fbgemm-xpu: adds performance benchmarking for the XPU jagged tensor operators

[aagalleg]

Adds a standalone jagged-sweep benchmark that verifies correctness against CPU before timing, patches the upstream FBGEMM jagged benchmark to run on XPU, commits PVC and BMG baselines with notes on interpreting them, and wires smoke runs into CI. Together these provide a reproducible performance reference for detecting regressions in the SYCL jagged kernels.

Operator(s)

Benchmarked (forward and backward; previously implemented):

• fbgemm::jagged_to_padded_dense — copies jagged values into a padded dense tensor
• fbgemm::jagged_2d_to_dense — 2D variant of jagged→dense conversion
• fbgemm::dense_to_jagged — gathers valid rows of a dense tensor into jagged form
• fbgemm::jagged_dense_elementwise_add_jagged_output — adds dense to jagged values, jagged output

Changes

• jagged-sweep benchmark in src/fbgemm_xpu/bench/jagged_sweep.py: sweeps B ∈ {32…2048} × max_len ∈ {16…1024} × {float32, bfloat16, float16}, D = 128 (DLRM v3 hstu_transducer_embedding_dim); every shape is verified against the CPU implementation before timing; emits CSV with a metadata header
• Timing helper in src/fbgemm_xpu/bench/bench_utils.py: XPU-event timing with last-level-cache flush and a GPU lead (torch.xpu._sleep spin) so the timed window never measures host submission
• XPU skip handling for non-ported upstream benchmark cases in src/fbgemm_xpu/bench/skips.py
• Patch 0002-Add-XPU-support-to-fbgemm-jagged-benchmark.patch enabling the upstream jagged_tensor_benchmark.py device case on XPU
• Committed baselines: bench/baselines/pvc-max1550/jagged_tensor.csv and bench/baselines/bmg-b60/jagged_tensor.csv, plus bench/baselines/README.md documenting methodology, byte-count formulas, and known caveats (effective vs. physical bandwidth, latency floors, BMG memory compression)
• CI: smoke-runs jagged-sweep (correctness gate) and the patched upstream benchmark on every run; validates the patch touches only fbgemm_gpu/bench/
• Docs: CONTRIBUTING.md benchmarking section and README.md updates
• Tests in tests/test_jagged_sweep.py, tests/test_bench_utils.py, tests/test_bench_skips.py

Testing

pytest -rsf packages/fbgemm-xpu/tests/   # includes test_jagged_sweep, test_bench_utils, test_bench_skips
python -m fbgemm_xpu.bench.jagged_sweep --smoke --output /tmp/jagged_sweep_smoke.csv
python fbgemm_gpu/bench/jagged_tensor_benchmark.py device --batch-size 8 --max-len 16   # after patch

jagged-sweep compares every forward output and input gradient against the CPU implementation with torch.testing.assert_close before timing; the run fails and writes no CSV on any mismatch. All 576 sweep cases passed on both devices.

Baseline run-to-run variation: PVC median 0.7% (max 5.5%), BMG median 0.1% (max 4.9%).

Hardware Validated

• Intel Data Center GPU Max 1550 (PVC), one tile (ZE_AFFINITY_MASK=0)
• Intel Arc Pro B60 Graphics (BMG), one device, 4-CPU pod with OMP_NUM_THREADS=4

[dvrogozh]

I suggest to drop comments in the .csv file and have only actual data here. This way Github will actually be able to render the .csv file for us in a nice column way. It will actually become reviewable. If you need to save comments, then create a README.md in the folder and describe what each .csv file contains there.