P4 to Git Change 1313559 by lmoriche@lmoriche_opencl_dev on 2016/09/12 22:58:58

SWDEV-94610 - Don't use the -cl-denorms-are-zero, but instead set the fp32/fp64 denorms with the target features +fp32-denormals and +fp64-denormals. fp64-denormals is always set, fp32-denormals in only set if device >= gfx900 and -cl-denorms-are-zero is not set.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocprogram.cpp#33 edit


[ROCm/clr commit: 7239172265]

projects/clr/rocclr/runtime/device/rocm/rocprogram.cpp

@@ -1438,9 +1438,18 @@ HSAILProgram::codegenOptions(amd::option::Options* options)
 {
     std::string optionsStr;
 
+#if defined(WITH_LIGHTNING_COMPILER)
+    bool fp32Denormals = !options->oVariables->DenormsAreZero
+        && dev().deviceInfo().gfxipVersion_ >= 900;
+
+    optionsStr.append(" -Xclang -target-feature -Xclang ");
+    optionsStr.append(fp32Denormals ? "+" : "-")
+        .append("fp32-denormals,+fp64-denormals");
+#else // !defined(WITH_LIGHTNING_COMPILER)
     if (dev().deviceInfo().gfxipVersion_ < 900) {
         optionsStr.append(" -cl-denorms-are-zero");
     }
+#endif // !defined(WITH_LIGHTNING_COMPILER)
 
     //check if the host is 64 bit or 32 bit
     LP64_ONLY(optionsStr.append(" -m64"));