Releases: diku-dk/futhark
0.18.1
0.17.3
Added
- Improved parallelisation of
futhark bench
compilation.
Fixed
-
Dataset generation for test programs now use the right
futhark
executable (#1133). -
Really fix NaN comparisons in interpreter (#1070, again).
-
Fix entry points with a parameter that is a sum type where
multiple constructors contain arrays of the same statically known
size. -
Fix in monomorphisation of types with constant sizes.
-
Fix in in-place lowering (#1142).
-
Fix tiling inside multiple nested loops (#1143).
0.17.2
Added
-
Obscure loop optimisation (#1110).
-
Faster matrix transposition in C backend.
-
Library code generated with CUDA backend can now be called from
multiple threads. -
Better optimisation of concatenations of array literals and
replicates. -
Array creation C API functions now accept
const
pointers. -
Arrays can now be indexed (but not sliced) with any signed integer
type (#1122). -
Added --list-devices command to OpenCL binaries (#1131)
-
Added --help command to C, CUDA and OpenCL binaries (#1131)
Removed
-
The integer modules no longer contain
iota
andreplicate
functions. The top-level ones still exist. -
The
size
module type has been removed from the prelude.
Changed
- Range literals may no longer be produced from unsigned integers.
Fixed
-
Entry points with names that are not valid C (or Python)
identifiers are now pointed out as problematic, rather than
generating invalid C code. -
Exotic tiling bug (#1112).
-
Missing synchronisation for in-place updates at group level.
-
Fixed (in a hacky way) an issue where
reduce_by_index
would use
too much local memory on AMD GPUs when using the OpenCL backend.
0.16.4
Added
-
#[unroll]
attribute. -
Better error message when writing
a[i][j]
(#1095). -
Better error message when missing "in" (#1091).
Fixed
-
Fixed compiler crash on certain patterns of nested parallelism
(#1068, #1069). -
NaN comparisons are now done properly in interpreter (#1070).
-
Fix incorrect movement of array indexing into branches
if
s
(#1073). -
Fix defunctorisation bug (#1088).
-
Fix issue where loop tiling might generate out-of-bounds reads
(#1094). -
Scans of empty arrays no longer result in out-of-bounds memory
reads. -
Fix yet another defunctionalisation bug due to missing
eta-expansion (#1100).
0.16.3
Added
-
random
input blocks forfuthark test
andfuthark bench
now
support floating-point literals, which must always have either an
f32
orf64
suffix. -
The
cuda
backend now supports the-d
option for executables. -
The integer modules now contain a
ctz
function for counting
trailing zeroes.
Fixed
-
The
pyopencl
backend now works with OpenCL devices that have
multiple types (most importantly, oclgrind). -
Fix barrier divergence when generating code for group-level
colletive copies in GPU backend. -
Intra-group flattening now looks properly inside of branches.
-
Intra-group flattened code versions are no longer used when the
resulting workgroups would have less than 32 threads (with default
thresholds anyway) (#1064).
0.16.2
Added
futhark autotune
: added--pass-option
.
Fixed
-
futhark bench
: progress bar now correct when number of runs is
less than 10 (#1050). -
Aliases of arguments passed for consuming parameters are now
properly checked (#1053). -
When using a GPU backend, errors are now properly cleared.
Previously, once e.g. an out-of-bounds error had occurred, all
future operations would fail with the same error. -
Size-coercing a transposed array no longer leads to invalid code
generation (#1054).
0.16.1
Added
-
Incremental flattening is now performed by default. Use
attributes to constrain and direct the flattening if you have
exotic needs. This will likely need further iteration and
refinement. -
Better code generation for
reverse
(and the equivalent explicit
slice). -
futhark bench
now prints progress bars. -
The
cuda
backend now supports similar profiling as theopencl
option, although it is likely slightly less accurate in the
presence of concurrent operations. -
A preprocessor macro
FUTHARK_BACKEND_foo
is now defined in
generated header files, where foo is the name of the backend
used. -
Non-inlined functions (via
#[noinline]
) are now supported in GPU
code, but only for functions that exclusively operate on
scalars. -
futhark repl
now accepts a command line argument to load a
program initially. -
Attributes are now also permitted on declarations and specs.
-
futhark repl
now has a:nanbreak
command (#839).
Removed
-
The C# backend has been removed (#984).
-
The
unsafe
keyword has been removed. Use#[unsafe]
instead.
Changed
-
Out-of-bounds literals are now an error rather than a warning.
-
Type ascriptions on entry points now always result in opaque types
when the underlying concrete type is a tuple (#1048).
Fixed
-
Fix bug in slice simplification (#992).
-
Fixed a typer checker bug for tracking the aliases of closures
(#995). -
Fixed handling of dumb terminals in futhark test (#1000).
-
Fixed exotic monomorphisation case involving lifted type
parameters instantiated with functions that take named parameters
(#1026). -
Further tightening of the causality restriction (#1042).
-
Fixed alias tracking for right-operand operator sections (#1043).
0.15.8
Added
-
Warnings for overflowing literals, such as
1000 : u8
. -
Futhark now supports an attribute system, whereby expressions can
be tagged with attributes that provide hints or directions to the
compiler. This is an expert-level feature, but it is sometimes
useful.
0.15.7
Added
-
Faster index calculations for very tight GPU kernels (such as the
ones corresponding to 2D tiling). -
scan
with vectorised operators (e.g.map2 (+)
) is now faster
in some cases. -
The C API has now been documented and stabilized, including
obtaining profiling information (although this is still
unstructured).
Fixed
-
Fixed some cases of missing fusion (#953).
-
Context deinitialisation is now more complete, and should not leak
memory (or at least not nearly as much, if any). This makes it
viable to repeatedly create and free Futhark contexts in the same
process (although this can still be quite slow).
0.15.6
Added
-
Binary operators now act as left-to-right sequence points with
respect to size types. -
futhark bench
now has more colourful and hopefully readable
output. -
The compiler is now about 30% faster for most nontrivial programs.
This is due to parallelising the inlining stage, and tweaking the
default configuration of the Haskell RTS. -
futhark dataset
is now about 8-10x faster.