Leandro Lacerda 08ff017fb0
[libc] Improve GPU benchmarking (#153512)
This patch improves the GPU benchmarking in this way:

* Replace `rand`/`srand` with a deterministic per-thread RNG seeded by
`call_index`: reproducible, apples-to-apples libc vs vendor comparisons.
* Fix input generation: sample the unbiased exponent uniformly in
`[min_exp, max_exp]`, clamp bounds, and skip `Inf`, `NaN`, `-0.0`, and
`+0.0`.
* Fix standard deviation: use an explicit estimator from sums and
sums-of-squares (`sqrt(E[x^2] − E[x]^2)`) across samples.
* Fix throughput overhead: subtract a loop-only baseline inside
NVPTX/AMDGPU timing backends so `benchmark()` gets cycles-per-call
already corrected (no `overhead()` call).
* Adapt existing math benchmarks to the new RNG/timing plumbing (plumb
`call_index`, drop `rand/srand`, clean includes).
* Correct inter-thread aggregation: use iteration-weighted pooling to
compute the global mean/variance, ensuring statistically sound `Cycles
(Mean)` and `Stddev`.
* Remove `Time / Iteration` column from the results table: it reported
per-thread convergence time (not per-call latency) and was
redundant/misleading next to `Cycles (Mean)`.
* Remove unused `BenchmarkLogger` files: dead code that added
maintenance and cognitive overhead without providing functionality.

---

## TODO (before merge)

* [ ] Investigate compiler warnings and address their root causes.
* [x] Review how per-thread results are aggregated into the overall
result.

## Follow-ups (future PRs)

* Add support to run throughput benchmarks with uniform (linear) input
distributions, alongside the current log2-uniform scheme.
* Review/adjust the configuration and coverage of existing math
benchmarks.
* Add more math benchmarks (e.g., `exp`/`expf`, others).
2025-08-15 11:00:17 -05:00
..

Libc mem* benchmarks

This framework has been designed to evaluate and compare relative performance of memory function implementations on a particular machine.

It relies on:

  • libc.src.string.<mem_function>_benchmark to run the benchmarks for the particular <mem_function>.
  • libc-benchmark-analysis.py3 a tool to process the measurements into reports.

Benchmarking tool

Setup

cd llvm-project
cmake -B/tmp/build -Sllvm -DLLVM_ENABLE_PROJECTS='clang;clang-tools-extra;libc' -DCMAKE_BUILD_TYPE=Release -DLIBC_INCLUDE_BENCHMARKS=Yes -G Ninja
ninja -C /tmp/build libc.src.string.<mem_function>_benchmark

Note: The machine should run in performance mode. This is achieved by running:

cpupower frequency-set --governor performance

Usage

The benchmark can run in two modes:

  • stochastic mode returns the average time per call for a particular size distribution, this is the default,
  • sweep mode returns the average time per size over a range of sizes.

Each benchmark requires the --study-name to be set, this is a name to identify a run and provide label during analysis. If stochastic mode is being used, you must also provide --size-distribution-name to pick one of the available MemorySizeDistribution's.

It also provides optional flags:

  • --num-trials: repeats the benchmark more times, the analysis tool can take this into account and give confidence intervals.
  • --output: specifies a file to write the report - or standard output if not set.

Stochastic mode

This is the preferred mode to use. The function parameters are randomized and the branch predictor is less likely to kick in.

/tmp/build/bin/libc.src.string.memcpy_benchmark \
    --study-name="new memcpy" \
    --size-distribution-name="memcpy Google A" \
    --num-trials=30 \
    --output=/tmp/benchmark_result.json

The --size-distribution-name flag is mandatory and points to one of the predefined distribution.

Note: These distributions are gathered from several important binaries at Google (servers, databases, realtime and batch jobs) and reflect the importance of focusing on small sizes.

Using a profiler to observe size distributions for calls into libc functions, it was found most operations act on a small number of bytes.

Function % of calls with size ≤ 128 % of calls with size ≤ 1024
memcpy 96% 99%
memset 91% 99.9%
memcmp1 99.5% ~100%

1 - The size refers to the size of the buffers to compare and not the number of bytes until the first difference.

Sweep mode

This mode is used to measure call latency per size for a certain range of sizes. Because it exercises the same size over and over again the branch predictor can kick in. It can still be useful to compare strength and weaknesses of particular implementations.

/tmp/build/bin/libc.src.string.memcpy_benchmark \
    --study-name="new memcpy" \
    --sweep-mode \
    --sweep-max-size=128 \
    --output=/tmp/benchmark_result.json

Analysis tool

Setup

Make sure to have matplotlib, pandas and seaborn setup correctly:

apt-get install python3-pip
pip3 install matplotlib pandas seaborn

You may need python3-gtk or similar package to display the graphs.

Usage

python3 libc/benchmarks/libc-benchmark-analysis.py3 /tmp/benchmark_result.json ...

When used with multiple trials Sweep Mode data the tool displays the 95% confidence interval.

When providing with multiple reports at the same time, all the graphs from the same machine are displayed side by side to allow for comparison.

The Y-axis unit can be changed via the --mode flag:

  • time displays the measured time (this is the default),
  • cycles displays the number of cycles computed from the cpu frequency,
  • bytespercycle displays the number of bytes per cycle (for Sweep Mode reports only).

Under the hood

To learn more about the design decisions behind the benchmarking framework, have a look at the RATIONALE.md file.