ARM SVE

Purpose

Guide agents through ARM Scalable Vector Extension (SVE/SVE2) programming: vector-length agnostic (VLA) code, predicate registers, SVE intrinsics via <arm_sve.h>, runtime vector length with svcnt, compiler flags, platform differences (Graviton3, Apple M4), and GDB debugging of SVE registers.

When to Use

• Writing high-performance SIMD on AArch64 servers (AWS Graviton3/4)
• Porting fixed-width NEON code to length-agnostic SVE
• Using predicate masks for loop tails instead of separate cleanup loops
• Auto-vectorizing with GCC/Clang -march=armv9-a+sve2
• Debugging SVE register state in GDB on hardware with SVE support
• Exploiting SVE2 dot product and crypto extensions

Workflow

1. SVE vs NEON

2. Predicate and VLA concepts

SVE registers
├── Z0–Z31  — scalable vector data registers
└── P0–P15  — predicate (mask) registers

Vector Length (VL) — determined at runtime per CPU
svcntb() → bytes per vector
svcntw() → 32-bit elements per vector

Code written once runs at full width on any SVE-capable CPU.

3. SVE intrinsics example

#include <arm_sve.h>
#include <stddef.h>

void saxpy_sve(float *y, const float *x, float alpha, size_t n) {
    svbool_t pg = svwhilelt_b32(0, n);
    size_t i = 0;

    do {
        svfloat32_t vx = svld1_f32(pg, &x[i]);
        svfloat32_t vy = svld1_f32(pg, &y[i]);
        vy = svmla_n_f32_x(pg, vy, vx, alpha);  // y += alpha * x
        svst1_f32(pg, &y[i], vy);

        i += svcntw();  // advance by vector length in 32-bit elements
        pg = svwhilelt_b32(i, n);
    } while (svptest_any(svptrue_b32(), pg));
}

gcc -march=armv9-a+sve2 -O3 -o saxpy saxpy.c

4. Key intrinsics

5. Loop tail handling

// SVE handles tails via predicates — no separate scalar epilogue
for (size_t i = 0; i < n; ) {
    svbool_t pg = svwhilelt_b32(i, n);
    // ... vector ops with pg ...
    i += svcntw();
}

Contrast with NEON: often needs scalar cleanup for n % 4 != 0.

6. Compiler auto-vectorization

# GCC vectorization remarks
gcc -march=armv9-a+sve2 -O3 -fopt-info-vec -o app app.c

# Clang
clang -march=armv9-a+sve2 -O3 -Rpass=vectorize -o app app.c

#pragma omp simd  // may use SVE when available
for (int i = 0; i < n; i++)
    c[i] = a[i] + b[i];

7. Platform differences

# Check SVE on Linux
grep -i sve /proc/cpuinfo          # "sve" or "sve2" in Features
# Or: cat /sys/devices/system/cpu/cpu0/regs/identification/id_aa64pfr0_el1

8. SVE2 extras

SVE2 adds integer dot product, crypto, and bitwise operations:

#include <arm_sve.h>

svint32_t dot = svdot_s32(svptrue_b32(),
    svld1_s8(pg, a), svld1_s8(pg, b));

Use for ML inference kernels on Graviton.

9. GDB debugging

gcc -g -march=armv9-a+sve2 -o saxpy saxpy.c
gdb ./saxpy

(gdb) break saxpy_sve
(gdb) run
(gdb) p $z0          # print SVE vector register
(gdb) p $p0          # print predicate register
(gdb) info registers z0 z1 p0

Requires GDB 10+ with SVE support and SVE-capable hardware.

10. NEON → SVE2 migration

Migration checklist
├── Replace fixed loops (i += 4) with svcntw() strides
├── Add svwhilelt predicates for tails
├── Use _x (merging) vs _z (zeroing) predicated ops intentionally
└── Test on multiple VL hardware or use QEMU sve-max-vq

# QEMU SVE emulation
qemu-aarch64 -cpu max ./saxpy

Common Problems

Related Skills

• skills/low-level-programming/assembly-arm — AArch64 assembly and NEON
• skills/low-level-programming/simd-intrinsics — general SIMD concepts
• skills/platform/apple-silicon — Apple M-series specifics
• skills/compilers/gcc — -march flags
• skills/compilers/clang — vectorization remarks
• skills/low-level-programming/cpu-cache-opt — memory layout for SIMD