P4 to Git Change 1193228 by xcui@merged_opencl_jxcwin on 2015/09/22 18:52:47

SWDEV-59579 - resubmit the changelist 1193161. refactory the Coare-grained SVM and  fine grain buffer  SVM code path, so that if the device SVM running on supports fine grain system, then the SVM API operation will be on system memory, no need to go through GPU backend. In addition, added support for PX system with CZ on windows 10, which  supports SVM fine grain system.

	code review:
	http://ocltc.amd.com/reviews/r/8530/
	precheckin:
	http://ocltc.amd.com:8111/viewModification.html?modId=58913&personal=true&buildTypeId=&tab=vcsModificationBuilds&show_all_builds=true

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_svm.cpp#15 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#527 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.hpp#152 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#382 edit


[ROCm/clr commit: 9379424ce3]

projects/clr/opencl/api/opencl/amdocl/cl_svm.cpp

@@ -746,41 +746,46 @@ RUNTIME_ENTRY(cl_int, clEnqueueSVMMap, (
     }
     amd::HostQueue& hostQueue = *queue;
     size_t offset = 0;
+    amd::Memory * svmMem = NULL;
+    if ((queue->device()).isFineGrainedSystem()) {
+        //leave blank on purpose for FGS no op
+    }
+    else {
+        svmMem = amd::SvmManager::FindSvmBuffer(svm_ptr);
+        if (NULL != svmMem) {
+            //make sure the context is the same as the context of creation of svm space
+            if (hostQueue.context() != svmMem->getContext()) {
+                LogWarning("different contexts");
+                return CL_INVALID_CONTEXT;
+            }
 
-    //make sure the context is the same as the context of creation of svm space
-    amd::Memory * svmMem = amd::SvmManager::FindSvmBuffer(svm_ptr);
-    if (NULL != svmMem) {
-        if (hostQueue.context() != svmMem->getContext()) {
-           LogWarning("different contexts");
-            return CL_INVALID_CONTEXT;
-        }
-
-        offset = static_cast<address>(svm_ptr) - static_cast<address>(svmMem->getSvmPtr());
-        if (offset < 0 || offset + size > svmMem->getSize()) {
-            LogWarning("wrong svm address ");
-            return CL_INVALID_VALUE;
-        }
-        amd::Buffer* srcBuffer = svmMem->asBuffer();
-
-        amd::Coord3D    srcSize(size);
-        amd::Coord3D    srcOffset(offset);
-        if (NULL != srcBuffer) {
-            if (!srcBuffer->validateRegion(srcOffset, srcSize)) {
+            offset = static_cast<address>(svm_ptr) - static_cast<address>(svmMem->getSvmPtr());
+            if (offset < 0 || offset + size > svmMem->getSize()) {
+                LogWarning("wrong svm address ");
                 return CL_INVALID_VALUE;
             }
-        }
+            amd::Buffer* srcBuffer = svmMem->asBuffer();
 
-        // Make sure we have memory for the command execution
-        device::Memory* mem = svmMem->getDeviceMemory(queue->device());
-        if (NULL == mem) {
-            LogPrintfError("Can't allocate memory size - 0x%08X bytes!",
-                svmMem->getSize());
-            return CL_OUT_OF_RESOURCES;
-        }
-        // Attempt to allocate the map target now (whether blocking or non-blocking)
-        void* mapPtr = (queue->device()).allocMapTarget(*svmMem, srcOffset, srcSize, map_flags);
-        if (NULL == mapPtr || mapPtr != svm_ptr) {
-            return CL_OUT_OF_RESOURCES;
+            amd::Coord3D    srcSize(size);
+            amd::Coord3D    srcOffset(offset);
+            if (NULL != srcBuffer) {
+                if (!srcBuffer->validateRegion(srcOffset, srcSize)) {
+                    return CL_INVALID_VALUE;
+                }
+            }
+
+            // Make sure we have memory for the command execution
+            device::Memory* mem = svmMem->getDeviceMemory(queue->device());
+            if (NULL == mem) {
+                LogPrintfError("Can't allocate memory size - 0x%08X bytes!",
+                    svmMem->getSize());
+                return CL_OUT_OF_RESOURCES;
+            }
+            // Attempt to allocate the map target now (whether blocking or non-blocking)
+            void* mapPtr = (queue->device()).allocMapTarget(*svmMem, srcOffset, srcSize, map_flags);
+            if (NULL == mapPtr || mapPtr != svm_ptr) {
+                return CL_OUT_OF_RESOURCES;
+            }
         }
     }
 
@@ -881,17 +886,19 @@ RUNTIME_ENTRY(cl_int, clEnqueueSVMUnmap, (
         return CL_INVALID_COMMAND_QUEUE;
     }
     amd::HostQueue& hostQueue = *queue;
-
-    //check if the ptr is in the svm space
-    amd::Memory * svmMem = amd::SvmManager::FindSvmBuffer(svm_ptr);
-    // Make sure we have memory for the command execution
-    if (NULL != svmMem) {
+    amd::Memory * svmMem = NULL;
+    if (!(queue->device()).isFineGrainedSystem()) {
+        //check if the ptr is in the svm space
+        svmMem = amd::SvmManager::FindSvmBuffer(svm_ptr);
         // Make sure we have memory for the command execution
-        device::Memory* mem = svmMem->getDeviceMemory(queue->device());
-        if (NULL == mem) {
-            LogPrintfError("Can't allocate memory size - 0x%08X bytes!",
-                svmMem->getSize());
-            return CL_INVALID_VALUE;
+        if (NULL != svmMem) {
+            // Make sure we have memory for the command execution
+            device::Memory* mem = svmMem->getDeviceMemory(queue->device());
+            if (NULL == mem) {
+                LogPrintfError("Can't allocate memory size - 0x%08X bytes!",
+                    svmMem->getSize());
+                return CL_INVALID_VALUE;
+            }
         }
     }