[Hexagon] Add a test to show how to use multi input async dma pipelin… #13110
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 |
areusch
added
needs-triage
| sch.compute_at(cache_read_block_b, no) | ||
| sch.fuse(*sch.get_loops(cache_read_block_b)[1:]) | ||
|
|
||
| cache_read_block_c = sch.cache_write(compute_block, 0, "global.vtcm") |
| b, | ||
| size_a, | ||
| expected_output, | ||
| use_async_copy=0, |
There was a problem hiding this comment.
do you want to enable async_copy by default?
There was a problem hiding this comment.
No, not that it mattered much since other runs dont mark those stages as async anyway though. Just felt more correct to have it as off.
nverke
force-pushed
the
multi_input_async_dma
branch
2 times, most recently
from
8eb3278 to
7617c4e
Compare
…ule generated conv2d.
nverke
force-pushed
the
multi_input_async_dma
branch
from
6d9b5cb to
66127f4
Compare
apache#13110) * [Hexagon] Add a test to show how to use multi input async dma pipelining. * updates to variable naming and removal of 4.19MB test that fails on 888 devices. * [Hexagon] Add test cases for adding async dma pipleining to metaschedule generated conv2d. * Add tests for actual conv2d sourced from resnet50 metascheduling. * [Hexagon] Change logs. * Add test to show effect of increased buffer sizes for pipelining. * skip tests in CI. * lint
apache#13110) * [Hexagon] Add a test to show how to use multi input async dma pipelining. * updates to variable naming and removal of 4.19MB test that fails on 888 devices. * [Hexagon] Add test cases for adding async dma pipleining to metaschedule generated conv2d. * Add tests for actual conv2d sourced from resnet50 metascheduling. * [Hexagon] Change logs. * Add test to show effect of increased buffer sizes for pipelining. * skip tests in CI. * lint
| B_i8x128 = B[0, 0:128] | ||
| B_i32x32: T.int32x32 = T.reinterpret(B_i8x128, dtype="int32x32") | ||
| C[0:32] = T.call_llvm_pure_intrin( | ||
| 4217, |
There was a problem hiding this comment.
This number doesn't mean anything---those IDs are generated on every build. What intrinsic was this intended to represent?
There was a problem hiding this comment.
yeah we should replace this with T.llvm_lookup_intrinsic_id("llvm.hexagon.V6.vrmpyub.acc.128B") @nverke
There was a problem hiding this comment.
Ahh interesting thought that they were tied to each intrin, will update accordingly!
…ing.
Just adding an example to show https://github.com/apache/tvm/pull/13037/files addition working for these tests as well.
cc @adstraw