Skip to content

Release 0.7.5 - #20

Merged
amasat01 merged 28 commits into
mainfrom
w075
Nov 10, 2026
Merged

amasat01 merged 28 commits into
mainfrom
w075

Conversation

@amasat01

@amasat01 amasat01 commented Nov 6, 2026 •

Copy link
Copy Markdown
Owner

This section covers what changes from 0.7.3. aether-dsc moves with the C++ version. The public API is unchanged; the banded mode (double precision on GPUs with weak or no FP64) grows.

Added

  • More banded functions: tan, exp2, expm1, log2, log10, log1p, sinh, cosh, tanh, asinh, acosh and atanh, each with a stated per-op ULP bound and a golden corpus behind it.
  • Banded expression chains. a * b + c on banded values keeps the result unrestored between ops and normalises only where a consumer reads the parts in order (a division, a root, a comparison, a store). Where an op would need a normalised value and has none, the overload is deleted and the error names the fix.
  • Banded headers compile under NVRTC, and aether-dsc ships them in its payload.
  • The install page has a Best practices section: keep aether-dsc at the family release.
  • pow vectorises on aarch64 with GCC. Under AETHER_HOST_VECTOR_MATH, aether::math::pow has the aarch64 vector variants GCC asks for, faithfully rounded (error < 1 ULP) and checked by make faithful-gate, which now runs on aarch64. Every other function stays libm's there, and clang keeps libm.
  • The C++ test suite builds and runs on aarch64, Linux and macOS: the x86-only packet-math tests report skipped there, and so do the few whose reference needs a long double wider than double (Apple arm64).

Changed

  • Banded storage handles values outside the cell's range by label. A magnitude at or above 2^126 is stored as signed infinity, one below 2^-94 as signed zero, and infinities and NaN keep their own words; it is one range test and a select, with no branch.
  • Banded inverse trigonometry is tighter: a one-float sector tangent with a matched atan phase, exact 1 - |x| in asin and acos.
  • CI takes the release version from the top heading of CHANGELOG.md; the label is only checked against it. Memo digests are narrowed to real inputs, and the sdist-install job needs no pip cache.

Fixed

  • Banded headers build without warnings under C++23 with libc++ (Apple clang): the deprecated denorm style in the banded numeric_limits no longer warns.
  • The build finds OpenMP under Apple clang with Homebrew's libomp: the OpenMP flags pass as separate arguments, and libomp's headers are on the include path.
  • Device allocation works on GPUs without memory pools (cudaDevAttrMemoryPoolsSupported == 0, common on virtualised GPUs): a stream-ordered Chunk::allocate falls back to plain cudaMalloc there, the chunk reports streamOrdered() == false, and allocating during a stream capture throws an error that names the missing pools.
  • Assign with a long unrolled fold builds under clang: the fold is chunked so clang's 256-operand limit is not hit.

…bel is only checked in its own workflow, siblings build at their heading version, no tag in the test digests
…scalar-multiplier fma, dominated hi sums, exact 1-ax in asin/acos
@amasat01 amasat01 changed the title 0.7.5 (draft): banded emulation, lean CI, one-pipeline release Release 0.7.5 Nov 10, 2026
@amasat01
amasat01 marked this pull request as ready for review November 10, 2026 15:37
@amasat01
amasat01 merged commit 7826fe1 into main Nov 10, 2026
29 of 31 checks passed
@amasat01
amasat01 deleted the w075 branch November 10, 2026 16:52
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant