Add SDXL conv shapes, extra iree flags option, tool to plot roofline percentages #19
Conversation
convbench/conv_bench.py
Outdated
| # Compute percentage of the roofline. | ||
| tflops_map = { | ||
| "f32": 653.7, | ||
| "f16": 1307.4, | ||
| "bf16": 1307.4, | ||
| "f8E4M3FNUZ": 2614.9, | ||
| "i8": 2614.9, | ||
| } |
There was a problem hiding this comment.
This is for SPX, right? For CPX you need a different set of numbers. Also, where did you find the reference numbers?
There was a problem hiding this comment.
Also, should we put this in some common file used by all scripts?
There was a problem hiding this comment.
The convbench scripts still have some hardcoded values specific to SPX MI300X, so I didn't bother trying to multiplex based on the target yet. I'm leaving that as a follow-up. I can leave a TODO comment here, though.
There was a problem hiding this comment.
Yeah these are just hardcoded values I pulled from here: https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/data-sheets/amd-instinct-mi300x-platform-data-sheet.pdf
|
You will need to change the other calls for shark_conv in the github workflows file for the CI to progress. Oh, you made that change. Trying a re-run |
.github/workflows/run_bench.yml
Outdated
| python convbench/shark_conv.py --roofline results/iree_gemm_tk.csv --plot results/iree_gemm_tk.png | ||
| python convbench/shark_conv.py --roofline results/iree_gemm.csv,results/iree_gemm_tk.csv,results/iree_attention.csv,results/iree_conv.csv --plot results/combined.png | ||
| python convbench/conv_bench.py --roofline results/iree_conv.csv --plot results/iree_conv_i8.png --dtype i8 | ||
| python convbench/conv_bench.py --roofline results/iree_conv.csv --plot results/iree_conv_f32.png --dtype f32 |
There was a problem hiding this comment.
Now that we are only using a subset of conv tests in CI, we don't need this dtype f32 call which is causing CI failure
Signed-off-by: Max Dawkins <max.dawkins@gmail.com>
Signed-off-by: Max Dawkins <max.dawkins@gmail.com>
Signed-off-by: Max Dawkins <max.dawkins@gmail.com>
shark_conv.pytoconv_bench.pyto match gemm and attention formats