fix(pto): infer matmul tile config by target arch by HecreReed · Pull Request #260 · zhangstevenunity/PTOAS

HecreReed · 2026-03-13T08:14:46Z

Summary

add a dedicated pto-infer-tile-config pass that derives LEFT/RIGHT/ACC tile config from pto.target_arch
run the pass before PTOViewToMemref so emitted pointer_cast/bind_tile metadata matches the target ISA on both A3 and A5
add A3/A5 regression tests as .pto inputs that lock the generated Tile<...> templates for matmul operands/results

PTO-ISA expects different canonical LEFT tile layout on A3 vs A5, while RIGHT/ACC also need fixed fractal/layout combinations
keeping this inference in PTO IR removes the need for users to hand-thread BLayout/SLayout into matmul tiles
this also prevents low-level lowering from carrying stale default configs into generated C++

ninja -C build-llvm19 ptoas
build-llvm19/tools/ptoas/ptoas --pto-arch=a3 test/basic/matmul_tile_config_infer_a3.pto | /opt/homebrew/opt/llvm@19/bin/FileCheck test/basic/matmul_tile_config_infer_a3.pto
build-llvm19/tools/ptoas/ptoas --pto-arch=a5 test/basic/matmul_tile_config_infer_a5.pto | /opt/homebrew/opt/llvm@19/bin/FileCheck test/basic/matmul_tile_config_infer_a5.pto

fix(pto): infer matmul tile config by target arch

01f8be4

HecreReed force-pushed the codex/auto-infer-matmul-tile-config branch from 9481541 to 01f8be4 Compare March 13, 2026 08:18

HecreReed added 2 commits March 13, 2026 18:24

fix(pto): sync inferred tile config across call boundaries

f451126

fix(pto): handle external func declarations in emitc

d277885

HecreReed marked this pull request as ready for review March 14, 2026 06:13

fix(pto): allow inferred tile config syntax

1e9ef1f