NNPACK

Author: Ludovic HENRY ludovic.henry@qti.qualcomm.com Date: 2026-06-17 Scope: RISC-V (riscv64/linux) support status for NNPACK Audience: Technical leadership, resource allocation strategy Verification policy: Every claim is cross-referenced to a primary upstream source. Items that could not be verified against a second source are marked [NEEDS VERIFICATION].

1. Project Overview

NNPACK (github.com/Maratyszcza/NNPACK) is an accelerated neural network inference library built around hand-tuned SIMD micro-kernels for convolution (FFT-based and Winograd), GEMM/GEMV, pooling, and activation functions. Its primary contribution was demonstrating that FFT-based convolution with batching could outperform direct convolution on x86 and ARM hardware in 2015-2016. It is written in C99/C++11 with architecture-specific backends generated via PeachPy (x86-64) or implemented as NEON C intrinsics (ARM/ARM64).

Governance. There is no formal governance structure. No MAINTAINERS, OWNERS, or CODEOWNERS file exists. The repository is a solo research project by Marat Dukhan (GitHub: Maratyszcza), created during his PhD at Georgia Institute of Technology (HPC Garage lab). There is no foundation affiliation, no steering committee, and no TSC. Decisions are made unilaterally by the owner. The project is not a member of the RISE Project, the Linux Foundation, or any other foundation.

Corporate sponsors. Marat Dukhan (Georgia Tech / Facebook AI Research) is the sole dominant contributor with 245 of approximately 260 total commits. Facebook AI Research provided production use and guidance (Nicolas Vasilache, Soumith Chintala, Andrew Tulloch); Andrew Tulloch contributed 1 commit. The US National Science Foundation funded the original work via Award Number 1339745. There is no formal corporate maintainer role.

Status. The project is effectively unmaintained. The last meaningful commit by Marat Dukhan was April 2020. A community typo-fix PR (#223) was opened July 2025 and has received no maintainer response. Issue creation is restricted. There are 44 open issues, 4 open PRs, and no GitHub releases have ever been published. Marat Dukhan moved active development to XNNPACK at Google, which is the designated successor and is actively maintained.

Community stance on new architecture ports. The README states that MIPS and MIPS64 are “not supported, and we have no plans to add it” while noting a “pull request would be welcome, though.” No RISC-V pull request has been filed. With the project in maintenance-only mode and no active maintainer responding to any issues or PRs, the de facto stance is that new ports are not being pursued.

2. Port History and Upstreaming Timeline

Date Event Source
2015 Initial release, x86-64 (AVX2) and ARM (NEON) only GitHub repo
2016-2017 ARM64 (AArch64 NEON) support added GitHub repo
2020 (approx.) Last meaningful commit by Marat Dukhan GitHub repo
2024-05-27 psimd dependency archived read-only github.com/Maratyszcza/psimd
2025-07 Community PR #223 (typo fix) opened; no maintainer response github.com/Maratyszcza/NNPACK/pull/223
2026-06 Zero RISC-V issues, PRs, or commits in repository GitHub issues/PR search

No RISC-V port has been attempted at any point. There is no first RISC-V commit, no tracking issue, no contributor working on it.

3. Upstream Support Tier

NNPACK has no formal tier policy. The recognized processor list in CMakeLists.txt implicitly defines the support tiers:

  • Tier 1 (fully supported): x86-64 (AVX2/FMA3), ARM (NEON), ARM64 (AArch64 NEON)
  • Tier 2 (portable fallback): Any processor that CMake does not recognize triggers either psimd or scalar backend via configure.py, but the CMakeLists.txt processor guard issues a fatal error for unrecognized processors (see Section 5).
  • riscv64: Not a recognized processor. Build aborts at configure time without patching.
Dimension amd64 arm64 riscv64
Hand-tuned backend Yes (PeachPy/AVX2) Yes (NEON C + .S asm) None
Runtime CPU detection Yes (cpuinfo, AVX2/FMA3) Yes (cpuinfo, NEON/FP16) None
CI coverage Yes (.travis.yml, scalar+psimd) No No
Official binary packages None (no releases) None (no releases) None
CMake recognizes processor Yes Yes No – fatal error
Maintainer acknowledgment Yes Yes No

4. Technical Architecture and RISC-V-Specific Subsystems

NNPACK’s architecture-specific code is organized into backend directories under src/:

Directory Architecture Implementation type File count
src/x86_64-fma/ x86-64 Hand-tuned PeachPy assembly + AVX2/FMA3 intrinsics 37
src/neon/ ARM 32/64-bit C intrinsics + 3 hand-written AArch32 .S files; runtime FP16/FMA detection 20
src/psimd/ Generic portable fallback Portable SIMD via psimd header library 30
src/scalar/ Emscripten only Pure C scalar 25
src/ref/ Reference Pure C reference implementations varies
riscv64 Not present None 0

Per-component RISC-V status:

Component amd64 arm64 riscv64
FFT convolution (8x8, 16x16) Full (PeachPy asm) Full (NEON C) Missing
Winograd convolution (8x8-3x3) Full (PeachPy asm) Full (NEON C) Missing
GEMM / GEMV (BLAS) Full (PeachPy asm) Full (NEON C + .S asm) Missing
ReLU / softmax / pooling Full (AVX2 C) Full (NEON C) Missing
Runtime CPU feature detection Full (cpuinfo, AVX2/FMA3) Full (cpuinfo, NEON/FP16) Missing
RVV (RISC-V Vector) kernels N/A N/A Missing

The #ifdef __riscv preprocessor guard appears zero times in the repository. No NNP_BACKEND_RISCV macro is defined. No .S assembly file, no RVV intrinsic (vfloat32m1_t, vle32_v, vle32_v, etc.), no Zba/Zbb/Zbc extension reference, and no arch/riscv/ directory exist anywhere in the tree.

If the CMakeLists.txt processor guard is patched (see Section 5), NNPACK on riscv64 would fall through to the psimd backend: a generic 128-bit portable SIMD abstraction with no hardware vector acceleration. Performance on riscv64 would be equivalent to running on any unoptimized scalar-class platform. The psimd dependency was archived read-only on 2024-05-27, so no RVV acceleration will ever land in that path.

5. Build System, Cross-Compilation, and Toolchain

Build system. CMake (minimum 2.8.12) plus Ninja, with architecture backends selected via configure.py and CMakeLists.txt. C99 and C++11 are required.

Critical blocker. CMakeLists.txt at line 60 contains a hard FATAL_ERROR for any processor not in its recognized list:

ELSEIF(NOT NNPACK_TARGET_PROCESSOR MATCHES
  "^(i686|x86_64|armv5te|armv7-a|armv7l|armv7|armv7s|aarch64|arm64|arm64e)$")
  MESSAGE(FATAL_ERROR
    "Unrecognized NNPACK_TARGET_PROCESSOR = ${NNPACK_TARGET_PROCESSOR}")

A native or cross-compiled build targeting riscv64 aborts at CMake configure time:

CMake Error: Unrecognized NNPACK_TARGET_PROCESSOR = riscv64

Required patch. Change FATAL_ERROR to WARNING to allow the build to proceed to the psimd/scalar backend. No workaround exists without patching the CMakeLists.txt.

Cross-compilation command after patching:

cmake -G Ninja \
  -DCMAKE_SYSTEM_NAME=Linux \
  -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
  -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc \
  -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++ \
  -DNNPACK_BACKEND=scalar \
  -DNNPACK_BUILD_TESTS=OFF \
  ..
ninja

Required -D flags:

Flag Value Reason
NNPACK_BACKEND scalar or psimd auto hits FATAL_ERROR; x86-64 and neon backends are not viable
NNPACK_BUILD_TESTS OFF Tests depend on cpuinfo and pthreadpool with their own riscv64 status
CMAKE_SYSTEM_NAME Linux The check for ^(Darwin|Linux|Android)$ must pass
CMakeLists.txt patch Required FATAL_ERROR is hardcoded; no cmake override bypasses it

Optional flags: -DNNPACK_INFERENCE_ONLY=ON reduces the source set by approximately half (removes training/backward pass code). -DNNPACK_CONVOLUTION_ONLY=ON further narrows scope. -DNNPACK_CUSTOM_THREADPOOL=ON allows supplying an external thread pool.

Toolchain requirements. No explicit GCC/Clang minimum is documented. C99 and C++11 are the language standards. Any current distribution-packaged riscv64-linux-gnu-gcc (GCC 10+) is sufficient.

QEMU. Not documented in the repository. Standard usage: QEMU_LD_PREFIX=/usr/riscv64-linux-gnu qemu-riscv64-static ./test/nnpack-tests. No QEMU-specific CI configuration exists.

Dockerfiles. None in the repository.

cmake/ directory contents. Only dependency download scripts: DownloadCpuinfo, DownloadFP16, DownloadFXdiv, DownloadPSimd, DownloadPThreadPool, DownloadPeachPy, DownloadGoogleTest, DownloadOpcodes, DownloadEnum, DownloadSix. No riscv64 toolchain file.

6. Feature Coverage and Gap Analysis vs arm64 and amd64

Feature amd64 arm64 riscv64
FFT convolution Full Full Not available (build fails without patch)
Winograd convolution Full Full Not available
GEMM/GEMV Full Full Not available
Pooling / activation Full Full Not available
Runtime CPU detection Full Full Not available
RVV acceleration N/A N/A Not available
nnp_initialize() return nnp_status_success nnp_status_success Would return nnp_status_unsupported_hardware on riscv64 even after patching, because src/init.c contains no RISC-V cpuinfo check
Build without patching Yes Yes No – CMake FATAL_ERROR

Functional gap. NNPACK cannot be used on riscv64 at all without patching the build system. Even after patching, nnp_initialize() returns nnp_status_unsupported_hardware because src/init.c has no code path for a RISC-V processor. All library functions return error codes without executing.

Performance gap. Even if the initialization gap were closed, the psimd fallback provides generic 128-bit portable SIMD with no RVV acceleration. No benchmark data for NNPACK on riscv64 exists anywhere. The TVM build report [NEEDS VERIFICATION - single source] for the SpacemiT K1 (RV64GCVB) from apache/tvm#17508 (November 2024) noted “libraries like MKL and NNPACK disabled” – NNPACK was excluded entirely rather than benchmarked.

Security hardening gaps. Data not available: no security hardening flags (stack protector, CFI, shadow call stack) are documented or referenced anywhere in the repository for any architecture.

Floating-point semantics. Data not available: no documentation or tests address NaN propagation or floating-point consistency across architectures.

7. CI/CD Infrastructure

The only CI configuration in the repository is .travis.yml at the repository root. Its full content:

language: c
compiler: clang
install:
  - git clone https://github.com/ninja-build/ninja.git /tmp/ninja
  - cd /tmp/ninja && git checkout release && python configure.py --bootstrap
  - mkdir -p $HOME/.local/bin && install -m 755 /tmp/ninja/ninja $HOME/.local/bin/ninja
  - export PATH=$HOME/.local/bin:$PATH
  - pip install --user git+https://github.com/Maratyszcza/PeachPy
  - pip install --user git+https://github.com/Maratyszcza/confu
before_script:
  - confu setup
  - python ./configure.py --toolchain=clang --backend=$BACKEND
  - ninja
script:
  - ninja smoketest
env:
  - BACKEND=psimd
  - BACKEND=scalar

No .github/workflows/ directory exists (GitHub API returns HTTP 404 for the path). No GitLab CI, Jenkinsfile, or Cirrus CI configuration exists. Travis CI targets x86-64 (the Travis CI host) only. The Travis CI badge in the README points to a service that is functionally defunct for open-source projects.

CI dimension amd64 arm64 riscv64
CI system exists Yes (Travis CI) No No
Backend tested psimd, scalar None None
Hardware tested x86-64 (Travis host) None None
QEMU testing No No No
RISE runners No No No
Release-blocking CI No No No

8. Distribution and Release Status

Channel riscv64 available? Notes
GitHub Releases No – no releases at all GitHub API returns empty array
PyPI No Only nnpack-0.1.0-py2-none-any.whl and .tar.gz from 2017-03-07; no native binary, pure-Python stub
RISE wheel builder No Proxies to PyPI; no separate packages
Ubuntu 24.04 (Noble) No NNPACK not packaged; XNNPACK is packaged as libxnnpack0/libxnnpack-dev for riscv64
Debian No Accepted into experimental (source + amd64 only); flagged “not part of any Debian distribution”; no riscv64 build record
Arch Linux RISC-V No Package not present in archriscv.felixc.at

What a user must do to get a working binary on riscv64. There is no path. NNPACK cannot be built for riscv64 without (a) patching CMakeLists.txt to remove the FATAL_ERROR on unrecognized processors, and (b) patching src/init.c to add a RISC-V cpuinfo code path so that nnp_initialize() does not return nnp_status_unsupported_hardware. Even after both patches, the library would use the psimd fallback with no hardware acceleration. The practical answer for riscv64 neural network acceleration is XNNPACK, which is packaged as libxnnpack0 in Ubuntu 24.04 for riscv64.

9. Dependencies

Summary Table

Dependency Role riscv64 build riscv64 test riscv64 release Notes
cpuinfo CPU feature detection Partial No (except QEMU CI merged via PR #219) No PR #397 (open, June 2026) adds full ISA + cache detection; unreviewed
pthreadpool Thread pool Likely (pure POSIX) No riscv64 CI N/A (header+source) Zero riscv64 issues; untested
FP16 Half-precision conversion Yes (scalar fallback) PR #45 open (June 19, 2026): riscv64 QEMU CI, 7/7 passing, not yet merged N/A (header-only) No RVV acceleration
FXdiv Fixed-point integer division Likely (arithmetic only) No riscv64 CI N/A (header-only) Recognized arch list excludes riscv64 but no SIMD involved
psimd Portable SIMD fallback Unknown No riscv64 CI Archived 2024-05-27 NNPACK’s riscv64 fallback path; archived means no RVV support ever
PeachPy x86-64 assembly code generator N/A N/A N/A Invoked only for x86-64 target; not involved on riscv64
googletest Unit testing (build-time only) Full Tested upstream Released for riscv64 No blocker

Dependency Deep-Dives

cpuinfo. cpuinfo is the most significant dependency for RISC-V enablement. The version bundled with NNPACK is a fork tracking pytorch/cpuinfo. A basic struct compiles for riscv64 (PR #190 merged), and QEMU CI was added (PR #219 merged). PR #397 (open, June 2026) adds full ISA extension detection (Zba, Zbb, Zbc, Zbs, V, Zvfh, etc.), vendor/uarch identification, and cache topology detection, but it is unreviewed and awaiting hardware validation on real silicon. Critically, NNPACK’s own CMakeLists.txt does not include riscv64 in its recognized processor list, so cpuinfo results would not be used even if the detection worked correctly.

psimd. psimd is the portable SIMD library that serves as NNPACK’s fallback backend for all non-x86/non-ARM targets. It was archived read-only on 2024-05-27. This means the only fallback path for riscv64 is permanently frozen at the portable 128-bit SIMD level with no possibility of RVV acceleration being added. The psimd path does compile on any POSIX system [NEEDS VERIFICATION - the archive and lack of CI means this has not been tested on riscv64 in any documented context].

FP16. The FP16 library’s scalar fallback path functions correctly on riscv64 under QEMU according to PR #45 (7/7 tests passing), but this PR was opened June 19, 2026 and is not yet merged. No native RVV fp16 acceleration exists.

pthreadpool. pthreadpool is a pure POSIX threading library with no architecture-specific code. It should compile and function correctly on riscv64 but has no documented test coverage for that architecture.

11. Known Bugs and Active Issues

The repository has 44 open issues. None reference RISC-V. The full open issue list relevant to correctness and builds:

Issue Title Category Severity
#222 Build fails: missing fp16/psimd.h when compiling psimd/blas/shdotxf.c Build bug High – blocks psimd backend
#221 “Unsupported hardware on supported CPU” Architecture detection High
#219 FP16 python module error at make Build bug Medium
#218 SIGFPE with nosmt kernel parameter Correctness bug (x86) High (x86-specific)
#216 “Unsupported hardware” error Architecture detection High
#212 Unsupported hardware on MacBook Pro 15 (late 2012) Architecture detection Medium
#211 Use CPack for packaging Enhancement Low
#209 Section address out of range for architecture x86_64 Build/arch bug Medium
#207 Unsupported Hardware on VM with compatible CPU Architecture detection High
#203 ModuleNotFoundError: No module named ‘peachpy.x86_64.avx’ Build bug Medium
#202 Build failed: cos_npi_over_8 not available in common Build bug Medium
#156 Cache/blocking sizes hardcoded for non-x86 targets Performance bug Medium – affects ARM and any future port

Issue #222 directly blocks the psimd backend that riscv64 would rely on: the psimd header include path is broken in the build. This would need to be fixed in addition to the CMakeLists.txt processor guard patch for a riscv64 build using the psimd backend to succeed.

Issue #156 (opened October 2018, no response) is relevant context: non-x86 architectures use hardcoded cache and blocking sizes rather than dynamic cpuinfo-based detection. Any riscv64 port would face the same problem, since cpuinfo does not yet have stable riscv64 cache topology detection.

Correctness bugs specific to riscv64: None filed, because no riscv64 work has been attempted.

12. Objections and Upstream Blockers

Stated objections. None explicitly for RISC-V. The project has no active maintainer to raise or resolve objections.

Technical blockers:

  1. CMakeLists.txt FATAL_ERROR on unrecognized processors. Requires a one-line patch to proceed.
  2. src/init.c has no RISC-V code path. nnp_initialize() returns nnp_status_unsupported_hardware without patching.
  3. psimd (the fallback backend) is archived. No future RVV support can be added to it. A riscv64 port with hardware acceleration would require writing a new src/rvv/ backend from scratch.
  4. cpuinfo riscv64 detection is incomplete (PR #397 open, unreviewed, no hardware validation).
  5. Issue #222 breaks the psimd backend build.
  6. The project has no active maintainer. Any patches submitted have a realistic acceptance probability near zero given the 4 open PRs (from 2021 to 2025) with no maintainer response.

Organizational blockers. No corporate sponsor is maintaining NNPACK. RISE Project has no involvement. Marat Dukhan (the sole maintainer) is at Google working on XNNPACK. There is no path to getting patches reviewed or merged absent forking the project.

Acceptance probability. Effectively zero for upstream acceptance. The project is in read-only maintenance mode.

13. Investment Analysis

NNPACK is a legacy project with no viable path to riscv64 support upstream. Its successor, XNNPACK, already supports riscv64 in Ubuntu 24.04 and has active RVV kernel development. Any investment in NNPACK for RISC-V would be spent on a dead-end rather than the active replacement. The analysis below covers what would be required if there were a business reason to use NNPACK specifically (e.g., an existing deployment that cannot be migrated to XNNPACK).

RISE has not funded any work on NNPACK and has no known plans to do so.

13.1 Functional Enablement

To reach a functional (non-accelerated) riscv64 build:

  • Patch CMakeLists.txt to remove the FATAL_ERROR on unrecognized processors (1 line).
  • Patch src/init.c to add a RISC-V cpuinfo code path and set psimd function pointers.
  • Fix issue #222 (broken psimd include path) so the psimd backend compiles.
  • Validate that pthreadpool, FP16, FXdiv, and psimd all compile for riscv64 (mostly trivial given their portable nature).
  • Verify nnp_initialize() succeeds and all compute functions return correct results under QEMU.

13.2 Performance Optimization

To add RVV (RISC-V Vector) acceleration:

  • Write a new src/rvv/ backend directory with RVV micro-kernels for the critical paths: FFT convolution, Winograd convolution, GEMM/GEMV, pooling.
  • This mirrors the existing 37-file src/x86_64-fma/ or 20-file src/neon/ backends in scope.
  • Add riscv64 cpuinfo detection to src/init.c to select RVV paths at runtime based on the V extension.
  • This work is only meaningful if done on XNNPACK instead – which already has RVV infrastructure in place.

13.3 CI/CD Infrastructure

  • Add a .github/workflows/riscv64.yml using QEMU (riscv64) to run the smoketest suite.
  • This requires the functional enablement work first.
  • Travis CI is defunct for open-source; GitHub Actions would be the replacement.

13.4 Ecosystem Enablement

NNPACK has no significant dependent package ecosystem requiring separate enablement (it is a C library consumed directly by a small number of frameworks, primarily PyTorch as a legacy fallback path). Section 10 is omitted.

13.5 Summary Table

Area Work Item Effort (person-weeks) Owner Priority
Functional CMakeLists.txt processor guard patch 0.1 Any Critical (prerequisite)
Functional src/init.c RISC-V cpuinfo code path 0.5 Any Critical (prerequisite)
Functional Fix issue #222 psimd include path 0.2 Any Critical (prerequisite)
Functional Validate psimd/scalar backend on riscv64 under QEMU 0.5 Any Critical
CI/CD Add .github/workflows/riscv64.yml (QEMU) 0.5 Any High
Performance Write src/rvv/ backend (FFT, Winograd, GEMM, pooling) 12-16 RISC-V SIMD specialist Low – invest in XNNPACK instead
Performance cpuinfo riscv64 cache detection (review PR #397) 1 cpuinfo maintainer Medium

Recommendation. Do not invest in NNPACK riscv64. The functional enablement work (approximately 1.3 person-weeks) produces a non-accelerated library on a dead project with no upstream acceptance path. The performance optimization work (12-16 person-weeks) duplicates effort that has already been done in XNNPACK. Direct any RISC-V neural network inference acceleration investment to XNNPACK, which has existing RVV infrastructure, active maintainership (Google), Ubuntu 24.04 packaging as libxnnpack0, and an active riscv64 CI.

14. Updates

No updates yet – initial report dated 2026-06-17.

15. References