P4 to Git Change 1502648 by gandryey@gera-ocl-lc on 2018/01/11 13:59:40

SWDEV-116219 - Allocating memory on GPU is 2x slower on win10 RS than on win7
	- Remove 512MB limit for resource cache to avoid expensive KMD memory allocations in Win10 RS

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpusettings.cpp#358 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#40 edit

rocclr/runtime/device/gpu/gpusettings.cpp

@@ -419,7 +419,6 @@ bool Settings::create(const CALdeviceattribs& calAttr, bool reportAsOCL12Device,
   } else {
     resourceCacheSize_ = std::max((calAttr.localRAM / 8) * Mi, GPU_RESOURCE_CACHE_SIZE * Mi);
   }
-  resourceCacheSize_ = std::min(resourceCacheSize_, 512 * Mi);
 #endif
 
   // Override current device settings

rocclr/runtime/device/pal/palsettings.cpp

@@ -400,17 +400,12 @@ bool Settings::create(const Pal::DeviceProperties& palProp,
     }
   }
 
-
   if (apuSystem_ &&
       ((heaps[Pal::GpuHeapLocal].heapSize + heaps[Pal::GpuHeapInvisible].heapSize) < (150 * Mi))) {
     remoteAlloc_ = true;
   }
 
-// Save resource cache size
-#ifdef ATI_OS_LINUX
-  // Due to EPR#406216, set the default value for Linux for now
-  resourceCacheSize_ = GPU_RESOURCE_CACHE_SIZE * Mi;
-#else
+  // Update resource cache size
   if (remoteAlloc_) {
     resourceCacheSize_ = std::max((heaps[Pal::GpuHeapGartUswc].heapSize / 8),
                                   (uint64_t)GPU_RESOURCE_CACHE_SIZE * Mi);
@@ -419,8 +414,6 @@ bool Settings::create(const Pal::DeviceProperties& palProp,
         std::max(((heaps[Pal::GpuHeapLocal].heapSize + heaps[Pal::GpuHeapInvisible].heapSize) / 8),
                  (uint64_t)GPU_RESOURCE_CACHE_SIZE * Mi);
   }
-  resourceCacheSize_ = std::min(resourceCacheSize_, 512 * Mi);
-#endif
 
 #if defined(WITH_LIGHTNING_COMPILER)
   switch (palProp.gfxLevel) {