CHANGELOG.md

Release Notes

{#release-1-6-0} 1.6.0 (2026-08-24)

This release adds finer control over compilation and execution with static array strides, unchecked memory accesses, programmatic dependent launch, and portable TileIR bytecode export. It expands the supported Python subset with user-defined context managers, dataclass methods, break in non-static loops, and integer divmod. Autotuning, JAX interoperability, diagnostics, and floating-point behavior are also improved.

CTK 13.4 features

• Add Programmatic Dependent Launch support with {py:func}ct.grid_dependency_control_launch_dependents() <cuda.tile.grid_dependency_control_launch_dependents>, {py:func}ct.grid_dependency_control_wait() <cuda.tile.grid_dependency_control_wait>, and the programmatic_dependent_launch option for {py:func}ct.launch() <cuda.tile.launch>.
• Add a check_bounds option to {py:func}ct.load() <cuda.tile.load>, {py:func}ct.store() <cuda.tile.store>, and the {py:meth}TiledView.load() <cuda.tile.TiledView.load> and {py:meth}TiledView.store() <cuda.tile.TiledView.store> methods. It defaults to True. Setting it to False declares that every tile element is within the array bounds and skips bounds checking for faster access.
• Add {py:func}ct.insert() <cuda.tile.insert> and {py:meth}Tile.insert() <cuda.tile.Tile.insert> to create a new tile by replacing a subtile of a larger tile. This operation is the inverse of {py:func}ct.extract() <cuda.tile.extract>. The original tile remains unchanged.
• Add float8_e5m3fnu scale dtype support for FP4 block-scaled MMA on Rubin (sm_107).

Features

• Add support for specializing array stride dimensions to their launch-time values, making them compile-time constants inside the kernel. Use {py:class}ct.ArrayAnnotation <cuda.tile.ArrayAnnotation> as Annotated metadata and list the dimensions to specialize, for example, Annotated[ct.Array, ct.ArrayAnnotation(static_stride_dims=(0, 1))]. Once any static stride is specified, the dispatcher stops automatically inferring stride == 1 for all dimensions, so list the contiguous dimension explicitly if it should remain a compile-time constant.
• Add {py:func}ct.ensure_constant() <cuda.tile.ensure_constant>, which statically asserts that a value is a compile-time constant.
• Add IEEE rounding-mode support to floating-point astype() conversions.
• Add a propagate_nan option to {py:func}ct.min() <cuda.tile.min>, {py:func}ct.max() <cuda.tile.max>, {py:func}ct.argmin() <cuda.tile.argmin>, {py:func}ct.argmax() <cuda.tile.argmax>, {py:func}ct.minimum() <cuda.tile.minimum>, and {py:func}ct.maximum() <cuda.tile.maximum>. By default, NaN values are ignored. When enabled, a NaN propagates, min/max/minimum/maximum return NaN and argmin/argmax return the index of the first NaN.

Python Features

• Add support for break in non-static for loops.
• Add support for user-defined context managers created with @contextlib.contextmanager.
• Add support for user-defined dataclass methods: __post_init__(), __call__(), __getitem__(), __setitem__(), __repr__(), and __str__().
• Add {py:func}ct.divmod() <cuda.tile.divmod>, as well as support for built-in divmod(), for integer inputs only.
• Allow basic symbolic expressions to be constructed inside {py:func}ct.static_eval() <cuda.tile.static_eval>.

Enhancements

• An array dimension whose stride is inferred to be one is no longer encoded as a static stride in the array's type unless it is explicitly selected with ArrayAnnotation(static_stride_dims=...). This allows more permissive type unification in control flow.
• Allow {py:func}ct.tune.exhaustive_search() <cuda.tile.tune.exhaustive_search> to accept a function for its kernel argument that maps each configuration to the kernel to tune. Passing a fixed kernel continues to work as before.
• Allow {py:func}export_kernel() <cuda.tile.compilation.export_kernel> to export TileIR bytecode without specifying gpu_code when output_format="tileir_bytecode". This requires bytecode version 13.3 or later.
• Specialize the sliced axis produced by {py:meth}Array.slice() <cuda.tile.Array.slice> to a compile-time constant dimension when both start and stop are compile-time constants.
• On TileIR 13.4 or later, lower x ** y and ct.pow(x, y) to integer-exponent FPowI when x is an unrestricted float and y is an integer whose value range fits in a signed 32-bit integer. Other cases continue to promote the exponent to floating point and use FPowF.
• Improve the diagnostic for a function whose return type differs across control-flow paths.
• Change the second element of each entry in {py:attr}TuningResult.failures <cuda.tile.tune.TuningResult.failures> from an exception class name to the exception class itself.

Bug Fixes

• Make {py:func}ct.argmin() <cuda.tile.argmin> and {py:func}ct.argmax() <cuda.tile.argmax> ignore NaN values by default, consistent with ct.min() and ct.max().
• Fix {py:func}ct.launch() <cuda.tile.launch> compiling kernels for CUDA device 0 rather than for the device being launched on. Kernels are now compiled for the launch device's architecture. Devices with the same architecture can still share a compiled kernel.
• Fix an AttributeError when compiler crash dumps are enabled with CUDA_TILE_ENABLE_CRASH_DUMP=1.
• Fix an internal compiler error when a runtime loop conditionally updates a loop-carried scalar.

Tooling

• Add cutile-cache log, a command-line interface for inspecting compilation history. Cache metadata includes mangled kernel names, compiler versions, compilation dates and durations, and, with TileIR 13.4, tileiras compiler remarks.

Ecosystem

• Add support for tuple parameters in the JAX FFI integration.

{#release-1-5-0} 1.5.0 (2026-07-08)

This release adds finer control over kernel specialization by making selected array shape dimensions compile-time constants and declaring scalar divisibility assumptions. Autotuning is faster and can isolate hanging kernels. The supported Python subset now includes tuple comprehensions, tuple-valued kernel arguments, enums, dictionaries, and variadic keyword parameters.

Features

• Add {py:func}ct.assume_divisible_by(x, divisor) <cuda.tile.assume_divisible_by>, a compiler hint that declares an integer scalar to be divisible by a constant. The compiler propagates this fact through arithmetic, which can prove the alignment of derived indices and pointer offsets and enable wider memory operations.
• Add support for specializing array shape dimensions to their launch-time values, making them compile-time constants inside the kernel. Use {py:class}ct.ArrayAnnotation <cuda.tile.ArrayAnnotation> as Annotated metadata and list the dimensions to specialize, for example, Annotated[ct.Array, ct.ArrayAnnotation(static_shape_dims=(0, -1))].
• Add the single_run_timeout_sec argument to {py:func}ct.tune.exhaustive_search() <cuda.tile.tune.exhaustive_search> to prevent a hanging kernel from stalling the entire search.
• Add support for passing Python tuples as kernel arguments. Elements may be arrays, scalars, lists, or nested tuples. Annotations may apply to the entire tuple or individual elements: ct.Constant[tuple[int, float]] makes the whole tuple compile-time constant, while tuple[ct.Constant[int], float] makes only the first element constant.

Python features

• Add support for tuple comprehensions.
• Add support for the in and not in operators on tuples.
• Add limited support for dictionaries, variadic keyword parameters in user-defined functions (for example, def foo(**kwargs)), and dictionary unpacking (for example, foo(x, **y)).
• Add comparison and constructor support for Python enum.Enum inside kernels. Enum members can be compared with == and != and constructed from a constant value, such as Color(0).
• Add support for printing dataclass instances.
• Allow frozen dataclass instances to be used as globals in device code.

Bug Fixes

• Fix a bug in automatic propagation of divisibility when an assumed variable created in a block is used outside that block.
• Fix a bug where strictly typed numeric constants (for example, ct.uint32(-3), ct.int8(300), and ct.float16(0.2)) were not converted according to their dtype. Out-of-range integers are now wrapped to the dtype's range, and floating-point values are rounded or clamped according to the dtype's precision and range.
• Honor wrappers installed on user-defined functions, including wrappers added by decorators that use functools.wraps().
• Reject repeated axes in the order arguments of load(), store(), and num_tiles().

Enhancements

• Extend {py:func}ct.arange() <cuda.tile.arange> with optional start and step arguments: ct.arange(size, start=0, step=1, dtype=...). size must be a constant integer, while start and step may be dynamic values. For example, ct.arange(8, start=7, step=-1, dtype=ct.int32) creates a reversed range.
• Improve exhaustive search performance by stopping early for slow configurations.

ABI Changes

• Introduce {py:class}calling convention v2 <cuda.tile.compilation.CallingConvention> for kernels with static shape annotations or tuple arguments.

{#release-1-4-0} 1.4.0 (2026-05-26)

This release highlights compatibility with CTK 13.3 which introduces Hopper(sm_90) GPU support, block-scaled MMA, new float dtypes, pack/unpack operations, atomic store operations, load and store with advanced indexing, tiled view with gapped or overlapped tile access, and support for large arrays and scalars. It also adds support for Python 3.14 including free-threading, star "*" expression, and frozen dataclasses, as well as integration with JAX.

CTK 13.3 features

• Support Hopper (sm_90 family) GPUs.
• Add float8_e8m0fnu and float4_e2m1fn restricted float dtype.
• Add {py:func}ct.mma_scaled() <cuda.tile.mma_scaled> operation for block-scaled matrix multiply-accumulate.
• Add {py:func}ct.pack_to_bytes() <cuda.tile.pack_to_bytes> operation that flattens a tile and reinterprets its raw bytes as a 1D uint8 tile; {py:func}ct.unpack_from_bytes() <cuda.tile.unpack_from_bytes> is the inverse of {py:func}ct.pack_to_bytes() <cuda.tile.pack_to_bytes>.
• Add {py:func}ct.load_advanced_indexing() <cuda.tile.load_advanced_indexing> and {py:func}ct.store_advanced_indexing() <cuda.tile.store_advanced_indexing> for gathering/scattering along one dimension while slicing on other dimensions.
• Add {py:meth}atomic_store_add <cuda.tile.TiledView.atomic_store_add> and more atomic methods on {py:class}TiledView <cuda.tile.TiledView> for performing element-wise atomic read-modify-write operations on a tiled view at a given tile index.
• Add support for {py:meth}Array.tiled_view() <cuda.tile.Array.tiled_view> with traversal_steps for creating a tiled space with overlapped or spaced tiles.
• {py:class}ByTarget <cuda.tile.ByTarget> now accepts a default value that applies to all architectures (e.g. ByTarget(sm_100=8, default=2)), allowing generated TileIR bytecode to be independent of the GPU architecture.
• Add {py:func}ct.astile() <cuda.tile.astile> for creating a tile from a scalar (yielding a 0-d tile) or a (possibly nested) tuple of scalars whose nesting determines the tile's shape.
• Support large arrays (>2B elements):
  • New {py:data}ct.IndexedWithInt64 <cuda.tile.IndexedWithInt64> annotation for array kernel parameters whose shape or stride values exceed the range of a 32-bit integer. Arrays without the annotation continue to use int32 for shape and stride.
  • New {py:data}ct.ScalarInt64 <cuda.tile.ScalarInt64> annotation that forces a scalar integer kernel parameter to be inferred as int64 instead of the default int32.
• Add use_fast_acc option to {py:func}ct.mma() <cuda.tile.mma> to enable fast accumulation mode for FP8 inputs (float8_e4m3fn, float8_e5m2) on Hopper GPUs.
• Add a new kernel hint num_worker_warps to {py:class}ct.kernel <cuda.tile.kernel>.
• {py:func}ct.atomic_add() <cuda.tile.atomic_add> now supports bfloat16 operands on Hopper (sm_90) and newer architectures.
• Optional rounding_mode parameter for {py:func}ct.exp() <cuda.tile.exp> (supports RoundingMode.FULL and RoundingMode.APPROX for f32).

Python features

• Add Python 3.14 and Python 3.14t (free-threading) support.
• Add support for frozen dataclasses.
• Add support for passing a tuple as a starred function argument, e.g. foo(a, *b, c).
• Add support for variadic parameters in user-defined functions, e.g. def foo(*args).

Enhancements

• {py:func}ct.extract() <cuda.tile.extract> now raises a compile-time TileTypeError when a constant index is out of bounds for the tile grid. Dynamic indices are unaffected.
• {py:func}ct.floordiv() <cuda.tile.floordiv> and the // operator now support floating-point operands.

Ecosystem

• Add support for JAX via {py:func}ct.jax.cutile_call() <cuda.tile.jax.cutile_call>.

{#release-1-3-0} 1.3.0 (2026-04-20)

Features

• Add API for ahead-of-time compilation and export via {py:func}compilation.export_kernel() <cuda.tile.compilation.export_kernel>.
  See the {doc}Compilation and Export </compilation> section for more details.
• Add API for autotuning via {py:func}tune.exhaustive_search() <cuda.tile.tune.exhaustive_search> and the following helpers:
  • Add API {py:meth}kernel.replace_hints() <cuda.tile.kernel.replace_hints> to get a new kernel with updated hints.
  • Add API function {py:func}compiler_timeout() <cuda.tile.compiler_timeout> for temporarily setting the timeout on the tileiras compiler.
    See the {ref}Autotuning <autotuning> section for more details.
• Add API {py:meth}Array.tiled_view() <cuda.tile.Array.tiled_view> to create a tiled view of an array with a fixed tile shape and padding mode.

Enhancements

• Add support for specifying memory_order and memory_scope on cuda.tile.load and cuda.tile.store operations.
• Improve print() to handle tuple and nested fstring.

Bug Fixes

• Fix a bug where restricted float dtype with simple reduce and scan did not raise proper TileTypeError.

ABI Changes

• Change kernel ABI convention to omit parameters annotated with cuda.tile.Constant.

{#release-1-2-0} 1.2.0 (2026-03-05)

CTK 13.2 features

• Support Ampere and Ada (sm80 family) GPUs.
• Support pip install cuda-tile[tileiras] to use tileiras from Python environment without system-wide CTK installation.
• Add ct.atan2(y, x) operation for computing the arctangent of y/x.
• Add optional rounding_mode parameter for ct.tanh(), supporting RoundingMode.FULL and RoundingMode.APPROX.
• Compiling FP8 operations for sm80 family GPUs will raise TileUnsupportedFeatureError.
• Setting opt_level=0 on ct.kernel is no longer required for ct.printf() and ct.print().

Features

• Add ct.static_iter keyword that enables compile-time for loops.
• Add ct.static_assert keyword that can be used to assert that a condition is true at compile time.
• Add ct.static_eval keyword that enables compile-time evaluation using the host Python interpreter.
• Add ct.scan() for custom scan.
• Add ct.isnan().
• Add print() and ct.print() that supports python-style print and f-strings.
• Add optional mask parameter to ct.gather() and ct.scatter() for custom boolean masking.
• Operator + can now be used to concatenate tuples.
• Support unpacking nested tuples (e.g., a, (b, c) = t) and using square brackets for unpacking (e.g., [a, b] = 1, 2).
• Add bytecode-to-cubin disk cache to avoid recompilation of unchanged kernels. Controlled by CUDA_TILE_CACHE_DIR and CUDA_TILE_CACHE_SIZE.

Bug Fixes

• Fix a bug where nan != nan returns False.
• Fix "potentially undefined variable $retval" error when a helper function returns after a while loop that contains no early return.
• Fix the missing column indicator in error messages when the underlined text is only one character wide.
• Add a missing check for unpacking a tuple with too many values. For example, a, b = 1, 2, 3 now raises an error instead of silently discarding the extra value.
• Fix a bug where the promoted dtype of uint16 and uint64 was incorrectly set to uint32.

Enhancements

• Erase the distinction between scalars and zero-dimensional tiles. They are now completely interchangeable.
• ~x for const boolean x will raise a TypeError to prevent inconsistent results compared to ~x on a boolean Tile.
• Add TileUnsupportedFeatureError to the public API.

{#release-1-1-0} 1.1.0 (2026-01-30)

Features

• Add support for nested functions and lambdas.
• Add support for custom reduction via ct.reduce().
• Add Array.slice(axis, start, stop) to create a view of an array sliced along a single axis. The result shares memory with the original array (no data copy).

Bug Fixes

• Fix reductions with multiple axes specified in non-increasing order.
• Fix a bug when pattern matching (FusedMultiplyAdd) attempts to remove a value that is used by the new operation.

Enhancements

• Allow assignments with type annotations. Type annotations are ignored.
• Support constructors of built-in numeric types (bool, int, float), e.g., float('inf').
• Lift the ban on recursive helper function calls. Instead, add a limit on recursion depth. Add a new exception class TileRecursionError, thrown at compile time when the recursion limit is reached during function call inlining.
• Improve error messages for type mismatches in control flow statements.
• Relax type checking rules for variables that are assigned a different type depending on the branch taken: it is now only an error if the variable is used afterwards.
• Stricter rules for potentially-undefined variable detection: if a variable is first assigned inside a for loop, and then used after the loop, it is now an error because the loop may take zero iterations, resulting in a use of an undefined variable.
• Include a full cuTile traceback in error messages. Improve formatting of code locations; include function names, remove unnecessary characters to reduce line lengths.
• Delay the loading of CUDA driver until kernel launch.
• Expose the TileError base class in the public API.
• Add ct.abs() for completeness.

{#release-1-0-1} 1.0.1 (2025-12-18)

Bug Fixes

• Fix a bug in hash function that resulted in potential performance regression for kernels with many specializations.
• Fix a bug where an if statement within a loop can trigger an internal compiler error.
• Fix SliceType __eq__ comparison logic.

Enhancements

• Improve error message for ct.cat().
• Support is not None comparison.

{#release-1-0-0} 1.0.0 (2025-12-02)

Initial release.