Repository navigation
Description
CUDA 13.5 is out (https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html), so a released tileiras now accepts Tile IR bytecode v13.5. Upstream, cutile-python main has moved V_13_5 out of "dev only" and into _SUPPORTED_VERSIONS (NVIDIA/cutile-python@9e8ef5e), its tileiras extra now pins cuda-toolkit<13.6, and the v1.8.0 release notes list the 13.5 features: foreign calls (in a separate cuda-tile-preview package), FP8 on Ada, and the wider num_worker_warps range. cuTile.jl emits v13.1 to v13.4 (SUPPORTED_BYTECODE_VERSIONS in src/bytecode/types.jl), and CUDA_Tile_jll stops at 13.4.1. The first three tasks change the encoding of every kernel, so they are needed before anything can target v13.5. The fourth turns v13.5 on.
Tasks
-
Function header. At v13.5,
add_function!insrc/bytecode/writer.jlsets only0x02in an entry's flags byte; the0x04"has hints" bit is gone. Every entry must carry anOptimizationHintsattribute, soencode_entry_hintscan no longer returnnothingand should emit an empty one instead. Every function's header then ends with an empty taggedArrayattribute forarg_attrs(tag0x06, count 0). NVIDIA/cutile-python@b8c4772 -
DISubprogramflags. At v13.5,subprogram!insrc/bytecode/writer.jlappends a trailing varint flags field, currently always 0.DebugAttrTableneeds to know the bytecode version for this. NVIDIA/cutile-python@b8c4772 -
inboundsonAtomicRedViewTkoOp. At v13.5,encode_AtomicRedViewTkoOp!insrc/bytecode/encodings.jlencodes a dense boolinboundsarray aftermode, asLoadViewTkoOpandStoreViewTkoOphave done since v13.4. It should throw if any entry istruebelow v13.5. Upstream always passes all-false; acheck_boundskeyword onct.atomic_store_*(src/language/atomics.jl) could come later. v13.5 also letsatomic_red_view_tkotake views that carry apadding_value; the "Tile IR atomic views cannot carry padding" comment andPaddingMode.Undeterminedinmake_atomic_tile_view(src/language/atomics.jl) only hold below v13.5. NVIDIA/cutile-python@b8c4772 -
Enable v13.5. Once the encoding tasks above are done:
- Add 13.5 to
SUPPORTED_BYTECODE_VERSIONS.test/bytecode.jlcurrently asserts thatTypeTable(; version=v"13.5")throws; flip that. - Build
CUDA_Tile_jll13.5 on Yggdrasil. TheCUDA_Tile_jll = "13.3"compat already admits it. - Emit v13.5 by default when the detected
tileirasis 13.5. - Update
docs/src/man/compatibility.md,docs/src/installation.mdandNEWS.md.
- Add 13.5 to
-
num_worker_warpsfrom 1 to 32. At v13.5, accept any power of two from 1 to 32. Below v13.5, keep allowing only 4 and 8.- Relax the
val in (4, 8)check invalidate_hint(src/language/types.jl). - Add the version check next to the existing v13.3 check in
encode_entry_hints(src/bytecode/writer.jl). - Update the table in
docs/src/man/performance.md.
- Relax the
-
Multimem view ops. Add encoders for
MultimemLoadReduceViewTkoOp(opcode 123),MultimemReduceViewTkoOp(124) andMultimemStoreViewTkoOp(125) tosrc/bytecode/encodings.jl, plus theMultimemReductionKindenum (add, min, max, and, or, xor, umin, umax). Upstream has only the encoders, with no user API yet. A usable API would also need the newptr_attrvaluesunicast(peer access over NVLink) andmulticast(requires NVSwitch with NVLink SHARP). The release notes list support onsm_100,sm_103andsm_107. NVIDIA/cutile-python@b8c4772 -
FP8 on Ada. With
tileiras13.5,Float8_E4M3FNandFloat8_E5M2casts andmmawork on sm_89. Upstream lowers its per-dtype minimum SM from 90 to 89 at v13.5 (NVIDIA/cutile-python@db66c13). cuTile.jl has no per-architecture dtype check, so any failure before 13.5 would have come fromtileiras. This is likely just a docs line plus a test on sm_89. -
Foreign calls: bytecode. Add the encoder for
PreviewCallOp(opcode 16386) and its preview attribute tags:PreviewExprInt(0xD1),PreviewConcreteType(0xDB) andPreviewExprConstStr(0xE4). These tags are varint-encoded as two bytes each. The existingAttributeTags are all pushed as single bytes, so this needs new handling. When a module uses preview ops, bit 0 of the version tag in the header must be set (write_header!). NVIDIA/cutile-python@66964ac -
Foreign calls: frontend. Add an intrinsic in the style of upstream's
cuda.tile_preview.foreign_call, gated on v13.5. It calls a symbol from a.tileliband takes these arguments:- the library path and version
- constant parameters (integers, strings and element types)
- tile or scalar inputs
- the output types
Supporting work:
- In
src/compiler/transform/token_order.jl, order the call as a store to all memory that also acquires. - In
run_tileiras(src/launch.jl), pass--preview=simtand-l <tilelib>. Add the tilelib paths and versions tocubin_cache_fields. - Upstream's
foreign_pointerdtype andArray.from_foreignexpose an array's base pointer to foreign code (NVIDIA/cutile-python@fc66868).TileArrayalready has aptrfield. This may only needPtr{T}operands, plus building aTileArrayfrom(ptr, sizes, strides)inside a kernel. - Optional: initialize and finalize NVSHMEM for kernels that call
nvshmem_*ornvshmemx_*symbols (NVIDIA/cutile-python@3fc0653).
Other upstream changes since v1.6.0
These changes are not tied to v13.5.
- Port the NVFP4 scaled matmul sample (NVIDIA/cutile-python@4dfd274) and the fused quantize-and-
mma_scaledkernel (NVIDIA/cutile-python@84391af) toexamples/. Both need Blackwell and v13.4.examples/has no block-scaled matmul yet.
Assisted by: Opus 5.5
Activity
maleadt commentedon Nov 6, 2026
Bunch of these were done in #348