Riscv Rvv

Scope

• RVV 1.0 C intrinsics and the vector-length-agnostic execution model.
• Instruction forms, policies, data types.
• Memory access patterns (contiguous / strided / segmented / gather / slide).
• Reduction, softmax-style normalization, NTT-style modular arithmetic, CRC folding.
• Perf counter usage with rdinstret / rdcycle.
• 275-idea optimization catalog split by topic, derived from peer-reviewed papers and open-source kernels.

Quick Template

#include <riscv_vector.h>
#include <stddef.h>

void my_kernel(float *dst, const float *src, size_t n) {
    size_t avl = n;
    while (avl > 0) {
        size_t vl = __riscv_vsetvl_e32m1(avl);
        vfloat32m1_t v = __riscv_vle32_v_f32m1(src, vl);
        // ... compute ...
        __riscv_vse32_v_f32m1(dst, v, vl);
        src += vl; dst += vl; avl -= vl;
    }
}

Reference files (load on demand)

Load only what the current task needs. Progressive disclosure — keep the context small.

Core RVV reference

• references/core-concepts.md — VL/AVL, VLMAX, tail/mask policies, naming, memory access, reductions, conversions, widen/narrow, masks, bit ops, config macros.
• references/patterns.md — algorithmic templates: vector add, transpose, reduction, softmax, NTT/polynomial mult, CRC, microbenches, perf counters.
• references/intrinsics.md — cheat sheet by category (loads/stores, arithmetic, reductions, conversions, permute, bit ops).
• references/examples.md — 7 worked kernels (vec add, transpose, softmax, NTT, reduction, CRC, microbenches) with host/kernel interaction.
• references/template.md — step-by-step template for authoring a new RVV kernel.
• references/glossary.md — AVL, VL, VLMAX, SEW, LMUL, tail/mask policies, Zvbc.

Optimization catalog (275 ideas)

Start with the overview, then load only the topic file that matches the task.

• references/optimizations/overview.md — entry index. Hardware targets, top quick-wins on K1, cross-topic routing table. Read this first when picking among ideas.
• references/optimizations/generic.md — base 56 generic tactics (A–F across vectorization, numerics, memory hierarchy, ILP, parallelism, domain kernels). Always-applicable RVV idioms; start here on any new kernel.
• references/optimizations/compiler.md — G1–G35. LLVM/GCC backends, MLIR/IREE/TVM autotuning, SVE→RVV / NEON→RVV / x86→RVV port, build flags, PGO, LLM-driven codegen. Load when setting up the build or porting from another SIMD ISA.
• references/optimizations/microarch.md — H1–H32. Per-board tuning (K1/X280/SG2042/Ara2), PMU/roofline profiling, energy + DVFS, NUMA, hugepages, io_uring, RT scheduling, autotuner persistence. Load when measuring, chasing the last 10–30 %, or working multi-core/NUMA.
• references/optimizations/nn-llm.md — I1–I35. FlashAttention, Flash-Decoding, GQA/MLA, KV quantization (KIVI/ZipCache/RotateKV/XQuant), AWQ/SmoothQuant/SqueezeLLM, MoE dispatch, Mamba/RWKV scans, speculative decoding, sparse attention, DejaVu, SwiGLU/GELU, NOVA Winograd, V-Seek recipes.
• references/optimizations/crypto-codecs.md — J1–J35. Zk* (AES/SHA/SM*/ GHASH), Zvk* vector crypto, Zb* bitmanip, post-quantum NTT, bitsliced AES, BLAKE3/xxHash3/CRC32C, LZ4/zstd/Huffman/Parquet/simdjson/UTF-8, hash join, packet classify, JPEG IDCT.
• references/optimizations/scientific.md — K1–K40. Sparse LA (SELL-C, HYB, SpGEMM, SpTRSV), stencil/CFD, FEM/multigrid, Krylov, FFT (Stockham, Bluestein), N-body (FMM, Barnes-Hut), MD (LJ, PME), Monte Carlo, BFS / triangle, double-double, BLIS GEMM, batched ERI, Morton/octree.
• references/optimizations/alignment.md — L1–L46. Striped Smith-Waterman (Farrar/Zhao), Wozniak anti-diagonal, SWIPE inter-sequence, WFA, ksw2, edlib (Myers), Suzuki-Kasahara difference recurrence, libgaba adaptive band, HMMER3 MSV/Viterbi, PSSM, Pair-HMM, block aligner, Hirschberg, GASAL2, K1/X280/Ara2 hardware-specific tuning.

Loading Guidance

Routing for RVV kernel work

• Starting a new kernel → template.md + patterns.md for the target algorithm.
• Reading or modifying existing RVV code → core-concepts.md + intrinsics.md.
• Picking between strided / gather / segmented → core-concepts.md §5 and patterns.md §2.
• NTT / Barrett / modular → patterns.md §5.
• CRC / carry-less multiply → patterns.md §6 + intrinsics.md §Bit Ops.
• Unfamiliar term → glossary.md.

Routing for optimization questions

Always begin with optimizations/overview.md, then load the topic file (or files) that match the task. The overview's "cross-topic routing" table is the single source of truth.