## Summary - add the RK3588 RKNPU backend/runtime foundation for the Relax/TIR path - add the residual-MLP MVP demo, native runner, and sweep tooling - add benchmark/probe references and the publishable hero demo docs
## Stack 1. `rknpu: add backend and runtime foundations` 2. `rknpu: add residual mlp mvp demo tools` 3. `rknpu: add benchmark and probe references` 4. `rknpu: add mvp demo documentation` ## Demo Scope - residual 2-D MLP with ReLU - real submit on RK3588 - bound params with `main(x)` runtime entry - native C++ runner for publishable latency numbers - pretty-printed compare artifact showing: - multi-submit non-fused - one-submit non-fused - one-submit fused ## Hero Artifact - `docs/arch/rknpu_mvp_hero_demo.md` - `docs/arch/rknpu_index.md` ## Notes - this branch is intentionally narrower than the full local RKNPU stack - it is rebuilt as a small stacked series on top of `origin/main` You can view, comment on, or merge this pull request online at: https://github.com/apache/tvm/pull/18916 -- Commit Summary -- * rknpu: add backend and runtime foundations * rknpu: add residual mlp mvp demo tools * rknpu: add benchmark and probe references * rknpu: add mvp demo documentation -- File Changes -- M CMakeLists.txt (3) M cmake/config.cmake (6) A cmake/modules/contrib/RKNPU.cmake (37) A docs/arch/rknpu_fusion_matrix_reference.md (132) A docs/arch/rknpu_hardware_block_guide.md (224) A docs/arch/rknpu_index.md (37) A docs/arch/rknpu_mvp_hero_demo.md (335) A docs/arch/rknpu_mvp_scope.md (95) A docs/arch/rknpu_overhead_reference.md (136) A docs/arch/rknpu_performance_reference.md (74) A docs/arch/rknpu_runtime_reuse_reference.md (218) A python/tvm/relax/backend/contrib/rknpu/__init__.py (310) A python/tvm/relax/backend/contrib/rknpu/codegen.py (2732) A python/tvm/relax/backend/contrib/rknpu/npu_core/__init__.py (32) A python/tvm/relax/backend/contrib/rknpu/npu_core/_cna_regcfg.py (847) A python/tvm/relax/backend/contrib/rknpu/npu_core/_dpu_regcfg.py (1012) A python/tvm/relax/backend/contrib/rknpu/npu_core/abstract.py (509) A python/tvm/relax/backend/contrib/rknpu/npu_core/alignment.py (14) A python/tvm/relax/backend/contrib/rknpu/npu_core/handles.py (30) A python/tvm/relax/backend/contrib/rknpu/npu_core/hardware.py (293) A python/tvm/relax/backend/contrib/rknpu/npu_core/layout.py (50) A python/tvm/relax/backend/contrib/rknpu/npu_core/lut_tables.py (364) A python/tvm/relax/backend/contrib/rknpu/npu_core/regcmd_gen.py (1010) A python/tvm/relax/backend/contrib/rknpu/patterns.py (475) A python/tvm/relax/backend/contrib/rknpu/schedule_pretty.py (249) A python/tvm/relax/backend/contrib/rknpu/tir_path.py (4278) A python/tvm/relax/backend/contrib/rknpu/transforms.py (256) A src/runtime/contrib/rknpu/rknpu_device.cc (364) A src/runtime/contrib/rknpu/rknpu_device.h (272) A src/runtime/contrib/rknpu/rknpu_runtime.cc (5244) A tools/rknpu_driver_probe.py (328) A tools/rknpu_mlp_mvp_gate.sh (42) A tools/rknpu_mlp_mvp_sweep.py (213) A tools/rknpu_performance_reference.py (2023) A tools/rknpu_tir_mlp_mvp_demo.py (914) A tools/rknpu_vm_cpp_runner.cc (405) A tools/rknpu_vm_cpp_runner.sh (33) -- Patch Links -- https://github.com/apache/tvm/pull/18916.patch https://github.com/apache/tvm/pull/18916.diff -- Reply to this email directly or view it on GitHub: https://github.com/apache/tvm/pull/18916 You are receiving this because you are subscribed to this thread. Message ID: <apache/tvm/pull/[email protected]>
