Download presentation
Presentation is loading. Please wait.
1
An Arithmetic Structure for Test Data Horizontal Compression Marie-Lise FLOTTES, Regis POIRIER, Bruno ROUZEYRE Laboratoire d’Informatique, de Robotique et de Microelectronique de Montpellier, France DATE ‘04 Laboratory of Reliable Computing Department of Electrical Engineering National Tsing Hua University Hsinchu, Taiwan
2
2 Reference An Arithmetic Structure for Test Data Horizontal Compression Marie-Lise FLOTTES, Regis POIRIER, Bruno ROUZEYRE DATE ‘04 Test Data Compression Using Dictionaries with Fixed-Length Indices Lei Li and Krishnendu Chakrabarty VTS ‘03 An Efficient Test Vector Compression Scheme Using Selective Huffman Coding Abhijit Jas, Jayabrata Ghosh-Dastidar, Mon-Eng Ng and Nur A. Touba IEEE Transaction on CAD of IC and System, June 2003 Improving compression ratio, area overhead, and test application time for SOC test data compression / decompression P. T. Gonciari, B. Al-Hashimi and N.Nicolici DAT ‘02
3
3 Outline Instruction Lossless Compression algorithm Compression principle and De-compressor Architecture Main Issue Compression Issue Timing Issue Experimental Result Compare & Conclusion
4
4 Why need compression ? Higher circuit densities and a large number of embedded cores will lead the increase of test data volume, which in turn leads to an increase in testing time. Transmitting test patterns and handshaking between cores and ATE will waste a lot testing time.
5
5 How to reduce testing cost ? ATE has limited pin counts and testing time is equal to the testing cost. More patterns transmit to core, more testing time Reduce testing cost a. Reduce testing time - compaction b. Reduce testing pin counts - compression n m Channel n < m
6
6 Compression Method (1) Easy compression → Hard decompression (2) Hard compression → Easy decompression Compression Method Lossless Lossy Compression can be executed through software, therefore the dominate area overhead is the circuit of decompression.
7
7 Lossless Compression Algorithm Run-Length Coding Frequency-Directed Run-Length (FDR) & Extend FDR Huffman Dictionary Based Main idea of above algorithms More common data, shorter bits to represent
8
8 Run-Length Coding First bit represents the data is ‘1’ or ‘0’. Next m bits represent the runs of ‘1’ or ‘0’. 111111000000011111100000 m=3 1110011111100011 101010 m=3 100100011001000110010001 Compression Ratio :
9
9 FDR & EFDR Source : VTS ‘99
10
10 Example of Huffman Tree Source : VTS ‘99
11
11 Simplified Huffman Tree Source : vlsitsa ‘01
12
12 Variable-Length Input Huffman Compression Source :DAT ‘02
13
13 Example 12 14 8 137 11154 52 13 1610 9 6 Source : VTS ‘03 1110001111100011 0100011001000110
14
14 Example(1/2) 137 1115124 528 1413 1610 9 6 11 52 13 16 9 6 maximum degree Sub graph 52 13 16 9 6 5 13 16 6 Clique
15
15 Example(2/2) 137 1115124 528 1413 1610 9 6 5 13 16 6 More this clique 137 1115124 28 14 10 9 New Graph
16
16 Example Result Obtain 4 cliques : {5,6,13,16}, {2,8,14}, {3,4,7}, {1,11} 12 words are encoded, and there are else 4 different words un-coded. So total need transmit (1+2)*12+(1+8)*4=72 bits. If no compression, we need transmit 8*16=128 bits. Therefore, after compression, reducing about 43.75% bits needed to transmit to core.
17
17 Parameter Definition M : # of ATE channel N : # of scan chains in CUT F : # of FF in each scan chain Initial test sequence { V 1, V 2, … } Numerous difference D i D i = V i+1 -V i Source : DATE ‘04
18
18 M-to-N Horizontal De-compressor Source : DATE ‘04 M pins in ATE N scan chains in CUT M < N
19
19 De-compressor Structure Source : DATE ‘04 Overhead Adder N bits-M SRA Shifts Reg./Accumulator
20
20 SRA Operation Mode Parallel Mode ( Shift =0 ) Parallel data-in from adder Transfer test pattern to scan chains in CUT Semi-Parallel Mode ( Shift =1 ) Store test pattern from ATE SRA likes many scan chains
21
21 5bits-3 SRA Source : DATE ‘04 Test Time Evaluate Parallel mode One clock cycle Semi-Parallel mode clock cycle
22
22 Shift-in Test Pattern M test pins in ATE Max difference is 2 M Compressible pattern & Un-compressible pattern Compressible pattern shift-in time Un-compressible pattern shift-in time
23
23 Compression Issue Reduce D max can reduce # of test pins M Importance : MSB > LSB Source : DATE ‘04
24
24 Compression Issue with Don’t Care Don’t care bit (X) can be 1 or 0 More don’t care bits, higher compression ratio Vk0: 1 0 1 x 0 Vk1: 0 x x 1 0 Vk2: x x 1 x x Vk3: 1 x 0 x x Vk4: 0 x 1 x 0 1 4 1 4 2 Vk0: x 0 0 1 1 Vk1: 1 x 0 x 0 Vk2: x x x 1 x Vk3: x x x 0 1 Vk4: x x 0 1 0 4 4 2 1 1 Vk0: 1 0 0 1 1 Vk1: 1 1 0 1 0 Vk2: 1 1 1 1 1 Vk3: 0 0 1 0 1 Vk4: 0 1 0 1 0 Dk0: 0 0 1 1 1 Dk1: 0 0 1 0 1 Dk2: 0 0 1 1 0 Dk3: 0 0 1 0 1 3 bits
25
25 Timing Issue Test pattern transmit time Total : Larger P comp, smaller test pattern transmit time Test pattern transmit time is shorter
26
26 Experimental Results (1) Source : DATE ‘04 ISCAS89 Benchmark
27
27 Experimental Results (2) ISCAS89 Benchmark : S9234 Version 1 6 scan chains & Length 42 Version 2 10 scan chains & Length 25 Version 3 10 scan chains & Length 25 6-to-10 compressor
28
28 Experimental Results (3) Source : DATE ‘04
29
29 Conclusion A useful & simple method for reducing test pin A low overhead Adder & Nbits-M SRA No impact the fault coverage compressible & un-compressible pattern Problem Not use the don’t care bit adequately
30
30 Conclusion (cont.) Combine the SRA and 1st scan chain … Scan chain SRA … Scan chain SRA
31
31 Combination Combine this approach with dictionary based compression 137 1115124 528 1413 1610 9 6 Dictionary Based CompressionArithmetic Compression Grouped pattern Compressed Pattern Un-compressed Pattern
Similar presentations
© 2024 SlidePlayer.com Inc.
All rights reserved.