Posits4TorcHA

Companion Python Package for Posits4Torch to be installed on an Adaptive Computing Platform (ACP) such as a Kria KV260 Vision Starter Kit / K26 System-on-Module (SOM). Posits4TorcHA enables Hardware Acceleration (HA) of General Matrix Multiplication (GEMM) with Posits through a PYNQ Overlay for a AXI Direct Memory Access (DMA)-compliant 2D Posit Multiply-Accumulate (MAC) Unit Array.

Installation

On an ACP such as a Kria KV260 Vision Starter Kit / K26 SOM:

1. Install Python Productiviy for ZYNQ (PYNQ) (see Getting Started with PYNQ or Kria-PYNQ);
2. Install Cython;

pip install Cython

3. Install Posits4TorcHA;

git clone https://github.com/gabriel-gubert/Posits4TorcHA.git
cd Posits4TorcHA
pip install .

4. (Optional) Install Posits4Torch.

git clone --recursive https://github.com/gabriel-gubert/Posits4Torch.git
cd Posits4Torch
source install.sh

Usage

Standalone

import numpy as np
from Posits4TorcHA import Accel

Part = 'KV260'
R = 8
C = 8
N = 8
Es = 2
QSize = 128
Depth = 512
num_threads = 4

accel = Accel(Part, R, C, N, Es, QSize, Depth, num_threads)

A = np.zeros([2, 2], dtype = np.uint8) + 72 # A = [[2.0, 2.0], [2.0, 2.0]]
B = np.zeros([2, 2], dtype = np.uint8) + 72 # B = [[2.0, 2.0], [2.0, 2.0]]
 
Y = accel.GEMM(A, B.T) # Y = [[8.0, 8.0], [8.0, 8.0]]

Standalone w/ Posits4Torch

import numpy as np
from Posits4TorcHA import Accel
from Posits4Torch.Utilities import gettype, astype, frombin, tobin

Part = 'KV260'
R = 8
C = 8
N = 8
Es = 2
QSize = 128
Depth = 512
num_threads = 4

accel = Accel(Part, R, C, N, Es, QSize, Depth, num_threads)

A = tobin(astype(np.zeros([2, 2]) + 2.0, gettype(N, Es))) # A = [[2.0, 2.0], [2.0, 2.0]]
B = tobin(astype(np.zeros([2, 2]) + 2.0, gettype(N, Es))) # B = [[2.0, 2.0], [2.0, 2.0]]

Y = astype(frombin(accel.GEMM(A, B.T), gettype(N, Es)), np.double) # Y = [[8.0, 8.0], [8.0, 8.0]]

HTTP-Server

python3 HTTPServer.py [--host <host>] [--port <port>] [--verbose] [--report_timings]

Supported Configurations

Posits4TorcHA offers off-the-shelf PYNQ Overlay implementations specifically synthesized for the AMD Kria KV260 Vision Starter Kit. Currently supported configurations for the hardware accelerator are described in Table 1. PYNQ Overlays for hardware platforms other than the AMD Kria KV260 Vision Starter Kit can be synthesized using the Vivado Design Suite from AMD along with the project files, which include VHDL and/or Verilog descriptions for both the AXI4-Stream-compliant Posit MAC Unit wrapper and the AXI DMA-compliant 2D Posit MAC Unit Array, in conjunction with their respectives parameterized IP packages and the top-level block diagram for the overlay.

Table 1: Supported Hardware Accelerator Configurations.

Part Name (Part)	Rows (R)	Columns (C)	Posit Precision (N)	Exponent Bit-length (Es)	Quire Size (QSize)	FIFO Depth (Depth)
KV260	4	4	8	2	128	8
KV260	4	4	8	2	128	16
KV260	4	4	8	2	128	32
KV260	4	4	8	2	128	64
KV260	4	4	8	2	128	128
KV260	4	4	8	2	128	256
KV260	8	8	6	2	128	512
KV260	8	8	7	2	128	512
KV260	8	8	8	2	128	8
KV260	8	8	8	2	128	16
KV260	8	8	8	2	128	32
KV260	8	8	8	2	128	64
KV260	8	8	8	2	128	128
KV260	8	8	8	2	128	256
KV260	8	8	8	2	128	512