futhark

Added

• The cuda and hip backends now generate faster code for scans
  that have been fused with maps that internally produce arrays.
  Work by Anders Holst and Christian Påbøl Jacobsen.

• f16.ldexp, f32.ldexp, f64.ldexp, corresponding to the
  functions in the C math library.

Fixed

• Incorrect data dependency information for scatter and vjp could
  cause invalid simplification.

• Barrier divergence in certain complicated kernels that contain both
  bounds checks and intragroup scans.

Commits

• fbf6a39: Remove duplicate typeprec (#2046). (Troels Henriksen)

Commits

• d003046: futhark dev: support --backend=hip. (Troels Henriksen)

Commits

• 4d63955: Onwards! (Troels Henriksen)
• 2dadd9d: Add CHANGELOG.md and README.md to extra-doc-files (#2044) (ˌbodʲɪˈɡrʲim) #2044

Added

• FutharkScript now has a $loadbytes builtin function for reading
  arbitrary bytes into Futhark programs.

• futhark profile can now process reports produced by the C API
  function futhark_context_report().

• futhark profile now also produces a timeline of events.

Fixed

• futhark literate no longer fails if the final line is a directive
  without a trailing newline.

• Parser now allows arbitrary patterns in function parameters and
  let bindings, although the type checker will reject any that are
  refutable (#2017).

• Avoid generating invalid code in cases where deduplicated sum types
  are exposed through entry points (#1960).

• A bug in data dependency analysis for histogram operations would
  mistakenly classify some loop parameters as redundant, leaving to
  code being removed.

Commits

• 96c8d92: Oops, use the right metric here. (Troels Henriksen)

Commits

• b4a4991: Disable AMD tests until KU-IT resurrects our machine. (Troels Henriksen)
• bc134be: No linebreak here. (Troels Henriksen)
• fix for multi-value scatters. (Troels Henriksen)

Commits

• c4acd2a: futhark profile: also show min and max. (Troels Henriksen)

Commits

• fix case for SegOps. (Troels Henriksen)

Commits

• 052da80: Document new feature. (Troels Henriksen)

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.

Commits

• 20b2ca2: Remember to synchronise after copying, before writing. (Troels Henriksen)

Commits

• e083ed7: Minor style fixes. (Troels Henriksen)
• 30c57f7: Not everything free in a size should be bound. (Troels Henriksen)
• 97ef688: Fix #2040. (Troels Henriksen)

Commits

• a57fc5d: Make WriteReturns simpler, yet more flexible. (#2039) (Troels Henriksen) #2039

Commits

• afbb1f1: These are better now. (Troels Henriksen)

Commits

• ba17feb: Fix #2038. (Troels Henriksen)

Commits

• 22d7dc2: Rename Patterns module to Pat. (Troels Henriksen)

Commits

• 3d970b0: Fix style issues. (Troels Henriksen)
• 70fb1bb: Document this. (Troels Henriksen)

Commits

• a419e0a: Onwards! (Troels Henriksen)
• d7308a4: Newer cabal snapshot. (Troels Henriksen)
• 6096315: Newer Nix snapshot. (Troels Henriksen)