P4 to Git Change 1577357 by gandryey@gera-w8 on 2018/07/06 17:44:59

SWDEV-158017 - CL_DEVICE_GLOBAL_FREE_MEMORY_AMD doesn't work correctly on PAL backend
	- Adjust system memory calculation for APU systems

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#594 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.hpp#89 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#95 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.hpp#26 edit

rocclr/runtime/device/gpu/gpudevice.cpp

@@ -1812,10 +1812,12 @@ bool Device::globalFreeMemory(size_t* freeMemory) const {
   if (settings().apuSystem_) {
     if (settings().viPlus_) {
       // for viPlus_, OCL is using remote instead remoteUSWC to avoid extra copy
-      freeMemory[TotalFreeMemory] += memInfo.agpMemAvailableCacheableBytes / Ki;
+      freeMemory[TotalFreeMemory] += (memInfo.agpMemAvailableCacheableBytes -
+        resourceCache().lclCacheSize() + resourceCache().cacheSize()) / Ki;
       freeMemory[LargestFreeBlock] += memInfo.agpCacheableLargestFreeBlockBytes / Ki;
     } else {
-      freeMemory[TotalFreeMemory] += memInfo.agpMemAvailableBytes / Ki;
+      freeMemory[TotalFreeMemory] += (memInfo.agpMemAvailableBytes -
+        resourceCache().lclCacheSize() + resourceCache().cacheSize()) / Ki;
       freeMemory[LargestFreeBlock] += memInfo.agpLargestFreeBlockBytes / Ki;
     }
   }

rocclr/runtime/device/gpu/gpuresource.hpp

@@ -449,6 +449,9 @@ class ResourceCache : public amd::HeapObject {
   //! Destroys cache
   bool free(size_t minCacheEntries = 0);
 
+  //! Returns the size of all memory, stored in the cache
+  size_t cacheSize() const { return cacheSize_; }
+
   //! Returns the size of local memory, stored in the cache
   size_t lclCacheSize() const { return lclCacheSize_; }
 

rocclr/runtime/device/pal/paldevice.cpp

@@ -1695,16 +1695,18 @@ bool Device::globalFreeMemory(size_t* freeMemory) const {
   freeMemory[LargestFreeBlock] = static_cast<size_t>(std::max(local, invisible) / Ki);
 
   if (settings().apuSystem_) {
-    Pal::gpusize uswc = allocedMem[Pal::GpuHeapGartUswc];
-    if (uswc >= heaps_[Pal::GpuHeapGartUswc].heapSize) {
-      uswc = 0;
+    Pal::GpuHeap heap = settings().viPlus_ ? Pal::GpuHeapGartCacheable: Pal::GpuHeapGartUswc;
+    Pal::gpusize sysMem = allocedMem[heap];
+    if (sysMem >= heaps_[heap].heapSize) {
+      sysMem = 0;
     } else {
-      uswc = heaps_[Pal::GpuHeapGartUswc].heapSize - uswc;
+      sysMem = heaps_[heap].heapSize - sysMem +
+        resourceCache().cacheSize() - resourceCache().lclCacheSize();
     }
-    uswc /= Ki;
-    freeMemory[TotalFreeMemory] += static_cast<size_t>(uswc);
-    if (freeMemory[LargestFreeBlock] < uswc) {
-      freeMemory[LargestFreeBlock] = static_cast<size_t>(uswc);
+    sysMem /= Ki;
+    freeMemory[TotalFreeMemory] += static_cast<size_t>(sysMem);
+    if (freeMemory[LargestFreeBlock] < sysMem) {
+      freeMemory[LargestFreeBlock] = static_cast<size_t>(sysMem);
     }
   }
 

rocclr/runtime/device/pal/palresource.hpp

@@ -558,6 +558,9 @@ class ResourceCache : public amd::HeapObject {
   //! Returns true if cache was freed and false if cache is already empty.
   bool free(size_t minCacheEntries = 0);
 
+  //! Returns the size of all memory, stored in the cache
+  size_t cacheSize() const { return cacheSize_; }
+
   //! Returns the size of local memory, stored in the cache
   size_t lclCacheSize() const { return lclCacheSize_; }