COMPMID-618: Fix mismatches in CLDepthConvert Oclgrind
Out-of-bounds float to integer conversion is implementation defined.
Oclgrind converts to S32 and truncated while GPU converts S32 and
clamps. We force to always SATURATE for float to int conversion.
Change-Id: I82be9e8cdcc49b32adb8c0da064542b63f891666
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/91512
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
diff --git a/src/core/CL/kernels/CLDepthConvertKernel.cpp b/src/core/CL/kernels/CLDepthConvertKernel.cpp
index c43884a..b213207 100644
--- a/src/core/CL/kernels/CLDepthConvertKernel.cpp
+++ b/src/core/CL/kernels/CLDepthConvertKernel.cpp
@@ -90,7 +90,8 @@
if(input_size > output_size)
{
kernel_name += "_down";
- build_opts.insert((policy == ConvertPolicy::WRAP) ? "-DWRAP" : "-DSATURATE");
+ // Down conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined
+ build_opts.insert(((policy == ConvertPolicy::WRAP) && !is_data_type_float(input->info()->data_type())) ? "-DWRAP" : "-DSATURATE");
}
else
{