[SYCL][HIP] Required aspect fp16 is not supported on the device

[zjin-lcf]

Running a SYCL program that contains "half" data types on a MI100 GPU shows the error: Required aspect fp16 is not supported on the device.

Please see a reproducer. Can you reproduce the error using a recent build of the SYCL compiler with HIP support ? Thanks.

https://github.com/zjin-lcf/HeCBench/tree/master/f16sp-sycl

[npmiller]

Yes, that's expected, half support for HIP is currently disabled, we need to look into it, see related comment:

• [SYCL][LIBCLC] Add sqrt for doubles for amdgcn-amdhsa #4223 (comment)

[zjin-lcf]

I am confused that the tahiti architecture includes GPUs of the vega20 architecture when browsing the CMakefileList.txt in the libclc.

# Setup aliases
set( cedar_aliases palm sumo sumo2 redwood juniper )
set( cypress_aliases hemlock )
set( barts_aliases turks caicos )
set( cayman_aliases aruba )
set( tahiti_aliases pitcairn verde oland hainan bonaire kabini kaveri hawaii
        mullins tonga iceland carrizo fiji stoney polaris10 polaris11 )

# Support for gfx9 was added in LLVM 5.0 (r295554)
if( ${LLVM_VERSION} VERSION_GREATER "4.99.99" )
        set( tahiti_aliases ${tahiti_aliases} gfx900 gfx902 )
endif()

# Support for Vega12 and Vega20 was added in LLVM 7 (r331215)
if( ${LLVM_VERSION} VERSION_GREATER "6.99.99" )
        set( tahiti_aliases ${tahiti_aliases} gfx904 gfx906 )
endif()

[npmiller]

Half support for HIP was added here:

• [SYCL][HIP][libclc] Wire up AMD half support #11626

So this should be working fine now

[zjin-lcf]

Thank you for the support.

[zjin-lcf]

@npmiller

I got the same error on an MI100 device. Can you reproduce the error using a recent build of the SYCL compiler with HIP support ?

clang version 18.0.0 (https://github.com/intel/llvm.git dfd64b6)
Target: x86_64-unknown-linux-gnu
Thread model: posix

[npmiller]

Oooh, yes, the general support for fp16 was merged but adding the aspect to the HIP plugin hasn't:

• Enable fp16 runtime support for hip oneapi-src/unified-runtime#988

So it will say it won't support it but it actually does, if you remove the aspect check from the program it should work fine.

[zjin-lcf]

There's no aspect check in user programs. Thanks.

[zjin-lcf]

Running the HIP and SYCL programs below shows the following timing results on an MI100 GPU

HeCBench/src/f16max-sycl

$ make HIP=yes run
./main 100
Average kernel execution time 5605.638184 (us)
fp16_hmax2 PASS
Average kernel execution time 5768.554199 (us)
cd ..fp16_hmax PASS

HeCBench/src/f16max-hip

$ make run
hipcc -std=c++14 -Wall -O3 -c main.cu -o main.o
hipcc -std=c++14 -Wall -O3 main.o -o main
./main 100
Average kernel execution time 3172.998535 (us)
fp16_hmax2 PASS
Average kernel execution time 3339.958496 (us)
fp16_hmax PASS