Skip to content

build.md

Latest commit

 

History

History
105 lines (82 loc) · 6.13 KB

build.md

File metadata and controls

105 lines (82 loc) · 6.13 KB
Raw

Build the accelerator

简体中文

The main repository contains the acceleration addon and the transformations used to build its 47 selected Ada kernels. Supply the NR runtime from your local RenoDX DLSS5 installation; the build extracts the needed PTX locally, applies the recipes, compiles the addon, tests its routing, and produces an installable ZIP.

Requirements

  • Windows x64, Python 3.11 or later (standard library only).
  • Visual Studio C++ desktop build tools with a Windows SDK.
  • CUDA Toolkit with ptxas.exe and cuobjdump.exe. The reproduction below used CUDA 13.3; generated PTX uses ISA version 9.3 and targets sm_89.
  • NGX SDK headers: point -NgxInclude to the directory containing nvsdk_ngx_params.h and nvsdk_ngx_defs.h, available with the NVIDIA DLSS SDK.
  • The local community nvngx_dlssnr.dll containing the supported NR kernels.

One build command

Open PowerShell in the repository or extracted source archive:

powershell.exe -NoProfile -ExecutionPolicy Bypass -File tools/build_local.ps1 `
  -RuntimePath 'D:\Games\YourGame\nvngx_dlssnr.dll' `
  -NgxInclude 'D:\SDK\ngx\include' `
  -CudaBin 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\bin'

Use the actual runtime and SDK locations on your machine. -CudaBin can be omitted when CUDA_PATH points to the toolkit. MSVC is discovered automatically; an x64 Native Tools shell also works.

The final line prints the location of build/local-<timestamp>/Faster-DLSS5-Ada.zip. Extract that ZIP and run Start.cmd to install over the working community base. Compilation itself does not install anything or modify the runtime supplied as input. Players receiving a complete ZIP only need the normal installer.

Each build uses a new directory. It contains extracted inputs, transformed PTX, 47 cubins, compiler logs, generated dispatch headers, the compiled addon, test results and the final ZIP. These are local build outputs.

Where the acceleration code lives

Source Responsibility
src/integration/ada_nr_addon.cpp Runtime integration, loading the local images, replacement dispatch and Evaluate forwarding.
src/optimization/dynamic_contract.h Native qualification, parameter and launch checks, dynamic dimensions, Pre/Post guards and fallback.
config/network.json Scalar ABI facts, 156-call sequence, 145 interior routes and selected shape specializations.
config/kernel-recipes.json The 47 selected recipes: pairing, register limits, block bounds, shared spilling and scalar folds.
kernel_pipeline.py Applies each recipe and validates its input contract.
ptx_ada_semantic.py Ada asynchronous-copy and completion adaptation; component reduction and fence lowering.
ptx_qkv512.py, ptx_projection_pairs.py Move proven fragment loads closer to their consumers in QKV and projection kernels.
ptx_ffn256_n64.py, ptx_boundary_n64.py Pair N64 computation groups for selected FFN and boundary kernels.
ptx_defer_shared.py Shorten shared-load live ranges across independently verified instructions.
ptx_shape_specialize.py Fold selected scalar parameters while retaining the declared dynamic dimensions.

The technical report explains the resource tradeoffs and timing evidence. The recipe table contains the settings selected for the acceleration bundle.

Build individual components

For development, use separate output directories:

python tools/extract_local_ptx.py --runtime 'D:\Games\YourGame\nvngx_dlssnr.dll' --output build/input --cuobjdump "$env:CUDA_PATH/bin/cuobjdump.exe"
python tools/build_kernels.py --ptx-dir build/input --output build/kernels --ptxas "$env:CUDA_PATH/bin/ptxas.exe"
powershell.exe -NoProfile -ExecutionPolicy Bypass -File tools/build_addon.ps1 -NgxInclude 'D:\SDK\ngx\include' -Bundle build/kernels -Output build/addon
python tools/package_accelerator.py --addon build/addon/ada-nr.addon64 --bundle build/kernels --output build/Faster-DLSS5-Ada.zip

build_kernels.py --only integrated-12.cubin builds one recipe for inspection; repeat --only to select more. Packaging needs all 47 images. --jobs controls parallel compilation (default 4).

Recipe matching checks entry parameters, copy sizes, instruction dependencies and the regions being transformed. It does not require a whole-runtime hash match. A change to those contracts produces a specific build error so the corresponding recipe can be updated. Build receipts record input/output identities for traceability; the addon is compiled against the images produced by that build.

Reproduction and checks

The source build extracted 15 PTX modules from the tested community runtime and rebuilt 47/47 cubins byte for byte identically to the retained acceleration bundle with CUDA 13.3. This establishes reproduction of the selected GPU code. The addon compiled from this repository passed synthetic native-route tests and the assembly audit that checks Evaluate forwarding remains a tail jump.

See the reproduction receipt and verification record. Existing GPU correctness and timing measurements remain documented separately in the benchmark and technical reports.

python -m unittest discover -s tests -p test_*.py
powershell.exe -NoProfile -File tests/manager.Tests.ps1
python tools/package_preview.py

The source archive includes our addon, transforms, metadata and tests. Vendor runtime, PTX, SASS, weights and generated GPU binaries are obtained or produced locally; see package contents.