Write Simple Post about Non-Branching Max

[dhermes]

https://gist.github.com/dhermes/c79846c6074b938b2e10

Also

• show assembly generated
• discuss CUDA branching issues (BE BRIEF)
• discuss what CUDA does to avoid branching for max/min

---

Maybe give a tiny refresher on IEEE:

import struct

def f(val):
    return ''.join([bin(ord(c))[2:].zfill(8) for c in struct.pack('>d', val)])

def g(bits):
    z = [bits[i:i + 8] for i in xrange(0, 64, 8)]
    z = [int(c, 2) for c in z]
    z = ''.join([chr(c) for c in z])
    return struct.unpack('>d', z)[0]

def h(s, e, m):
    return g(s + e + m)

h('0', '1' * 11, '1' + '0' * 51)

[dhermes]

objdump -d test

[dhermes]

https://gist.github.com/dhermes/f17fc85999f79ae2f304

[dhermes]

• sign_bit_simple_mask.c
  • https://godbolt.org/g/ZgYvMu
  • https://godbolt.org/g/9PPTbC (optimized)
• sign_bit_via_char_bytes.c
  • https://godbolt.org/g/EAoktt
  • https://godbolt.org/g/C8n7MZ (optimized)

https://github.com/mattgodbolt/gcc-explorer is awesome

[dhermes]

Also, a simple "optimized" fabs: https://godbolt.org/g/q2lryx