P4 to Git Change 1217843 by gandryey@gera-w8 on 2015/12/04 16:33:36

SWDEV-79445 - Enable CL_AMD_FP16 extension and allow to force CL_KHR_FP16

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#264 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpusettings.cpp#337 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#247 edit


[ROCm/clr commit: 05ac0307b7]

projects/clr/rocclr/runtime/device/device.hpp

@@ -113,6 +113,7 @@ enum OclExtensions {
     ClKhrMipMapImageWrites,
     ClKhrIlProgram,
     ClAMDLiquidFlash,
+    ClAmdFp16,
     ClExtTotal
 };
 
@@ -156,6 +157,7 @@ OclExtensionsString[] = {
     "cl_khr_mipmap_image_writes ",
     "",
     (IS_MAINLINE || IS_LINUX) ? "" : "cl_amd_liquid_flash ",
+    "cl_amd_fp16 ",
     NULL
 };
 

projects/clr/rocclr/runtime/device/gpu/gpusettings.cpp

@@ -200,6 +200,12 @@ Settings::create(
         // Keep this false even though we have support
         // singleFpDenorm_ = true;
         viPlus_ = true;
+        if (GPU_FORCE_FP16) {
+            enableExtension(ClKhrFp16);
+        }
+        else {
+            enableExtension(ClAmdFp16);
+        }
         // Fall through to CI ...
     case CAL_TARGET_KALINDI:
     case CAL_TARGET_SPECTRE:

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

@@ -167,7 +167,9 @@ release(bool, GPU_MIPMAP, true,                                               \
 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")                \
+        "Force denorm for single precision: -1 - don't force, 0 - disable, 1 - enable") \
+release(bool, GPU_FORCE_FP16, false,                                          \
+        "Force CL_KHR_FP16 extension")                                        \
 debug(bool, OCL_FORCE_CPU_SVM, false,                                         \
         "force svm support for CPU")                                          \
 debug(uint, OCL_SET_SVM_SIZE, 4096,                                           \