[RISC-V] Vectorize float32 minimum reductions with RVV

[ww8191201-coder]

[RISC-V] Vectorize float32 minimum reductions with RVV

Motivation

Single-axis minimum reductions over long rows are a common tensor operation. The current sequential CPU reduction processes those rows one element at a time, which limits throughput on RVV hardware. A specialized vector reduction improves this case while preserving the general reduction machinery for more complex shapes and operators.

Implementation

• Add an RVV specialization for float32 minimum reductions.
• Enable the fast path only for a single reduced axis, identity mapping, no index output, and sufficiently long rows.
• Support non-unit input strides with vector strided loads.
• Preserve NaN handling and select the first matching minimum when equal values are present, including signed-zero cases.
• Keep the existing sequential reduction kernel as the fallback for all other reductions.

Testing

The patch was tested against commit b84609d3fc73d20929c114eab95faaa56e6c5ede on a RISC-V host with eight SpacemiT A100 cores (CPUs 0-7), GCC 14.3.0, and Linux 6.18.3.

Apply and build:

git apply patch-opt.diff
make -j8

The full build completed successfully. Functional operator tests were not run as part of this performance-validation pass; the results below are performance measurements, not a numerical-correctness claim.

The benchmark used MXNet's benchmark.opperf.utils.benchmark_utils.run_performance_test with the native profiler, float32 CPU tensors, 50 warm-up iterations, and 300 measured iterations. Each result below is from an independent invocation with OMP_NUM_THREADS=8, OMP_DYNAMIC=FALSE, and taskset -c 0-7.

operator: mx.nd.min
input:    data=(4096, 1024), axis=1
metric:   avg_time_forward_min (ms)

Performance

Lower is better.

Build	Run 1	Run 2	Run 3	Median
Baseline	3.9407 ms	3.9122 ms	3.9528 ms	3.9407 ms
After git apply patch-opt.diff	2.0157 ms	2.0267 ms	2.0153 ms	2.0157 ms

The median latency is reduced by 48.85%, corresponding to a 1.96x speedup. The patched run-to-run range was 0.57% of its median.