oneapi-agent-toolkit

GPU test: softmax_hecbench

• Result: ok: PASS on NVIDIA GeForce RTX 4090 (cuda:0, nvidia): sycl_verify_vs_reference: PASS: 1/1 case(s) match (worst max abs err 9.69e-08, atol 1e-06); sycl_verify_vs_native: PASS: 1/1 case(s) match (worst max abs err 7.45e-09, atol 1e-06)
• Device: NVIDIA GeForce RTX 4090 (cuda:0), vendor nvidia
• SYCL target flags: -fsycl-targets=nvidia_gpu_sm_89

step	status	summary
reference_build	ok	exit 0: g++ -O2 -std=c++17 -I/tmp/tmp.s2UTLv4Bj4/task /tmp/tmp.s2UTLv4Bj4/task/softmax_ref.cpp -o build/reference
native_build	ok	exit 0: /usr/local/cuda/bin/nvcc -O2 -std=c++17 -arch=native -I/tmp/tmp.s2UTLv4Bj4/task /tmp/tmp.s2UTLv4Bj4/task/softmax.cu -o build/native
native_verify_vs_reference	ok	PASS: 1/1 case(s) match (worst max abs err 9.69e-08, atol 1e-06)
sycl_compile	ok	compiled 1 source(s) -> /tmp/tmp.s2UTLv4Bj4/out/work/build/sycl (0 warning(s))
sycl_verify_vs_reference	ok	PASS: 1/1 case(s) match (worst max abs err 9.69e-08, atol 1e-06)
sycl_verify_vs_native	ok	PASS: 1/1 case(s) match (worst max abs err 7.45e-09, atol 1e-06)
sycl_benchmark	ok	10 run(s) each; cuda:0: median 300.07 ms, p95 331.45 ms
native_benchmark	ok	10 run(s) each; none: median 334.13 ms, p95 361.46 ms

Limitations:

• Benchmarks are single-process wall times plus program-reported kernel times; treat them as rough.

This site is open source. Improve this page.