P4 to Git Change 1063414 by gandryey@gera-dev-w7 on 2014/08/06 11:30:27

ECR #304775 - Add OPENCL_VERSION key to force OCL1.2

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#453 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpusettings.cpp#274 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#209 edit

rocclr/runtime/device/gpu/gpudevice.cpp

@@ -744,7 +744,8 @@ Device::create(CALuint ordinal)
     appProfile_.init();
 
     // Open GSL device
-    if (!open(ordinal, appProfile_.enableHighPerformanceState(), appProfile_.reportAsOCL12Device())) {
+    if (!open(ordinal, appProfile_.enableHighPerformanceState(),
+        (appProfile_.reportAsOCL12Device() || (OPENCL_VERSION < 200)))) {
         return false;
     }
 

rocclr/runtime/device/gpu/gpusettings.cpp

@@ -291,7 +291,7 @@ Settings::create(
                 force32BitOcl20_ = true;
                 oclVersion_ = reportAsOCL12Device ? OpenCL12 : XCONCAT(OpenCL,XCONCAT(OPENCL_MAJOR,OPENCL_MINOR));
             }
-            if (hsail_) {
+            if (hsail_ || (OPENCL_VERSION < 200)) {
                 oclVersion_ = OpenCL12;
             }
             numComputeRings_ = 8;

rocclr/runtime/utils/flags.hpp

@@ -156,6 +156,8 @@ release(bool, AMD_DEPTH_MSAA_INTEROP, false,                                  \
         "Enable depth stencil and MSAA buffer interop")                       \
 release(bool, AMD_THREAD_TRACE_ENABLE, false,                                 \
         "Enable thread trace extension")                                      \
+release(uint, OPENCL_VERSION, 200,                                            \
+        "Force opencl verison")                                               \
 release(bool, ENVVAR_HSA_POLL_KERNEL_COMPLETION, false,                       \
         "Determines if Hsa runtime should use polling scheme")                \
 release(bool, HSA_LOCAL_MEMORY_ENABLE, false,                                 \