Pytorch/rocm7.1 internal testing hstu drop 1#881
Open
scxiao wants to merge 21 commits into
Open
Conversation
This PR generalize flags in translate_to_asm function and adds new variant to HIPOptions.schedule_hint for iterative-ilp scheduler.
…rch/rocm7.1_internal_testing_hstu_temp
See details here triton-lang#8145 --------- Co-authored-by: Vinayak Gokhale <Vinayak.Gokhale@amd.com> Co-authored-by: Dewei Wang <Dewei.Wang@amd.com>
…rch/rocm7.1_internal_testing_hstu_drop_1
) On `GFX9`, padding alone is not enough to avoid shared memory bank conflicts when using direct-to-LDS loads. For example, a `global_load_lds_dwordx4` instruction loads `64 threads × 16 bytes = 1024 bytes` at once, so padding can only be inserted at 1024-byte boundaries in LDS. If the tensor’s rows are smaller than the transfer size, we can’t insert padding between them, since they are loaded by the same instruction and written contiguously to shared memory. For instance, if each row is 256 bytes, the layout would look like: ``` [[row0], [row1], [row2], [row3], [padding], [row4], [row5], [row6], [row7], [pading], [row8] ...] ``` However, to avoid bank conflicts when reading the data, we would actually want padding between row0 and row1, to shift elements from the same column to different banks. To solve this, we need to rearrange the rows in the shared memory layout so that padding separates contiguous rows of the tensor. In this example, by padding after every 4 rows, we could permute rows like this: ``` [[row0], [row16], [row32], [row48], [padding] [row1], [row17], [row33], [row49], [padding], [row2], [row18] ...] ``` This allows us to load the data conflict free. This PR adds an optional linear component to `PaddedSharedEncoding` that allows us to permute the rows in LDS. There will be follow-up PRs adding support for them in the stream pipeliner and when lowering direct-to-lds loads on AMD.
Stop requiring local_load to have a DotOperand layout attached when trying to generate ds_read_tr instructions. This simplifies the code and allows loading of more complex cases such as local_load + tt.trans.
…rch/rocm7.1_internal_testing_hstu_drop_3
…cm7.1_internal_testing_hstu_drop_2
) This is a very old bug. The problem can be explained by following simple statement, where the fatPtrs is a simple wrapper of a hashmap. fatPtrs[toKey] = fatPtrs.at(fromKey) time 0: RHS's at() function return a reference to an existing element. time 1: There is no element associated with toKey, so the LHS needs to resize its internal bucket in order to accommodate the new entry. When resize, it invalidates the reference it took in the step 1 The fix is to change the map's value type from an aggregate type ValueT to std::unique_ptr<ValueT>. With this change, when map resizes, while unique_ptr may change, the raw pointer it maintains is unchanged, and hence we can get a "sticky" pointer for the data it maps to.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
9 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The core Triton is a small number of people, and we receive many PRs (thank
you!). To help us review your code more quickly, if you are a new
contributor (less than 3 PRs merged) we ask that you complete the following
tasks and include the filled-out checklist in your PR description.
This PR is to force to turn on buffer ops for triton kernels.
Complete the following tasks before sending your PR, and replace
[ ]with[x]to indicate you have done them.I am not making a trivial change, such as fixing a typo in a comment.
I have written a PR description following these
rules.
I have run
pre-commit run --from-ref origin/main --to-ref HEAD.Select one of the following.
/testforlittests/unittestfor C++ tests/python/testfor end-to-end testsFILL THIS IN.Select one of the following.
littests.littests I have added follow these best practices,including the "tests should be minimal" section. (Usually running Python code
and using the instructions it generates is not minimal.)