host and device results mismatch in nbody

[zjin-lcf]

I find that the device and host results from the nbody programs in SYCL and OpenMP (icpx) do not match when the error bound is 1e-3. The program is https://github.com/zjin-lcf/HeCBench/tree/master/src/nbody-sycl.

The results match for the CUDA, HIP and OpenMP programs (nvc++ and amdclang++).

The issue can be reproduced on Intel and other vendors' GPUs.

On an Intel GPU

nbody-omp$ make run
icpx  -std=c++17 -Wall -I../nbody-cuda -O3 -fiopenmp -fopenmp-targets=spir64 -D__STRICT_ANSI__ -c main.cpp -o main.o
icpx  -std=c++17 -Wall -I../nbody-cuda -O3 -fiopenmp -fopenmp-targets=spir64 -D__STRICT_ANSI__ -c GSimulation.cpp -o GSimulation.o
icpx  -std=c++17 -Wall -I../nbody-cuda -O3 -fiopenmp -fopenmp-targets=spir64 -D__STRICT_ANSI__ main.o GSimulation.o -o main
./main 16000 10
===============================
 Initialize Gravity Simulation

# Total Energy        : 14571.5
# Total Time (s)      : 0.118287
# Average Performance : 1616.15 +- 331.716
===============================

FAIL

On an NVIDIA GPU

make CUDA=yes CUDA_ARCH=sm_90 CC=icpx run -B
icpx  -std=c++17 -Wall -I../nbody-cuda -fsycl --gcc-toolchain= -fsycl-targets=nvptx64-nvidia-cuda -Xsycl-target-backend --cuda-gpu-arch=sm_90 -O3 -DUSE_GPU -c main.cpp -o main.o
icpx  -std=c++17 -Wall -I../nbody-cuda -fsycl --gcc-toolchain= -fsycl-targets=nvptx64-nvidia-cuda -Xsycl-target-backend --cuda-gpu-arch=sm_90 -O3 -DUSE_GPU -c GSimulation.cpp -o GSimulation.o
icpx  -std=c++17 -Wall -I../nbody-cuda -fsycl --gcc-toolchain= -fsycl-targets=nvptx64-nvidia-cuda -Xsycl-target-backend --cuda-gpu-arch=sm_90 -O3 -DUSE_GPU main.o GSimulation.o -o main
fatbinary warning : option 'image' has been deprecated
fatbinary warning : option 'image' has been deprecated
./main 16000 10
===============================
 Initialize Gravity Simulation

# Total Energy        : 14571.5
# Total Time (s)      : 0.0262945
# Average Performance : 2891.07 +- 253.893
===============================

FAIL