dymokomi/luce-render
Source of release 0.1.0
0.1.0 / tests / gpu / overhead.lucb
## How long a dispatch costs in luce-gpu with no work in it: passes of N
## dependent one-workgroup dispatches of the camera kernel with no pixels, timed on the GPU and on the CPU. Printed,
## not checked.
import memory
from luce_std import clock
from luce_std import math
from luce_gpu import gpu
from luce_render import render
pub func measure(device: gpu.Device) -> !:
var kernel = try gpu.Kernel.create(device, render.camera_comp_words, render.camera_comp_msl, render.camera_comp_group, render.camera_comp_bindings)
defer kernel.destroy()
var queues = try gpu.Buffer.allocate(device, 4096)
defer queues.destroy()
var constants = try gpu.Buffer.allocate(device, 4096)
defer constants.destroy()
var bindings: gpu.Binding[16] = [gpu.Binding(), gpu.Binding(), gpu.Binding(), gpu.Binding(), gpu.Binding(), gpu.Binding(), gpu.Binding(), gpu.Binding(),
gpu.Binding(), gpu.Binding(), gpu.Binding(), gpu.Binding(), gpu.Binding(), gpu.Binding(), gpu.Binding(), gpu.Binding()]
for at in 0..<(usize)16:
if render.camera_comp_bindings[at] == 1: bindings[at] = gpu.Binding(buffer = queues if at == 6 else constants)
var block: u8[104]
for at in 0..<(usize)104: block[at] = 0
for count in [10, 100, 400]:
var best: f64 = 1e9
for round in 0..<(usize)5:
_ = round
let start = clock.now()
var pass = try gpu.Compute.begin(device)
defer pass.close()
for at in 0..<count:
_ = at
try pass.dispatch(kernel, bindings, block, 1)
let recorded = start.elapsed().seconds()
try pass.submit()
try pass.wait()
let seconds = start.elapsed().seconds()
best = math.min(best, seconds)
if round == 4: print(f"overhead: {count} dispatches: {seconds * 1e6 / (f64)count} us each wall, recording {recorded * 1e6 / (f64)count} us each")
_ = best