[Hexagon] Add E2E test demonstrating how to apply blocked layout schedule to conv2d via metaschedule #13180
Conversation
converting input and output activation to Hexagon's blocked layout.
See TODO, this may also disable auto scheduling for non-convolution ops.
|
Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment. Generated by tvm-bot |
|
@tvm-bot rerun |
tests/python/contrib/test_hexagon/metaschedule_e2e/test_resnet50_int8.py
Outdated
Show resolved
Hide resolved
| strategy="replay-trace", | ||
| builder=get_hexagon_local_builder(), | ||
| runner=get_hexagon_rpc_runner(hexagon_launcher, number=20), | ||
| # TODO(csullivan): Configrm the below is accurate |
There was a problem hiding this comment.
cc @zxybazh so by using ScheduleFn here, MS won't be looking at non conv2d blocks?
There was a problem hiding this comment.
By using ScheduleFn here, the search space is solely generated by the schedule function over the whole workload. In this case, if you're using _schedule_packed_8x8x32_conv2d's schedule_fn to construct the ScheduleFn space generator, it will only look at conv2d blocks because the function doesn't transform any of the non-conv2d blocks right?
There was a problem hiding this comment.
@csullivan so shall we remove the TODO comment (which also has a typo lol)?
There was a problem hiding this comment.
:) Yes, but one more question, if one utilizes a ScheduleFn to define a search space, but also sets custom sch_rules to be applied per block as,
space=ms.space_generator.ScheduleFn(
schedule_conv2d_for_tune,
sch_rules=[some set of rules] , # <<<<
postprocs=[],
mutator_probs={},
),
what will be the interaction between the scheduling function and the rules?
There was a problem hiding this comment.
For ScheduleFn I don't think there will be any interaction between the scheduling function and the rules, they will only be used for PostOrderApply space generator.
There was a problem hiding this comment.
@csullivan can you remove the TODO comment?
| def schedule_fn(sch, conv2d_block: Optional[BlockRV] = None) -> bool: | ||
| if conv2d_block == None: | ||
| try: | ||
| conv2d_block = sch.get_block("conv2d_NCHWc_int8") |
There was a problem hiding this comment.
My PR #13206 has has_block(sch, block_name) utility, which can remove this hack. I'll update this code after it is merged.
|
Thanks for patience, let us hold off on merging this one for a bit as there are a few issues that arise at the end of a tuning run with this test:
|
|
Hi Chris, regarding (2) does the worker hang when the search space that has only one sample, and will not terminate by itself even after 5 iterations? |
|
Thanks @zxybazh for following the thread closely. Looks like (2) was a red herring and is not reproducible for me any longer. Furthermore, (1) is coming in a follow up PR and this test is skipped so no harm merging this now. |
…dule to conv2d via metaschedule (apache#13180) Demonstrates the use of a custom search space generation from a manual written schedule. In this there are no degrees of freedom so tuning will not improve the schedule currently, but the test demonstrates the schedule-time layout transformation of the conv2d primfunc to the 8x8x32 packed layout for convolution ops. * Add test that demonstrates applying a custom TIR schedule to E2E model. * Add example scheduling that demonstrates converting input and output activation to Hexagon's blocked layout. * PR feedback: Remove latency measurement and profiling report. * Update comment to reflect PR discussion summary, remove TODO. * Use ScheduleFn space generator to disable any autotuning.
…dule to conv2d via metaschedule (apache#13180) Demonstrates the use of a custom search space generation from a manual written schedule. In this there are no degrees of freedom so tuning will not improve the schedule currently, but the test demonstrates the schedule-time layout transformation of the conv2d primfunc to the 8x8x32 packed layout for convolution ops. * Add test that demonstrates applying a custom TIR schedule to E2E model. * Add example scheduling that demonstrates converting input and output activation to Hexagon's blocked layout. * PR feedback: Remove latency measurement and profiling report. * Update comment to reflect PR discussion summary, remove TODO. * Use ScheduleFn space generator to disable any autotuning.
Demonstrates the use of a custom search space generation from a manual written schedule. In this there are no degrees of freedom so tuning will not improve the schedule currently, but the test demonstrates the schedule-time layout transformation of the conv2d primfunc to the 8x8x32 packed layout for convolution ops.