[CUDA] Failed to build crossEntropy with error in backend: Cannot select: t38: v2f16 = bitcast t37

[skosnits]

Describe the bug
[CUDA] Failed to build crossEntropy with error in backend: Cannot select: t38: v2f16 = bitcast t37

To Reproduce

1. Get SYCL with CUDA support (GSG)
2. Build attached source file: crossEntropy_for_external.cpp.txt

clang++ -fsycl -fsycl-unnamed-lambda -fsycl-targets=nvptx64-nvidia-cuda-sycldevice crossEntropy_for_external.cpp -o crossEntropy.run  -lOpenCL

Error message:

clang++: warning: argument 'nvptx64-nvidia-cuda-sycldevice' is deprecated, use 'nvptx64-nvidia-cuda' instead [-Wdeprecated]
warning: linking module 'remangled-l64-signed_char.libspirv-nvptx64--nvidiacl.bc': Linking two modules of different target triples: 'remangled-l64-signed_char.libspirv-nvptx64--nvidiacl.bc' is 'nvptx64-unknown-nvidiacl' whereas 'crossEntropy_for_external.cpp' is 'nvptx64-nvidia-cuda'
 [-Wlinker-warnings]
1 warning generated.
fatal error: error in backend: Cannot select: t38: v2f16 = bitcast t37
  t37: f32,ch = CopyFromReg t0, Register:f32 %7
    t36: f32 = Register %7
In function: _ZZZ15loss_fwd_kernelIN2cl4sycl6detail9half_impl4halfEENS1_5eventERNS1_5queueEPT_PlSA_S9_S9_S9_S9_ENKUlRNS1_7handlerEE_clESC_ENKUlNS1_7nd_itemILi3EEEE_clESF_
llvm-foreach:
clang++: error: clang frontend command failed with exit code 70 (use -v to see invocation)
clang version 15.0.0 (https://github.com/intel/llvm.git 8ec97550b7a84e5d2483f4891059f80abd8db39c)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: <path to built sycl>
clang++: note: diagnostic msg: Error generating preprocessed source(s).

Environment

• OS: Linux
• Target device and vendor: Nvidia GPU
• DPC++ version: https://github.com/intel/llvm.git 8ec9755
• CUDA version:

nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.57.02    Driver Version: 470.57.02    CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+

[zjin-lcf]

I didn't observe compile/run errors using HIP backend.

[bader]

This looks like a bug in NVPTX LLVM target back-end.

[jchlanda]

Would it not be an UB to typecast between float and 2xhalf, like the example here does?
Saying that, I think the compiler should do better than hard crash on it. Submitted a patch against the mainline here: https://reviews.llvm.org/D124171

[zjin-lcf]

The cuda compiler supports the conversion.

[jchlanda]

I've run it on a titan machine with DPCPP just now:

# ./crossEntropy.run
Matrix size(bs * W * H) = 128 * 81 * 8732
LossNLL_FWD CPU time(ms)=. 2.01997
LossNLL_FWD GPU time(ms)=. 2.00132
FWDBandWidth = 179.963433 (GB / s), 386722816.000000
LossNLL_BWD CPU time(ms)=. 1.82208
LossNLL_BWD GPU time(ms)=. 1.80568
BWDBandWidth = 214.169755 (GB / s), 386722816.000000

[jchlanda]

This has been merged into sycl branch in: 76d1f5e