Example of how to capture CPU counters with ETW on Windows, perf on Linux or kperf on Apple.
Using ETW needs somewhat recently updated Windows 10 or 11. Not sure about exact version.
Currently tested on:
| import torch | |
| import torch._inductor.config | |
| import time | |
| torch._inductor.config.triton.cudagraphs = False | |
| torch.set_float32_matmul_precision('high') | |
| def bench(f, name=None, iters=100, warmup=5, display=True, profile=False): | |
| for _ in range(warmup): | |
| f() |
| // | |
| // KITTY compression algorithm, by snowcat | |
| // converted to C, by odzhan | |
| // 2019-12-07 | |
| // | |
| #include <stdio.h> | |
| #include <stdlib.h> | |
| #include <stdint.h> | |
| #include <time.h> |
If you don't care about the explanation, scroll down to find the code, it's 50 some odd lines and written by someone who doesn't know any better. You have been warned.
This is a very simple proof of concept jitting RPN calculator implemented in python. Basically, it takes the source code, tokenizes it via whitespace, and asks itself one simple question: am I looking at a number or not?
First, let's talk about the underlying program flow. Pretend that you are a shoe connoisseur with a tiny desk. You may only have two individual shoes on that desk at any one time, but should you ever purchase a new one or get harassed by an unruly shoe salesman without realizing that you have the power to say no (or even maybe?), you can always sweep aside one of the two shoes on the desk (the one on the right, because you're a lefty and you feel that the left side is always superior) onto the messy floor, put the other shoe on the right hand side, and then place your newly acquired shoe in
