Runtime support for CPU hoist ops #2152
Draft
Conversation
This file contains 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
vwellsTT
changed the title
vwellsTT
added a commit
that referenced
this pull request
Feb 19, 2025
### Problem description This PR adds support for CPU hoisting changes to our TTNNToFlatbuffer translation. ### What's changed This PR introduces new flatbuffer fields for CPU hoisted ops. It hooks up the TTNNToFlatbuffer pass to properly utilize passes introduced in other PRs s.t. a CPU dylib can be embedded in our flatbuffer. Note: this PR is pretty much impossible to test standalone, unfortunately. 2 following PRs will be strictly dependent on this PR merging: 1. #2148 this PR adds earlier support for hoisting in the TTNNPipeline for TTIRToTTNN. However, this generates IR that TTNNToFlatbuffer pass cannot parse until this PR lands. (So this TTNNToFlatbuffer PR is useless without TTIRToTTNN PR, but TTIRToTTNN breaks ttmlir-translate without this PR 😄 ). 2. #2152 Runtime PR, which will add support for actually executing new flatbuffers. This is obviously dependent on this PR + its flatbuffer changes as well.
vwellsTT
added a commit
that referenced
this pull request
Feb 20, 2025
### Problem description This PR adds support for CPU hoisting changes to our TTNNToFlatbuffer translation. ### What's changed This PR introduces new flatbuffer fields for CPU hoisted ops. It hooks up the TTNNToFlatbuffer pass to properly utilize passes introduced in other PRs s.t. a CPU dylib can be embedded in our flatbuffer. Note: this PR is pretty much impossible to test standalone, unfortunately. 2 following PRs will be strictly dependent on this PR merging: 1. #2148 this PR adds earlier support for hoisting in the TTNNPipeline for TTIRToTTNN. However, this generates IR that TTNNToFlatbuffer pass cannot parse until this PR lands. (So this TTNNToFlatbuffer PR is useless without TTIRToTTNN PR, but TTIRToTTNN breaks ttmlir-translate without this PR 😄 ). 2. #2152 Runtime PR, which will add support for actually executing new flatbuffers. This is obviously dependent on this PR + its flatbuffer changes as well.
vwellsTT
force-pushed
the
vwells/cpu_hoist_runtime
branch
from
1ea72a8 to
2fbe384
Compare
vwellsTT
force-pushed
the
vwells/cpu_hoist_runtime
branch
from
2fbe384 to
d947859
Compare
vwellsTT
commented
Feb 21, 2025
| @@ -36,21 +37,6 @@ static std::string asJson(void const *fbb, uint8_t const *binarySchema, | |||
| return text; | |||
| } | |||
|
|
|||
| static std::vector<uint32_t> | |||
There was a problem hiding this comment.
Moving this to a common header because I need it for my cpu.cpp file
vwellsTT
commented
Feb 21, 2025
| @@ -141,7 +140,12 @@ class ProgramContext { | |||
| const std::unordered_map<uint32_t, ::ttnn::Tensor *> &liveTensors, | |||
| const std::vector<uint32_t> &programInputs, | |||
| const std::vector<uint32_t> &programOutputs, | |||
| ::ttnn::MeshDevice *parentMesh); | |||
| const DylibHandleMap *programDylibs, ::ttnn::MeshDevice *parentMesh) | |||
There was a problem hiding this comment.
imo this ctor should be defined in header, since 1. it's trivial and 2. other ctors are defined in header here (via= default admittedly)
vwellsTT
commented
Feb 21, 2025
| sizes[j] = ins->Get(i)->desc()->shape()->Get(j); | ||
| } | ||
| std::vector<int64_t> strides = common::calculateStride(sizes); | ||
| int64_t *sizes_and_strides = new int64_t[2 * rank]; |
There was a problem hiding this comment.
Don't like using manual new [] calls, but this is annoying because we can't a use an std::vector in this function (since memory needs to outlive this func). I guess we could pass in a 2D vector (one for each of input tensors) we exists in the calling func instead, but that's sort of ugly too imo
vwellsTT
commented
Feb 21, 2025
| @@ -0,0 +1,24 @@ | |||
| // SPDX-FileCopyrightText: (c) 2024 Tenstorrent AI ULC | |||
There was a problem hiding this comment.
Not sure if some existing header here would be more appropriate than a standalone for this--would've said it should maybe go in workarounds.h, but I don't think we have any similar funcs there
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Problem description
Add runtime support for executing hoisted CPU ops.
What's changed
Add mechanism for reading + opening dylibs from flatbuffer, executing CPU ops (by executing proper function within a given dylib).
Checklist