Add scan (prefix sum) operations support #39
Conversation
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
This commit enables reduce_sum and reduce_max operations on all numeric types, extending beyond the previous float-only support. ## Infrastructure Changes ### Bytecode Layer - Added IntegerIdentityOp struct with signed/unsigned handling - Added encode_tagged_int! for integer identity encoding - Added mask_to_width function with zigzag encoding for signed types - Added encode_identity! dispatch for FloatIdentityOp and IntegerIdentityOp - Refactored ReduceIdentity → IdentityOp for extensibility ### Compiler Layer - Refactored emit_reduce! to use dispatch-based approach - Added operation_identity dispatch for add/max operations - Added encode_reduce_body dispatch for float and integer operations - Removed T <: AbstractFloat constraints from intrinsics ### Language Layer - Removed type constraints from reduce_sum and reduce_max in operations.jl ## Test Coverage ### Codegen Tests - Added FileCheck tests for Int32/UInt32 reduce_sum and reduce_max - Verifies correct IR generation (addi, maxi instructions) ### Execution Tests - Factory pattern for easy extension (makeReduceKernel, cpu_reduce) - Tests 10 types: Int8, Int16, Int32, Int64, UInt16, UInt32, UInt64, Float16, Float32, Float64 - Tests 2 operations: reduce_sum, reduce_max - CPU verification for all test cases - Type-appropriate input ranges to prevent overflow ## Files Changed - src/bytecode/encodings.jl: Fix IdentityOp type annotation - src/bytecode/writer.jl: Integer identity infrastructure - src/compiler/intrinsics.jl: Import identity types - src/compiler/intrinsics/core.jl: Dispatch-based reduce implementation - src/cuTile.jl: Export identity types - src/language/operations.jl: Remove type constraints - test/codegen.jl: Add integer reduction codegen tests - test/execution.jl: Add extendable execution tests ## Extensibility The infrastructure is designed for easy extension: - Add new reduce operations by defining operation_identity and encode_reduce_body methods - Add new types by adding to TEST_TYPES array and appropriate data generation
- Constrain T <: Number in reduce_sum/reduce_max signatures for type safety - Ensures only numeric types can be used with reduction operations
This commit adds support for scan (parallel prefix sum) operations to cuTile, based on the IntegerReduce branch and commit 0c9ab90. Key changes: - Added encode_ScanOp! to bytecode encodings for generating ScanOp bytecode - Added encode_scan_identity_array! to reuse existing identity encoding - Added scan intrinsic implementation using operation_identity from IntegerReduce - Added scan() and cumsum() public APIs with proper 1-indexed to 0-indexed axis conversion - Added comprehensive codegen tests for scan operations - Added scankernel.jl example demonstrating CSDL scan algorithm Features: - Supports cumulative sum (cumsum) for float and integer types - Supports both forward and reverse scan directions - Reuses FloatIdentityOp and IntegerIdentityOp from IntegerReduce - Uses operation_identity function for cleaner identity value creation - 1-indexed axis parameter (consistent with reduce operations) - Preserves tile shape (scan is an element-wise operation along one dimension) Tests: - All 142 codegen tests pass (including 6 new scan tests) - Scankernel.jl example runs successfully with CSDL algorithm - Clarify that it demonstrates device-side scan operation - Add note that test might occasionally fail (race condition in phase 2 loop) Minor comment improvements in scankernel.jl example - Clarify that it demonstrates device-side scan operation - Add note that test might occasionally fail (race condition in phase 2 loop)
Contributor
Author
|
This depends on #37 PR. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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 commit adds support for scan (parallel prefix sum) operations to cuTile,
based on the IntegerReduce branch and commit 0c9ab90.
Key changes:
Features:
Tests:
All 142 codegen tests pass (including 6 new scan tests)
Scankernel.jl example runs successfully with CSDL algorithm
Clarify that it demonstrates device-side scan operation
Add note that test might occasionally fail (race condition in phase 2 loop)
Minor comment improvements in scankernel.jl example