futhark

Added

• Memory short circuiting, a major new optimisation by Philip
  Munksgaard that avoids copies by constructing values in-place.

Fixed

• assert was sometimes optimised away by CSE.

• futhark literate now handles type abbreviations in entry points
  (#1750).

• Handling of non-inlined functions in GPU code. Still very restricted.

• Uniqueness checking bug (#1751).

• Simplification bug (#1753).

• A bug related to optimisation of scalar code migrated to GPU.

• Memory optimisation bug for top-level constants (#1755).

• Handling of holes in defunctionalisation (again).

• A few cases where optimisation (safely but perhaps confusingly)
  removed bounds checks (#1758).

• Futhark now works on Windows again (#1734). This support remains
  very flaky and not well tested.

• Type inference of field projection (#1762).

Fixed

• Non-server executables neglected to synchronise before printing results (#1731).

• Fixed handling of holes in defunctionalisation (#1738).

• Ascription of higher-order modules (#1741).

• Fixed compiler crash when attempting to tile irregular parallelism (#1739).

• Incorrect regularity checking in loop interchange (#1744).

• Pattern match exhaustiveness of bools (#1748).

• Improper consumption checking of nonlocal bindings (#1749).

Added

• futhark repl is now delightfully more colourful.

• futhark repl no longer prints scalar types with type suffixes (#1724).

• futhark pyopencl executables now accept --build-option.

• New functions: f16.nextafter, f32.nextafter, f64.nextafter,
  matching the ones in the C math library.

• futhark literate now prints directives in the output exactly as
  they appear in the source.

Fixed

• Diagnostics will no longer contain control codes when output is a
  file.

• CLI executables now fail when printing an opaque instead of
  producing garbage. This improves handling of some incorrect uses of
  auto output (#1251).

Removed

• Removed prelude functions reduce_stream, map_stream,
  reduce_stream_per, and reduce_map_per.

Fixed

• Various fixes to scalar migration (#1721).

Fixed

• Corrupted OOM error messages.

• Excessive internal fragmentation for some programs (#1715).

Fixed

• replicate was broken for sizes that didn't fit in i32.

• Transposition of empty arrays in interpreter (#1700).

• Exotic simplification error (#1309).

• Rare race condition could lead to leaking of error message memory in
  multicore and ispc backends (#1709).

• Overzealous aliasing for built-in overloaded types (#1711).

Added

• New fusion engine by Amar Topalovic and Walter Restelli-Nielsen.
  Fuses more, which is good for some programs and bad for others.
  Most programs see no change. This is mainly a robust foundation for
  further improvements.

• New experimental backend: ispc. By Louis Marott Normann,
  Kristoffer August Kortbæk, William Pema Norbu Holmes Malling, and
  Oliver Bak Kjersgaard Petersen.

• New prelude functions: hist, spread. These function as
  non-consuming variants of reduce_by_index and scatter.

• Using == to compare arrays is now deprecated.

• New command: futhark tokens. You probably don't care about this one.

• In the C API, opaque types that correspond to tuples or records can
  now have their elements inspected and be created from elements
  (#1568).

• New server protocol commands: types, fields, entry_points.

• Tuples and records can now be passed from FutharkScript to Futhark
  entry points (#1684).

Fixed

• Sometimes missing cases in match expressions were not detected.

• A defective simplification rule could in very rare cases lead to
  infinite recursion in the compiler (#1685).

• Some broken links in futhark doc output (#1686).

• Incorrect checking of whether a function parameter is consumable
  based on its type (#1687).

• Missing aliases for functions that return multiple aliased values
  (#1690).

• new/values functions for GPU backends are now properly
  asynchronous (#1664). This may uncover bugs in application code.

Added

• Somewhat simplified the handling of "uniqueness types" (which is a
  term we are moving away from). You should never see * in
  non-function types, and they are better thought of as effect
  indicators.

• futhark literate: prints tracing output (and other logging
  messages that may occur) when run with -v (#1678).

• Entry points can now be any valid Futhark identifier.

Fixed

• futhark test -C was broken.

• futhark_context_free() for the GPU backends neglected to free some
  memory used for internal bookkeeping, which could lead to memory
  leaks for processes that repeatedly create and destroy contexts
  (#1676).

• FutharkScript now allows ' in names.

• futhark lsp now handles warnings in programs that also have errors.

Added

• The CUDA backend now supports compute capability 8.6 and 8.7.

• Philip Børgesen has implemented a new optimisation for GPU backends
  that migrates scalar work to the GPU, in order to reduce
  synchronisation. This results in major speedup for some programs.

• String literals are now allowed in input blocks.

• Experimental and undocumented support for automatic differentiation,
  available on the secret menu.

• Assertions and attributes are now ignored when used as size
  expressions. E.g. iota (assert p n) 0 now has size n.

• futhark test only runs the interpreter if passed -i.

• futhark literate now shows progress bars when run with -v.

Fixed

• futhark lsp is now better at handling multiple files (#1647).

• Incorrect handling of local quantification when determining type
  equivalence in during module type ascription (#1648).

• Incorrect checking of liftedness when instantiating polymorphic
  functions during module type ascription.

• Tightened some restrictions on the use of existential sizes that
  could otherwise lead to compiler crashes (#1650). This restriction
  is perhaps a bit too might and it may be possible to loosen it in
  the future.

• Another defunctorisation bug (#1653). Somehow we find these every
  time Martin Elsman writes a nontrivial Futhark program.

• futhark bench: convergence phase now does at least --runs runs.

• Errors and warnings no longer awkwardly mixed together in console output.

• Slightly better type errors for ambiguous sizes (#1661).

• Better type errors for module ascription involving nested modules
  (#1660).

• futhark doc: some formatting bugs.

• futhark doc didn't notice all local module types (#1666).

• Missing consumption check in optimisation could lead to ICE (#1669).

Added

• New math functions: f16.erf, f32.erf, f64.erf.

• New math functions: f16.erfc, f32.erfc, f64.erfc.

• New math functions: f16.cbrt, f32.cbrt, f64.cbrt.

Fixed

• Variables being indexed now have correct source spans in AST.

• futhark lsps hover information now contains proper range information.

• futhark query and futhark lsp incorrectly thought size
  parameters had type i32.

• futhark doc put documentation for prelude modules in the wrong
  location (which also led to messed-up style sheets).

Added

• Sun Haoran has implemented unnamed typed holes, with syntax ???.

• Sun Haoran has implemented the beginnings of a language server:
  futhark lsp. A VSCode language extension is available on the
  marketplace, but the language server should work with any editor.

• Crucial new command: futhark thanks.

• The GPU backends now support a caching mechanism for JIT-compiled
  code, significantly improving startup times. Use the
  futhark_context_config_set_cache_file() in the C API, the
  --cache-file option on executables, or the --cache-extension
  option on futhark test and futhark bench. These also work for
  the non-GPU backends, but currently have no effect. (#1614)

• Aleksander Junge has improved futhark bench such that it
  intelligently chooses how many runs to perform (#1335).

Fixed

• Incomplete simplification would cause some instances of nested
  parallelism to require irregular allocations (#1610).

• Missing alias checking for a simplification rule related to in-place
  updates (#1615, #1628).

• Incorrect code generation for certain copies of transposed arrays
  (#1627).

• Fusion would mistakenly try to treat some loops with irregular sizes
  (#1631).

• Memory annotation bug for non-inlined functions (#1634).

Added

• Slightly better parse errors (again).

• futhark literate now supports a file: option in :img and
  :video directives (#1491).

Fixed

• Improved hoisting of size computations. This could cause some
  regular nested parallel programs to run into compiler limitations,
  as if they were irregular.

• Rare code generation bug for histograms (#1609).

Added

• futhark check-syntax: check syntactic validity of a program
  file, without type-checking.

• Parsing multi-file programs is now parallelised, making it
  slightly faster.

• Reloading a large program in futhark repl is now faster, as long
  as not too many of its files have been modified (#1597).

Fixed

• Mistake in occurs check could cause infinite loop in type checker
  for programs with type errors (#1599).

Added

• futhark bench now explicitly notes when a tuning file is not
  present.

• futhark bench, futhark test and friends are now better at
  handling fatally terminating programs (e.g. segmentation faults).

• Generated C code is now a lot smaller for large programs, as error
  recovery has been more centralised (#1584).

Fixed

• Some bugs in checking for local memory capacity for particularly
  exotic generated code.

• Insufficient hoisting of allocation sizes led to problems with
  memory expansion in rare cases (#1579).

• Conversion of floating-point NaNs and infinities to integers now
  well defined (produces zero) (#1586).

• Better handling of OOM for certain short-lived CPU-side allocations (#1585).

Added

• API functions now return more precise error codes in some cases.

• Out-of-memory errors contain more information.

Fixed

• Yet another in-place lowering issue (#1569).

• Removed unnecessary bounds checks in register tiling, giving about
  1.8x speedup on e.g. matrix multiplication on newer NVIDIA GPUs.

• A parser bug erroneously demanded whitespace in some type
  expressions (#1573).

• Some memory was not being freed correctly when shutting down OpenCL
  and CUDA contexts, which could lead to memory leaks in processes
  that created and freed many contexts.

• An incorrect copy-removal in some exotic cases (#1572).

• 'restore'-functions might perform undefined pointer arithmetic when
  passed garbage.

Fixed

• A size inference bug in type checking of loops (#1565).

• Exotic flattening bug (#1563).

• Segmented reduce_by_index with fairly small histogram size would
  use vastly more memory than needed.

Added

• Parse errors now list possible expected tokens.

• Lexer errors now mention the file.

Fixed

• Overloaded number literals cannot be sum types (#1557).

• Defective GPU code generation for vectorised non-commutative
  operatators (#1559).

• Excessive memory usage for some programs (#1325).

Added

• New functions: reduce_by_index_2d, reduce_by_index_3d.

• Manifests now contain compiler version information.

Fixed

• Allocation insertion pass bug (#1546).

• An exotic bug involving TLS and dynamically loading code generated
  by the multicore backend.

• Unconstrained ambiguous types now default to () (#1552). This
  should essentially never have any observable impact, except that
  more programs will type check.

• Double buffering compiler crash (#1553).

Added

• Top-level value definitions can (and should) now be declared with
  with def, although let still works.

• New tool: futhark defs, for printing locations of top-level
  definitions.

Changed

• def is now a reserved word.

Fixed

• Contrived intra-group code versions with no actual parallelism would
  be given a group size of zero (#1524).

Added

• futhark repl now allows Ctrl-c to interrupt execution.

Fixed

• Alias tracking of sum types.

• Proper checking that a function declared to return a unique-typed
  value actually does so.

• Faulty uniqueness checking and inference for lambdas (#1535).

• Monomorphisation would duplicate functions under rare circumstances
  (#1537).

• Interpreter didn't check that the arguments passed to unflatten
  made sense (#1539).

• futhark literate now supports a $loaddata builtin function for
  passing datasets to Futhark programs.