We have implemented a library (lib_nn) of efficient neural network functions developed to maximize the performance and minimize the memory footprint of neural network inference on XMOS xcore.ai.
- xTIMEcomposer Tools Version 15.0.0 or later
Only XS3 based microcontrollers are supported with this library. The previous generation XS1 and XS2 based microcontrollers are not supported.
XS3 based microcontrollers, like xcore.ai, have a vector unit with 256 bit wide registers and can operate in 8bit, 16bit or 32bit integer mode.
This document assumes familiarity with the XMOS xCORE architecture, the XMOS tool chain, the 'C' programming language, and neural network concepts.
For an XCore build: mkdir -p build_xcore cd build_xcore cmake -DCMAKE_TOOLCHAIN_FILE=../etc/xmos_toolchain.cmake .. make
For an x86 build: mkdir -p build_x86 cd build_x86 cmake .. make
The table below gives a quick overview of the APIs in lib_nn. Unless otherwise noted, all kernels below operate on signed 8-bit input and output tensors. The following symbols are used:
- Cin - The number of input channels
- Cout - The number of output channels
- Kh - The kernel, filter or pool height
- Kw - The kernel, filter or pool width
- Sh - The stride height
- Sw - The stride width
For full documentation of each API function, please refer to the description in the lib_nn/api/nn_operator.h header file.
| Group | API | VPU Optimized | Constraints | Comments |
|---|---|---|---|---|
| Convolution | ||||
| conv2d_deep | Yes | Cin % 4 = 0, Cout % 4 = 0 | ||
| conv2d_shallowin | Yes | Cin % 4 = 0, Cout % 4 = 0, Cin * Kw = 32 | ||
| conv2d_1x1 | Yes | Cin % 4 = 0, Cout % 4 = 0, Kh = Kw = 1, Sh = Sw = 1 | ||
| conv2d_depthwise | Yes | Cin % 4 = 0, Cout % 4 = 0, Cin = Cout | ||
| Fully Connected | ||||
| fully_connected_8 | Yes | Cin % 4 = 01 | Output is 8-bit | |
| fully_connected_16 | Yes | Cin % 4 = 01 | Output is 16-bit | |
| Pooling | ||||
| maxpool2d | Yes | Cin % 4 = 0 | ||
| avgpool2d | Yes | Cin % 4 = 0 | ||
| avgpool2d_global | Yes | Cin % 4 = 0 | ||
| Argmax | ||||
| argmax_16 | No | Input is rank-1 | Input is 16-bit | |
| Activations | ||||
| lookup8 | No | None | Logistic (sigmoid), tanh & ReLU activation functions can be implemented using a look-up table mapping 8-bit inputs to 8-bit outputs | |
| Misc | ||||
| add_elementwise | Yes | None | ||
| requantize_16_to_8 | Yes | None | Reduces the bit depth of a vector with 16-bit elements to a vector of 8-bit elements |
1It is possible to relax this constraint. See the documentation for the API function.