Changelog

Source: NEWS.md

stablehlo (development version)

Breaking changes

• A Shape is now represented as an integer.
• shape.Shape was removed.

Features

• Inference functions now refuse a dimension attribute that contains a missing value, instead of letting it surface as R’s missing value where TRUE/FALSE needed. GatherDimensionNumbers() and ScatterDimensionNumbers() check their dimension vectors the same way.

• infer_types_convolution() refuses a padding that takes away more than a spatial dimension holds. Such a shape made XLA’s own inference abort the process; negative padding that only empties a dimension stays legal. It also refuses a zero-sized kernel spatial dimension, which would otherwise infer a non-empty result from an empty window.

• Dimension attributes are checked before they are coerced, so a whole number outside the integer range (3e9, Inf) is reported as the value the caller passed instead of the NA as.integer() made of it. Affects hlo_convolution(), hlo_broadcast_in_dim(), hlo_dynamic_slice() and hlo_empty().

• 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.

Bug fixes

• hlo_pad() refused negative edge padding whose magnitude exceeded rank(operand) rather than the size of the dimension it applied to, so pad(tensor<10xf32>, low = -5) was rejected although its result is a tensor<5xf32>. Padding that genuinely empties a dimension is still refused, now naming the arguments involved.
• infer_types_concatenate() rejects inputs of different rank. Its (C2) check compared the shapes without the concatenation dimension, which a shorter shape passed, so concatenate(tensor<2x3x4>, tensor<2x3>, dimension = 2) inferred a result with an unknown size along that dimension.
• infer_types_if() rejects a branch that declares inputs.
• infer_types_while() checks its body’s inputs and not only its outputs.
• The short assembly form (%0 = stablehlo.<op> %a : <type>) is emitted only for ops that actually allow it.
• hlo_triangular_solve() now rejects operands that are not of floating-point type, as required by the StableHLO spec.
• infer_types_slice() rejects a stride of 0. The spec’s (C4) is 0 < strides, but the check read 0 <= strides.
• infer_types_dynamic_slice(), infer_types_dynamic_update_slice() and infer_types_gather() reject start_indices that are not of integer type, as the spec requires. A float one used to reach MLIR and come back as a raw parse error.
• Corrected some checks in the inference functions.
• Added some missing checks in the inference functions.
• infer_types_reduce(), infer_types_reduce_window(), infer_types_scatter() and infer_types_sort() check their region’s arguments. Only the region’s outputs were read, so a body with the wrong arity, a non-scalar argument, or the wrong element type passed inference – a one-argument sort comparator returning an f32 was accepted and rendered.
• infer_types_reduce() accepts a body that accumulates into a wider element type. (C6) is is_promotable(element_type(inputs[i]), Ei), not an equality, so summing an i8 into an i32 is legal; it used to be refused.

stablehlo 0.4.0

Features

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

Breaking changes

• 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.

Performance

• 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.

Bug fixes

• 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.

stablehlo 0.3.0

Features

• 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.

Bug fixes

• hlo_reduce_window() now works on rank-1 inputs.

stablehlo 0.2.0

Features

• Added support for Modules

Bug fixes

• Constants +-Inf/NaN of dtype f64 are now correctly created.
• Fixed assembly format for select op with all boolean (i1) types.

Miscellaneous

• Use a simpler StableHLO string format for improved readability

stablehlo 0.1.0

• Initial release