[quidditch_snitch] Reintroduce tensor.microkernel op
#106
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
The op as it is currently used and specified returns the index of compute cores (in our current simulations 0 to exclusive 8) and has unspecified behavior on the DMA core. This will be used in a later PR to properly lower it when doing DMA code specialization.
Barriers may be inserted in a loop resulting from the lowering of a `scf.forall`. The lowering of `compute_core_index` unfortunately returns `num_compute_cores + 1` which is outside the specified range of the operation and leads to such loops and their barriers being skipped by the DMA core. As the pass already assumes non-divergent control flow, we can fix this by specializing `compute_core_index` to any integer in the defined range of the op.
1b20d75 previously removed the `tensor.microkernel` operation as it at the time seemed not worth the extra code. Since then, we noted that microkernels execute in a more asynchronous manner due to Snitch's asynchronous FPU requiring the use of an explicit `microkernel_fence` operation. Optimizing the placement of these is easier done in tensor land, making the operation more worth it. Additionally, more experience in bufferization lead to simplifying its implementation by restricting `microkernel_yield` to only tensor operations. A tensor counterpart of `microkernel_fence` called `sync_tensor` has also been added which makes a result tensor of a `tensor.microkernel` operation available. It bufferizes to `microkernel_fence` and its placement could be further optimized in the future. The conservative placement of `microkernel_fence` operations was also removed from `speicalize-dma-code` leading to less barriers and `microkernel_fence`s.
zero9178
force-pushed
the
tensor-microkernel-op
branch
from
0b4285b to
50406dc
Compare
# Conflicts: # codegen/tests/Dialect/Snitch/Transforms/specialize-dma-code.mlir
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.
1b20d75 previously removed the
tensor.microkerneloperation as it at the time seemed not worth the extra code.Since then, we noted that microkernels execute in a more asynchronous manner due to Snitch's asynchronous FPU requiring the use of an explicit
microkernel_fenceoperation. Optimizing the placement of these is easier done in tensor land, making the operation more worth it. Additionally, more experience in bufferization lead to simplifying its implementation by restrictingmicrokernel_yieldto only tensor operations.A tensor counterpart of
microkernel_fencecalledsync_tensorhas also been added which makes a result tensor of atensor.microkerneloperation available. It bufferizes tomicrokernel_fenceand its placement could be further optimized in the future. The conservative placement ofmicrokernel_fenceoperations was also removed fromspeicalize-dma-codeleading to less barriers andmicrokernel_fences.