composable_kernel

mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 18:17:44 +00:00

Files

ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 8134f1d6f4 [CK_TILE] fix pk_fp4 compilation for non-gfx950 GPUs (#2983 )

See build error log from
https://github.com/ROCm/composable_kernel/issues/2271#issuecomment-3150218542

This PR make vector element access constexpr-safe by avoiding operator[] on
ext_vector_type(2) and replace those sites in the pk_fp4 conversions so they
can be used in constant expressions, as The operator[] on ext_vector_type(2)
isn't allowed in constant expressions, which caused "constexpr function never
produces a constant expression" with a note at x[0]. Using `bit_cast` to a
trivial array representation keeps it constexpr-compatible.

Signed-off-by: Hollow Man <hollowman@opensuse.org>

[ROCm/composable_kernel commit: fb66b4f5e4]

2025-10-09 07:43:41 -07:00

bfloat16.hpp

[CK_TILE] Fixing Type Conversions in PassThroughPack8 (#2769 )

2025-09-29 13:34:47 +03:00

e8m0.hpp

fix wrong nan producion. (#2640 )

2025-08-14 15:12:31 +08:00

float8.hpp

[CK_TILE] Add gtests for FMHA (#2744 )

2025-09-10 08:06:14 +05:00

half.hpp

[CK TILE] GEMM with packed i4 (#1885 )