Changelog • stablehlo

stablehlo (development version)

New hlo_optimization_barrier().
hlo_return() accepts a token, e.g. to hand one on from a while body, and two token ValueTypes compare equal.
Bugfix: format_double() rounds an f32 constant to single precision once, instead of rounding a 9-digit decimal of it a second time.

A Shape is now represented as an integer.
shape.Shape was removed.
The tensor generics and DataType now come from xlamisc, which absorbed tengen; stablehlo no longer depends on tengen.

CustomOpBackendConfig() now accepts ConstantAttr items, so a custom call can carry array-valued attributes (what an XLA FFI handler decodes as Span<const T>) and not just scalars, booleans and strings.
hlo_custom_call() gained an output_operand_aliases argument, built with the new OutputOperandAlias(). XLA then hands the handler the same buffer for the aliased operand and result, which is what lets an in-place kernel avoid a copy.

Added hlo_convolution()
hlo_dot_general() gained a precision_config argument, which is either NULL or one or two of "DEFAULT", "HIGH" and "HIGHEST".

Adopted tengen’s enum-style DataType. The BooleanType(), FloatType(), IntegerType() and UIntegerType() constructors were removed, use dtype() / as_dtype() instead.
The package now requires R >= 4.4.0.

The package was optimized w.r.t. runtime performance. This was achieved by reducing the number of classes that are used internally when creating Funcs. The hlo_<*> user API remains unaffected.
The hlo_* builders of common ops gained an output_types argument. When the output types are known ahead of time (e.g. from a lowering that ran type inference at trace time), passing them skips redundant inference and its input validation.

emit width-correct hex for f64 NaN/Inf constants
hlo_reduce_window() now validates the shape of padding against its declared type instead of its R representation.

Added support for CHLO ops, a higher-level companion dialect to stableHLO that is lowered to stableHLO during compilation. New ops:
- Inverse trig: hlo_acos(), hlo_asin(), hlo_atan().
- Hyperbolic: hlo_cosh(), hlo_sinh(), hlo_acosh(), hlo_asinh(), hlo_atanh().
- Gamma family: hlo_lgamma(), hlo_digamma(), hlo_polygamma().
- Error / Bessel / misc: hlo_erf(), hlo_erfc(), hlo_erf_inv(), hlo_bessel_i1e(), hlo_square().
- Float predicates: hlo_is_inf(), hlo_is_pos_inf(), hlo_is_neg_inf().
- Selection: hlo_top_k() returning the top-k values and their indices along the last dimension.
OpName() and new_Op() gain a dialect argument (default "stablehlo") to support ops from other MLIR dialects.