P4 to Git Change 1115252 by skudchad@skudchad_test_win_opencl2 on 2015/01/26 12:25:46

EPR #403782 - IOMMU2/SVM
	- Detect SVM Fine Grain system in IOL and OpenCL runtime
	- Changes to convey Allocation attribute to IOL

	ReviewBoardURL = http://ocltc.amd.com/reviews/r/6685/diff/

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.cpp#203 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.hpp#78 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuscsi.cpp#31 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpusettings.cpp#299 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpusettings.hpp#89 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gslbe/src/include/cal/cal.h#31 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gslbe/src/include/cal/calcl.h#31 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gslbe/src/rt/GSLDevice.cpp#104 edit

rocclr/runtime/device/gpu/gpuresource.cpp

@@ -406,6 +406,7 @@ Resource::create(MemoryType memType, CreateParams* params, bool heap)
         // Fall through ...
     case RemoteUSWC:
     case Remote:
+    case Shader:
     case BusAddressable:
     case ExternalPhysical:
         // Fall through to process the memory allocation ...
@@ -437,6 +438,15 @@ Resource::create(MemoryType memType, CreateParams* params, bool heap)
             else if (memoryType() == RemoteUSWC) {
                 desc.type = GSL_MOA_MEMORY_AGP;
             }
+            else if (memoryType() == Shader){
+                if(dev().settings().svmFineGrainSystem_) {
+                    desc.isAllocExecute = true;
+                }
+                // force to use remote memory for HW DEBUG or use
+                // local memory once we determine if FGS is supported
+                memType = (!dev().settings().enableHwDebug_) ? Local : RemoteUSWC;
+                cal_.type_ = memType;
+            }
             else if (memoryType() == BusAddressable){
                 desc.type = GSL_MOA_MEMORY_CARD_BUS_ADDRESSABLE;
             }

rocclr/runtime/device/gpu/gpuresource.hpp

@@ -137,7 +137,8 @@ public:
         BusAddressable,     //!< resource is a bus addressable memory
         ExternalPhysical,   //!< resource is an external physical memory
         D3D9Interop,        //!< resource is a D3D9 memory object
-        Scratch             //!< resource is scratch memory
+        Scratch,            //!< resource is scratch memory
+        Shader,             //!< resource is a shader
     };
 
     //! Resource map flags

rocclr/runtime/device/gpu/gpuscsi.cpp

@@ -175,11 +175,8 @@ HSAILKernel::aqlCreateHWInfo(const void* shader, size_t shaderSize)
     codeSize_ = codeEndAddress - codeStartAddress;
     code_ = new gpu::Memory(dev(), amd::alignUp(codeSize_, gpu::ConstBuffer::VectorSize));
 
-    // force to use remote memory for HW DEBUG
-    Resource::MemoryType resMemType = (!dev().settings().enableHwDebug_) ? Resource::Local : Resource::RemoteUSWC;
-
     // Initialize kernel ISA code
-    if ((code_ != NULL) && code_->create(resMemType)) {
+    if ((code_ != NULL) && code_->create(Resource::Shader)) {
         address cpuCodePtr = static_cast<address>(code_->map(NULL, Resource::WriteOnly));
         // Copy only amd_kernel_code_t
         memcpy(cpuCodePtr, codeStartAddress, codeSize_);

rocclr/runtime/device/gpu/gpusettings.cpp

@@ -424,6 +424,8 @@ Settings::create(
         svmAtomics_ = true;
     }
 
+    svmFineGrainSystem_ = calAttr.isSVMFineGrainSystem;
+
     // Enable some platform extensions
     enableExtension(ClAmdDeviceAttributeQuery);
 

rocclr/runtime/device/gpu/gpusettings.hpp

@@ -69,11 +69,12 @@ public:
             uint    hsail_: 1;          //!< Enables HSAIL compilation
             uint    stagingWritePersistent_: 1; //!< Enables persistent writes
             uint    svmAtomics_: 1;     //!< SVM device atomics
+            uint    svmFineGrainSystem_: 1;     //!< SVM fine grain system support
             uint    apuSystem_: 1;      //!< Device is APU system with shared memory
             uint    asyncMemCopy_: 1;   //!< Use async memory transfers
             uint    hsailDirectSRD_: 1; //!< Controls direct SRD for HSAIL
             uint    useDeviceQueue_: 1; //!< Submit to separate device queue
-            uint    reserved_: 2;
+            uint    reserved_: 1;
         };
         uint    value_;
     };

rocclr/runtime/device/gpu/gslbe/src/include/cal/cal.h

@@ -239,7 +239,8 @@ typedef struct CALdeviceattribsRec {
     CALuint64  vaEnd;                               /**< VA end address */
     bool       isWorkstation;                       /**< Whether Device is a Workstation/Server part */
     CALuint    numOfVpu;                            /**< number of vpu in the device*/
-    bool       isOpenCL200Device;                    /**< the flag to mark if the device is OpenCL 200*/
+    bool       isOpenCL200Device;                   /**< the flag to mark if the device is OpenCL 200 */
+    bool       isSVMFineGrainSystem;                /**< check if SVM finegrainsystem */
 } CALdeviceattribs;
 
 /** CAL device status */

rocclr/runtime/device/gpu/gslbe/src/include/cal/calcl.h

@@ -146,6 +146,7 @@ typedef struct CALresourceDescRec {
     mcaddr          vaBase;
     gslMemObjectAttribSection section;
     CALuint         minAlignment;
+    bool            isAllocExecute;
 } CALresourceDesc;
 
 typedef enum CALresallocsliceviewflagsRec {

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

@@ -156,6 +156,7 @@ CALGSLDevice::getAttribs_int(gsl::gsCtx* cs)
     m_attribs.vaEnd   = static_cast<CALuint64>(m_adp->pAsicInfo->vaEnd);
     m_attribs.numOfVpu = m_adp->pAsicInfo->numberOfVPU;
     m_attribs.isOpenCL200Device = m_adp->pAsicInfo->bIsOpen2Device;
+    m_attribs.isSVMFineGrainSystem = m_adp->pAsicInfo->svmFineGrainSystem;
 }
 
 void
@@ -733,6 +734,7 @@ CALGSLDevice::resAlloc(const CALresourceDesc* desc) const
     attribs.location = desc->type;
     attribs.vaBase   = desc->vaBase;
     attribs.section = desc->section;
+    attribs.isAllocExecute = desc->isAllocExecute;
     attribs.minAlignment = desc->minAlignment;
 
     //!@note GSL asserts with tiled 1D images of any type.
@@ -837,6 +839,7 @@ CALGSLDevice::resAllocView(gslMemObject res, gslResource3D size, CALdomain offse
     );
     attribs.bytePitch = bytePitch;
     attribs.section = res->getAttribs().section;
+    attribs.isAllocExecute = res->getAttribs().isAllocExecute;
 
     // Need to get the alignment info from hwl.
     // Not sure hwl is correct though. Linear aligned 256b, tiled 8kb according to the address library.