P4 to Git Change 1213547 by skudchad@skudchad_test_win_opencl2 on 2015/11/20 18:45:03

SWDEV-77172 - IOMMUv2 changes for Windows 10 (Part 2)
	- Fix SVM FGS malloc arguments being passed to kernel
	- Move setting allocation attributes to runtime
	- Fix some bugs from my previous checkins

	ReviewBoardURL = http://ocltc.amd.com/reviews/r/9035/diff/

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.cpp#311 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.cpp#232 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.hpp#86 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#393 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gslbe/src/rt/GSLDevice.cpp#150 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gslbe/src/rt/backend.h#11 edit

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(true)) {
+                    else if (!dev().isFineGrainedSystem(true)) {
                         return NULL;
                     }
                     break;

rocclr/runtime/device/gpu/gpuresource.cpp

@@ -87,8 +87,9 @@ Resource::Resource(
     cal_.buffer_    = true;
     cal_.imageArray_ = false;
     cal_.imageType_  = 0;
-    cal_.SVMRes_ = false;
+    cal_.skipRsrcCache_ = false;
     cal_.scratch_ = false;
+    cal_.isAllocSVM_ = false;
     cal_.isAllocExecute_ = false;
 }
 
@@ -130,8 +131,9 @@ Resource::Resource(
     cal_.buffer_        = false;
     cal_.imageArray_    = false;
     cal_.imageType_     = imageType;
-    cal_.SVMRes_ = false;
+    cal_.skipRsrcCache_ = false;
     cal_.scratch_ = false;
+    cal_.isAllocSVM_ = false;
     cal_.isAllocExecute_ = false;
 
     switch (imageType) {
@@ -348,21 +350,31 @@ Resource::create(MemoryType memType, CreateParams* params)
     desc.vaBase = 0;
     desc.minAlignment = 0;
     desc.isAllocExecute = false;
+    desc.isAllocSVM = false;
     desc.section = GSL_SECTION_REGULAR;
     if (NULL != params && NULL != params->owner_) {   //make sure params not NULL
         mcaddr svmPtr = reinterpret_cast<mcaddr>(params->owner_->getSvmPtr());
         desc.vaBase = (svmPtr == 1)? 0:svmPtr;
-        cal_.SVMRes_ = (svmPtr != 0);
+        // Dont cache coarse\fine grain svm resource as these may not be released
+        // and allocations may fail since there is limited space for coarse\fine grainbuffers
+        cal_.skipRsrcCache_ = (svmPtr != 0);
         desc.section = (svmPtr != 0) ? GSL_SECTION_SVM : GSL_SECTION_REGULAR;
 
         if (params->owner_->getMemFlags() & CL_MEM_SVM_ATOMICS) {
             desc.section = GSL_SECTION_SVM_ATOMICS;
         }
+
+        if (dev().settings().svmFineGrainSystem_ &&
+            (desc.section == GSL_SECTION_SVM ||
+             desc.section == GSL_SECTION_SVM_ATOMICS)) {
+                cal_.isAllocSVM_ = desc.isAllocSVM = true;
+        }
     }
 
     if (memType == Shader){
         if(dev().settings().svmFineGrainSystem_) {
             cal_.isAllocExecute_ = desc.isAllocExecute = true;
+            cal_.isAllocSVM_ = desc.isAllocSVM = true;
         }
         // force to use remote memory for HW DEBUG or use
         // local memory once we determine if FGS is supported
@@ -596,6 +608,11 @@ Resource::create(MemoryType memType, CreateParams* params)
             return false;
         }
 
+        if(dev().settings().svmFineGrainSystem_) {
+            cal_.isAllocExecute_ = desc.isAllocExecute = true;
+            cal_.isAllocSVM_ = desc.isAllocSVM = true;
+        }
+
         gslResource = dev().resAlloc(&desc);
         if (gslResource != 0) {
             calRes = true;
@@ -1893,7 +1910,7 @@ ResourceCache::addCalResource(
          (desc->type_ == Resource::Remote) ||
          (desc->type_ == Resource::RemoteUSWC)) &&
          (size < cacheSizeLimit_) &&
-         !desc->SVMRes_) {
+         !desc->skipRsrcCache_) {
         // Validate the cache size limit. Loop until we have enough space
         while ((cacheSize_ + size) > cacheSizeLimit_) {
             removeLast();
@@ -1921,7 +1938,7 @@ ResourceCache::findCalResource(Resource::CalResourceDesc* desc)
     size_t  size = getResourceSize(desc);
 
     // Early exit if resource is too big
-    if (size >= cacheSizeLimit_ || desc->SVMRes_) {
+    if (size >= cacheSizeLimit_ || desc->skipRsrcCache_) {
         //! \note we may need to free the cache here to reduce memory pressure
         return ref;
     }
@@ -1939,6 +1956,7 @@ ResourceCache::findCalResource(Resource::CalResourceDesc* desc)
             (entry->format_ == desc->format_) &&
             (entry->flags_ == desc->flags_) &&
             (entry->mipLevels_ == desc->mipLevels_) &&
+            (entry->isAllocSVM_  == desc->isAllocSVM_) &&
             (entry->isAllocExecute_  == desc->isAllocExecute_)) {
                 ref = it.second;
                 delete it.first;

rocclr/runtime/device/gpu/gpuresource.hpp

@@ -173,8 +173,9 @@ public:
                 uint    imageArray_     : 1;    //!< GSL resource is an array of images
                 uint    buffer_         : 1;    //!< GSL resource is a buffer
                 uint    tiled_          : 1;    //!< GSL resource is tiled
-                uint    SVMRes_         : 1;    //!< SVM flag to the cal resource
                 uint    scratch_        : 1;    //!< Scratch buffer
+                uint    skipRsrcCache_  : 1;    //!< Skip caching of a cal resource
+                uint    isAllocSVM_     : 1;    //!< SVM resource attribute
                 uint    isAllocExecute_ : 1;    //!< SVM resource allocation attribute for shader\cmdbuf
             };
             uint    state_;

rocclr/runtime/device/gpu/gpuvirtual.cpp

@@ -3232,6 +3232,7 @@ VirtualGPU::processMemObjectsHSA(
                 // Clear memory dependency state
                 const static bool All = true;
                 memoryDependency().clear(!All);
+                continue;
             }
         }
         else {
@@ -3270,6 +3271,7 @@ VirtualGPU::processMemObjectsHSA(
                     // Clear memory dependency state
                     const static bool All = true;
                     memoryDependency().clear(!All);
+                    continue;
                 }
             }
 

rocclr/runtime/device/gpu/gslbe/src/rt/GSLDevice.cpp

@@ -635,6 +635,7 @@ CALGSLDevice::resAlloc(const CALresourceDesc* desc) const
     attribs.location = desc->type;
     attribs.vaBase   = desc->vaBase;
     attribs.section = desc->section;
+    attribs.isAllocSVM = desc->isAllocSVM;
     attribs.isAllocExecute = desc->isAllocExecute;
     attribs.minAlignment = desc->minAlignment;
 
@@ -744,6 +745,7 @@ CALGSLDevice::resAllocView(gslMemObject res, gslResource3D size, size_t offset,
     );
     attribs.bytePitch = bytePitch;
     attribs.section = res->getAttribs().section;
+    attribs.isAllocSVM = res->getAttribs().isAllocSVM;
     attribs.isAllocExecute = res->getAttribs().isAllocExecute;
 
     // Need to get the alignment info from hwl.

rocclr/runtime/device/gpu/gslbe/src/rt/backend.h

@@ -149,6 +149,7 @@ typedef struct CALresourceDescRec {
     mcaddr          vaBase;
     gslMemObjectAttribSection section;
     CALuint         minAlignment;
+    bool            isAllocSVM;
     bool            isAllocExecute;
 } CALresourceDesc;