DNN Accelerator Generator

Analytical model and auto-scheduler

https://github.com/xuanyoya/CNN-blocking/tree/dev

Designs in the paper

Name	Dataflow	Dimension	PE Number	RF Size	Mem Size
OS4	X	1D	4	32 B	32 KB
OS8	X	1D	8	64 B	64 KB
WS16	C K	2D	16	64 B	32 KB

How to run catapult

https://docs.google.com/document/d/1NrQi8JJyspAChkBYfLLc_DQbHuIAQlz9DGFKn6Y0wls/edit?ts=5bbc0010

Description of HLS files

conv_systolic_packed_OS_v5 - This folder has OS16 design

• conv_systolic_packed_OS_v5/conv_ref.cpp - Gold model for convolution
• conv_systolic_packed_OS_v5/conv_ref.h - Header file for conv_ref.cpp
• conv_systolic_packed_OS_v5/Stencil_catapult.h - Library for stencil data type
• conv_systolic_packed_OS_v5/double_buffer.cpp - This is different in the two folders. This one seems to have 16 outputs - it corresponds to the memory BW required by the systolic array.
• conv_systolic_packed_OS_v5/hls_target.cpp - Top level design file input to HLS
• conv_systolic_packed_OS_v5/hls_target.h - Header file for hls_target.cpp. Define the hardware attributes (such as PE number) and layer size. Changing the parameters in this file can generate OS4 or OS8 designs.
• conv_systolic_packed_OS_v5/tb_hls_target.cpp - Top level testbench - verifies if the design in hls_target.cpp is same as conv_ref.cpp for randomly generated test vectors
• conv_systolic_packed_OS_v5/setup.tcl - HLS directives
• conv_systolic_packed_OS_v5/concat_rtl.v - Verilog generated by HLS

conv_systolic_packed_v13 - This folder has WS16 design

• conv_systolic_packed_v13/Stencil_catapult.h - Same as the file above folder
• conv_systolic_packed_v13/conv.h - Same as hls_target.h above
• conv_systolic_packed_v13/conv_ref.cpp - Same as the file in above folder
• conv_systolic_packed_v13/conv_ref.h
• conv_systolic_packed_v13/double_buffer.cpp
• conv_systolic_packed_v13/tb_gemm_systolic.cpp
• conv_systolic_packed_v13/catapult_gemm_systolic.cpp - Same as hls_target.cpp above
• conv_systolic_packed_v13/concat_rtl.v