P4 to Git Change 1095130 by xcui@merged_opencl_jxcwin on 2014/11/07 18:09:55

EPR #408459 - changed the implementation of svmAlloc, so that the first device can create amd::Memory object, and the rest of devices only added gpu memory to it. This is part of changes for mgpu support for svmalloc
	code review:
	http://ocltc.amd.com/reviews/r/6245/
	precheckin testing results:
	http://ocltc.amd.com:8111/viewModification.html?modId=43136&personal=true&buildTypeId=&tab=vcsModificationTests

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/cpu/cpudevice.hpp#88 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#233 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#479 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.hpp#133 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa/hsadevice.cpp#87 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa/hsadevice.hpp#44 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsadevice.cpp#19 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsadevice.hpp#6 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/context.cpp#34 edit


[ROCm/clr commit: efbedb25be]

projects/clr/rocclr/runtime/device/cpu/cpudevice.hpp

@@ -175,7 +175,7 @@ public:
         return workerThreadsAffinity_;
     }
     //! host memory alloc
-    virtual void* svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags) const
+    virtual void* svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags, void* svmPtr) const
     {
         return NULL;
     }

projects/clr/rocclr/runtime/device/device.hpp

@@ -1624,7 +1624,7 @@ public:
     /**
      * @copydoc amd::Context::svmAlloc
      */
-    virtual void* svmAlloc(Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags) const = 0;
+    virtual void* svmAlloc(Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags, void* svmPtr) const = 0;
     /**
      * @copydoc amd::Context::svmFree
      */

projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp

@@ -2467,7 +2467,7 @@ Device::hostFree(void* ptr, size_t size) const
 }
 
 void*
-Device::svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags) const
+Device::svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags, void* svmPtr) const
 {
     alignment = std::max(alignment, static_cast<size_t>(info_.memBaseAddrAlign_));
 
@@ -2476,25 +2476,34 @@ Device::svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_me
     alignment =  (alignment < vmBigK) ? vmBigK : alignment;
 
     size = amd::alignUp(size, alignment);
+    amd::Memory* mem = NULL;
+    if (NULL == svmPtr) {
+        //create a hidden buffer, which will allocated on the device later
+        mem = new (context)amd::Buffer(context, flags, size, reinterpret_cast<void*>(1));
+        if (mem == NULL) {
+            LogError("failed to create a svm mem object!");
+            return NULL;
+        }
+
+        if (!mem->create(NULL, false)) {
+            LogError("failed to create a svm hidden buffer!");
+            mem->release();
+            return NULL;
+        }
+        gpu::Memory* gpuMem = getGpuMemory(mem);
+        //add the information to context so that we can use it later.
+        amd::SvmManager::AddSvmBuffer(mem->getSvmPtr(), mem);
 
-    //create a hidden buffer, which will allocated on the device later
-    amd::Memory* mem = new (context) amd::Buffer(context, flags, size, reinterpret_cast<void*>(1));
-    if (mem == NULL) {
-        LogError("failed to create a svm mem object!");
-        return NULL;
     }
-
-    if (!mem->create(NULL, false)) {
-        LogError("failed to create a svm hidden buffer!");
-        mem->release();
-        return NULL;
+    else {
+        //find the existing amd::mem object
+        mem = amd::SvmManager::FindSvmBuffer(svmPtr);
+        if (NULL == mem) {
+            return NULL;
+        }
+        gpu::Memory* gpuMem = getGpuMemory(mem);
     }
 
-    gpu::Memory* gpuMem = getGpuMemory(mem);
-
-    //add the information to context so that we can use it later.
-    amd::SvmManager::AddSvmBuffer(mem->getSvmPtr(), mem);
-
     return mem->getSvmPtr();
 }
 

projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp

@@ -128,7 +128,7 @@ public:
     //! Get GPU device settings
     const gpu::Settings& settings() const
         { return reinterpret_cast<gpu::Settings&>(*settings_); }
-    virtual void* svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags) const {return NULL;}
+    virtual void* svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags, void* svmPtr) const { return NULL; }
     virtual void svmFree(void* ptr) const {return;}
 
 protected:
@@ -547,7 +547,7 @@ public:
     virtual void* hostAlloc(size_t size, size_t alignment, bool atomics = false) const;
 
     //! SVM allocation
-    virtual void* svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags) const;
+    virtual void* svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags, void* svmPtr) const;
 
     //! Free host SVM memory
     void hostFree(void* ptr, size_t size) const;

projects/clr/rocclr/runtime/device/hsa/hsadevice.cpp

@@ -884,7 +884,7 @@ Device::hostFree(void* ptr, size_t size) const
 }
 
 void*
-Device::svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags) const
+Device::svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags, void* svmPtr) const
 {
     bool atomics = (flags & CL_MEM_SVM_ATOMICS) != 0;
     return hostAlloc(size, alignment, atomics);

projects/clr/rocclr/runtime/device/hsa/hsadevice.hpp

@@ -149,7 +149,8 @@ public:
         amd::Context& context,          //!< The context used to create a buffer
         size_t size,                    //!< size of svm spaces
         size_t alignment,               //!< alignment requirement of svm spaces
-        cl_svm_mem_flags flags          //!< flags of creation svm spaces
+        cl_svm_mem_flags flags,         //!< flags of creation svm spaces
+        void* svmPtr                    //!< existing svm pointer for mGPU case
         ) const {
             ShouldNotReachHere();
             return NULL;
@@ -333,7 +334,7 @@ public:
 
     virtual void hostFree(void* ptr, size_t size = 0) const;
 
-    virtual void* svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags = CL_MEM_READ_WRITE) const;
+    virtual void* svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags = CL_MEM_READ_WRITE, void* svmPtr = NULL) const;
 
     virtual void svmFree(void* ptr) const;
 

projects/clr/rocclr/runtime/platform/context.cpp

@@ -307,11 +307,8 @@ Context::svmAlloc(size_t size, size_t alignment, cl_svm_mem_flags flags)
 
         for (const auto& dev : svmAllocDevice_) {
             if (dev->type() == CL_DEVICE_TYPE_GPU) {
-                tempPtr = dev->svmAlloc(*this, size, alignment, flags);
-                if (dev == svmAllocDevice_.front()) {
-                    svmPtrAlloced = tempPtr;
-                }
-                if ((svmPtrAlloced != tempPtr) || (NULL == tempPtr)) {
+                svmPtrAlloced = dev->svmAlloc(*this, size, alignment, flags, svmPtrAlloced);
+                if (svmPtrAlloced == NULL) {
                     return NULL;
                 }
             }