PTO Kernels

Scope

workloads/pto_kernels is the LinxISA PTO tile-kernel workspace for kernel sources, tooling, and host-side validation flows.

Upstream

• Repository: https://github.com/LinxISA/PTO-Kernel
• Merge-back target branch: main

What This Submodule Owns

• PTO kernel sources (kernels/)
• Public PTO headers (include/)
• Kernel/tooling/test scaffolding (tools/, tests/, docs/)

Canonical Build and Test Commands

Run from /Users/zhoubot/linx-isa/workloads/pto_kernels.

cmake -S . -B build
cmake --build build -j"$(sysctl -n hw.ncpu 2>/dev/null || nproc)"
ctest --test-dir build --output-on-failure

# Add -DPTO_BENCHMARK_WORKBOOK=/path/to/benchmark_master_list_completed.xlsx
# to enable the workbook-backed manifest test.

cmake -S . -B build-linx -DPTO_ENABLE_LINX_CROSS=ON
cmake --build build-linx --target pto_linx_contracts

LinxISA Integration Touchpoints

• Submodule pinned by the superproject for PTO kernel bring-up
• Consumed by Linx tile flow contracts and integration tests
• Tooling and artifacts integrated with LinxISA AVS/runtime workflows

Related Docs

• /Users/zhoubot/linx-isa/docs/project/navigation.md
• /Users/zhoubot/linx-isa/docs/bringup/
• /Users/zhoubot/linx-isa/workloads/pto_kernels/docs/
• Pages dashboard: https://linxisa.github.io/pto-kernel/ (published from docs/)
• DeepSeek TileKernels provenance and source-to-PTO mapping: docs/upstream/deepseek-tilekernels.json

DeepSeek TileKernels lane

The kernels/upstream/deepseek/ lane ports all 37 kernel source files from deepseek-ai/TileKernels main at the exact commit recorded in the provenance manifest. The port is freestanding and has no Python, PyTorch, CUDA, or TileLang runtime dependency. Kernel data paths are loop-free C++ compositions of named PTO ISA v0.57 intrinsics. Scalar iteration is limited to tile-grid control in the shared helper; element data paths remain PTO operations. The physical carrier is 32x32 row-major, while runtime dimensions drive rectangular tail masks (LB0=ValidCol, LB1=ValidRow, LB2=physical Col) and row stride. The QEMU suite covers both the 32x32 standard profile and non-multiple tail shapes, including a multi-tile 35x37 transpose.

Run the host contracts locally:

cmake -S . -B build
cmake --build build
ctest --test-dir build --output-on-failure

From the LinxISA superproject, compile and execute the standard-input suite:

python3 avs/qemu/run_tests.py --suite deepseek_tilekernels --verbose

Tidy-Up Status

• done: split kernel sources by operation family under kernels/attention, kernels/matmul, kernels/normalization, kernels/elementwise, kernels/layout, kernels/indexing, kernels/routing, kernels/decode, and kernels/memory
• done: replaced flat-path discovery with kernels/catalog.txt for CMake, benchmark manifest tooling, parity tooling, asm tooling, and QEMU suite wiring
• done: added shared runtime and environment setup headers in include/common/runtime/kernel_env.hpp, include/common/runtime/kernel_shapes.hpp, and include/common/runtime/kernel_tiling.hpp
• done: moved representative benchmarked tiled kernels onto parameterized shape and tiling policy instead of per-file hardcoded tile constants
• done: kept full-profile shapes large by default and preserved reduced smoke shapes for QEMU
• done: revalidated host tests, Linx asm contracts, and host-vs-QEMU parity after the refactor
• in_progress: migrate the remaining non-benchmarked kernels to the shared shape and tiling headers so every kernel family uses the same configuration surface

Status Panel

• site source: docs/index.html, docs/assets/site.css, docs/assets/site.js
• generated data: docs/data/status.json
• refresh command: python3 tools/build_status_site.py --workbook /path/to/benchmark_master_list_completed.xlsx
• manifest command: python3 tools/benchmark_manifest.py --workbook /path/to/benchmark_master_list_completed.xlsx
• the workbook is always an explicit input; repository tools never search a home directory for it