Presenter : Ching-Hua Huang
2013/9/16
Visibility Enhancement for Silicon DebugCited count : 62
Yu-Chin Hsu; Furshing Tsai; Wells Jong; Ying-Tsai ChangNovas Software, San Jose, CA Design Automation Conference (DAC), 2006 43rd ACM/IEEE
National Sun Yat-sen University
Embedded System Laboratory
Several emerging Design-for-Debug (DFD) methodologies are addressing silicon debug by making internal signal values and other data observable. Most of these methodologies require the instrumentation of on-chip logic for extracting the internal register data from in situ silicon. Unfortunately, lack of visibility of the combinational network values impedes the ability to functionally debug the silicon part.
Visibility enhancement techniques enable the virtual observation of combinational nodes with minimal computational overhead. These techniques also cover the register selection analysis for DFD and multi-level design abstraction correlation for viewing values at the register transfer level (RTL).
Abstract
2
Experimental results show that visibility enhancement techniques can leverage a small amount of extracted data to provide a high amount of computed combinational signal data. Visibility enhancement provides the needed connection between data obtained from the DFD logic and HDL simulation-related debug systems.
Abstract (Cont.)
3
4
Introduction
The process of IC designs can be broken into two stages
Key factors which make silicon debug difficult include: Limited silicon signal data visibility Silicon errors may either be functional bugs or physical defects Silicon debug data is usually associated with a gate-level netlist
|--------First stage------||----Second stage---|
Decreasing Increasing
Introduction (Cont.)
5
Manufacturing Test To check the physical defects Based on DFT and ATPG
System Validation To ensure the silicon operates according to specification Based on DFD logic
。Many DFD implementations leverage DFT circuitry.。DFD is the access mechanism that provides in-situ scan register
visibility Data to Debug Flow
Scancontrol
Combinationallogic
Combinationallogic
DFTDFT
DFDDFD
Real-timeclock &reset
control
What’s the problem
6
Finding and analyzing the problems in silicon is dreadfully tedious and time consuming. DFD provide very limited signal access and visibility into a device. DFT allows observation of states in scan chain, but it is still limited by
only the register signals.
The data extracted from device is usually mapped to gate-level netlist. Designers normally implement the design at the behavioral- level.
This paper propose a methodology to resolve the above issues.
Related Work
7
[1] [2] User Wants and
Needs
The time required to move from prototype to volume production is increasing
This paper:Visibility Enhancement for Silicon Debug
[5] DFT and ATPG
A common methodology utilized in approximately 82% of designs
[8] DFD
[4][16] More sophisticatedDFD methodologies
[17] A Reconfigurable Design-for-Debug
Infrastructure
Add internal monitoring and breakpoints
Extract values to be observedduring run real-time observation
Around five percent of all designs contain some type of DFD
Proposed method
8
Proposed method This methodology includes three main steps:
Essential Signal Analysis: 。Analyze design (RTL, gate) and provide optimal
(minimal and sufficient) set of signals to be observed
Data expansion: 。Using the knowledge of the design function to
computes the missing data. Gate-to-RTL Correlation:
。Map the gate level signals back to RTL
Essential Signal Analysis
9
ESA starts by inferring the HDL description ◦ It recursively traverses inferred logic netlist, analyzes the
logic equations for each functional block.
The ESA technology can also be applied in regression simulation.◦ For regression simulation
1.ESA analyzes the RTL or netlist code to find the storage elements, memory elements, and the primary inputs.
2.Only dumping the essential signals 3.Full visibility can be achieved using the data expansion
technology◦ The regression simulation methodology can reduce the total
simulation times for debug.
Data expansion
10
To obtain the maximal visibility of a device.◦ The quality of essential signal analysis is closely related to
the capability of data expansion.
Translate Data for the HDL Domain◦ The captured data must be translate to properly format for
debugging tools◦ There are a number of steps to this process:
Temporal transformation Correlate time or cycle to the captured data from silicon world to virtual
world Signal mapping
The data must be associated with the signal names in the HDL design Format transformation
The data is made available in standard VCD format or Novas FSDB format
Data expansion (Cont.)
11
Expand Captured Data◦ Computes the missing data by using the knowledge of the
design function.
There are two primary differences with the simulation◦ Data expansion where only the needed value will be
computed is a must for users.◦ The time-processing is not necessary.
Data expansion
Data expansion
11
00
00
11
11
00
11
11
00
00
11
11
??
? ?
?
?
?
?
??
?
??
01
1
1
0
00
10 0 1
01
11
1
11
11
00
00
11
11
11
00
00
11
11
00
Cycle N Cycle N+1 Cycle N Cycle N+1
Gate-to-RTL Correlation
12
In general, Silicon signal data is usually easy to assign to gate-level signals
Gate-level netlist and values are difficult for RTL designers to understand and debug◦ One-to-one correspondence is not always the case after
synthesis transformations
Correlation
0100101110010110 11010110001101101000011111000000
Correlation
Gate-level
RTL
Silicon extracted easy difficult
Before the experiment…
13
Essential Signal Analysis How many Percentage of signals can reach the full
visibilty ? As described in ESA, the regression simulation can
reduce the total simulation times ?
Gate-to-RTL Correlation If synthesized an RTL design into gates both with and
without preserving hierarchy. Is this difference will effect our debug ?
Essential signal analysis technology
14
Working scope
# of signals
# of essential signals
% of signals
Case 1 138842 13190 9.5%
Case 2 6808801 992248 14.6%
Case 3 4568074 173544 3.8%
Case 1 Case 2
Simulation time
FSDB filesize
Simulationtime
FSDB file size
Full dump
506 sec. 290 MB 6 hr. 39 min.
2.1 GB
No dump 132 sec. N/A 2 hr. 12 min.
N/A
ES dump 174 sec. 25 MB 2 hr. 29 min.
61.7 MB
Effectiveness of essential signal analysis technology
The regression simulation can reduce the total simulation
15
Gate to RTL correlation Effectiveness of the gate to RTL correlation technology
Count Percentage
Total 990
Can be mapped 207 21%
1 statement 684 69%
2-3 statements 54 5.4%
More than 3 statements 21 2.1%
Other (floating, scan chain …)
24 2.5%
Count Percentage
Total 952
Can be mapped 180 19%
1 statement 307 32%
2-3 statements 194 20%
More than 3 statements 213 22%
Other (floating, scan chain …)
58 6%
With preserving hierarchy
Without preserving hierarchy
It is useful for silicon debug.
It will make the silicon debug difficultly.
Conclusions and My comments
16
Conclusions These technologies provide the ability to
。Analyze the essential signals 。Expand the captured data 。Correlation to present gate-level data at the RTL
This experiments show 。The proposed technology is effective。It greatly enhance the comprehension of silicon prototype
operation
My comments This paper was cited from many papers which proposed the
approaches of silicon debug. It discuss about the Gate-to-RTL Correlation
。Some issues are similar with my work.