From 0e8b2ec0ba4a6e46446179a9cc4cc541a435265d Mon Sep 17 00:00:00 2001 From: foreman Date: Thu, 21 Aug 2014 18:12:12 -0400 Subject: [PATCH] P4 to Git Change 1069045 by gandryey@gera-dev-w7 on 2014/08/21 17:55:18 EPR #404714 - [CQE OCL][2.0][DTB]Opencl1.2 WF Conf. Math test failedon Pitcairn and Oland due to CL#1065597 - Add a new MapCacheLock monitor to separate the map cache from the global lock Affected files ... ... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#456 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.hpp#131 edit [ROCm/clr commit: 18b88ee0950cd0104d3567f0fdcb6c4a96917c8d] --- .../clr/rocclr/runtime/device/gpu/gpudevice.cpp | 15 +++++++++++---- .../clr/rocclr/runtime/device/gpu/gpudevice.hpp | 1 + 2 files changed, 12 insertions(+), 4 deletions(-) diff --git a/projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp b/projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp index 84858beb3a..6eb98503df 100644 --- a/projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp +++ b/projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp @@ -366,6 +366,7 @@ Device::Device() , lockAsyncOpsForInitHeap_(NULL) , vgpusAccess_(NULL) , scratchAlloc_(NULL) + , mapCacheOps_(NULL) , xferRead_(NULL) , xferWrite_(NULL) , vaCacheAccess_(NULL) @@ -427,6 +428,7 @@ Device::~Device() delete lockAsyncOpsForInitHeap_; delete vgpusAccess_; delete scratchAlloc_; + delete mapCacheOps_; delete vaCacheAccess_; delete vaCacheList_; @@ -801,6 +803,11 @@ Device::create(CALuint ordinal) return false; } + mapCacheOps_ = new amd::Monitor("Map Cache Lock", true); + if (NULL == scratchAlloc_) { + return false; + } + vaCacheAccess_ = new amd::Monitor("VA Cache Ops Lock", true); if (NULL == vaCacheAccess_) { return false; @@ -2190,8 +2197,8 @@ Device::findMemoryFromVA(const void* ptr, size_t* offset) const amd::Memory* Device::findMapTarget(size_t size) const { - // Must be serialised. Global async is too conservative - amd::ScopedLock lk(*lockAsyncOps_); + // Must be serialised for access + amd::ScopedLock lk(*mapCacheOps_); amd::Memory* map = NULL; size_t minSize = 0; @@ -2250,8 +2257,8 @@ Device::findMapTarget(size_t size) const bool Device::addMapTarget(amd::Memory* memory) const { - // Must be serialised. Global async is too conservative - amd::ScopedLock lk(*lockAsyncOps_); + // Must be serialised for access + amd::ScopedLock lk(*mapCacheOps_); //the svm memory shouldn't be cached if (!memory->canBeCached()) { diff --git a/projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp b/projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp index c5657e4cb2..a3ca8f92f3 100644 --- a/projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp +++ b/projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp @@ -610,6 +610,7 @@ private: amd::Monitor* lockAsyncOpsForInitHeap_; //!< Lock to serialise all async ops on initialization heap operation amd::Monitor* vgpusAccess_; //!< Lock to serialise virtual gpu list access amd::Monitor* scratchAlloc_; //!< Lock to serialise scratch allocation + amd::Monitor* mapCacheOps_; //!< Lock to serialise cache for the map resources XferBuffers* xferRead_; //!< Transfer buffers read XferBuffers* xferWrite_; //!< Transfer buffers write