[TIR][Schedule] Derive Nonnegative Bounds from Shape Var #15210
Merged
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
|
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 |
3 tasks
This PR enhance the arithmetic analysis used in compute-at to further help symbolic bound simplification. Previously, when a variable `n` appears in the shape of an input buffer `T.Buffer((n * 32), "float32")`, we could safely assume that `n` is nonnegative as it is part of the shape. This could help us simplify some bounds during scheduling as well as lowering. For example, for integers `n` and `bx` where `bx` has a symbolic bound `[0, 32 * n)`, if `n` is nonnegative, we could simplify the following expressions to True: ``` 0 <= floordiv(bx, n) < 32 0 <= floormod(bx, n) < n ``` This PR depends on apache#15193 to provide an interface that hints analyzer.
junrushao
force-pushed
the
feature/2023-07-02/compute-at-symbolic-bound
branch
from
1052628 to
17faf25
Compare
yzh119
approved these changes
Jul 3, 2023
tqchen
approved these changes
Jul 3, 2023
junrushao
added a commit
to junrushao/tvm
that referenced
this pull request
Jul 4, 2023
This PR enhances Decode-GEMV rule with the following changes: - Normalize the GEMV iter domain to S-R-C via transform-block-layout. This would help with further analysis and scheduling, in cases for example, when there was no spatial loop in the original reduction block. - Get rid of the ad hoc iter type analysis, including the logic calling into a TVM packed func `tir.schedule.GetLoopIterType` using `tvm._ffi.get_global_func`. - Split out the logic for two separate cases of scheduling, where the innermost dimension is spatial or reduction. - Introduces `suggest_threads_per_block` to guess the threads to be allocated each threadblock. This helps avoid the previous case where dlight allocates 256 threads for a workload whose degree of parallelism is only 128. - Misc improvements. This rest of the changes are split out to separate PRs that are already merged to main. - [x] Pass the hints to arithmetic analyzer that shape variables should be positive ones (apache#15210) - [x] Eliminate unnecessary block predicate generation - should be provable via affine analysis (apache#15193) - [x] Shrink local memory allocation if only one element `X[threadIdx.x]` is used (apache#15207)
junrushao
added a commit
to junrushao/tvm
that referenced
this pull request
Jul 4, 2023
This PR enhances Decode-GEMV rule with the following changes: - Normalize the GEMV iter domain to S-R-C via transform-block-layout. This would help with further analysis and scheduling, in cases for example, when there was no spatial loop in the original reduction block. - Get rid of the ad hoc iter type analysis, including the logic calling into a TVM packed func `tir.schedule.GetLoopIterType` using `tvm._ffi.get_global_func`. - Split out the logic for two separate cases of scheduling, where the innermost dimension is spatial or reduction. - Introduces `suggest_threads_per_block` to guess the threads to be allocated each threadblock. This helps avoid the previous case where dlight allocates 256 threads for a workload whose degree of parallelism is only 128. - Misc improvements. This rest of the changes are split out to separate PRs that are already merged to main. - [x] Pass the hints to arithmetic analyzer that shape variables should be positive ones (apache#15210) - [x] Eliminate unnecessary block predicate generation - should be provable via affine analysis (apache#15193) - [x] Shrink local memory allocation if only one element `X[threadIdx.x]` is used (apache#15207)
junrushao
added a commit
to junrushao/tvm
that referenced
this pull request
Jul 4, 2023
This PR enhances Decode-GEMV rule with the following changes: - Normalize the GEMV iter domain to S-R-C via transform-block-layout. This would help with further analysis and scheduling, in cases for example, when there was no spatial loop in the original reduction block. - Get rid of the ad hoc iter type analysis, including the logic calling into a TVM packed func `tir.schedule.GetLoopIterType` using `tvm._ffi.get_global_func`. - Split out the logic for two separate cases of scheduling, where the innermost dimension is spatial or reduction. - Introduces `suggest_threads_per_block` to guess the threads to be allocated each threadblock. This helps avoid the previous case where dlight allocates 256 threads for a workload whose degree of parallelism is only 128. - Misc improvements. This rest of the changes are split out to separate PRs that are already merged to main. - [x] Pass the hints to arithmetic analyzer that shape variables should be positive ones (apache#15210) - [x] Eliminate unnecessary block predicate generation - should be provable via affine analysis (apache#15193) - [x] Shrink local memory allocation if only one element `X[threadIdx.x]` is used (apache#15207)
junrushao
added a commit
to junrushao/tvm
that referenced
this pull request
Jul 4, 2023
This PR enhances Decode-GEMV rule with the following changes: - Normalize the GEMV iter domain to S-R-C via transform-block-layout. This would help with further analysis and scheduling, in cases for example, when there was no spatial loop in the original reduction block. - Get rid of the ad hoc iter type analysis, including the logic calling into a TVM packed func `tir.schedule.GetLoopIterType` using `tvm._ffi.get_global_func`. - Split out the logic for two separate cases of scheduling, where the innermost dimension is spatial or reduction. - Introduces `suggest_threads_per_block` to guess the threads to be allocated each threadblock. This helps avoid the previous case where dlight allocates 256 threads for a workload whose degree of parallelism is only 128. - Misc improvements. This rest of the changes are split out to separate PRs that are already merged to main. - [x] Pass the hints to arithmetic analyzer that shape variables should be positive ones (apache#15210) - [x] Eliminate unnecessary block predicate generation - should be provable via affine analysis (apache#15193) - [x] Shrink local memory allocation if only one element `X[threadIdx.x]` is used (apache#15207)
junrushao
added a commit
to junrushao/tvm
that referenced
this pull request
Jul 5, 2023
This PR enhances Decode-GEMV rule with the following changes: - Normalize the GEMV iter domain to S-R-C via transform-block-layout. This would help with further analysis and scheduling, in cases for example, when there was no spatial loop in the original reduction block. - Get rid of the ad hoc iter type analysis, including the logic calling into a TVM packed func `tir.schedule.GetLoopIterType` using `tvm._ffi.get_global_func`. - Split out the logic for two separate cases of scheduling, where the innermost dimension is spatial or reduction. - Introduces `suggest_threads_per_block` to guess the threads to be allocated each threadblock. This helps avoid the previous case where dlight allocates 256 threads for a workload whose degree of parallelism is only 128. - Misc improvements. This rest of the changes are split out to separate PRs that are already merged to main. - [x] Pass the hints to arithmetic analyzer that shape variables should be positive ones (apache#15210) - [x] Eliminate unnecessary block predicate generation - should be provable via affine analysis (apache#15193) - [x] Shrink local memory allocation if only one element `X[threadIdx.x]` is used (apache#15207)
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.
This PR enhance the arithmetic analysis used in compute-at to further help symbolic bound simplification.
Previously, when a variable
nappears in the shape of an input bufferT.Buffer((n * 32), "float32"), we could safely assume thatnis nonnegative as it is part of the shape. This could help us simplify some bounds during scheduling as well as lowering.For example, for integers
nandbxwherebxhas a symbolic bound[0, 32 * n), ifnis nonnegative, we could simplify the following expressions to True:This PR depends on #15193 to provide an interface that hints analyzer.