GitLab

* Init support for sm120

* fmt

* resolve comments

* unify mma gemm

* fmt

---------
Co-authored-by: LeiWang1999 <leiwang1999@outlook.com>

[Enhancement] Add eviction policy support for TMA operations, enhance CUDA codegen, and introduce new pass config (#690)

* Enhance TMA and barrier handling in CUDA code generation

- Updated `CodeGenTileLangCUDA` to support eviction policies for TMA operations, allowing for more flexible memory management.
- Introduced a new `CacheHintSm90` enum to define eviction strategies in `copy_sm90.h`.
- Modified TMA load/store functions to accept eviction policies, improving performance on different architectures.
- Enhanced `TmaBarrierCollector` and `TmaBarrierRewriter` to account for SIMT copies, ensuring correct barrier insertion.
- Refactored thread synchronization logic to utilize barrier IDs, improving the efficiency of partial thread synchronization.
- Updated Python interface for `copy` and `c2d_im2col` to include optional eviction policy parameters, enhancing usability.

* update shuffle and elect optimization

* fix bug

* fix bug

* fix potential bug

* lint fix

* lint fix

* update shuffle_elect template

* fix bug

* fix bug

* fix template

* lint and fix

* fix typo

* gemm_with_stride sm89

* fix offset issue

* bug fix

* format

* sm80 support

* add sm90

* add testing

* format

* add static_assert for wgmma

* Enhance error message for inner_box_dim validation in LowerBulkCopy

* lint fix

---------
Co-authored-by: LeiWang1999 <leiwang1999@outlook.com>

* use more efficient bf16 type related conversion

* update macro

* [Enhancement] Improve memory access condition checks in GlobalMemChecker

- Updated the condition checks in the GlobalMemChecker to utilize symbolic bounds in the CanProve method, enhancing the accuracy of memory access validations.
- This change ensures that both upper and lower bound conditions are evaluated with improved proof strength, contributing to more robust memory access analysis.

* lintfix

* [Enhancement] Add legality checks for shared memory and global range in LowerBulkCopy

- Implemented checks to ensure that the shared memory range and global range are legal during the bulk copy operation.
- Added assertions to validate that the extents of global and shared ranges match, improving the robustness of memory access validation in the LowerBulkCopy function.

* [Refactor] Update barrier and clear operations in warp specialization examples

- Replaced `mbarrier_wait_parity` and `mbarrier_arrive` with `barrier_wait` and `barrier_arrive` for improved clarity and consistency in synchronization.
- Adjusted the order of `clear` operations for local fragments in `example_warp_specialize_gemm_copy_1_gemm_0` to enhance parallel execution efficiency.

* [Enhancement] Implement thread partial synchronization and improve shared memory allocation handling

- Added support for thread partial barrier synchronization in CUDA, allowing for more flexible thread management.
- Enhanced the `MergeSharedMemoryAllocations` function to accept alignment bytes, improving memory allocation efficiency based on target requirements.
- Updated the `Lower` methods in `Copy` and `Fill` classes to include conditional predicates for thread execution, ensuring better control over thread behavior.
- Refactored the `print` function to include warp group and warp IDs for more detailed debugging output.
- Improved the handling of dynamic shared memory allocations in the `LowerAndLegalize` function to align with target-specific requirements.

* [Enhancement] Add support for disabling TMA in Copy operations

- Introduced a new `disable_tma` parameter in the `Copy` class to control thread memory access behavior.
- Updated the `Lower` method to conditionally execute bulk copy operations based on the `disable_tma` flag.
- Enhanced the `copy` function to accept the `disable_tma` argument, allowing for more flexible memory copy operations.
- Improved handling of `coalesced_width` to ensure it defaults to -1 when not provided, enhancing robustness in memory operations.

* [Refactor] Clean up whitespace and formatting in multiple files

- Removed unnecessary blank lines and adjusted line breaks for improved code readability in `example_mla_decode.py`, `example_warp_specialize_gemm_copy_gemm_0_1.py`, `phase.py`, and `copy.py`.
- Ensured consistent formatting across functions to enhance maintainability and clarity of the codebase.

* [Enhancement] Refactor flash attention implementation for improved performance and configurability

- Split the shared memory allocations for query and key-value pairs to optimize memory usage.
- Introduced command-line arguments for batch size, number of heads, and dimensions, enhancing flexibility in running the example.
- Updated kernel execution parameters to improve thread management and synchronization.
- Enhanced the overall structure of the flash attention function for better readability and maintainability.

* fix

* Update layout inference in ParallelOp to account for thread bounds; remove debug print in OptimizeForTarget

* Refactor barrier handling and update example configurations

- Replaced commented-out barrier creation with new barrier allocation in GEMM example.
- Updated kernel configuration in warp specialization example to include async copy settings.
- Enhanced barrier management in the phase optimization process to improve synchronization handling.
- Introduced new barrier allocation function for better memory management in shared contexts.

* Refactor barrier handling in LowerAndLegalize and OptimizeForTarget

- Reintroduced barrier lowering in OptimizeForTarget to enhance synchronization.
- Removed commented-out barrier lowering in LowerAndLegalize for cleaner code.
- Added exit() call in OptimizeForTarget to halt execution after barrier lowering.

* Enhance CMake configuration and clean up example scripts

- Enabled compile command export in CMakeLists.txt for better build integration.
- Removed unnecessary print statement in the warp specialization example.
- Cleaned up commented-out code in GEMM example for improved readability.
- Updated barrier handling in shared memory allocation transformations for better synchronization.

* Refactor barrier handling in warp specialization examples

- Replaced commented-out mbarrier code with new barrier allocation using T.alloc_barrier for improved synchronization.
- Updated barrier wait and arrive calls to align with the new allocation method across multiple example scripts.
- Enhanced code readability by removing unnecessary comments and ensuring consistent barrier management.

* Update lower_shared_barrier.cc

* Update phase.py

* Update warp specialization example and Cython wrapper

- Removed commented-out pass configuration options in the warp specialization example for clarity.
- Added functionality to write the generated kernel source to a file named "kernel.cu".
- Enhanced Cython wrapper to support boolean type conversion for improved type handling.

* Add storage synchronization call in shared barrier transformation

- Introduced a new evaluation statement to call the TVM storage sync function with "shared" as an argument, enhancing synchronization in the shared barrier handling process.

* remove debug files

* Remove kernel source output to file in warp specialization example

* remove comments

* Refactor tensor handling and update test execution in TileLang

- Changed `Buffer` to `Tensor` in `customize.py` for better type consistency.
- Updated `mbarrier_wait_parity` and `mbarrier_arrive` functions in `builtin.py` to use `tir.BufferLoad` instead of `BufferLoad`.
- Commented out the main testing function in `test_tilelang_language_reshape.py` and replaced it with a direct call to `run_reshape_smem` for streamlined testing.
- Removed unnecessary NVCC compiler flags in `libgen.py` to reduce verbosity.

* Update test_tilelang_language_reshape.py

* [Enhancement] Add argument simplification option to StmtSimplifier

- Introduced a new `simplify_arguments` flag in the `StmtSimplifier::Apply` method to control argument simplification behavior.
- Updated the `Simplify` function to accept the new flag, allowing for enhanced flexibility in the simplification process.
- Adjusted the `LowerAndLegalize` and `_Simplify` functions to utilize the new argument, ensuring consistent behavior across the codebase.
- Added comments to clarify the purpose of the new flag and its impact on simplification logic.

* lint fix

* [Enhancement] Improve layout inference and reduce operation handling

- Updated `ParallelOp::InferLayout` to check for pure buffer stores, enhancing layout inference logic.
- Modified `ReduceOp::Lower` to include all threads in the AllReduce operation, improving performance on specific architectures.
- Added a TODO comment in `AllReduce` to consider merging synchronization barriers for optimization.

* lint fix

* [Enhancement] Add input validation for GEMM parameters

- Introduced checks to ensure that the dimensions M and N are divisible by their respective warp sizes (kMPerWarp and kNPerWarp) in the Gemm::ComputeWarpPartition method.
- Added informative error messages to assist in debugging when the input parameters do not meet the required conditions.

* bug fix

* [experimental] add a draft gemm_sp

* [3rdparty] bump cutlass to v3.9.3

* [lint] run format.sh

* [chore] rebase

* [chore] use abs path

* [gemm_sp] add metadata layout

* [ci] add more example

* [lint] run format.sh

* [chore] polish

* [chore] move gemm_sp to experimental

* [chore] polish

* [lint] run format.sh

* [Enhancement] Improve bulk copy handling and update GEMM sparse tensor test

* Added a warning log for unsupported non-swizzled global layouts in the bulk copy operation, ensuring fallback to normal copy.
* Refactored the GEMM sparse tensor test by removing unnecessary imports and simplifying the kernel compilation process.
* Updated the test to directly call the `run_gemm_sp` function, enhancing clarity and functionality.

* Implement Test

* [Enhancement] Update GEMM SP and SM89 templates for improved functionality

* Refactored GEMM SP computation to enhance warp partitioning logic, ensuring compatibility with Hopper architecture.
* Updated layout inference to support new WGMMA conditions and improved error messaging for unsupported targets.
* Modified SM89 templates to utilize new MMA atom structures, enhancing performance and compatibility with fp8 types.
* Added conditional inclusion for GEMM SP header based on CUDA architecture version.

* lint fix

* [gemm_sp] support more layout and data types

* Enhancement: sync T.gemm_sp's layout inference with T.gemm

* Enhancement: support more block_k in compress util

* [Enhancement] enable block_k=64

* [Lint] run format.sh

* [Enhancement] compressor support more dtype

* Enhancement: enable block_K=32

* [Lint] format.sh

* [Fixbug] fix shape

* Refactor: sync gemm

* [Enhancement] enable transpose

* [Enhancement] enable fp8_e4m3

* [Enhancement] enable int8

* [Lint] run format.sh

* [Benchmark] add gemm_sp benchmark

* [Example] fix 256 threads hang

* [CI] fix ci

* [Chore] resolve gemini feedback

* [Benchmark] increase search space

* [Lint] format

* [CI] skip sparse tensor core related tests as only sm90 is supported

* [CI] pass local run

* Update gemm_sm89.h

* lint fix

* lint fix

* [Enhancement] Add support for sparse GEMM and initialize CUDA architecture flags

- Introduced a new boolean flag `enable_sparse_gemm_` to control the inclusion of sparse GEMM functionality in CUDA code generation.
- Updated the `Finish` method to conditionally include the sparse GEMM header based on the new flag.
- Implemented logic in `VisitStmt_` to enable sparse GEMM when the corresponding external call is detected.
- Added a function to initialize the `TORCH_CUDA_ARCH_LIST` environment variable based on the target compute version, enhancing compatibility with PyTorch.
- Refactored the initialization function into the appropriate module and ensured it is called in the sparse utilities module.

* Update test_compress_utils.py

---------
Co-authored-by: LeiWang1999 <leiwang1999@outlook.com>
Co-authored-by: Lei Wang <34334180+LeiWang1999@users.noreply.github.com>

- Replaced the use of `tiled_mma.accumulate_ = GMMA::ScaleOut::Zero` with a call to `clear(acc)` for better clarity and maintainability in the accumulation logic.
- This change enhances the readability of the code by standardizing the approach to clearing accumulation values across multiple sections of the file.

* [Feature] Add Quarter Bank Swizzle Layout and Update GEMM Layout Logic

- Introduced a new `makeQuarterBankSwizzleLayout` function for layout swizzling of 32 bytes.
- Updated `makeGemmABLayout` to include an `enable_padding` parameter, allowing for conditional layout selection between padded and quarter bank swizzle layouts.
- Adjusted layout inference in GEMM operations to utilize the new quarter bank swizzle layout when appropriate.
- Enhanced bulk copy operations to recognize and handle the new layout type, improving memory access patterns.

* lint fix

* [Refactor] Update GEMM Layout Functions and Inference Logic

- Removed the `enable_padding` parameter from `makeGemmABLayout` to simplify its signature.
- Introduced `makeGemmABLayoutHopper` for enhanced layout handling specific to Hopper architecture.
- Updated layout inference in GEMM operations to utilize the new `makeGemmABLayoutHopper` function, improving clarity and maintainability in layout selection.
- Adjusted related layout functions to ensure consistent behavior across different architectures.

* [Refactor] Remove tf32 Casting Logic from GEMM Templates

- Eliminated the `cast_float_to_tf32` function from `gemm_sm80`, `gemm_sm89`, and `gemm_sm90` templates to streamline the code.
- Removed conditional casting logic for float32 to tfloat32 conversion, enhancing clarity and maintainability.
- Updated relevant sections in GEMM operations to reflect the removal of casting, ensuring consistent behavior across templates.
- Adjusted tensor view handling to improve performance and accuracy in matrix operations.

* Update bulk_copy.cc

* Fix profiler initialization in GEMM test by removing TensorSupplyType argument for improved flexibility.

* Add tf32 casting functionality to GEMM templates

- Introduced a `cast_float_to_tf32` function to convert float32 values to tfloat32 format across gemm_sm80, gemm_sm89, and gemm_sm90 templates.
- Implemented conditional casting in relevant sections of the GEMM operations to ensure compatibility with tfloat32 types.
- Enhanced the handling of tensor views to support the new casting logic, improving performance and accuracy in matrix operations.

* lint fix

* Refactor tfloat32 casting logic in GEMM templates

- Replaced the `is_tfloat32` boolean with `need_tfloat32_cast` to improve clarity and accuracy in determining when to cast float32 to tfloat32.
- Updated relevant sections in `gemm_sm80`, `gemm_sm89`, and `gemm_sm90` to utilize the new casting logic, enhancing compatibility with tfloat32 types.
- Ensured consistent application of casting across tensor views, improving performance and correctness in matrix operations.

* Refactor GEMM template functions for improved readability

- Simplified the function signature of `body_rs` in both `gemm_sm80` and `gemm_sm90` templates for better clarity.
- Adjusted the casting logic in `gemm_sm90` to ensure consistent application of `cast_float_to_tf32` across tensor views, enhancing performance and maintainability.

* Enhance tf32 casting logic in GEMM templates

- Updated the `cast_float_to_tf32` function in `gemm_sm80`, `gemm_sm89`, and `gemm_sm90` to conditionally apply the casting only if the input is finite, improving robustness.
- Simplified the `need_tfloat32_cast` logic to clarify the conditions under which tfloat32 casting is required, enhancing code readability and maintainability.

* Refactor GEMM template functions and layout inference logic

- Removed the `cast_float_to_tf32` function from `gemm_sm90` and updated the `body_sr` function to streamline the casting process for tensor views, enhancing code clarity and maintainability.
- Improved layout inference in `layout_inference.cc` by adding checks for the layout map's definition, ensuring robustness in handling layout annotations.
- Simplified the handling of layout maps in the `annotate_layout` function, allowing for more flexible layout definitions and error handling.

* [wip] feat: add nvrtc backend

* [wip] fix: handle out_idx

* [wip] refactor: move lib logic to libgen

* feat: cache for nvrtc backend

* fmt: run format

* fix: handle cuda bindings import error

* fix: handle cuda bindings import error

* fix: handle cuda bindings import error

* fix: handle cuda bindings import error

* fix: get kernel source

* refactor: speedup pyimport

* Improve error handling for missing cuda-python dependency in nvrtc backend. Raise ImportError with detailed installation instructions instead of logging a warning.

* Enhance nvrtc backend error handling by introducing a flag to check for cuda-python availability. Raise ImportError with detailed installation instructions during initialization if the nvrtc backend is unavailable, improving user experience and clarity.

* Update README.md to include recent NVRTC Backend addition, highlighting reduced compilation time for CUDA templates.

* fix tl_templates

* ensure CUDA context

---------
Co-authored-by: LeiWang1999 <leiwang1999@outlook.com>

* [Enhancement] Add atomic addition functions for FLOAT16x2 and FLOAT16x4 in CUDA

* Introduced `AtomicAddx2` and `AtomicAddx4` functions for performing atomic addition operations on double-width float types in CUDA.
* Updated `customize.py` to include the new `atomic_addx4` function for external calls.
* Modified `__init__.py` to export the new atomic addition function, ensuring accessibility in the module.

* lint fix

* [Refactor] Enhance GEMM Warp Partitioning Logic and Introduce Buffer Remapping (#516)

* Improved the warp partitioning logic in `Gemm::ComputeWarpPartition` to better accommodate various GEMM policies, including FullRow, FullCol, and Square, ensuring optimal performance based on matrix dimensions.
* Introduced a new `RemapBufferRewriter` class to handle buffer reference updates and padding annotations during statement transformations, enhancing memory access safety and clarity.
* Updated the `OptimizeForTarget` function to include a new step for configuring index bitwidth, improving the overall optimization process.
* Refactored existing code to utilize constants for warp sizes, enhancing maintainability and readability.
* Added checks to ensure correct warp allocation and padding map handling, improving robustness in memory management strategies.

* [Refactor] Update ConfigIndexBitwidthRewriter to Support Auto-Check Feature

* Modified the constructor of `ConfigIndexBitwidthRewriter` to include an `auto_check` parameter, allowing for dynamic bitwidth adjustments based on input conditions.
* Enhanced the `VisitExpr_` methods to apply the new auto-check logic, ensuring that integer types are upgraded to 64 bits when necessary, or to a specified index bitwidth otherwise.
* Updated the `ConfigIndexBitwidth` pass to determine the index bitwidth based on the presence of configuration, improving flexibility in handling different scenarios.

* Add dynamic matrix multiplication example and corresponding test

* Introduced `example_dynamic.py` to demonstrate dynamic matrix multiplication using TileLang and PyTorch, including a main function for execution and performance profiling.
* Added `test_example_dynamic.py` to validate the functionality of the dynamic matrix multiplication example.
* The example includes detailed parameter configurations and checks against PyTorch's implementation for correctness.

* lint fix

* Add get_num_sms function to retrieve the number of streaming multiprocessors on the CUDA device

* Implemented the `get_num_sms` function in `cuda_driver.py` to return the count of streaming multiprocessors for a specified CUDA device.
* Updated the `__init__.py` file to include the new function in the module exports.

* lint fix

* Add global barrier state and expectation handling in CUDA code generation

* Introduced `vid_global_barrier_state_` and `vid_global_barrier_expect_` to manage global barrier synchronization in the CUDA code generator.
* Updated `Finish` method to declare the global barrier state if needed.
* Implemented handling for `EvaluateNode` to initialize the barrier expectation.
* Removed unnecessary extern declaration for the global barrier state in `PrintStorageSync` method.
* Enhanced CUDA FP8 type definitions for better alignment and structure.

* Enhance CUDA FP8 type handling and debug printing

* Updated `cuda_fp8.h` to replace NVidia's FP8 types with Cute's FP8 types for better compatibility and structure.
* Added specializations for `debug_print_var` and `debug_print_buffer_value` functions to support the new FP8 types, improving debugging capabilities for these data types.
* Updated `debug.h` to include the new `cuda_fp8.h` header for access to the FP8 type definitions.

* Refactor CUDA code generation to remove unnecessary managed qualifier for global barrier state

* Updated the `Finish` method in `codegen_cuda.cc` to declare the global barrier state without the `__managed__` qualifier, simplifying the declaration.
* Added a new `sync_global` function in `builtin.py` to synchronize all threads in a block, enhancing synchronization capabilities in the TileLang framework.

* Remove deprecated CUDA kernel and Python script for FP8 E4M3 casting

* Deleted the `cast_to_fp8_e4m3_kernel` CUDA kernel implementation and its corresponding Python script, streamlining the codebase by removing unused components related to FP8 E4M3 type casting.
* This cleanup enhances maintainability and reduces potential confusion regarding obsolete code.

* lint fix

* [Refactor] Enhance GEMM Warp Partitioning Logic and Introduce Buffer Remapping (#516)

* Improved the warp partitioning logic in `Gemm::ComputeWarpPartition` to better accommodate various GEMM policies, including FullRow, FullCol, and Square, ensuring optimal performance based on matrix dimensions.
* Introduced a new `RemapBufferRewriter` class to handle buffer reference updates and padding annotations during statement transformations, enhancing memory access safety and clarity.
* Updated the `OptimizeForTarget` function to include a new step for configuring index bitwidth, improving the overall optimization process.
* Refactored existing code to utilize constants for warp sizes, enhancing maintainability and readability.
* Added checks to ensure correct warp allocation and padding map handling, improving robustness in memory management strategies.

* [Refactor] Update ConfigIndexBitwidthRewriter to Support Auto-Check Feature

* Modified the constructor of `ConfigIndexBitwidthRewriter` to include an `auto_check` parameter, allowing for dynamic bitwidth adjustments based on input conditions.
* Enhanced the `VisitExpr_` methods to apply the new auto-check logic, ensuring that integer types are upgraded to 64 bits when necessary, or to a specified index bitwidth otherwise.
* Updated the `ConfigIndexBitwidth` pass to determine the index bitwidth based on the presence of configuration, improving flexibility in handling different scenarios.

* Add dynamic matrix multiplication example and corresponding test

* Introduced `example_dynamic.py` to demonstrate dynamic matrix multiplication using TileLang and PyTorch, including a main function for execution and performance profiling.
* Added `test_example_dynamic.py` to validate the functionality of the dynamic matrix multiplication example.
* The example includes detailed parameter configurations and checks against PyTorch's implementation for correctness.

* lint fix

* Add get_num_sms function to retrieve the number of streaming multiprocessors on the CUDA device

* Implemented the `get_num_sms` function in `cuda_driver.py` to return the count of streaming multiprocessors for a specified CUDA device.
* Updated the `__init__.py` file to include the new function in the module exports.

* lint fix

* Add global barrier state and expectation handling in CUDA code generation

* Introduced `vid_global_barrier_state_` and `vid_global_barrier_expect_` to manage global barrier synchronization in the CUDA code generator.
* Updated `Finish` method to declare the global barrier state if needed.
* Implemented handling for `EvaluateNode` to initialize the barrier expectation.
* Removed unnecessary extern declaration for the global barrier state in `PrintStorageSync` method.
* Enhanced CUDA FP8 type definitions for better alignment and structure.

* [Refactor] Enhance GEMM warp partitioning logic for improved performance and flexibility

* Updated the warp partitioning logic in `Gemm::ComputeWarpPartition` to better handle various GEMM policies, including FullRow, FullCol, and Square.
* Implemented checks to dynamically adjust warp allocation based on matrix dimensions, ensuring optimal performance.
* Introduced a new `SelectCopy` template to streamline memory access patterns in CUDA templates, enhancing compatibility across different architectures.
* Refactored the Python `GemmWarpPolicy` class to align with the updated C++ logic, improving clarity and maintainability in warp allocation strategies.

* [Refactor] Optimize matrix multiplication parameters and performance in quickstart example

* Updated thread count in the kernel context from 256 to 128 to enhance performance.
* Increased block sizes for matrix dimensions (M, N, block_M, block_N) to 1024 and 128 respectively, improving computational efficiency.
* Adjusted the pipeline stages in the GEMM loop from 0 to 3 for better parallel execution.
* Cleaned up comments for clarity and corrected a typo in the memory copy comment.

* [Refactor] Simplify Copy type selection in OperandTraits for improved clarity

* Replaced the conditional Copy type definition with a new SelectCopy template in OperandTraits, enhancing readability and maintainability of the code.
* This change streamlines the logic for selecting memory copy patterns based on matrix dimensions and warp configurations.

* [Enhancement] Improve GEMM layout function and documentation

* Added detailed documentation for the makeGemmABLayout function, explaining parameters and layout selection strategies.
* Updated the layout selection logic to use mat_continuous consistently, enhancing clarity and correctness in memory layout calculations.
* Adjusted the InferLayout method to reflect changes in the layout function, ensuring accurate matrix dimension handling for transposed cases.

* lint fix

* [Refactor] Update GEMM layout and operand traits for improved CUDA compatibility

* Adjusted the InferLayout method in gemm.cc to include trans_A in fragment creation, enhancing layout inference for transposed matrices.
* Updated OperandTraits in gemm_sm89.h and gemm_sm90.h to change the Copy type from SM75_U16x4_LDSM_N to SM75_U16x4_LDSM_T, optimizing memory access patterns for different warp configurations.
* Enhanced static assertions in gemm_sm90.h to clarify requirements for num_warp_m, ensuring compatibility with Hopper architecture.

* [Refactor] Clean up formatting in GEMM implementation and CUDA templates

* Simplified the formatting of the fragment creation in the InferLayout method of gemm.cc for better readability.
* Adjusted the static assertion message in gemm_sm90.h to enhance clarity regarding the num_warp_m requirement for Hopper architecture.

* Remove debug print statement from block_sparse_attn_triton.py and implement a timeout handler in autotuner for function execution. This enhances the robustness of the autotuner by allowing it to handle timeouts gracefully.

* Enhance the autotuner module by adding a timeout handler for function execution, improving robustness in handling long-running tasks. This change includes the introduction of a custom TimeoutException and updates to the run_with_timeout function for better signal management.

* Add merge shared memory allocations pass and related configurations

- Introduced a new pass for merging shared memory allocations in GPU kernels, allowing for more efficient memory usage.
- Registered configuration options for debugging and controlling the merging behavior.
- Updated relevant files to integrate the new pass into the TileLang engine and transform modules.
- Adjusted import paths and added documentation for the new functionality.

* Reduce num_stages parameter in GEMM functions from 3 to 1 for improved performance in test_tilelang_kernel_gemm.py

* Update Copy type in OperandTraits for GEMM templates to use conditional selection based on num_warp_n. This change enhances memory access patterns for different configurations in CUDA kernels.

* lint fix

* Update Copy type in OperandTraits for GEMM templates to use SM75_U16x4_LDSM_T and SM75_U16x8_LDSM_T for improved memory access patterns across CUDA architectures.

* Remove debug print statement from block_sparse_attn_triton.py and implement a timeout handler in autotuner for function execution. This enhances the robustness of the autotuner by allowing it to handle timeouts gracefully.

* Enhance the autotuner module by adding a timeout handler for function execution, improving robustness in handling long-running tasks. This change includes the introduction of a custom TimeoutException and updates to the run_with_timeout function for better signal management.

* Add merge shared memory allocations pass and related configurations

- Introduced a new pass for merging shared memory allocations in GPU kernels, allowing for more efficient memory usage.
- Registered configuration options for debugging and controlling the merging behavior.
- Updated relevant files to integrate the new pass into the TileLang engine and transform modules.
- Adjusted import paths and added documentation for the new functionality.

* Reduce num_stages parameter in GEMM functions from 3 to 1 for improved performance in test_tilelang_kernel_gemm.py

* Update Copy type in OperandTraits for GEMM templates to use conditional selection based on num_warp_n. This change enhances memory access patterns for different configurations in CUDA kernels.

* lint fix

* [Refactor] Enhance TMA barrier validation and support for additional architectures (#463)

* Updated the TMA barrier validation in `inject_tma_barrier.cc` to check for non-empty `barrier_id_to_range_` before raising an error for missing `create_list_of_mbarrier`.
* Refactored architecture checks in `phase.py` to utilize a new constant `SUPPORTED_TMA_ARCHS`, allowing for easier updates and improved readability in the target architecture validation logic.

* [Feature] Implement fast integer power operation and related API

* Added a new math operation `tl.power_of_int` in `math.cc` for efficient integer exponentiation.
* Introduced a corresponding Python API `pow_of_int` in `tir/op.py` to facilitate usage in TileLang.
* Enhanced `common.h` with a template function for integer power calculations.
* Updated documentation to reflect the new functionality and usage examples.

* [Feature] Add TILELANG_CHECK_LAST_ERROR macro for improved error handling in CUDA and HIP

* Introduced TILELANG_CHECK_LAST_ERROR macro to streamline error checking for kernel launches in both CUDA and HIP.
* Updated kernel launch code in wrapper.py to utilize the new macro, enhancing readability and maintainability.
* This change improves error reporting by providing detailed messages when kernel execution fails.

* [Refactor] Standardize error message formatting in TILELANG_CHECK_LAST_ERROR macro

* Updated the TILELANG_CHECK_LAST_ERROR macro in both CUDA and HIP implementations to ensure consistent formatting of error messages.
* Enhanced readability by aligning the error message structure across different platforms, improving maintainability of error handling code.

* [Enhancement] Improve layout inference accuracy in ParallelOp (#441)

* Added logic to use non-replicated buffers as source buffers for more accurate layout inference.
* Enhanced comments to clarify the rationale behind buffer selection in layout inference process.

* [Enhancement] Add error handling macros and refactor loop partitioning logic

* Introduced TILELANG_CHECK macro for improved error handling in CUDA and HIP code, providing detailed error messages for kernel launches.
* Enhanced loop partitioning logic to handle fragment buffers more effectively, ensuring correct replication based on thread extent.
* Added logging for thread range in PlanLoopPartition to aid in debugging and performance analysis.
* Updated pass configuration management to streamline vectorization control in the optimization process.

* lint fix

* remove debug print

* [Enhancement] Improve error handling in layout inference and update profiler type in tests

* Added a detailed error message in the layout inference for local.fragment to clarify the requirement for trans_B.
* Updated the profiler type in the cumulative sum test from TensorSupplyType.One to TensorDistributionType.Randn for better profiling accuracy.

* lint fix

* [Refactor] Update OperandTraits to include num_warp_n parameter

* Modified OperandTraits templates across gemm_sm80.h, gemm_sm89.h, and gemm_sm90.h to include an additional num_warp_n parameter for improved flexibility in layout and copy operations.
* Adjusted Copy type selection based on the new parameter to enhance performance and adaptability in various scenarios.

* lint fix

* [Refactor] Update DispatchInstruction templates to include N parameter

* Modified DispatchInstruction templates in gemm_sm80.h, gemm_sm89.h, and gemm_sm90.h to include an additional N parameter, enhancing flexibility in tile size calculations.
* Adjusted MMA_Group definitions to use std::min for improved handling of warp sizes, ensuring better performance and adaptability in various scenarios.

* [Feature] Implement CumSum operation in TileLang

* Added CumSumOp class for cumulative sum operations, including argument validation and lowering logic.
* Introduced CumSum2D template for CUDA, supporting both forward and reverse cumulative sums.
* Created tests for CumSum functionality in shared memory and fragment contexts.
* Updated language interface to include cumsum operation, enhancing the reduction capabilities of TileLang.
* Refactored reduce.py to support cumsum functionality with appropriate memory allocation and copying mechanisms.

* lint fix

* [Enhancement] Introduce logical operations `any_of` and `all_of` for buffer checks

- Added new logical operations `any_of` and `all_of` to the TileLang language interface, allowing users to check conditions across buffer elements.
- Implemented corresponding intrinsic calls for CUDA, enhancing the functionality of the TileLang framework.
- Updated the `allocate.py` to handle boolean types correctly in shared memory allocations.
- Introduced tests for the new logical operations to ensure correctness and performance.
Co-authored-by: Zhiwen Mo <zhiwen.mo25@ic.ac.uk>

* lint fix

---------
Co-authored-by: Zhiwen Mo <zhiwen.mo25@ic.ac.uk>

* [bug] fix T.abs on float16

* [lint] lint

* [Enhancement] Add support for CUDA architecture 8.9 in GEMM template

- Introduced conditional inclusion of "gemm_sm89.h" for CUDA architectures 8.9 and above, enhancing compatibility with newer hardware.
- This change ensures that the GEMM template can leverage optimizations specific to the 8.9 architecture, improving performance for users with compatible GPUs.

* lintfix

* [Refactor] Clean up includes in gemm_sm89.h

- Removed duplicate inclusion of "common.h" and added "cuda_fp8.h" for improved clarity and organization.
- This change enhances the maintainability of the code by ensuring that header files are included only once and in a logical order.

- Added conditional compilation for BFLOAT16 atomic operations to ensure compatibility with CUDA architectures greater than 7.5.
- Improved code clarity by organizing the AtomicAdd functions and adding relevant comments for better understanding.

* [Refactor] Improve flash attention example and layout comparison logic

- Removed unnecessary annotation for `lse_local_split` in the flash attention example to streamline the code.
- Updated the handling of `lse_local_split` to utilize parallel processing for better performance.
- Refactored kernel compilation and profiling logic to enhance clarity and maintainability in the flash attention example.
- Added a condition in `FragmentNode::IsEqual` to handle broadcast cases, improving the robustness of layout comparisons.

* lint fix

* [Enhancement] Add support for shared memory scope in Fill operation

- Introduced handling for `shared.dyn` and `shared` memory scopes in the Fill operation.
- Implemented parallel operation and layout inference for improved performance in shared memory scenarios.
- Updated thread loop partitioning and vectorization logic to accommodate new memory scope handling.

* [Refactor] Remove deprecated decorator and enhance Cython kernel handling

- Removed the deprecated decorator from the main module and added a new implementation in the utils module for better organization.
- Introduced a pointer map in the Cython kernel adapter to manage pointer arguments, improving runtime shape resolution.
- Updated the Cython kernel wrapper to utilize the new pointer map for handling kernel arguments.
- Enhanced error checking in the tensor utility functions to ensure static shapes are enforced.
- Added a new proxy module for buffer and tensor handling, streamlining the interface for TIR programs.

* [Feature] Add matrix multiplication test and kernel implementation

- Introduced a new test file `test_tilelang_language_ptr.py` that implements a matrix multiplication function using TileLang's primitives.
- The `matmul_test` function defines a kernel for performing tile-level GEMM operations with customizable block sizes and data types.
- Added a `run_matmul` function to compile and execute the kernel, along with a test function to validate the implementation.
- Updated the `proxy.py` file to enhance type handling for buffer and tensor proxies, ensuring compatibility with TIR programs.
- Minor formatting improvements in `deprecated.py` for better readability.

* lint fix

* [Refactor] Update tensor creation in matrix multiplication test

- Replaced `T.Tensor.from_ptr` with `T.make_tensor` in `matmul_test` for improved clarity and consistency.
- Updated imports in `__init__.py` to include `make_tensor`.
- Added `make_tensor` function in `proxy.py` to streamline tensor creation from pointers.

* [Refactor] Update tensor definitions across multiple files

- Replaced instances of `T.Tensor` with updated tensor definitions in various benchmark and example files to enhance consistency and clarity.
- Adjusted tensor shapes and types in functions related to matrix multiplication, attention mechanisms, and other operations.
- Improved documentation in README and example files to reflect changes in tensor usage.

* lint fix

* [Refactor] Update tensor types in attention and matrix multiplication examples

- Replaced instances of `T.Tensor` with `T.SharedTensor` and `T.FragmentTensor` in various attention and matrix multiplication functions to improve consistency and clarity.
- Adjusted tensor definitions in benchmark and example files to align with the new tensor types.
- Enhanced the overall structure and readability of the code by standardizing tensor usage across multiple files.

* lint fix

* [Refactor] Update tensor types in GEMM example and test files

- Replaced instances of `T.Tensor` with `T.LocalTensor` and `T.Buffer` in the GEMM example and related test functions to improve consistency and clarity.
- Enhanced the overall structure of the code by standardizing tensor usage across multiple files, aligning with recent updates in tensor definitions.

* [Refactor] Update tensor usage in customize.py

- Replaced instances of `T.Tensor` with `T.Buffer` in the `reshape` and `view` functions to enhance consistency with recent tensor definitions.
- Improved code clarity by standardizing buffer usage across the file.

* [Refactor] Update tensor types in test_tilelang_transform_annotate_device_regions.py

- Replaced instances of `T.Tensor` with `T.Buffer` in the `before` and `expected` methods of the `TestAnnotateThreadExtent` and `TestAnnotateDeviceScope` classes to enhance consistency with recent tensor definitions.
- Improved code clarity by standardizing buffer usage across the test file.

* [Refactor] Update tensor types to SharedBuffer and FragmentBuffer

- Replaced instances of `T.SharedTensor` and `T.FragmentTensor` with `T.SharedBuffer` and `T.FragmentBuffer` across multiple benchmark, example, and test files to enhance consistency with recent tensor definitions.
- Improved code clarity and structure by standardizing buffer usage in attention and matrix multiplication functions.

* [Refactor] Introduce Tensor alias for Buffer in proxy.py

- Added a new alias `Tensor` for `Buffer` in `proxy.py` to facilitate JIT compilation, ensuring that inputs and outputs are mapped with `torch.Tensor`.
- This change enhances clarity and consistency in tensor usage across the codebase.

* [Refactor] Revamp cache management and enhance documentation in env.py and proxy.py

- Replaced global cache functions with a CacheState class to improve encapsulation and management of kernel caching.
- Updated the `from_ptr` method in BufferProxy and BaseTensorProxy classes to include detailed docstrings for better clarity on parameters and return values.
- Enhanced class docstrings across various proxy classes to provide clearer descriptions of their purpose and functionality, improving overall code documentation.

* [Refactor] Update imports in __init__.py for tir compatibility

- Added imports for `prim_func` and `tir.op` to enhance compatibility with the upstream tir script.
- Marked imports with `# noqa: F401` to suppress linting warnings for unused imports, indicating future removal once compatibility is achieved.

* lint fix

* [Refactor] Update imports in tir.ir.py for improved compatibility

- Removed unused import of `PrimExpr` from `tvm.script.ir_builder.tir` and replaced it with the correct import from `tvm.tir`.
- Added import for `tir.ir` in `__init__.py` to enhance module accessibility and maintain compatibility with upstream changes.

* [Refactor] Update function calls in tir.ir.py to return values

- Modified the `serial`, `parallel`, `vectorized`, `unroll`, `thread_binding`, and `grid` functions to return the results of their respective calls to `_ir` methods, enhancing clarity and ensuring proper value propagation.

* bugfix

* [Enhancement] Add support for uint16 data type in TLCUDASourceWrapper

- Introduced the "uint16" mapping to the type dictionary in the TLCUDASourceWrapper class, expanding the range of supported data types for CUDA operations.

* bugfix

* [Update] Sync subproject commit and modify CUDA atomic add functions

- Updated the subproject commit for TVM to edd35139a0481e9359aa269e3e50450b95ba2f5a.
- Commented out the CUDA capability check in the example convolution script to prevent execution errors.
- Refactored atomic add functions for BFLOAT16 in common.h to include a conditional compilation directive for improved compatibility with CUDA architectures.

* remove llvm build

* [Refactor] Update kernel compilation and profiling in examples

- Replaced `tilelang.lower` with `tilelang.compile` in multiple example scripts to streamline kernel compilation.
- Updated profiling calls to utilize the new `get_profiler` method, enhancing performance measurement consistency.
- Adjusted assertions and benchmarking methods to align with the new profiling structure across various examples, ensuring correctness and clarity in performance evaluations.

* lint fix

* License Update

* [Refactor] Improve code formatting and documentation in CUDA header and HIP runtime files

- Adjusted formatting in `cuda.h` for better readability, including alignment of comments and struct fields.
- Cleaned up whitespace and improved comment clarity in `rt_mod_hip.cc` to enhance code maintainability.

* [Refactor] Enhance formatting and clarity in CUDA header and HIP runtime files

- Improved comment alignment and readability in `cuda.h`.
- Cleaned up whitespace and formatting in `rt_mod_hip.cc` to enhance maintainability.

* lint fix

* lint fix

* lint fix

* lint fix

* fix

* License update

* [Enhancement] Update JITKernel to use artifact for kernel source

- Assigned the generated artifact to `self.artifact` for better management.
- Updated kernel source references to use `artifact.kernel_source` for consistency in execution backend handling.

* lint fix

* Add @tilelang.testing.requires_llvm decorator to vectorization tests

* Enhance setup.py and env.py for library management

- Added functionality to remove original files after copying in CMakeBuild.
- Updated TVM_LIBRARY_PATH in env.py to include the PyPI build library path for better integration.

* Refactor TVM_LIBRARY_PATH assignment for improved readability in env.py

* Refactor CMakeBuild file handling in setup.py

- Added a check to ensure the target library directory exists before copying .so files.
- Improved the logic for creating the target directory and copying files to enhance robustness.

* bugfix

* Rename BuildTLDebug to BuildTileLangCUDAWithoutCompile and update registration. Add @tilelang.testing.requires_llvm decorator to multiple tests for LLVM requirement.

* lint fix

* Enhance TileLang code generation by adding support for device code generation without compilation. Updated `host_codegen` and `device_codegen` functions to include new transformations and registration for `tilelang_hip_without_compile`. Refactored JIT kernel adapters to accommodate host and device modules, improving overall integration and flexibility.

* lint fix

* Add support for C target in device code generation

- Updated `device_codegen_without_compile` to include handling for the C target by registering the `tilelang_cpp` function.

* [Enhancement] Implement auto-clear cache feature based on environment variable

* Added TILELANG_CLEAR_CACHE environment variable to control cache clearing.
* Updated CI workflow to set TILELANG_CLEAR_CACHE during testing.
* Modified cache initialization to clear cache if TILELANG_CLEAR_CACHE is set to true.

* [Refactor] Update kernel invocation and import paths in tests and cache

* Changed kernel invocation in `test_tilelang_kernel_dequantize_gemm.py` to return the result.
* Updated import statements in `test_tilelang_kernel_int4_gemm_mma.py` to use `bitblas` instead of `tilelang`.
* Refactored paths for artifact and parameters in `kernel_cache.py` for better maintainability.

* [Refactor] Clean up whitespace and improve code formatting in kernel_cache.py

* Removed unnecessary blank lines and adjusted spacing for better readability in the KernelCache class.
* Enhanced overall code formatting to align with project standards.

* [Enhancement] Add bfloat16 test case and improve kernel caching logic

* Introduced a new test case for bfloat16 matrix multiplication in `test_tilelang_kernel_gemm_mma_intrinsic.py`.
* Updated `KernelCache` to handle multiple kernel source files and improve error handling during saving and loading.
* Refactored `JITKernel` to support instantiation from a database, enhancing flexibility in kernel management.
* Adjusted `CtypesKernelAdapter` and `CythonKernelAdapter` to utilize the new kernel loading mechanism from the database.
* Improved code formatting and readability across several files.

* lint fix

* Update bfloat16 matrix multiplication test case to use larger dimensions for improved coverage

* [Enhancement] Add zero initialization option to GEMM operations

- Introduced a new `zero_init` parameter to the GEMM function, allowing for optional zero initialization of the accumulator.
- Updated the GEMM implementation across various CUDA architectures to support the new parameter.
- Modified the Python interface for GEMM to include the `zero_init` argument, enhancing flexibility in kernel execution.
- Ensured compatibility with existing functionality while improving initialization control for performance optimization.

* rename zero_init to clear_accum

* lint

* Refactor GEMM and Bulk Copy operations to enhance layout handling and support for Hopper architecture

- Update `ComputeWarpPartition` to include a new parameter for Hopper WGMMA support.
- Modify layout checks in `LowerBulkCopy` to accommodate new GEMM layout types.
- Enhance layout inference logic in `InferLayout` for better compatibility with Hopper architecture.
- Include necessary header files for built-in operations and layout inference improvements.

* Refactor parameter formatting in CUDA matrix load functions for consistency

- Adjusted parameter alignment in `ptx_ldmatrix_x1`, `ptx_ldmatrix_x2`, `ptx_ldmatrix_x4`, and their transposed counterparts for improved readability.
- Added a blank line in `get_tensor_supply` function in `tensor.py` to enhance code clarity.

* Enhance tensor supply generation in `get_tensor_supply` function

- Introduced handling for unsigned integer and float8 tensor types, allowing for specific random tensor generation based on data type.
- Updated logic to return appropriate random tensors for different data types, improving flexibility and functionality of tensor supply generation.
- Refactored existing conditions for clarity and maintainability.

* Fix tensor supply generation logic in `get_tensor_supply` function

- Updated the variable reference from `tensor` to `param` to ensure correct handling of tensor data types.
- Improved the accuracy of unsigned integer and float8 checks for tensor supply generation, enhancing functionality and reliability.

* Enhance tensor supply checks in `get_tensor_supply` function

- Updated the logic for identifying unsigned integers and float8 types by using `removeprefix` on the dtype string, improving accuracy in tensor supply generation.
- Ensured better handling of tensor data types for more reliable random tensor generation based on the updated checks.

* Enhance KernelParam functionality and improve tensor supply checks

- Added methods `is_unsigned` and `is_float8` to the `KernelParam` class for better type identification of parameters.
- Updated the `get_tensor_supply` function to utilize the new methods, improving clarity and accuracy in tensor supply generation based on parameter types.

* add test for issue 101

* use ss_smem_selector from cutlass

* fix mismatch between smem layout and mma

* only fix for sm90

* Add CUDA requirements to GEMM thread tests

* lint fix

---------
Co-authored-by: Lei Wang <34334180+LeiWang1999@users.noreply.github.com>

* Enhance error message for constant size stack allocation in CUDA codegen. Include the actual constant size and buffer variable name in the error output for better debugging.

* Refactor GEMM and Bulk Copy operations to enhance layout handling and support for Hopper architecture

- Update `ComputeWarpPartition` to include a new parameter for Hopper WGMMA support.
- Modify layout checks in `LowerBulkCopy` to accommodate new GEMM layout types.
- Enhance layout inference logic in `InferLayout` for better compatibility with Hopper architecture.
- Include necessary header files for built-in operations and layout inference improvements.

* lint fix

* Remove unused builtin.h include directive

* Update include path for builtin.h

* upgrade cutlass to upstream v3.8.0

* Implement fp8 gemm and add example script

* Fix dtype retrieval with map_torch_type for fp8 inputs

* Disable vectorization of fp8 values

* Make MMA declaration compatible with cutlass 3.4.0+

* Add test for fp8 T.gemm

* fix indent

* fix indent

* Add copyright and license header

* Add copyright and license header

* lint fix

* Refactor matmul_nt and assert_matmul_correctness functions for improved readability by consolidating parameter definitions and adjusting formatting.

* clang format lint

---------
Co-authored-by: Lei Wang <34334180+LeiWang1999@users.noreply.github.com>
Co-authored-by: LeiWang1999 <leiwang1999@outlook.com>

- Introduce TMAStoreArrive and TMAStoreWait operations for CUDA TMA store synchronization
- Add new builtin operations in op/builtin.cc and op/builtin.h
- Implement TMAStoreSyncInjector to automatically inject TMA store synchronization calls
- Update CUDA codegen to support new TMA store synchronization intrinsics
- Add Python language bindings for new TMA store synchronization operations

* [Dev][Bugfix] Add RMS Normalization Kernels and Fix Reduce Bug

- Implement two RMS normalization implementations in TileLang:
  * `rms_norm_splitk`: Split-K reduction approach for large matrices
  * `rms_norm`: Full reduction kernel with simplified implementation
- Add reference implementation using PyTorch for validation
- Include performance benchmarking for both kernel variants
- Demonstrate flexible block size and matrix size configurations

* [Examples] Simplify RMS Normalization Kernel Compilation

- Remove commented-out code for split-K RMS normalization
- Simplify kernel compilation by removing explicit TMA lowering configuration
- Update copyright header to Tile-AI Corporation
- Streamline main script for RMS normalization example

* Fix debug print buffer template for unsigned char type

- Update debug_print_buffer_value template specialization for unsigned char
- Modify test_tilelang_debug_print.py to include additional dtype tests
- Add test case for uint8 dtype in debug print buffer function

* Refactor debug print buffer template formatting for unsigned char

- Improve code formatting for debug_print_buffer_value template specialization
- Adjust line breaks and indentation for better readability
- Maintain consistent code style with other template specializations

[Dev] Adjust computation logic to avoid precision loss when casting acc_s from float to float16 (#141)

- Remove redundant `acc_s_0` fragment in flash attention kernel
- Simplify memory copy and reduction operations
- Reorder memory copy and scaling steps for improved performance
- Add Hopper-specific synchronization method in CUDA reduce template
- Update reduce operation to use architecture-specific synchronization