CODEWITHSUNDEEP
c++armarm64neon
gaoshuzhendanteb
gaoshuzhendanteb

Reputation: 11

What is the difference between vfmaq_f32 and vmlaq_f32 in the neon instruction set, and the difference in running speed and accuracy

hello，What is the difference between vfmaq_f32 and vmlaq_f32 in the neon instruction set, and the difference in running speed and accuracy

On macOS ARM64, the code runs consistently

#include<arm_neon.h>
#include<iostream>
using namespace std;
int main(){
    float a = 12.3839467819;
    float b = 21.437678904;
    float c = 4171.42144;
    printf("%.17f\n",a);
    printf("%.17f\n",b);
    printf("%.17f\n",c);


    printf("%.17f\n",a+b*c);

    float32x4_t a_reg = vdupq_n_f32(a);
    float32x4_t b_reg = vdupq_n_f32(b);
    float32x4_t c_reg = vdupq_n_f32(c);
    float32x4_t res_reg = vfmaq_f32(a_reg, b_reg, c_reg);
    float res[4] = {0.f};
    vst1q_f32(res,res_reg);
    printf("%.17f\n",res[0]);


    res_reg = vmlaq_f32(a_reg, b_reg, c_reg);
    vst1q_f32(res,res_reg);
    printf("%.17f\n",res[0]);


    res_reg = vmulq_f32(b_reg, c_reg);
    res_reg = vaddq_f32(res_reg, a_reg);
    vst1q_f32(res,res_reg);
    printf("%.17f\n",res[0]);
    return 0;
}

Upvotes: -1

Views: 635

Answers (1)

nemequ
nemequ

Reputation: 17502

vfmaq_f32 defined as a single fused operation, whereas vmlaq_f32 can be implemented with a multiply then an accumulate.

Two explanations come to mind. First, at some point the fused version (the FMLA instruction) was possibly an optional instruction (I don't know when, and I'm a bit too lazy to dig through really old documentation). The second possibility, which seems more likely, is that the fused version may have, at some point, been a bit slower.

These days, it looks like compilers pretty much just compile both to the same instruction so it's effectively an alias, but you should probably still use vfmaq_f32 if you want accuracy, but vmlaq_f32 if you are more interested in speed.

Upvotes: 0

Related Questions