diff --git a/projects/clr/rocclr/runtime/platform/context.cpp b/projects/clr/rocclr/runtime/platform/context.cpp index b097ee7c1b..d015c29b93 100644 --- a/projects/clr/rocclr/runtime/platform/context.cpp +++ b/projects/clr/rocclr/runtime/platform/context.cpp @@ -46,11 +46,11 @@ Context::Context( std::swap(svmAllocDevice_.front(), svmAllocDevice_.back()); } - uint isFirstDeviceFGSEnabled = svmAllocDevice_.front()->isFineGrainedSystem(); + uint isFirstDeviceFGSEnabled = svmAllocDevice_.front()->isFineGrainedSystem(true); for (auto& dev : svmAllocDevice_) { //allocation on fine - grained system incapable device first if (isFirstDeviceFGSEnabled && (dev->type() == CL_DEVICE_TYPE_GPU) - && (!(dev->isFineGrainedSystem()))) { + && (!(dev->isFineGrainedSystem(true)))) { std::swap(svmAllocDevice_.front(), dev); break; }