vitis.rst

Vitis

Xilinx Vitis-AI is a platform for high-performance deep learning compression and inference on Xilinx FPGA device.

Quantization Scheme

8bit per-tensor symmetric linear quantization with power of two scales.

For weights/biases:

q = clamp(⌊x * s⌉, lb, ub)

For activations:

$$\begin{aligned} \begin{equation} \begin{aligned} q &= \mathtt{clamp}(\lceil x * s \rceil, lb, ub), \text{ where } x*s-\lfloor x*s\rfloor = 0.5 \text{ and } x < 0 \\ q &= \mathtt{clamp}(\lfloor x * s \rceil, lb, ub), \text{ else}. \end{aligned} \end{equation} \end{aligned}$$

where s is scaling factor to quantize a number from floating range to integer range, lb and ub are bounds of integer range, and [lb, ub] = [-128, 127].

Supported Operations

Vitis-AI provides a software stack for DPUs of Xilinx FPGA, including network quantization and DPU inference etc.. To apply MQBench to Vitis-AI, we support a backend and its related observer and fake quantizer which is proven to be aligned with DPU. For now, we have supported typical operations used in classification and detection jobs: The supported Operations are listed here:

• Conv2d
• ReLU
• MaxPooling2d
• Concat
• Input
• Add
• Resize(Interpolate)
• Flatten
• Gemm(Linear)
• GlobalAveragePool