ArrayFire v3.8.2 (CUDA, 64-bit Linux, build a9b6b0e)
Platform: CUDA Runtime 11.4, Driver: 495.29.05
[0] Quadro RTX 3000, 5935 MB, CUDA Compute 7.5

Max vs ragged max
                 M               max            ragmax
                 8        0.00241453        0.00872827
                16         0.0023666        0.00861896
                32        0.00237074        0.00865682
                64        0.00245299         0.0087447
               128        0.00240028        0.00879951
               256        0.00414862        0.00866826
               512        0.00701668        0.0096099

af::array *a, *b;
void doMax() {
  af::max(*a, 0).eval();
}
void doRMax() {
  af::array val;
  af::array idx;
  af::max(val, idx, *a, *b, 0);
}
void raggedMax(const double REPEAT = 20) {
  BENCH("Max vs ragged max\n");
  BENCH("M");
  BENCH("max");
  BENCH("ragmax");
  BENCH(std::endl);
  for (s = 8; s <= 512; s *= 2) {
    BENCH(s);
    a = new af::array(s, s);
    BENCH(1000 * af::timeit(doMax));
    af::array seqlen = af::constant((unsigned)s / 2, 1, s, u32); 
    b = &seqlen;
    BENCH(1000 * af::timeit(doRMax));
    delete a;
    BENCH(std::endl);
  }
}