forked from tinygrad/tinygrad
qazalandGitHub
1b3d00d6ac
* work * also no more no-color * reorder * update llama * sqtt readme * itertools * rm that * signals back
43 lines
1.4 KiB
Plaintext
43 lines
1.4 KiB
Plaintext
A command line tool for exploring the VIZ trace.
|
|
|
|
# Lightweight tracing
|
|
|
|
Supported on all backends.
|
|
|
|
Flags: VIZ=-1 to only save the trace to a file, VIZ=1 also launches a web server.
|
|
|
|
1. Set VIZ to -1 to save the trace.
|
|
2. Use `extra/viz/cli.py` to inspect the trace files.
|
|
|
|
## Inspect runtime profiling
|
|
|
|
Use `extra/viz/cli.py --profile` to list all traced devices.
|
|
|
|
List top slowest kernels on a device: `--profile --device "AMD" | head 10`
|
|
List samples of a kernel on a device: `--profile --device "AMD" --kernel E_3 | head 4`
|
|
|
|
## Inspect codegen and PatternMatcher
|
|
|
|
Use `extra/viz/cli.py --rewrites` to list all traced kernels.
|
|
|
|
List all codegen steps for a kernel: `--rewrites --kernel E_3`
|
|
Get source code: `--rewrites --kernel E_3 --select "View Source"`
|
|
Inspect a graph rewrite: `--rewrites --kernel E_3 --select "initial symbolic"`
|
|
|
|
# SQTT tracing
|
|
|
|
Supported on AMD for RDNA3 and RDNA4 (best) and CDNA (developing).
|
|
|
|
Flags: VIZ=-2 to save SQTT trace to a file. VIZ=2 also launches a web server. View other flags in tinygrad/runtime/ops_amd.py to configure SQTT as needed.
|
|
|
|
Use `extra/viz/cli.py --profile | grep SQTT` to view all available SQTT traces.
|
|
You can select a specific trace option with --device, Example workflow:
|
|
|
|
```bash
|
|
# Run amd_asm_matmul with VIZ=-2 to capture the trace
|
|
VIZ=-2 python extra/gemm/amd_asm_matmul.py
|
|
|
|
# View barriers
|
|
extra/viz/cli.py --profile --device "Exec kernel SQTT PKTS SE:0" | rg BARRIER | head -10
|
|
```
|