Google Git Google Git
Sign in
chromium/external/github.com/google/XNNPACK/ebd43ff35e32b74f6c070e0593cb6e6c9d77c590
commit
ebd43ff35e32b74f6c070e0593cb6e6c9d77c590
[log]
author
Frank Barchard <fbarchard@google.com>
Fri Oct 09 01:40:05 2026
committer
XNNPACK Team <xnnpack-github-robot@google.com>
Fri Oct 09 01:40:58 2026
tree
e8d9797ae199ecfbc9fd9b3d5e9ace002c2a26de
parent
0182e24d8afb5278d675831f09b7921d71919a37 [diff] 
Optimize AArch64 NEON `qs8`/`qu8 -> f32` (`TBL2 + FSUB`) and `bf16 -> f32` (`ZIP1`/`ZIP2` with zero) conversions in XNNPACK and YNNPACK.

```bash
bazel build -c opt //bench:vunary_bench //ynnpack/kernels/unary:bench //ynnpack/kernels/ternary:bench
bazel-bin/bench/vunary_bench --benchmark_filter="xnn_(qs8_f32|qu8_f32|bf16_f32)_vcvt_ukernel__neon_u32" --benchmark_min_time=100x
bazel-bin/ynnpack/kernels/unary/bench --benchmark_filter="convert_bf16_to_fp32_neon" --benchmark_min_time=100x
bazel-bin/ynnpack/kernels/ternary/bench --benchmark_filter="dequantize_(int8|uint8)_to_fp32_neon" --benchmark_min_time=100x
```

XNNPACK `vunary_bench` (`neon_u32`, `N=8192`):
```text
CPU Architecture              qs8 -> f32    qu8 -> f32   bf16 -> f32
--------------------------  ------------  ------------  ------------
ARM Neoverse-N1             +22.5% (5.3G) +31.8% (5.7G) +21.4% (11G)
ARM Neoverse-V2 (Axion)     +22.5% (5.3G) +31.7% (5.7G) +19.1% (11G)
NVIDIA Grace (Neoverse-V2)  +22.3% (11G)  +33.3% (12G)  -32.9% (17G)
Ampere Siryn                -11.1% (4.8G)  +0.8% (5.5G)   0.0% (12G)
ARM Cortex-A53              -17.1% (0.1G)  -4.3% (0.3G)  +2.1% (0.3G)
```

YNNPACK `unary` / `ternary` (`neon`, `N=4096`):
```text
CPU Architecture             int8 -> f32   uint8 -> f32  bf16 -> f32
--------------------------  ------------  ------------  ------------
ARM Neoverse-N1             +17.6% (3.2G) +26.5% (3.4G)  +8.5% (11G)
ARM Neoverse-V2 (Axion)     +17.7% (3.2G) +26.2% (3.4G)  +8.1% (11G)
NVIDIA Grace (Neoverse-V2)  +25.0% (6.8G) +35.7% (7.3G)  +7.4% (21G)
Ampere Siryn                -15.0% (3.2G)  -6.8% (3.6G)  +5.6% (12G)
ARM Cortex-A53               -9.2% (0.1G)  -8.5% (0.1G)  +0.1% (0.7G)
```

PiperOrigin-RevId: 996203284
16 files changed
tree: e8d9797ae199ecfbc9fd9b3d5e9ace002c2a26de
  1. .github/
  2. bench/
  3. build_config/
  4. build_overrides/
  5. cmake/
  6. doc/
  7. docker/
  8. gemm_compiler/
  9. gen/
  10. include/
  11. litert/
  12. scripts/
  13. src/
  14. test/
  15. third_party/
  16. tools/
  17. ynnpack/
  18. .bazelrc
  19. .clang-format
  20. .gitignore
  21. .gn
  22. BUILD.bazel
  23. BUILD.gn
  24. BUILD.md
  25. build_defs.bzl
  26. build_params.bzl
  27. build_srcs.bzl
  28. CMakeLists.txt
  29. CONTRIBUTING.md
  30. cpuinfo_windows_arm64_msvc.patch
  31. DEPS
  32. emscripten.bzl
  33. generated_file.bzl
  34. LICENSE
  35. MODULE.bazel
  36. preamble.js.lds
  37. README.md
  38. register_extension_info.bzl
README.md

XNNPACK

XNNPACK is a highly optimized solution for neural network inference on ARM, x86, WebAssembly, and RISC-V platforms. XNNPACK is not intended for direct use by deep learning practitioners and researchers; instead it provides low-level performance primitives for accelerating high-level machine learning frameworks, such as TensorFlow Lite, TensorFlow.js, PyTorch, ONNX Runtime, ExecuTorch, and MediaPipe.

Supported Architectures

  • ARM64 on Android, iOS, macOS, Linux, and Windows
  • ARMv7 (with NEON) on Android
  • ARMv6 (with VFPv2) on Linux
  • x86 and x86-64 (up to AVX512) on Windows, Linux, macOS, Android, and iOS simulator
  • WebAssembly MVP
  • WebAssembly SIMD
  • WebAssembly Relaxed SIMD (experimental)
  • RISC-V (RV32GC and RV64GC)
  • Hexagon (with HVX)

Operator Coverage

XNNPACK implements the following neural network operators:

  • 2D Convolution (including grouped and depthwise)
  • 2D Deconvolution (AKA Transposed Convolution)
  • 2D Average Pooling
  • 2D Max Pooling
  • 2D ArgMax Pooling (Max Pooling + indices)
  • 2D Unpooling
  • 2D Bilinear Resize
  • 2D Depth-to-Space (AKA Pixel Shuffle)
  • Add (including broadcasting, two inputs only)
  • Subtract (including broadcasting)
  • Divide (including broadcasting)
  • Maximum (including broadcasting)
  • Minimum (including broadcasting)
  • Multiply (including broadcasting)
  • Squared Difference (including broadcasting)
  • Global Average Pooling
  • Channel Shuffle
  • Fully Connected
  • Abs (absolute value)
  • Bankers' Rounding (rounding to nearest, ties to even)
  • Ceiling (rounding to integer above)
  • Clamp (includes ReLU and ReLU6)
  • Convert (includes fixed-point and half-precision quantization and dequantization)
  • Copy
  • ELU
  • Floor (rounding to integer below)
  • HardSwish
  • Leaky ReLU
  • Negate
  • Sigmoid
  • Softmax
  • Square
  • Tanh
  • Transpose
  • Truncation (rounding to integer towards zero)
  • PReLU

All operators in XNNPACK support NHWC layout, but additionally allow custom stride along the Channel dimension. Thus, operators can consume a subset of channels in the input tensor, and produce a subset of channels in the output tensor, providing a zero-cost Channel Split and Channel Concatenation operations.

Performance

Mobile phones

The table below presents single-threaded performance of XNNPACK library on three generations of MobileNet models and three generations of Pixel phones.

ModelPixel, msPixel 2, msPixel 3a, ms
FP32 MobileNet v1 1.0X828688
FP32 MobileNet v2 1.0X495355
FP32 MobileNet v3 Large394244
FP32 MobileNet v3 Small121414

The following table presents multi-threaded (using as many threads as there are big cores) performance of XNNPACK library on three generations of MobileNet models and three generations of Pixel phones.

ModelPixel, msPixel 2, msPixel 3a, ms
FP32 MobileNet v1 1.0X432746
FP32 MobileNet v2 1.0X261828
FP32 MobileNet v3 Large221624
FP32 MobileNet v3 Small768

Benchmarked on March 27, 2020 with end2end_bench --benchmark_min_time=5 on an Android/ARM64 build with Android NDK r21 (bazel build -c opt --config android_arm64 :end2end_bench) and neural network models with randomized weights and inputs.

Raspberry Pi

The table below presents multi-threaded performance of XNNPACK library on three generations of MobileNet models and three generations of Raspberry Pi boards.

ModelRPi Zero W (BCM2835), msRPi 2 (BCM2836), msRPi 3+ (BCM2837B0), msRPi 4 (BCM2711), msRPi 4 (BCM2711, ARM64), ms
FP32 MobileNet v1 1.0X39193021147277
FP32 MobileNet v2 1.0X1987191794146
FP32 MobileNet v3 Large1658161673840
FP32 MobileNet v3 Small47450221315
INT8 MobileNet v1 1.0X2589128462924
INT8 MobileNet v2 1.0X149582302017

Benchmarked on Feb 8, 2022 with end2end-bench --benchmark_min_time=5 on a Raspbian Buster build with CMake (./scripts/build-local.sh) and neural network models with randomized weights and inputs. INT8 inference was evaluated on per-channel quantization schema.

Minimum build requirements

  • C11
  • C++17
  • Python 3

Publications

Ecosystem

Machine Learning Frameworks

Acknowledgements

XNNPACK is based on QNNPACK library. Over time its codebase diverged a lot, and XNNPACK API is no longer compatible with QNNPACK.