neon_benchs's Introduction

ARM NEON benchmarks

This is simple and dumb benchmarks written to explore ARM CPUs performance.

It implements simple alpha blending for premultiplied pixels format using different techniques.

extern void
opSourceOver_premul(
    uint8_t* restrict Rrgba, const uint8_t* restrict Srgba,
    const uint8_t* restrict Drgba, size_t len)
{
    size_t i = 0;
    for (; i < len*4; i += 4) {
        uint8_t Sa = 255 - Srgba[i + 3];
        Rrgba[i + 0] = DIV255(Srgba[i + 0] * 255 + Drgba[i + 0] * Sa);
        Rrgba[i + 1] = DIV255(Srgba[i + 1] * 255 + Drgba[i + 1] * Sa);
        Rrgba[i + 2] = DIV255(Srgba[i + 2] * 255 + Drgba[i + 2] * Sa);
        Rrgba[i + 3] = DIV255(Srgba[i + 3] * 255 + Drgba[i + 3] * Sa);
    }
}

Runing

$ make

Run default benchmarks, which include not vectorized and auto vectorized variants.

$ make arm

Run benchmarks for ARM target, which include NEON optimized versions.

$ make x86

Run benchmarks for x86 target, which include SSE and AVX optimized versions.

$ CC='clang' make arm

Use different C compiler. Default compiler is cc.

$ CFLAGS='-O3' make arm

Use different compiler parameters. Default is -Wall -O2.

Meaning

Each test runs 4 times and each run contains of 100 000 calls of opSourceOver_premul function, which gives 100MPx picture in total. The results are given in seconds, so, for example, if result is 0.160 s, the speed of conversion is 100 / 0.16 = 625 MPx/s.

Recommend Projects