This is a first step towards high level representation for fp8 types
that have been built in to hardware with near term roadmaps. Like the
BFLOAT16 type, the family of fp8 types are inspired by IEEE-754 binary
floating point formats but, due to the size limits, have been tweaked in
various ways in order to maximally use the range/precision in various
scenarios. The list of variants is small/finite and bounded by real
hardware.
This patch introduces the E5M2 FP8 format as proposed by Nvidia, ARM,
and Intel in the paper: https://arxiv.org/pdf/2209.05433.pdf
As the more conformant of the two implemented datatypes, we are plumbing
it through LLVM's APFloat type and MLIR's type system first as a
template. It will be followed by the range optimized E4M3 FP8 format
described in the paper. Since that format deviates further from the
IEEE-754 norms, it may require more debate and implementation
complexity.
Given that we see two parts of the FP8 implementation space represented
by these cases, we are recommending naming of:
* `F8M<N>` : For FP8 types that can be conceived of as following the
same rules as FP16 but with a smaller number of mantissa/exponent
bits. Including the number of mantissa bits in the type name is enough
to fully specify the type. This naming scheme is used to represent
the E5M2 type described in the paper.
* `F8M<N>F` : For FP8 types such as E4M3 which only support finite
values.
The first of these (this patch) seems fairly non-controversial. The
second is previewed here to illustrate options for extending to the
other known variant (but can be discussed in detail in the patch
which implements it).
Many conversations about these types focus on the Machine-Learning
ecosystem where they are used to represent mixed-datatype computations
at a high level. At that level (which is why we also expose them in
MLIR), it is important to retain the actual type definition so that when
lowering to actual kernels or target specific code, the correct
promotions, casts and rescalings can be done as needed. We expect that
most LLVM backends will only experience these types as opaque `I8`
values that are applicable to some instructions.
MLIR does not make it particularly easy to add new floating point types
(i.e. the FloatType hierarchy is not open). Given the need to fully
model FloatTypes and make them interop with tooling, such types will
always be "heavy-weight" and it is not expected that a highly open type
system will be particularly helpful. There are also a bounded number of
floating point types in use for current and upcoming hardware, and we
can just implement them like this (perhaps looking for some cosmetic
ways to reduce the number of places that need to change). Creating a
more generic mechanism for extending floating point types seems like it
wouldn't be worth it and we should just deal with defining them one by
one on an as-needed basis when real hardware implements a new scheme.
Hopefully, with some additional production use and complete software
stacks, hardware makers will converge on a set of such types that is not
terribly divergent at the level that the compiler cares about.
(I cleaned up some old formatting and sorted some items for this case:
If we converge on landing this in some form, I will NFC commit format
only changes as a separate commit)
Differential Revision: https://reviews.llvm.org/D133823
It allows finding all intervals that overlap with any given point.
At this time, it does not support any deletion or rebalancing
operations.
The IntervalTree is designed to be set up once, and then queried
without any further additions.
Reviewed By: psamolysov, probinson
Differential Revision: https://reviews.llvm.org/D125776
This patch deprecates llvm::empty as I've migrated all known uses of
llvm::empty(x) to x.empty().
Differential Revision: https://reviews.llvm.org/D134141
LLVM contains a helpful function for getting the size of a C-style
array: `llvm::array_lengthof`. This is useful prior to C++17, but not as
helpful for C++17 or later: `std::size` already has support for C-style
arrays.
Change call sites to use `std::size` instead.
Differential Revision: https://reviews.llvm.org/D133429
This change adds an assert to dyn_cast that the value passed-in is present. In the past, this relied on the isa_impl assertion (which still works in many cases) but which we can tighten up for a better QoI.
The PointerUnion change is because it seems like (based on the call sites) the semantics of the member dyn_cast are actually dyn_cast_if_present.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D133221
`llvm::all_equal` checks if all values in the given range are equal, i.e., there are no two elements that are not equal.
Similar to `llvm::all_of`, it returns `true` when the range is empty.
`llvm::all_equal` is intended to supersede `llvm::is_splat`, which will be deprecated and removed in future patches.
See the discussion thread for more details:
https://discourse.llvm.org/t/adt-is-splat-and-empty-ranges/64692.
Reviewed By: dblaikie, shchenz
Differential Revision: https://reviews.llvm.org/D132334
This patch implements Any::has_value for consistency with std::any in
C++17.
My plan is to deprecate Any::hasValue after migrating all of its uses
to Any::has_value. Since I am about to do so, this patch simply
replaces hasValue with has_value in the unit test instead of adding
tests for has_value.
Differential Revision: https://reviews.llvm.org/D132278
This patch makes `SmallSet::insert(const T &)` return
`std::pair<const_iterator, bool>` instead of
`std::pair<NoneType, bool>`. This will exactly match std::set's behavior
and make deduplicating items with SmallSet easier.
Reviewed By: dblaikie, lattner
Differential Revision: https://reviews.llvm.org/D131549
This patch implements Optional::transform for consistency with
std::optional::transform in C++23.
Note that the new function is identical to Optional::map. My plan is
to deprecate Optional::map after migrating all of its uses to
Optional::transform.
Differential Revision: https://reviews.llvm.org/D131829
As a follow-up of e8578968f6 which replaced the
callers to use the C++17 equivalents, remove the equivalents from
`STLForwardCompat.h` entirely and their corresponding tests.
Differential Revision: https://reviews.llvm.org/D131769
STLForwardCompat.h defines several utilities and type traits to mimic that of
the ones in the C++17 standard library. Now that LLVM is built with the C++17
standards mode, remove use of these equivalents in favor of the ones from the
standard library.
Differential Revision: https://reviews.llvm.org/D131717
This patch adds the ability to deconstruct the `value_type` returned by `llvm::enumarate` into index and value of the wrapping range. Main use case is the common occurence of using it during loop iteration. After this patch it'd then be possible to write code such as:
```
for (auto [index, value] : enumerate(container)) {
...
}
```
where `index` is the current index and `value` a reference to elements in the given container.
Differential Revision: https://reviews.llvm.org/D131486
Allow for `is_splat` to be used inline, similar to `is_contained`, e.g.,
```
if (is_splat({type1, type2, type3, type4}))
...
```
which is much more concise and less typo-prone than an equivalent chain of equality comparisons.
My immediate motivation is to clean up some code in the SPIR-V dialect that currently needs to either construct a temporary container or use `makeArrayRef` before calling `is_splat`.
Reviewed By: dblaikie
Differential Revision: https://reviews.llvm.org/D131289
std::iterator has been deprecated in C++17 and some standard library implementations such as MS STL or libc++ emit deperecation messages when using the class.
Since LLVM has now switched to C++17 these will emit warnings on these implementations, or worse, errors in build configurations using -Werror.
This patch fixes these issues by replacing them with LLVMs own llvm::iterator_facade_base which offers a superset of functionality of std::iterator.
Differential Revision: https://reviews.llvm.org/D131320
In SmallVector unit test, referring test suite members every time using 'this'
(due to how non-dependent name lookups works) is quite cumbersome and can be improved with aliases.
Some test cases already use this approach, in this change we expand that approach to whole test file.
Differential Revision: https://reviews.llvm.org/D129990
`SmallDenseMap` constructor with reserve gets an arbitrary `NumInitBuckets` value and passes it below to `init` method.
If `NumInitBuckets` is greater then `InlineBuckets`, then `SmallDenseMap` initializes to large representation passing `NumInitBuckets` below to `DenseMap` initialization. `DenseMap::initEmpty` method asserts that initial buckets count must be a power of 2.
Proposed solution is to update `NumInitBuckets` value in `SmallDenseMap` constructor till the next power of 2. It should satisfy both `DenseMap` preconditions and required minimum buckets count for reservation.
Reviewed By: atrick
Differential Revision: https://reviews.llvm.org/D129825
llvm::sort is beneficial even when we use the iterator-based overload,
since it can optionally shuffle the elements (to detect
non-determinism). However llvm::sort is not usable everywhere, for
example, in compiler-rt.
Reviewed By: nhaehnle
Differential Revision: https://reviews.llvm.org/D130406
Due to how operator== changed after operator<=> landing in C++20
(error: use of overloaded operator '==' is ambiguous (with operand
types 'const llvm::APSInt' and 'const unsigned long')),
CheckIntPartMin/CheckIntPartMax interfaces are now leveraging
APSInt::compareValues.
Differential Revision: https://reviews.llvm.org/D129567
In D94439, BumpPtrAllocator changed its implementation to use an empty base optimization for the underlying allocator.
This patch builds on that by extending its functionality to more classes as well as enabling the underlying allocator to be a reference type, something not currently possible as you can't derive from a reference.
The main place this sees use is in StringMaps which often use the default MallocAllocator, yet have to pay the size of a pointer for no reason.
Reviewed By: dblaikie
Differential Revision: https://reviews.llvm.org/D129206
Binary size of `clang` is trivial; namely, numerical value doesn't
change when measured in MiB, and `.data` section increases from 139Ki to
173 Ki.
Differential Revision: https://reviews.llvm.org/D128070
This patch adds has_value, value, value_or to llvm::Optional so that
llvm::Optional looks more like std::optional.
I will keep the existing functions while migrating their callers and
then remove them later.
Differential Revision: https://reviews.llvm.org/D128131
Stella Laurenzo