[RISC-V] Vectorize col2im accumulation in convolution backward

[ww8191201-coder]

[RISC-V] Vectorize col2im accumulation in convolution backward

Motivation

CPU convolution backward-data uses col2im to accumulate columns back into the input-gradient tensor. The inner loop performs a bounds check and scalar addition for every output column. For a fixed kernel position, the in-bounds columns form one contiguous interval, which allows the padding work to be skipped and the accumulation to be vectorized on RVV systems.

Implementation

• Compute the valid output-column interval outside the innermost col2im loop.
• Skip padding positions in blocks while advancing the column buffer consistently.
• Add a float32 RVV accumulation helper for dst += src; use unit-stride loads and stores where possible and strided accesses when required by the destination layout.
• Keep the existing scalar accumulation path for non-float types and builds without RVV.
• Make the patch self-contained by adding the required headers and local RVV feature guard. It does not depend on the separate im2col optimization.

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. Backward execution was enabled. Each result below is from an independent invocation with OMP_NUM_THREADS=8, OMP_DYNAMIC=FALSE, and taskset -c 0-7.

operator:    mx.nd.Convolution (backward)
inputs:      data=(32, 3, 256, 256), weight=(1, 3, 3, 3), bias=(1,)
parameters:  kernel=(3, 3), num_filter=1, layout=NCHW,
             pad=(0, 0), stride=(1, 1), dilate=(1, 1)
metric:      avg_time_backward_Convolution (ms)

Performance

Lower is better.

Build	Run 1	Run 2	Run 3	Median
Baseline	261.8669 ms	261.7465 ms	261.6667 ms	261.7465 ms
After git apply patch-opt.diff	194.4698 ms	195.1131 ms	192.8183 ms	194.4698 ms

The median latency is reduced by 25.70%, corresponding to a 1.35x speedup. The patched run-to-run range was 1.18% of its median.