0.25.7

Added

• futhark autotune now supports hip backend.

• Better parallelisation of scatter when the target is
  multidimensional (#2035).

Fixed

• Very large iotas now work.

• Lambda lifting in while conditions (#2038).

• Size expressions in local function parameters had an interesting
  interaction with defunctionalisation (#2040).

• The store command in server executables did not properly
  synchronise when storing opaque values, which would lead to
  use-after-free errors.

0.25.6

Added

• The various C API functions that accept strings now perform a copy,
  meaning the caller does not have to keep the strings alive.

• Slightly better lexer error messages.

• Fusion across slicing is now possible in some cases.

• New tool: futhark profile.

Fixed

• Inefficient locking for certain segmented histograms (#2024).

0.25.5

Added

• futhark repl now has a :format command. Work by Dominic
  Kennedy.

Fixed

• Textual floating-point numbers printed by executables now always
  print enough digits to not hide information. Binary output is
  unchanged.

• Invalid CSE on constants could crash the compiler (#2021).

0.25.4

Fixed

• Invalid simplification (#2015).

• Rarely occurring deadlock for fused map-scan compositions in CUDA
  backend, when a bounds check failed in the map function.

• Compiler and interpreter crash for tricky interactions of abstract
  types and sizes (#2016). Solved by banning such uses - in principle
  this could break code.

• Incomplete alias tracking could cause removal of necessary copies,
  leading to compiler crash (#2018).

0.25.3

Added

• pyopencl backend: compatibility with future versions of PyOpenCL.

• New backend: hip.

Fixed

• Exotic problems related to intra-group reductions with array
  operands. (Very rare in real code, although sometimes generated by
  AD.)

• Interpreter issue related to sizes in modules (#1992, #1993, #2002).

• Incorrect insertion of size arguments in in complex cases (#1998).

• Incorrect handling of match in lambda lifting (#2000).

• Regression in checking of consumption (#2007).

• Error in type checking of horisontally fused scatters could crash
  the compiler (#2009).

• Size-polymorphic value bindings with existential sizes are now
  rejected by type checker (#1993).

• Single pass scans with complicated fused map functions were
  insufficiently memory-expanded (#2023).

• Invalid short circuiting (#2013).

0.25.2

Added

• Flattening/unflattening as the final operation in an entry point no
  longer forces a copy.

• The opencl backend no longer always fails on platforms that do
  not support 64-bit integer atomics, although it will still fail if
  the program needs them.

• Various performance improvements to the compiler itself;
  particularly the frontend. It should be moderately faster.

Fixed

• Code generation for f16 literals in CUDA backend (#1979).

• Branches that return arrays differing in sign of their stride
  (#1984).

0.25.1

Added

• Arbitrary expressions of type i64 are now allowed as sizes. Work
  by Lubin Bailly.

• New prelude function resize.

Removed

• The prelude functions concat_to and flatten_to. They are often
  not necessary now, and otherwise resize is available.

Changed

• The prelude functions flatten and unflatten (and their
  multidimensional variants), as well as split, now have more
  precise types.

• Local and anonymous (lambda) functions that must return unique
  results (because they are passed to a higher order function that
  requires this) must now have an explicit return type ascription that
  declares this, using *. This is very rare (in practice
  unobserved) in real programs.

Fixed

• futhark doc produced some invalid links.

• flatten did not properly check for claimed negative array sizes.

• Type checker crash on some ill-typed programs (#1926).

• Some soundness bugs in memory short circuiting (#1927, #1930).

• Another compiler crash in block tiling (#1933, #1940).

• Global arrays with size parameters no longer have aliases.

• futhark eval no longer crashes on ambiguously typed expressions (#1946).

• A code motion pass was ignorant of consumption constraints, leading
  to compiler crash (#1947).

• Type checker could get confused and think unknown sizes were
  available when they really weren't (#1950).

• Some index optimisations removed certificates (#1952).

• GPU backends can now transpose arrays whose size does not fit in a
  32-bit integer (#1953).

• Bug in alias checking for the core language type checker (#1949).
  Actually (finally) a proper fix of #803.

• Defunctionalisation duplicates less code (#1968).

0.24.3

Fixed

• Certain cases of noninlined functions in multicore backend.

• Defunctionalisation of match where the constructors carry
  functions (#1917).

• Shape coercions involving sum types (#1918). This required
  tightening the rules a little bit, so some coercions involving
  anonymous sizes may now be rejected. Add expected sizes as needed.

• Defunctionalisation somtimes forgot about sizes bound at top level
  (#1920).

0.24.2

Added

• futhark literate (and FutharkScript in general) is now able to do
  a bit of type-coercion of constants.

Fixed

• Accumulators (produced by AD) had defective code generation for
  intra-group GPU kernel versions. (#1895)

• The header file generated by --library contained a prototype for
  an undefined function. (#1896)

• Crashing bug in LSP caused by stray undefined (#1907).

• Missing check for anonymous sizes in type abbreviations (#1903).

• Defunctionalisation crashed on projection of holes.

• Loop optimisation would sometimes remove holes.

• A potential barrier divergence for certain GPU kernels that fail
  bounds checking.

• A potential infinite loop when looking up aliases (#1915).

• futhark literate: less extraneous whitespace.

0.24.1

Added

• The manifest file now lists which tuning parameters are relevant for
  each entry point. (#1884)

• A command tuning_params has been added to the server protocol.

Changed

• If part of a function parameter is marked as consuming ("unique"),
  the entire parameter is now marked as consuming.

Fixed

• A somewhat obscure simplification rule could mess up use of memory.

• Corner case optimisation for mapping over iota (#1874).

• AD for certain combinations of map and indexing (#1878).

• Barrier divergence in generated code in some exotic cases (#1883).

• Handling of higher-order holes (#1885).