[RISC-V] Add an RVV log-softmax forward path

[ww8191201-coder]

[RISC-V] Add an RVV log-softmax forward path

Motivation

Log-softmax over a large contiguous last axis is dominated by row reductions and exponentiation. The generic CPU path does not generate an efficient RVV loop for this workload, leaving the RISC-V vector units underused. A dedicated float32 path substantially reduces the forward-pass latency while retaining the current implementation as a fallback.

Implementation

• Detect RVV compiler support in CMake, preferring rv64gcv_zfhmin_zvfhmin and falling back to rv64gcv.
• Add a contiguous float32 fast path for forward log-softmax with temperature == 1 and no length mask.
• Compute the row maximum, vectorized exponentials, and exponential sum with RVV; subtract the resulting log denominator in the output pass.
• Use range reduction and a polynomial approximation for the vector exponential.
• Pass the vector source and scalar seed to the RVV reduction intrinsics in the order required by the intrinsic API.
• Route unsupported layouts, data types, and parameter combinations through the existing CPU implementation.

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.log_softmax
input:    data=(1024, 1024), axis=-1
metric:   avg_time_forward_log_softmax (ms)

Performance

Lower is better.

Build	Run 1	Run 2	Run 3	Median
Baseline	9.9588 ms	9.6534 ms	9.6218 ms	9.6534 ms
After git apply patch-opt.diff	1.8245 ms	1.8008 ms	1.7956 ms	1.8008 ms

The median latency is reduced by 81.35%, corresponding to a 5.36x speedup. The patched run-to-run range was 1.60% of its median.