ARM Neon armv7 SIMD instruction with if comparison

Question

how to write neon code for the following loop:

float sfx[64], delta = 9.9e-5;
for(int i = 0; i < 64; i++) {
    if (sfx[i] < delta) {
        abq[i] = 1.0/delta;
    } else {
        abq[i] = 1.0/sfx[i];
    }
}

I try to use vbslq_f32, but I have to construct its parameter one by one. Why NEON doesn't provide a more convenient way to do the job? Is there any better way to do this?

float32x4_t vdelta = vdupq_n_f32((float)1.0/delta);
for(int i = 0; i < 64; i+=4) {
    float32x4_t vsfx = vld1q_f32((const float32_t*)(sfx+i));
    uint32x4_t vcon;
    vcon = vsetq_lane_u32((vgetq_lane_f32(vsfx,0)<delta), vcon, 0);
    vcon = vsetq_lane_u32((vgetq_lane_f32(vsfx,1)<delta), vcon, 1);
    vcon = vsetq_lane_u32((vgetq_lane_f32(vsfx,2)<delta), vcon, 2);
    vcon = vsetq_lane_u32((vgetq_lane_f32(vsfx,3)<delta), vcon, 3);

    float32x4_t vsfxdiv;
    vsfxdiv = vsetq_lane_f32((1.0/vgetq_lane_f32(vsfx,0)), vsfxdiv, 0);  
    vsfxdiv = vsetq_lane_f32((1.0/vgetq_lane_f32(vsfx,1)), vsfxdiv, 1);  
    vsfxdiv = vsetq_lane_f32((1.0/vgetq_lane_f32(vsfx,2)), vsfxdiv, 2);  
    vsfxdiv = vsetq_lane_f32((1.0/vgetq_lane_f32(vsfx,3)), vsfxdiv, 3);

    float32x4_t vabq = vblsq_f32(vcon, vsfxdiv, vdelta);
    vst1q_f32((abq+i), vabq);
}

Notlikethat Notlikethat · Accepted Answer · 2016-07-02T16:24:37

Indeed, it's a bit silly to defeat the point of vectorisation by performing operations one lane at a time. It's also mostly unnecessary.

This:

vcon = vsetq_lane_u32((vgetq_lane_f32(vsfx,0)<delta), vcon, 0);
vcon = vsetq_lane_u32((vgetq_lane_f32(vsfx,1)<delta), vcon, 1);
vcon = vsetq_lane_u32((vgetq_lane_f32(vsfx,2)<delta), vcon, 2);
vcon = vsetq_lane_u32((vgetq_lane_f32(vsfx,3)<delta), vcon, 3);

is just a silly way of doing this:

float32x4_t vdelta = vdupq_n_f32(delta);
// vector compare less than
vcon = vcltq_f32(vsfx, vdelta);

The division is a bit more awkward, as NEON has no division instruction, but we don't actually need general-purpose division; this reciprocal operation:

vsfxdiv = vsetq_lane_f32((1.0/vgetq_lane_f32(vsfx,0)), vsfxdiv, 0);
vsfxdiv = vsetq_lane_f32((1.0/vgetq_lane_f32(vsfx,1)), vsfxdiv, 1);
vsfxdiv = vsetq_lane_f32((1.0/vgetq_lane_f32(vsfx,2)), vsfxdiv, 2);
vsfxdiv = vsetq_lane_f32((1.0/vgetq_lane_f32(vsfx,3)), vsfxdiv, 3);

can instead be vectorised to this:

// reciprocal estimate; if precision isn't all that critical, this may suffice on its own
vsfxdiv = vrecpeq_f32(vsfx);
// otherwise, as a general rule of thumb, two Newton-Raphson iterations is
// probably sufficient for single-precision floats
vsfxdiv = vmulq_f32(vsfxdiv, vrecpsq_f32(vsfxdiv, vsfx));
vsfxdiv = vmulq_f32(vsfxdiv, vrecpsq_f32(vsfxdiv, vsfx));

Now, that said, this particular case can be simplified even further, when you consider that this:

if (sfx[i] < delta) {
    abq[i] = 1.0/delta;
} else {
    abq[i] = 1.0/sfx[i];
}

is simply this:

abq[i] = 1.0/max(delta, sfx[i]);

which means the explicit comparison and conditional select can be elided entirely and we end up with this:

float32x4_t vdelta = vdupq_n_f32(delta);
for(int i = 0; i < 64; i+=4) {
    float32x4_t vsfx, vabq;

    vsfx = vld1q_f32((const float32_t*)(sfx+i));
    vsfx = vmaxq_f32(vsfx, vdelta);

    vabq = vrecpeq_f32(vsfx);
    vabq = vmulq_f32(vabq, vrecpsq_f32(vabq, vsfx));
    vabq = vmulq_f32(vabq, vrecpsq_f32(vabq, vsfx));

    vst1q_f32((abq+i), vabq);
}

ARM Neon armv7 SIMD instruction with if comparison

1 Answers