Add unsafe aliased memory checking system#1079
Conversation
|
/build |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
There was a problem hiding this comment.
Greptile Overview
Greptile Summary
This PR introduces an opt-in compile-time system to detect unsafe memory aliasing between input and output tensors in MatX transform operations. When MATX_EN_UNSAFE_ALIAS_DETECTION is enabled, the framework checks whether operations that permute indices (like fftshift, transpose, reverse, matmul) write to the same memory they read from, throwing exceptions to prevent race conditions. The implementation adds a can_alias() trait that allows inherently safe operations (FFT, element-wise ops) to bypass checks, while dangerous patterns like (a = fftshift1D(a)).run() now throw matxInvalidParameter. The detection system propagates aliasing metadata through the operator tree via the new ALIASED_MEMORY capability, checking pointer ranges at runtime only when the flag is set. Core changes include: adding AliasedMemoryQueryInput to track memory ranges and permutation flags; implementing get_capability<ALIASED_MEMORY> across 25+ operator files to mark which operations permute indices; injecting aliasing checks in BaseOp::run() before transform execution; and adding comprehensive documentation in docs_input/basics/debug.rst.
Important Files Changed
| Filename | Score | Overview |
|---|---|---|
include/matx/core/capabilities.h |
/5 | Content not provided in context; likely defines OperatorCapability::ALIASED_MEMORY enum and related infrastructure |
include/matx/core/operator_options.h |
5/5 | Adds AliasedMemoryQueryInput struct to track aliasing metadata (permutation flag, prerun phase, memory range pointers) |
include/matx/core/type_utils_both.h |
2/5 | Implements can_alias() trait detection but has inverted logic: docstring says "without trait" while code checks FOR trait presence |
include/matx/core/tensor_impl.h |
3/5 | Implements ALIASED_MEMORY capability in tensor_impl_t but has potential underflow bug when Size(dim) == 0, and partial-overlap logic may be too aggressive |
include/matx/operators/base_operator.h |
3/5 | Injects aliasing checks into BaseOp::run() but has size-0 underflow risk in get_memory_range() and complex branching that needs careful validation |
include/matx/generators/logspace.h |
1/5 | Duplicate function definition for operator()(index_t idx) will cause compilation failure |
include/matx/operators/fftshift.h |
1/5 | Only FFTShift1DOp received ALIASED_MEMORY handling; FFTShift2DOp, IFFTShift1DOp, IFFTShift2DOp missing implementation |
test/00_operators/aliased_memory_test.cu |
2/5 | Test has inverted assertions: OverlappingSlices and NestedTransformAliasing comments contradict EXPECT_THROW/EXPECT_NO_THROW |
include/matx/operators/transpose.h |
4/5 | License header typo ('COpBRIGHT'); otherwise correctly implements ALIASED_MEMORY capability |
include/matx/operators/find_peaks.h |
3/5 | Multiple license header corruptions ('sum', 'COpBRIGHT', 'SHsum') alongside correct capability implementation |
include/matx/operators/conv.h |
4/5 | License typo ('COpBRIGHT'); removes duplicate get_capability_proc() correctly |
include/matx/operators/fft.h |
5/5 | Adds can_alias trait and ALIASED_MEMORY handler that returns false to allow safe in-place FFT operations |
include/matx/operators/chol.h |
4/5 | License typo ('COpBRIGHT'); correctly adds can_alias trait for cuSOLVER in-place support |
include/matx/operators/matmul.h |
3.5/5 | Sets permutes_input_output=true but matmul doesn't actually permute indices in the transpose sense—may be overly conservative |
CMakeLists.txt |
5/5 | Cleanly adds MATX_EN_UNSAFE_ALIAS_DETECTION build option following existing patterns |
docs_input/basics/debug.rst |
4/5 | Comprehensive documentation with typos ('unitialized'); correctly explains imperfect detection and false positive scenarios |
include/matx/operators/{reverse,hermitian,planar,interleaved,r2c,shift,stack,toeplitz}.h |
5/5 | All correctly implement ALIASED_MEMORY by setting permutes_input_output=true |
include/matx/operators/{argmax,argminmax,set}.h |
4/5 | Properly propagate capability checks; set.h has significant refactoring alongside aliasing additions |
include/matx/generators/{linspace,diag,alternate}.h & include/matx/operators/{index,interp}.h |
5/5 | Clean refactoring from get_capability_proc() to unified get_capability(InType&) signature |
include/matx/transforms/svd/svd_cuda.h |
5/5 | Defensive initialization of i_dspace to prevent undefined behavior |
include/matx/operators/binary_operators.h |
5/5 | Removes trailing whitespace—pure formatting cleanup |
test/00_operators/CMakeLists.txt |
5/5 | Adds aliased_memory_test.cu to build in correct alphabetical position |
Confidence score: 2/5
- This PR requires very careful review before merging due to multiple correctness issues that will cause immediate failures and potential runtime bugs
- Score reflects critical compilation errors (logspace.h duplicate function), incomplete operator coverage (fftshift variants), inverted test assertions (aliased_memory_test.cu), semantic confusion in trait detection (can_alias() docstring vs implementation mismatch), potential underflow bugs in memory range calculation (base_operator.h, tensor_impl.h when size=0), and widespread license header corruption across multiple files
- Pay close attention to
include/matx/generators/logspace.h(will not compile),include/matx/operators/fftshift.h(missing ALIASED_MEMORY for 3 of 4 operators),test/00_operators/aliased_memory_test.cu(assertions contradicting comments at lines 129-133 and 230-235),include/matx/core/type_utils_both.h(clarify intended trait semantics), and all files with license header typos (transpose.h,conv.h,chol.h,find_peaks.h)
Sequence Diagram
sequenceDiagram
participant User
participant Executor
participant SetOp
participant AliasChecker
participant Operator
participant CUDADevice
User->>Executor: call run(set_op)
Executor->>SetOp: Exec()
alt MATX_EN_UNSAFE_ALIAS_DETECTION enabled
SetOp->>AliasChecker: check_aliased_memory(lhs, rhs)
AliasChecker->>SetOp: get_memory_range(lhs)
Note over AliasChecker: Get first & last element pointers
AliasChecker->>Operator: get_capability<ALIASED_MEMORY>(rhs, mem_range)
Operator->>Operator: Check if memory overlaps
alt Memory overlaps detected
Operator-->>AliasChecker: true (aliased)
AliasChecker-->>SetOp: throw matxInvalidParameter
SetOp-->>User: Error: Aliased memory detected
else No overlap or safe aliasing
Operator-->>AliasChecker: false (not aliased)
AliasChecker-->>SetOp: continue
end
end
SetOp->>Operator: PreRun()
SetOp->>CUDADevice: Execute kernel
SetOp->>Operator: PostRun()
SetOp-->>User: Operation complete
Additional Comments (9)
-
include/matx/operators/transpose.h, line 21-24 (link)syntax: typo: 'COpBRIGHT' should be 'COPYRIGHT'
-
include/matx/generators/logspace.h, line 102-105 (link)logic: duplicate operator() method - this exact signature already exists at lines 73-88 with the same implementation
-
include/matx/operators/find_peaks.h, line 5 (link)syntax: typo: 'sum rights reserved' should be 'All rights reserved'
-
include/matx/operators/find_peaks.h, line 21 (link)syntax: typo: 'COpBRIGHT' should be 'COPYRIGHT'
-
include/matx/operators/find_peaks.h, line 24 (link)syntax: typo: 'SHsum' should be 'SHALL', and 'COpBRIGHT' should be 'COPYRIGHT'
-
include/matx/operators/conv.h, line 21 (link)syntax: Typo:
COpBRIGHTshould beCOPYRIGHT -
include/matx/operators/chol.h, line 21-24 (link)syntax: typo: 'COpBRIGHT' should be 'COPYRIGHT'
-
include/matx/operators/fftshift.h, line 251 (link)syntax: Parameter
tshould be passed by const reference, not by value -
include/matx/operators/fftshift.h, line 343 (link)syntax: Parameter
tshould be passed by const reference
35 files reviewed, 23 comments
There was a problem hiding this comment.
Greptile Overview
Greptile Summary
This review covers only the changes made since the last review, not the entire PR. The latest commit addresses a critical terminology error in the documentation. The text on line 67 of docs_input/basics/debug.rst incorrectly states the system has a "false negative rate of 0" when the surrounding context (lines 66, 79) and the system's actual behavior demonstrate it produces false positives, not false negatives. The unsafe aliasing detection system is explicitly described as "imperfect" (line 65) and line 79 provides an example labeled "False positive" where non-unity strides trigger incorrect warnings. The phrase "false negative rate of 0" contradicts this—it would mean the system never misses unsafe aliasing (which isn't claimed), when what's actually meant is the system avoids false negatives at the cost of false positives. This is a documentation-only change that corrects the technical description of the aliasing detection trade-offs to accurately reflect that the system errs on the side of over-reporting potential aliasing issues.
Important Files Changed
| Filename | Score | Overview |
|---|---|---|
| docs_input/basics/debug.rst | 3/5 | Corrected uninitialized spelling errors and added aliasing detection documentation with terminology confusion |
Confidence score: 3/5
- This documentation-only change is low-risk but contains a critical terminology error that could confuse users about the system's behavior
- Score reflects that line 67's "false negative rate of 0" statement directly contradicts the documented false positive behavior in the examples and system design philosophy
- The terminology error in docs_input/basics/debug.rst line 67 needs immediate correction before merge to avoid user confusion about detection trade-offs
1 file reviewed, 3 comments
cliffburdick
force-pushed
the
aliased_input_output
branch
from
5c3321f to
6f34044
Compare
|
/build |
There was a problem hiding this comment.
Greptile Overview
Greptile Summary
This PR implements a comprehensive unsafe aliased memory detection system for MatX to catch race conditions when transform operations write to the same tensor they read from while permuting indices. The system introduces a new ALIASED_MEMORY capability that propagates through the operator tree using an AliasedMemoryQueryInput structure containing memory range and permutation flags. Operators that reorder elements (fftshift, reverse, transpose, matmul, etc.) set permutes_input_output=true to signal unsafe aliasing when input/output memory overlaps. Element-wise operations and operators like FFT that support in-place execution can define a can_alias trait to opt out. The detection is opt-in via the MATX_EN_UNSAFE_ALIAS_DETECTION CMake flag to avoid performance overhead. The tensor_impl class computes memory footprints and checks for overlaps, distinguishing between complete overlaps (safe for contiguous non-permuting operations) and partial overlaps (always unsafe). Comprehensive unit tests validate the system catches unsafe patterns like (a = fftshift1D(a)).run() while allowing safe operations like (a = a + a).run().
Important Files Changed
| Filename | Score | Overview |
|---|---|---|
| include/matx/core/capabilities.h | 4/5 | Adds ALIASED_MEMORY and ELEMENT_WISE enum values to OperatorCapability with capability_attributes specialization and OR_QUERY mapping |
| include/matx/core/operator_options.h | 5/5 | Defines AliasedMemoryQueryInput struct with permutes_input_output flag and memory range pointers for aliasing checks |
| include/matx/core/type_utils_both.h | 4.5/5 | Adds can_alias() trait detection function to identify operators that explicitly allow input/output memory aliasing |
| include/matx/core/tensor_impl.h | 4/5 | Implements ALIASED_MEMORY capability check in tensor_impl by computing memory footprint and detecting overlaps |
| include/matx/operators/base_operator.h | 0/5 | File included in PR but no meaningful changes detected in the diff provided |
| include/matx/operators/matmul.h | 4/5 | Adds ALIASED_MEMORY capability handling marking both matrix operands as permuting input/output indices |
| include/matx/operators/transpose.h | 4/5 | Implements ALIASED_MEMORY check marking transpose operations as index-permuting |
| include/matx/operators/reverse.h | 5/5 | Adds aliasing detection by setting permutes_input_output=true for reverse operations |
| include/matx/operators/fftshift.h | 0/5 | File included in PR but no detailed changes provided in diff |
| include/matx/operators/shift.h | 4/5 | Adds ALIASED_MEMORY capability marking shift operations as index-permuting |
| include/matx/operators/hermitian.h | 3/5 | Implements aliasing detection for hermitian operator which swaps and conjugates dimensions |
| include/matx/operators/stack.h | 4/5 | Adds ALIASED_MEMORY handling marking stack operations as permuting via dimension insertion |
| include/matx/operators/interleaved.h | 4/5 | Marks interleaved layout conversion as index-permuting for aliasing detection |
| include/matx/operators/planar.h | 3/5 | Adds aliasing detection to planar layout conversion though classification as permuting may need validation |
| include/matx/operators/r2c.h | 4/5 | Marks R2C reconstruction operations as index-permuting for unsafe aliasing detection |
| include/matx/operators/fft.h | 2/5 | Adds can_alias trait to FFT operations and returns false for ALIASED_MEMORY but has incomplete implementation (only non-cuFFTDx branch, FFT2Op missing trait) |
| include/matx/operators/chol.h | 3/5 | Adds can_alias trait to Cholesky decomposition but uses type-alias instead of function pattern |
| include/matx/operators/argmax.h | 4/5 | Adds get_capability method to propagate capability queries through input operator chain |
| include/matx/operators/argminmax.h | 4/5 | Implements capability propagation for argminmax reduction operator |
| include/matx/operators/find_peaks.h | 2/5 | Updates capability signature but contains critical license header corruption typos |
| include/matx/operators/set.h | 0/5 | File included in PR but no meaningful changes detected |
| include/matx/operators/conv.h | 0/5 | File included in PR but no meaningful changes detected |
| include/matx/operators/toeplitz.h | 0/5 | File included in PR but no meaningful changes detected |
| include/matx/operators/interp.h | 0/5 | File included in PR but no meaningful changes detected |
| include/matx/operators/binary_operators.h | 5/5 | Trivial whitespace cleanup removing trailing spaces |
| include/matx/generators/diag.h | 5/5 | Updates get_capability signature to accept InType parameter for aliasing infrastructure |
| include/matx/generators/logspace.h | 0/5 | Refactoring appears incomplete with conflicting operator() definitions causing compilation ambiguity |
| include/matx/generators/linspace.h | 4.5/5 | Removes duplicate get_capability_proc method retaining standardized get_capability interface |
| include/matx/generators/alternate.h | 4/5 | Renames get_capability_proc to get_capability with InType parameter for aliasing system |
| include/matx/generators/index.h | 5/5 | Removes redundant get_capability_proc method leaving unified get_capability interface |
| include/matx/transforms/svd/svd_cuda.h | 5/5 | Defensive initialization of i_dspace variable to prevent undefined behavior |
| CMakeLists.txt | 5/5 | Adds MATX_EN_UNSAFE_ALIAS_DETECTION CMake option defaulting to OFF with corresponding compile definition |
| test/00_operators/CMakeLists.txt | 5/5 | Registers new aliased_memory_test.cu test file in alphabetical order |
| test/00_operators/aliased_memory_test.cu | 3/5 | Comprehensive aliasing detection tests with contradictory comments and test naming issues |
| docs_input/basics/debug.rst | 4.5/5 | Adds documentation for unsafe aliased memory checking system with examples and limitations |
Confidence score: 2/5
- This PR requires careful review before merging due to incomplete implementations, conflicting code patterns, and critical bugs that will prevent compilation
- Score reduced due to: 1) logspace.h has duplicate conflicting operator() definitions causing compilation failure, 2) FFT aliasing support incomplete (only non-cuFFTDx branch, FFT2Op missing can_alias trait), 3) find_peaks.h has corrupted license header, 4) test file has contradictory comments/expectations suggesting unclear specification, 5) multiple operators missing from aliasing detection (fftshift 2D variants per previous review), 6) tensor_impl overlap logic may be incorrect for non-contiguous tensors (previous review concern unresolved)
- Pay close attention to include/matx/generators/logspace.h (compilation failure), include/matx/operators/fft.h (incomplete implementation), include/matx/operators/find_peaks.h (license corruption), test/00_operators/aliased_memory_test.cu (spec ambiguity), and include/matx/core/tensor_impl.h (overlap logic validation needed)
Sequence Diagram
sequenceDiagram
participant User
participant BaseOp
participant SetOp
participant get_memory_range
participant check_aliased_memory
participant Operator
participant get_capability
User->>BaseOp: (a = op).run(exec)
BaseOp->>SetOp: Exec()
alt is_matx_set_op with transform and tensor_view
SetOp->>check_aliased_memory: check_aliased_memory(lhs, rhs, false)
check_aliased_memory->>get_memory_range: get_memory_range(lhs, is_prerun)
get_memory_range->>get_memory_range: Get first element address
get_memory_range->>get_memory_range: Get last element address
get_memory_range-->>check_aliased_memory: AliasedMemoryQueryInput
check_aliased_memory->>Operator: get_operator_capability<ALIASED_MEMORY>(rhs, mem_range)
Operator->>get_capability: get_capability<ALIASED_MEMORY>(in)
get_capability->>get_capability: Check if memory overlaps
alt Memory overlaps
get_capability-->>Operator: true
Operator-->>check_aliased_memory: true
check_aliased_memory->>BaseOp: MATX_THROW(matxInvalidParameter)
BaseOp-->>User: Exception
else No overlap
get_capability-->>Operator: false
Operator-->>check_aliased_memory: false
check_aliased_memory-->>SetOp: continue
SetOp->>SetOp: TransformExec()
SetOp-->>BaseOp: Success
BaseOp-->>User: Success
end
else Other set operations
SetOp->>check_aliased_memory: check_aliased_memory(lhs, rhs, true)
check_aliased_memory->>get_memory_range: get_memory_range(lhs, is_prerun)
get_memory_range-->>check_aliased_memory: AliasedMemoryQueryInput
check_aliased_memory->>Operator: get_operator_capability<ALIASED_MEMORY>(rhs, mem_range)
Operator->>get_capability: get_capability<ALIASED_MEMORY>(in)
get_capability-->>Operator: result
Operator-->>check_aliased_memory: result
alt Memory aliases
check_aliased_memory->>BaseOp: MATX_THROW(matxInvalidParameter)
BaseOp-->>User: Exception
else No aliasing
check_aliased_memory-->>SetOp: continue
SetOp->>SetOp: PreRun()
SetOp->>SetOp: ex.Exec()
SetOp->>SetOp: PostRun()
SetOp-->>BaseOp: Success
BaseOp-->>User: Success
end
end
Additional Comments (9)
-
include/matx/generators/logspace.h, line 73-88 (link)logic: This templated operator() is never invoked - line 104 non-templated operator() calls itself with DefaultCapabilities, but the templated version at line 74 expects a CapType parameter that is never used from the calling site. Should the non-templated operator() at line 102-105 call the templated version at line 74, or should the templated version be removed if it's not part of the capability system API?
-
include/matx/generators/logspace.h, line 102-105 (link)logic: Duplicate operator() function - both line 74 and line 102 define operator()(index_t idx) with different signatures. This will cause compilation ambiguity
-
include/matx/operators/fft.h, line 688 (link)logic: FFT2Op is missing the
can_aliastrait that FFTOp has on line 85. This inconsistency means FFT2 operations won't participate in the aliasing detection system properly. -
include/matx/operators/conv.h, line 21 (link)syntax: typo: 'COpBRIGHT' should be 'COPYRIGHT'
-
include/matx/operators/conv.h, line 24 (link)syntax: typo: 'COpBRIGHT' should be 'COPYRIGHT'
-
include/matx/generators/linspace.h, line 71-80 (link)style: get_capability removed get_capability_proc template parameter, but still takes unused InType parameter. This parameter isn't used anywhere in the function body.
-
include/matx/operators/find_peaks.h, line 5 (link)syntax: typo: 'sum' should be 'All'
-
include/matx/operators/find_peaks.h, line 21 (link)syntax: typo: 'COpBRIGHT' should be 'COPYRIGHT'
-
include/matx/operators/find_peaks.h, line 24 (link)syntax: typo: 'SHsum' should be 'SHALL', and 'COpBRIGHT' should be 'COPYRIGHT'
35 files reviewed, 20 comments
Implement detection for unsafe memory aliasing between input and output tensors in transform operations. Enabled via MATX_EN_UNSAFE_ALIAS_DETECTION flag. Includes can_alias() trait function and matmul aliasing unit test.
cliffburdick
force-pushed
the
aliased_input_output
branch
from
6f34044 to
503c666
Compare
|
/build |
There was a problem hiding this comment.
Greptile Overview
Greptile Summary
This PR implements unsafe memory aliasing detection for MatX transform operations, addressing issue #1025. The system detects when input and output tensors share memory and the operation permutes indices (creating race conditions).
Key Implementation Details:
- Capability System Integration: Adds
ALIASED_MEMORYcapability to the existing operator capability framework with OR_QUERY aggregation - Trait-Based Detection: Introduces
can_alias()trait function to identify transform operators that support aliasing checks - Memory Range Comparison: Computes first/last element addresses to detect overlapping memory regions between LHS and RHS
- Operator Marking: Transform operators that permute indices (fftshift, shift, reverse, transpose, matmul) set
permutes_input_output=trueflag - Compile-Time Flag: Feature is opt-in via
MATX_EN_UNSAFE_ALIAS_DETECTIONCMake flag
Major Changes:
- Core infrastructure:
type_utils_both.h,capabilities.h,tensor_impl.h,base_operator.h - Operator implementations: 20+ operators updated with aliasing capability checks
- Comprehensive test suite: 11 test cases covering safe/unsafe aliasing scenarios
- Documentation: User guide explaining the feature and its trade-offs
Issues Found:
Most issues have already been identified in previous comments. The implementation is sound overall, with a few areas needing attention around edge cases and consistency.
Confidence Score: 4/5
- This PR is safe to merge with minor fixes needed for test comments and documentation typos
- The core implementation is architecturally sound and well-integrated into the existing capability system. The feature is opt-in via compile flag, minimizing risk. However, several issues prevent a score of 5: (1) test comments contradict actual expectations in aliased_memory_test.cu:129-130 and :230-232, (2) multiple documentation typos, (3) inconsistent aliasing capability implementation across similar operators (e.g., FFTShift1D has it but IFFTShift variants don't), and (4) some edge case logic needs verification (non-contiguous tensor aliasing, zero-size dimensions). These are quality-of-life issues rather than critical bugs, but should be addressed before merge.
- test/00_operators/aliased_memory_test.cu (fix contradictory comments), docs_input/basics/debug.rst (fix typos and terminology), include/matx/operators/fftshift.h (add aliasing to IFFT variants for consistency)
Important Files Changed
File Analysis
| Filename | Score | Overview |
|---|---|---|
| include/matx/core/type_utils_both.h | 4/5 | Adds can_alias() trait function to detect transform ops with aliasing capability. Implementation correctly checks for both transform op type and can_alias trait presence. |
| include/matx/core/capabilities.h | 5/5 | Adds ALIASED_MEMORY capability enum and associated query infrastructure. Clean implementation with proper OR_QUERY aggregation logic. |
| include/matx/core/tensor_impl.h | 4/5 | Implements memory range detection for tensors using first/last element addresses. Logic handles complete vs partial overlaps, though complete overlap logic for non-contiguous tensors needs verification. |
| include/matx/operators/base_operator.h | 4/5 | Core aliasing detection logic with check_aliased_memory() and check_mtie_aliased_memory() functions. Properly queries operator tree for aliasing using capability system. |
| test/00_operators/aliased_memory_test.cu | 3/5 | Comprehensive test suite with positive and negative test cases. However, comments on lines 129-130 and 230-232 contradict the actual test expectations (already reported). |
| docs_input/basics/debug.rst | 3/5 | Adds documentation for aliasing detection feature. Contains typos ('unitialized' instead of 'uninitialized') and some terminology issues around false positives/negatives (already reported). |
Sequence Diagram
sequenceDiagram
participant User
participant SetOp as SetOp/Transform
participant BaseOp as BaseOp::check_aliased_memory
participant LHS as LHS Tensor
participant RHS as RHS Operator Tree
participant TensorImpl as TensorImpl::get_capability
User->>SetOp: (a = fftshift1D(a)).run()
SetOp->>BaseOp: check_aliased_memory(a, fftshift1D(a))
Note over BaseOp: If MATX_EN_UNSAFE_ALIAS_DETECTION enabled
BaseOp->>LHS: get_memory_range(a)
LHS-->>BaseOp: {start_ptr, end_ptr}
BaseOp->>RHS: get_operator_capability<ALIASED_MEMORY>(fftshift1D(a), mem_range)
RHS->>RHS: FFTShift1DOp::get_capability()
Note over RHS: Sets permutes_input_output=true
RHS->>TensorImpl: Recurse to child op (tensor a)
TensorImpl->>TensorImpl: Check if tensor memory overlaps
Note over TensorImpl: Compare start/end pointers<br/>Check complete vs partial overlap
TensorImpl-->>RHS: true (aliasing detected)
RHS-->>BaseOp: true (propagated via OR_QUERY)
alt Aliasing Detected
BaseOp->>BaseOp: throw matxInvalidParameter
BaseOp-->>User: Exception: "Aliased memory detected"
else No Aliasing
BaseOp-->>SetOp: Continue execution
SetOp->>SetOp: Execute transform
SetOp-->>User: Success
end
35 files reviewed, no comments
1 participant
Closes #1025
Implement detection for unsafe memory aliasing between input and output tensors in transform operations. Enabled via MATX_EN_UNSAFE_ALIAS_DETECTION flag. Includes can_alias() trait function and matmul aliasing unit test.