[Hexagon] Add test to show scheduling of resnet50 with async dma pipe… #13352
Conversation
|
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 |
| return tvm_build(mod, target=target) | ||
| else: | ||
| mod = RemoveWeightLayoutRewriteBlock(skip_ndarray_rewrite=True)(mod) | ||
| return tvm_build(mod, target=target) |
There was a problem hiding this comment.
Use meta_schdule.builder.default_build here.
There was a problem hiding this comment.
✅ Changed this to use the old strategy if pass context is not present.
| def dot_product_32x4_u8i8i32_vtcm_desc( | ||
| A: T.Buffer((4,), "uint8", offset_factor=1, scope="global.vtcm"), | ||
| B: T.Buffer((32, 4), "int8", offset_factor=1, scope="global.vtcm"), | ||
| C: T.Buffer((32,), "int32", offset_factor=1, scope="global.vtcm"), |
There was a problem hiding this comment.
Is it possible to parametrize scope to remove duplication?
tests/python/contrib/test_hexagon/metaschedule_e2e/test_resnet50_int8.py
Outdated
Show resolved
Hide resolved
|
|
||
|
|
||
| @tvm.testing.requires_hexagon | ||
| def test_async_dma_resnet50(hexagon_launcher): |
There was a problem hiding this comment.
I think it should be possible to remove dups with test_packed_8x8x32_resnet50
There was a problem hiding this comment.
Was able to unify a good amount of the code ✅
| debug_ex = session.get_graph_debug_executor( | ||
| hexagon_lowered.get_graph_json(), hexagon_lowered.lib | ||
| ) | ||
| print(debug_ex.profile(input_name=inp.copy())) |
There was a problem hiding this comment.
Do we want benchmark at L589 and profiling here, given that this test is not really tuning anything?
There was a problem hiding this comment.
I wanted to run these to see what the performance of the model is in summation after we use all of these schedules. Is there a better way to get that?
| sch.unroll(new_loops[-4]) | ||
| # TODO(nverke): Add compute optimizations here. | ||
| else: | ||
| # Handle case where kernel is 1x1 |
There was a problem hiding this comment.
I think if you set preserve_unit_loops = True in compute_at, you can get rid off this branch.
nverke
force-pushed
the
async_dma_resnet50
branch
from
476948e to
7980e3a
Compare
|
The changes to TIR broke the tensorization with this so those issues need to be fixed. For the time being rebased on top of another commit and made updates that were requested. |
|
Actually there was also a |
| sch.parallel(new_loops[4]) | ||
| sch.unroll(new_loops[5]) | ||
| # TODO(nverke): Add compute optimizations here. | ||
| sch.blockize(loop=oc_i) |
There was a problem hiding this comment.
Probably doesn't need this blockize (tensorize does blockize anyway)
There was a problem hiding this comment.
For some reason tensorization breaks when removing this...
nverke
force-pushed
the
async_dma_resnet50
branch
from
10ace0b to
c38809e
Compare
|
Was able to pull in the fix for vrmpy tensorization and rebase this onto mainline so should be ok to push now. |
…lines using metaschedule
nverke
force-pushed
the
async_dma_resnet50
branch
from
42da1b3 to
183cbc8
Compare
apache#13352) * [Hexagon] Add test to show scheduling of resnet50 with async dma pipelines using metaschedule * lint
…lines using metaschedule
This test uses a schedule function to run each conv2d in resnet50 in a pipeline async dma copies to vtcm and parallel HVX.