Spectral RTL Test Generation for Microprocessors

1
Spectral RTL Test Generation for Microprocessors

• Nitin Yogi and Vishwani D. Agrawal
• Auburn University
• Department of ECE
• Auburn, AL 36849

To be presented at the 20th International
Conference on VLSI Design, Jan. 2007, Bangalore
2
Outline

• Microprocessor testing Issues
• Problem and Approach
• Spectral analysis background
• RTL testing approach
• Ex. Processor PARWAN
• RTL DFT
• Experimental Results
• Future Work
• Conclusion

3
Microprocessor testing Issues

• Issues arising from Increased Design Complexity
• Increased Test Generation Complexity
• Viable Test Method RTL test generation
• Advantages
• Low testing complexity
• Early detection of testability issues
• Increased Demands on Testing
• Viable Test Method Functional at-speed tests
• Advantages
• Better defect coverage
• Detection of delay faults

4
Problem and Approach

• The problem is
• Develop an RTL-based ATPG method to generate
  functional at-speed tests.
• And our approach is
• Circuit characterization using RTL
• RTL test generation
• Analysis of information content and noise in RTL
  vectors.
• Test generation for gate-level implementation
• Generation of spectral vectors
• Fault simulation and vector compaction

5
RTL Faults
CombinationalLogic
Inputs
Outputs
RTL stuck-at fault sites
FF
FF
6
Walsh Functions and Hadamard Spectrum
w0

• Walsh functions form an orthogonal and complete
  set of basis functions that can represent any
  arbitrary bit-stream.
• Walsh functions are the rows of the Hadamard
  matrix.
• Example of Hadamard matrix of order 8

w1
w2
w3
Walsh functions (order 8)
1 1 1 1 1 1 1 1 1 -1 1 -1 1 -1 1 -1 1
1 -1 -1 1 1 -1 -1 1 -1 -1 1 1 -1 -1 1 1 1
1 1 -1 -1 -1 -1 1 -1 1 -1 -1 1 -1 1 1 1 -1
-1 -1 -1 1 1 1 -1 -1 1 -1 1 1 -1
H8
w4
w5
w6
w7
7
Characterizing a Bit-Stream

• A bit-stream is correlated with each row of
  Hadamard matrix.
• Highly correlated basis Walsh functions are
  retained as essential components and others are
  regarded as noise.

Bit stream to analyze
Correlating with Walsh functions by multiplying
with Hadamard matrix.
Bit stream
Spectral coeffs.
Essential component (others noise)
Hadamard Matrix
8
Test Vector Generation

• New spectral sets are generated by retaining
  essential components and perturbing noise.
• Bit-streams are generated from perturbed spectra
  by multiplying with Hadamard matrix.

Perturbation
Spectral components
Bits changed
Generation of test vectors by multiplying with
Hadamard matrix
Essential component retained
New bit stream
9
RTL Testing Approach (Circuit Characterization)

• RTL test generation
• Test vectors generated for RTL faults (PIs, POs
  and inputs - outputs of flip-flops.)
• Spectral analysis
• Test sequences for each input are analyzed using
  Hadamard matrix.
• Noise components are currently determined by a
  gradually increasing threshold level.

10
PARWAN processor
Reference Z. Navabi, Analysis and Modeling of
Digital Systems. New York McGraw-Hill, 1993.
11
Power Spectrum for interrupt signal
Essential components
Normalized Power
Noise components
Randomlevel(1/128)
Spectral Coefficients
12
Power Spectrum for ready signal
Examples of Essential components
Normalized Power
Examples of Noise components
Randomlevel(1/128)
Spectral Coefficients
13
Power Spectrum for DataIn5 signal
Examples of Essential components
Examples of Noise components
Normalized Power
Randomlevel(1/128)
Spectral Coefficients
14
Power Spectrum for DataIn4 signal
Examples of Essential components
Examples of Noise components
Normalized Power
Randomlevel(1/128)
Spectral Coefficients
15
Power Spectrum for random signal
Normalized Power
Randomlevel(1/128)
Spectral Coefficients
16
Selecting Minimal Vector Sequences Using ILP

• Fault simulation of new sequences
• A set of perturbation vector sequences V1, V2,
  .. , VM are generated.
• Vector sequences are fault simulated and faults
  detected by each is obtained.
• Compaction problem
• Find minimum set of vector sequences which cover
  all the detected faults.
• Minimize CountV1, ,VM to obtain compressed
  seq. V1, ,VC where V1, ,VC V1, ,
  VM CountV1, ,VC CountV1, ,VM Fault
  CoverageV1, ,VC Fault CoverageV1, ,VM
• Compaction problem is formulated as an Integer
  Linear Program (ILP) 1.

1 P. Drineas and Y. Makris, Independent Test
Sequence Compaction through Integer Programming,"
Proc. ICCD03, pp. 380-386.
17
RTL DFT

• Main goals of DFT
• Improved controllability
• Improved observability
• XOR tree as DFT
• Low area overhead
• Low performance penalty
• Hard-to-detect RTL faults targeted for DFT
• 35 observation points selected between datapath
  and controller.

XOR tree
To system output
Hard-to-detect RTL faults
18
Experimental Results
RTL characterization results
PARWAN processor
PARWAN processor with DFT
19
Experimental Results
Gate-level Fault Coverage results
20
Experimental Results
21
Experimental Results
22
Future Work

• 2-D Spectral Analysis
• Temporal correlation along the vectors
• Spatial correlation among the inputs
• Mathematically2D Spectrum (S) Hi Test Set
  Block (T) Hv

Circuit UnderTest
Test Set Block (T)
i Inputs
v Vectors
23
Power spectrum of a test data block for b14
circuit
Temporal frequencies
Spatial frequencies
24
Power spectrum of averaged 100 random test data
blocks
Temporal frequencies
Spatial frequencies
25
Future Work (cont.)

• Information Entropy
• Measure of information content
• Spectral entropy
• Measure of peakedness or flatness or spectrum
• Spectral entropy1 Es - ?pn x log2(pn) where
  pn xn2 / x2 xn spectral coefficient

1 Mester, R. Franke, U., "Spectral
entropy-activity classification in adaptive
transform coding," Selected Areas in
Communications, IEEE Journal on , vol.10,
no.5pp.913-917, Jun 1992
26
Conclusion

• Spectral RTL ATPG technique applied to PARWAN
  processor.
• The proposed ATPG method provides
• Good quality functional at-speed tests
• Lower test generation complexity
• Enables testability appraisal at RTL
• RTL based XOR tree as DFT was found to improve
  results.
• Future work
• 2-D Spectral analysis
• Spectral entropy

27

• Thank You !
• Questions ?