P4 to Git Change 1359979 by skudchad@skudchad_test_win_opencl2 on 2017/01/09 18:53:16

SWDEV-107271 - [OpenCL][GFXIP9 Bring up] add support for Raven(gfx901)
	- Fix USWC memory calculations
	- Try USWC allocations if local are unsuccessful

	ReviewBoardURL = http://ocltc.amd.com/reviews/r/12116/diff/

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#40 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palmemory.cpp#9 edit

rocclr/runtime/device/pal/paldevice.cpp

@@ -276,9 +276,11 @@ void NullDevice::fillDeviceInfo(
         (static_cast<cl_ulong>(std::min(GPU_MAX_HEAP_SIZE, 100u)) *
         static_cast<cl_ulong>(localRAM) / 100u);
 
+    uint uswcPercentAvailable = ((static_cast<cl_ulong>(heaps[Pal::GpuHeapGartUswc].heapSize) / Mi) > 1536 && IS_WINDOWS)
+                                ? 75 : 50;
     if (settings().apuSystem_) {
-        info_.globalMemSize_   +=
-            (static_cast<cl_ulong>(heaps[Pal::GpuHeapGartUswc].heapSize) * Mi * 75)/100;
+        info_.globalMemSize_ +=
+            (static_cast<cl_ulong>(heaps[Pal::GpuHeapGartUswc].heapSize) * uswcPercentAvailable) / 100;
     }
 
     // Find the largest heap form FB memory
@@ -289,7 +291,7 @@ void NullDevice::fillDeviceInfo(
 #if defined(ATI_OS_WIN)
     if (settings().apuSystem_) {
         info_.maxMemAllocSize_ = std::max(
-            (static_cast<cl_ulong>(heaps[Pal::GpuHeapGartUswc].heapSize) * Mi * 75)/100,
+            (static_cast<cl_ulong>(heaps[Pal::GpuHeapGartUswc].heapSize) * uswcPercentAvailable)/100,
             info_.maxMemAllocSize_);
     }
 #endif

rocclr/runtime/device/pal/palmemory.cpp

@@ -124,12 +124,45 @@ Memory::create(
     Resource::CreateParams* params)
 {
     bool    result;
-
+    uint allocAttempt = 0;
     // Reset the flag in case we reallocate the heap in local/remote
     flags_ &= ~HostMemoryDirectAccess;
 
-    // Create a resource in CAL
-    result = Resource::create(memType, params);
+    do {
+        // Create a resource in CAL
+        result = Resource::create(memType, params);
+        if (!result) {
+            size_t  freeMemory[2];
+            // if requested memory is greater than available then exit the loop
+            dev().globalFreeMemory(freeMemory);
+
+            // Local to Persistent
+            if (memoryType() == Local) {
+                // For dgpu freeMemory[0] reports a sum of visible+invisible fb
+                if (owner()->getSize() > (freeMemory[0] * Ki)) {
+                    break;
+                }
+                memType = Persistent;
+            }
+            // Don't switch to USWC if persistent memory was explicitly asked
+            else if ((allocAttempt > 0) && (memoryType() == Persistent)) {
+                memType = RemoteUSWC;
+            }
+            // Remote cacheable to uncacheable
+            else if (memoryType() == Remote) {
+                memType = RemoteUSWC;
+            }
+            else if (dev().settings().apuSystem_ && memoryType() == RemoteUSWC) {
+                if (owner()->getSize() > (freeMemory[0] * Ki)) {
+                    break;
+                }
+            }
+            else {
+                break;
+            }
+            allocAttempt++;
+        }
+    } while (!result);
 
     // Check if CAL created a resource
     if (result) {