P4 to Git Change 1189314 by gandryey@gera-w8 on 2015/09/09 16:57:25

ECR #304775 - Add  DISABLE_DEFERRED_ALLOC key

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/platform/memory.cpp#121 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#240 edit


[ROCm/clr commit: 1420c01d90]

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

@@ -282,12 +282,14 @@ Memory::create(void* initFrom, bool sysMemAlloc)
         return false;
     }
 
-    bool ok = true;
-
     const std::vector<Device*>& devices = context_().devices();
 
+    // Forces system memory allocation on the device,
+    // instead of device memory
+    forceSysMemAlloc_ = sysMemAlloc;
+
     // Create memory on all available devices
-    for (size_t i = 0; ok && i < devices.size(); i++) {
+    for (size_t i = 0; i < devices.size(); i++) {
         deviceAlloced_[devices[i]] = AllocInit;
 
         // Only GPU devices have device memory objects
@@ -295,13 +297,18 @@ Memory::create(void* initFrom, bool sysMemAlloc)
             deviceMemories_[i].ref_     = devices[i];
             deviceMemories_[i].value_   = NULL;
         }
+
+        if (DISABLE_DEFERRED_ALLOC) {
+            device::Memory* mem = getDeviceMemory(*devices[i]);
+            if (NULL == mem) {
+                LogPrintfError("Can't allocate memory size - 0x%08X bytes!",
+                    getSize());
+                return false;
+            }
+        }
     }
 
-    // Forces system memory allocation on the device,
-    // instead of device memory
-    forceSysMemAlloc_ = sysMemAlloc;
-
-    return ok;
+    return true;
 }
 
 bool

projects/clr/rocclr/runtime/utils/flags.hpp

@@ -166,6 +166,8 @@ release(bool, GPU_IFH_MODE, false,                                            \
         "1 = Enable GPU IFH (infinitely fast hardware) mode. Any other value keeps setting disabled.") \
 release(bool, GPU_MIPMAP, true,                                               \
         "Enables GPU mipmap extension")                                       \
+release(bool, DISABLE_DEFERRED_ALLOC, false,                                  \
+        "Disables deferred memory allocation on device")                      \
 release(int, AMD_GPU_FORCE_SINGLE_FP_DENORM, -1,                              \
         "Force denorm for single precision: -1 - don't force, 0 - disable, 1 - enable")                \
 debug(bool, OCL_FORCE_CPU_SVM, false,                                         \