P4 to Git Change 1397729 by kzhuravl@kzhuravl-fiji-ocllc on 2017/04/12 20:05:08

SWDEV-116535 - Report CL_FP_DENORM for single fp config for gfx9 for LC for rocm/pal and force denorms on based on AMD_GPU_FORCE_SINGLE_FP_DENORM for rocm.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#23 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#47 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocprogram.cpp#62 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocsettings.cpp#16 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocsettings.hpp#7 edit


[ROCm/clr commit: c7e04f3222]

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

@@ -68,10 +68,13 @@ Settings::Settings()
     const static size_t MaxPinnedXferSize = 32;
     pinnedXferSize_     = std::min(GPU_PINNED_XFER_SIZE, MaxPinnedXferSize) * Mi;
     pinnedMinXferSize_  = std::min(GPU_PINNED_MIN_XFER_SIZE * Ki, pinnedXferSize_);
+
+    // Don't support Denormals for single precision by default
+    singleFpDenorm_ = false;
 }
 
 bool
-Settings::create(bool fullProfile)
+Settings::create(bool fullProfile, int gfxipVersion)
 {
     customHostAllocator_ = false;
 
@@ -111,6 +114,14 @@ Settings::create(bool fullProfile)
     enableExtension(ClKhrDepthImages);
     supportDepthsRGB_ = true;
 
+#if defined(WITH_LIGHTNING_COMPILER)
+    switch (gfxipVersion) {
+    case 900:
+        singleFpDenorm_ = true;
+        break;
+    }
+#endif // WITH_LIGHTNING_COMPILER
+
     // Override current device settings
     override();
 
@@ -140,6 +151,19 @@ Settings::override()
     if (!flagIsDefault(GPU_PINNED_MIN_XFER_SIZE)) {
         pinnedMinXferSize_  = std::min(GPU_PINNED_MIN_XFER_SIZE * Ki, pinnedXferSize_);
     }
+
+    if (!flagIsDefault(AMD_GPU_FORCE_SINGLE_FP_DENORM)) {
+        switch (AMD_GPU_FORCE_SINGLE_FP_DENORM) {
+        case 0:
+            singleFpDenorm_ = false;
+            break;
+        case 1:
+            singleFpDenorm_ = true;
+            break;
+        default:
+            break;
+        }
+    }
 }
 
 } // namespace roc