xtsis-nn

xtsis-nn software library is a collection of efficient neural network kernels developed to maximize the performance and minimize the memory footprint of neural networks on XuanTie processors.

Documentation

• Architecture — Architecture, module organization, DSP optimization strategy, build system, and test framework
• API Reference — Complete public type definitions and function API reference

Supported Framework

The library follows the int8 and int16 quantization specification of TensorFlow Lite for Microcontrollers.

Architecture Support

Implementation Path	Description
Pure C	Generic implementation for all RISC-V processors
RISC-V P Extension DSP	Optimized for processors with P extension (zpn v0.9.4), auto-enabled via RISCV_MATH_DSP macro

Operator Support

Operator	C int8	C int16	C int4*	DSP int8	DSP int16	DSP int4*
Conv2D	Yes	Yes	Yes	Yes	Yes	Yes
DepthwiseConv2D	Yes	Yes	Yes	Yes	Yes	Yes
TransposeConv2D	Yes	No	No	Yes	No	No
GroupedConv2D	Yes	No	No	No	No	No
Fully Connected	Yes	Yes	Yes	Yes	Yes	Yes
Batch Matmul	Yes	Yes	No	Yes	Yes	No
Add	Yes	Yes	N/A	Yes	Yes	N/A
Mul	Yes	Yes	N/A	Yes	Yes	N/A
Minimum	Yes	No	N/A	Yes	No	N/A
Maximum	Yes	No	N/A	Yes	No	N/A
MaxPooling	Yes	Yes	N/A	No	Yes	N/A
AvgPooling	Yes	Yes	N/A	Yes	Yes	N/A
Softmax	Yes	Yes	N/A	Yes	No	N/A
Activation	Yes	Yes	N/A	Yes	Yes	N/A
LSTM	Yes	Yes	No	Yes	Yes	No
SVDF	Yes	No	No	Yes	No	No
Pad	Yes	No	N/A	No	No	N/A
Reshape	Yes	No	N/A	No	No	N/A
Transpose	Yes	No	N/A	No	No	N/A
Concatenation	Yes	No	N/A	No	No	N/A

* int4 weights + int8 activations

Getting Started

Prerequisites

• RISC-V GCC toolchain (with Xuantie e907 / P extension support)
• CMake >= 3.15

Building the Library

Using the bare-metal toolchain (riscv64-unknown-elf-gcc) for RTOS integration:

# Configure
cmake -S . -B build \
  -DCMAKE_SYSTEM_NAME=Generic \
  -DCMAKE_TRY_COMPILE_TARGET_TYPE=STATIC_LIBRARY \
  -DCMAKE_C_COMPILER=riscv64-unknown-elf-gcc \
  -DCMAKE_C_FLAGS="-mcpu=e907fdp -mabi=ilp32"

# Build
cmake --build build

The output libxtsis-nn.a can be linked into your RTOS project.

Testing

Note: Unit tests use QEMU user-mode emulation for functional verification, providing a convenient way to validate correctness without hardware. Although the bare-metal toolchain (riscv64-unknown-elf-gcc) is used for compilation, test executables are statically linked with newlib's Linux syscall stubs and run directly under qemu-riscv32 user-mode (not QEMU system-mode). This approach avoids the need for linker scripts, startup code, and UART drivers, making it easy to verify operator correctness in any development environment. The test binaries are not RTOS-style bare-metal ELFs — they behave like Linux executables solely for the purpose of functional verification. The xtsis-nn library itself has no OS dependency and can be linked into any RTOS or bare-metal project.

Additional Prerequisites for Testing

• QEMU (riscv32, with e907fdp cpu model)
• Python 3 (for generating test runners)

Running All Unit Tests

# 1. Configure
cmake -S Tests/UnitTest -B build-test \
  -DBUILD_XTSIS_NN_UNIT=ON \
  -DXTSIS_NN_PATH=$(pwd) \
  -DCMAKE_SYSTEM_NAME=Generic \
  -DCMAKE_TRY_COMPILE_TARGET_TYPE=STATIC_LIBRARY \
  -DCMAKE_C_COMPILER=riscv64-unknown-elf-gcc \
  -DCMAKE_CXX_COMPILER=riscv64-unknown-elf-g++ \
  -DCMAKE_C_FLAGS="-mcpu=e907fdp -mabi=ilp32 -static" \
  -DCMAKE_CXX_FLAGS="-mcpu=e907fdp -mabi=ilp32 -static" \
  -DCMAKE_CROSSCOMPILING_EMULATOR="qemu-riscv32;-cpu;e907fdp" \
  -DCMAKE_POLICY_VERSION_MINIMUM=3.5

# 2. Build all tests
cmake --build build-test --target xtsis_nn_unit_tests

# 3. Run all tests
ctest --test-dir build-test --output-on-failure

DSP Optimization Details

When the compiler targets a RISC-V processor with P extension support (__riscv + one of __riscv_dsp / __riscv_p / __riscv_zpn), the RISCV_MATH_DSP macro is automatically defined in Include/riscv_nn_math_types.h, enabling P extension DSP instruction acceleration:

DSP Instruction	Purpose	Wrapper Macro
kmada	Dual 16-bit multiply-accumulate	SMLAD
smbb16 / smtt16	16-bit half-word multiply	SMULBB / SMULTT
kmabb / kmatt	16-bit multiply-accumulate (single half-word)	SMLABB / SMLATT
smaqa	Signed int8×4 quad multiply-accumulate	SMAQA
smaqa.su	Signed×Unsigned int8×4 quad MAC	SMAQA_SU
kwmmul.u	Saturating doubling high-word multiply	KWMMUL_U
sunpkd820	Sign-extend byte to half-word (byte0, byte2)	SXTB16
sunpkd810	Sign-extend byte to half-word (byte0, byte1)	SUNPKD810
zunpkd810	Zero-extend byte to half-word	ZUNPKD810
add16	Parallel 16-bit addition	SADD16
pkbb16 / pkbt16 / pktt16 / pktb16	Half-word packing	PKHBT / PKHTB
kaddw	Saturating addition	QADD
smax8 / smin8	Parallel byte-level max/min	SMAX8 / SMIN8
smax16 / smin16	Parallel 16-bit max/min	SMAX16 / SMIN16

Unit Test Guide

The test framework is based on Unity. Test cases are located under Tests/UnitTest/TestCases/, with each operator having a corresponding test_riscv_* directory.