[TIRx][CUDA] Add a table-driven PTX dialect and retire tirx.ptx.* - #20103

Merged

jinhongyii merged 17 commits into

apache:mainfrom

spectrometerHBH:apache-pr-ptx-dialect

Aug 7, 2026

Commits on Aug 7, 2026

feat(tvmscript): extend low-level CUDA PTX intrinsics (#48)
spectrometerHBH
committed
a001466
feat(lower-tirx): improve CUDA lowering for recurrent KDA (#49)
spectrometerHBH
committed
dd4b5d4
feat(op): table-driven PTX dialect (T.ptxd) and tirx.ptx.* retirement (#47)

spectrometerHBH
and
claude
committed
d09932e
refactor(op): rename the PTX dialect namespace from ptxd to ptx (#51)
spectrometerHBH
committed
cab4cd9
fix(infra): give the NVSHMEM objects CUDA's cccl include directory (#52)
spectrometerHBH
committed
05f0a66
test(infra): load registry workloads from per-kernel config files
spectrometerHBH
committed
9cb54a9
test(infra): support both tirx-kernels workload layouts
spectrometerHBH
committed
5605fbb
fix(op): drop the duplicated cp.async.bulk entry and gate the table
spectrometerHBH
committed
ea1a467
refactor(op): order the instruction table by PTX ISA section
spectrometerHBH
committed
ccd6454
fix(op): make the T.ptx stub generator independent of ruff
spectrometerHBH
committed
0a2ef49
refactor(op): rename the ptx_dialect package to ptx
spectrometerHBH
committed
ae15abd
refactor(op): factor the operand role into direction and kind, pred into a dtype
spectrometerHBH
committed
7639ca7
refactor(op): derive the .b8 staging from the dtype table
spectrometerHBH
committed
4d76f19
feat(op): register the cp.async ignore-src lines
spectrometerHBH
committed
7c4342e
feat(op): register the cache_hint axis on ld, st, red and atom
spectrometerHBH
committed
025d8ed
feat(op): register the sink symbol as a per-call operand choice
spectrometerHBH
committed
226fd7a
feat(op): register the 256-bit sink lines, and stop multiplying the sink axis
spectrometerHBH
committed
0a52599

Navigation Menu

Uh oh!

[TIRx][CUDA] Add a table-driven PTX dialect and retire tirx.ptx.* - #20103

[TIRx][CUDA] Add a table-driven PTX dialect and retire tirx.ptx.*#20103
jinhongyii merged 17 commits into
apache:mainfrom
spectrometerHBH:apache-pr-ptx-dialect

Commits on Aug 7, 2026

feat(tvmscript): extend low-level CUDA PTX intrinsics (#48)

feat(lower-tirx): improve CUDA lowering for recurrent KDA (#49)

feat(op): table-driven PTX dialect (T.ptxd) and tirx.ptx.* retirement (#47)

refactor(op): rename the PTX dialect namespace from ptxd to ptx (#51)

fix(infra): give the NVSHMEM objects CUDA's cccl include directory (#52)

test(infra): load registry workloads from per-kernel config files

test(infra): support both tirx-kernels workload layouts

fix(op): drop the duplicated cp.async.bulk entry and gate the table

refactor(op): order the instruction table by PTX ISA section

fix(op): make the T.ptx stub generator independent of ruff

refactor(op): rename the ptx_dialect package to ptx

refactor(op): factor the operand role into direction and kind, pred into a dtype

refactor(op): derive the .b8 staging from the dtype table

feat(op): register the cp.async ignore-src lines

feat(op): register the cache_hint axis on ld, st, red and atom

feat(op): register the sink symbol as a per-call operand choice

feat(op): register the 256-bit sink lines, and stop multiplying the sink axis