How google/xnnpack works
XNNPACK is a low-level neural-network inference library that supplies optimized operators to higher-level frameworks. A typical graph-based workflow defines tensor values and operator nodes, builds a runtime, then reshapes, sets up, and executes the selected operators to produce output tensors. The repository also contains direct operator APIs, architecture-specific kernels, weight packing, and a LiteRT tensor backend. The README documents broad operator and architecture coverage; sampled code verifies the average-pooling path, while several wider runtime links remain unverified.
Main entry points: xnnpack.h XNNPACK API · subgraph.c Subgraph Definition · configs Kernel Configurations · average-pooling-nhwc.c Average Pooling
Components 17
Public API
- XNNPACK API
include/xnnpack.h - Experimental API
include/experimental.h
Graph Runtime
- Subgraph Definition
src/subgraph.c - Average Pool Node
src/subgraph/average-pooling-2d.c - Runtime Execution
src/xnnpack/subgraph.h
Operator Implementations
- Operator API
include/xnnpack.h - Average Pooling
src/operators/average-pooling-nhwc.c - Operator State
src/xnnpack/operator.h - Operator Coverage
src/operators
Kernels And Dispatch
- Kernel Configurations
src/configs - Dense And Quantized GEMM
src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-8x32-minmax-neoni8mm.c - Sparse Matrix Kernels
src/f16-spmm/gen/f16-spmm-32x1-minmax-neonfp16arith-pipelined.c - Weight Packing
src/reference/packing.cc
Framework Integrations
- LiteRT Tensor API
litert/tensor/arithmetic.h - LiteRT XNNPACK Backend
litert/tensor/backends/xnnpack/arithmetic.cc - LiteRT Runner
litert/tensor/runners/xnnpack/runner.cc
Other components
- ML Framework
Connections 16
- ML Framework → XNNPACK API: calls library
- ML Framework → Subgraph Definition: defines graph
- XNNPACK API → Subgraph Definition: defines nodes
src/subgraph.c - Experimental API → Subgraph Definition: extends graph API
src/subgraph.c - Subgraph Definition → Average Pool Node: dispatches node
src/subgraph/average-pooling-2d.c - Average Pool Node → Average Pooling: creates and sets up
src/subgraph/average-pooling-2d.c - XNNPACK API → Operator API: exposes operations
include/xnnpack.h - Operator API → Average Pooling: creates and runs
src/operators/average-pooling-nhwc.c - Runtime Execution → Operator State: manages lifecycle
- Average Pooling → Kernel Configurations: selects kernels
- Kernel Configurations → Dense And Quantized GEMM: dispatches kernels
- Kernel Configurations → Sparse Matrix Kernels: dispatches kernels
- Weight Packing → Dense And Quantized GEMM: prepares weights
- XNNPACK API → Operator Coverage: documents operators
src/operators/average-pooling-nhwc.c - LiteRT Tensor API → LiteRT XNNPACK Backend: delegates arithmetic
- LiteRT XNNPACK Backend → LiteRT Runner: executes backend