AlgorithmStatus2005FEB
- 1. Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 PowerPC Algorithm Development Update Bo Lin, CPD/NCSG FEB 2005
- 2. Slide 2 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 What’s New? • Vectorised RSA and modular exponetiation • AltiVec and Scalar CRC32 (IEEE 802.3 FCS) for wireless LAN • Other vendor’s information on AES128 • Re-arranged presentation Code available Code in planning
- 3. Slide 3 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Theme Key Messages: • PowerPC (core and AltiVec) is very efficient – beyond imagination. • Software solution can provide surprisingly good results. • The key is to develop algorithms best suitable for PowerPC. 2005 focus: • Vectorised RSA and modular exponentiation • Networking support – “dictionary”
- 4. Slide 4 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 The Impact of Innovative Algorithm Design By properly designing algorithms for PowerPC, performance can be dramatically increased. A good example is the EEMBC’s Convolutional Encoder (see www.eembc.com) where AltiVec enabled algorithm out-perform the “out-of-box” version 600 ~ 1000 times. Another example is that many popular crypto algorithms are very efficiently executed on PowerPC.
- 5. Slide 5 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Benchmark information All figures are tested on silicon Small variances are expected on different test platforms The figures on 8641(D) are expected better since AltiVec improvement on the device provides “out-of-order” execution Devices & Algorithms • G4 & AES, DES, 3DES, Kasumi, tdmCRC16, CRC32, RSA MPC7457 @ 1000MHz
- 6. Slide 6 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 What is benchmarked? • The cycle count of a function call. For example, we measure how many cycles are needed for G4 or other PowerPC processor to complete a function call such as AES128(data, cipher, key). • All CPU resources are used - the same convention of benchmark practice as other vendors.
- 7. Slide 7 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 POPI Guideline Please share figures with customers but not the whole set of slides.
- 8. Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Benchmark information on Encryption (Public key Cryptos)
- 9. Slide 9 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 vRSA Performance figures on G4 vRSA • Data block length: 1024 bits • Key length: 1024 bits • IEEE P1363 test vector Relative performance (cycle count) • Signing: 3,720,000 cycles Absolute performance on Test Setup • 3.72 ms per signing Other info: CRT, arbitrary length Test Setup: CPU: 7457 without L3 @1GHz Language: ANSI C Compiler: gcc 3.3 on Linux
- 10. Slide 10 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 1024 modular exponentiation on G4 Y = XE mod N • all parameters are of 1024 bits Relative performance (cycle count) • Modular exp: 12,850,000 cycles Absolute performance on Test Setup • 12.8 ms per mod exp Other info • E usually is 160 bits or 256 bits • Arbitrary length Test Setup: CPU: 7457 without L3 @1GHz Language: ANSI C Compiler: gcc 3.3 on Linux
- 11. Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Benchmark information on Encryption (Conventional Cryptos)
- 12. Slide 12 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 AES128 Performance figures on G4 AES128 • Data block length: 128 bits • Key length: 128 bits Relative performance (cycle count) • Encryption: 336 cycles per 128 bit data block • → 0.381 bits/cycle → 2.6 cycles/bit • Decryption: 335 cycles per 128 bit data block • → 0.321 bits/cycle → 2.6 cycles/bit Absolute performance on Test Setup • 380.9 Mbps for encryption Test Setup: CPU: 7457 without L3 @1GHz Language: ANSI C Compiler: gcc 3.3 on Linux
- 13. Slide 13 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 AES128 Benchmarked by Others (reference only) http://www.tcs.hut.fi/~helger/aes/rijndael.html • Relative performance (cycle count) on 7457@1.25 GHz with gcc Encryption: 385 cycles per 128 bit data block Decryption: 391 cycles per 128 bit data block • Relative performance (cycle count) on Pentium M@1.33 GHz with gcc Encryption: 348 cycles per 128 bit data block Decryption: 376 cycles per 128 bit data block http://www.eskimo.com/~weidai/benchmarks.html • Relative performance (cycle count) on Pentium 4 @ 2.1GHz with Microsoft Visual C++ .NET 2003 (whole program optimization, optimize for speed, P4 code generation Encryption: 542 cycles per 128 bit data block
- 14. Slide 14 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 AES192 Performance Figures on G4 AES192 • Data block length: 128 bits • Key length: 192 bits Relative performance (cycle count) • Encryption: 390 cycles per 128 bit data block • → 0.328 bits/cycle → 3.0 cycles/bit • Decryption: 387 cycles per 128 bit data block • → 0.331 bits/cycle → 3.0 cycles/bit Absolute performance on Test Setup • 328.2 Mbps for encryption Test Setup: CPU: 7457 without L3 @1GHz Language: ANSI C Compiler: gcc 3.3 on Linux
- 15. Slide 15 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 AES256 Performance Figures on G4 AES256 • Data block length: 128 bits • Key length: 256 bits Relative performance (cycle count) • Encryption: 444 cycles per 128 bit data block • → 0.288 bits/cycle → 3.5 cycles/bit • Decryption: 438 cycles per 128 bit data block • → 0.292 bits/cycle → 3.4 cycles/bit Absolute performance on Test Setup • 288.2 Mbps for encryption Test Setup: CPU: 7457 without L3 @1GHz Language: ANSI C Compiler: gcc 3.3 on Linux
- 16. Slide 16 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Kasumi Performance Figures on G4 Kasumi • Data block length: 64 bits • Key length: 128 bits Relative performance (cycle count) • Kasumi operation: 427 cycles per 64 bit data block • → 0.150 bits/cycle → 6.67 cycles/bit Absolute performance on Test Setup • 149.9 Mbps Test Setup: CPU: 7457 without L3 @1GHz Language: ANSI C Compiler: gcc 3.3 on Linux
- 17. Slide 17 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 DES Performance Figures on G4 DES • Data block length: 64 bits • Key length: 56 bits Relative performance (cycle count) • DES function: 423 cycles per 64 bit data block • → 0.151 bits/cycle → 6.61 cycles/bit Absolute performance on Test Setup • 151.1 Mbps for encryption Test Setup: CPU: 7457 without L3 @1GHz Language: ANSI C Compiler: gcc 3.3 on Linux
- 18. Slide 18 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 3DES Performance Figures on G4 3DES • Data block length: 64 bits • Key length: 112 bits Relative performance (cycle count) • 3DES function: 1214 cycles per 64 bit data block • → 0.52 bits/cycle → 18.9 cycles/bit Absolute performance on Test Setup • 52.9 Mbps for encryption Test Setup: CPU: 7457 without L3 @1GHz Language: ANSI C Compiler: gcc 3.3 on Linux
- 19. Slide 19 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Absolute Performance (Mbps) 0 50 100 150 200 250 300 350 400 G4/1000 AES128 AES192 AES256 Kasumi DES 3DES Tested performance (Mbps)
- 20. Slide 20 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Core & Compiler Efficiency (bits/cycle) 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 G4/1000 AES128 AES192 AES256 Kasumi DES 3DES Core & compiler efficiency
- 21. Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Benchmark Information on CRC
- 22. Slide 22 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 G4 Performance for CRC32 (IEEE 802.3 FCS) Frame Size cycles bits/cycle cycles bits/cycle 60 281 1.71 250 1.92 62 303 1.64 277 1.79 64 247 2.07 264 1.94 71 287 1.98 299 1.90 80 277 2.31 320 2.00 89 320 2.23 352 2.02 96 307 2.50 367 2.09 105 350 2.40 408 2.06 112 337 2.66 432 2.07 128 367 2.79 488 2.10 161 429 3.00 604 2.13 192 487 3.15 712 2.16 225 549 3.28 828 2.17 256 607 3.37 936 2.19 384 847 3.63 1384 2.22 449 970 3.70 1613 2.23 512 1088 3.76 1832 2.24 768 1568 3.92 2728 2.25 895 1864 3.84 3183 2.25 1020 2081 3.92 3610 2.26 1024 2048 4.00 3624 2.26 1514 3004 4.03 5354 2.26 1536 3008 4.09 5417 2.27 2048 3968 4.13 7208 2.27 4096 7808 4.20 14377 2.28 AltiVec-enabled Scalar * Frame Size is in BYTE * The coloured Frame Sizes are within the specification of IEEE 802.3 Absolute Asymptotical Performance on Test Setup: AltiVec: 4200 Mbps Scalar: 2280 Mbps Test Setup: CPU: 7457 without L3 @1GHz Language: ANSI C Compiler: gcc 3.3 on Linux
- 23. Slide 23 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 G4 Performance figures for multi-channel CRC16 tdmCRC16 • 16.35 Gbps for processing 16 channels simultaneously, AltiVec enabled • 16.35 bits/cycle → 0.061 cycles/bit BL: This function reads 16 equal- length packets arranged in columns and calculates all the 16 CRC16 simultaneously. One potential application will be CRC8 to calculate ATM’s cell header HEC’s. Even higher performance is expected. Other test case benchmark figures are available upon request Test Setup: CPU: 7457 without L3 @1GHz Language: ANSI C Compiler: gcc 3.3 on Linux
- 24. Slide 24 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Absolute Asymptotical Performance (Mbps) 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 G4/1000 tdmCRC16 vCRC32 sCRC32Tested performance (Mbps)
- 25. Slide 25 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Core & Compiler Efficiency (bits/cycle) 0 2 4 6 8 10 12 14 16 18 G4/1000 tdmCRC16 vCRC32 sCRC32Core & compiler efficiency
- 26. Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Other Information on Benchmark, Codes, etc
- 27. Slide 27 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Performance Comparison with scalar version routines: checksum() 0 500 1000 1500 2000 2500 32 448 864 1280 1696 2112 2528 2944 3360 3776 4192 4608 5024 5440 5856 6272 6688 7104 7520 7936 bytes checked MB/second AltiVec Linux assembly C Test setup: 7445 1GHz, MPC107 with 133MHz bus 36.5 bits/cycle More improvement possible
- 28. Slide 28 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Code for benchmarking and evaluation 1. Encryption 1. DES/3DES 2. AES128/192/256 3. Kasumi 4. RC5 5. RSA/modular exp 2. CRC’s 1. CRC32/24/16/12/8 3. EEMBC www.eembc.com 1. Telecom suite 2. Networking suite 3. Consumer suite 4. 3GPP TS25.212 (baseband symbol rate processing) 1. Report available upon request
- 29. Slide 29 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 AltiVec/e600 offering 1. Code for Web download – ready to use • Fully tested code and API complying with, say, ANSI C or Linux • http://www.freescale.com/altivec 2. Prototype code – customer evaluation & integration • Develop, benchmarked, and verified against a set of test vectors. • For example, the AES code can re-produce test vectors in FIPS 197 • IEEE FCS/CRC32 generates correct result against IEEE 802.11 • 1024 RSA generates IEEE P1363 test vectors • Some restrictions may apply. • For example, AltiVec-enabled code may require data buffer quad- ward aligned. • Documentation may not completed 3. Code in pipeline – under development and planned • Code committed to be in “prototype” state • Customer input welcome
- 30. Slide 30 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 What is in the AltiVec-enabled libc library? *Not ANSI C but present in many implementation (e.g. VxWorks) strlen strcmp strcpy strcmp strncpy memcpy memmove memcmp memchr bcopy* bzero*
- 31. Slide 31 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Other networking functions from Linux that can be done in AltiVec __copy_tofrom_user csum_partial csum_partial_copy_generic page_copy
- 32. Slide 32 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 How are the AltiVec-enabled libc libraries used? Freescale provides an archive (binary file) that can be linked with existing object files. Dhrystone example: ld dhry21a.o dhry21b.o c:/sw/libmoto/libmotovec.a c:/gcc/lib/gcc-lib/powerpc-eabisim2.95.2libgcc.a -( -lsim -lc ) -o gccBM.elf Put libmotovec.a on linker command line before the compiler’s libc library Binary library is eabi compliant, i.e. independent of compiler used. • Proven with gcc, Green Hills, Diab, Metaware, Metrowerks
- 33. Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 Planned Code (customer input welcome)
- 34. Slide 34 Freescale Semiconductor Internal Use Only. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004 s/w solution table lookup (“Dictionary”) performance estimate • Main feature • One underlying data structure to support mixed instructions consisting of lookup, LPM, range search, insert, delete • Key technology • Planar subdivided dictionaries • Linear hashing ax mod M under 100 cycles • For x with variable length (say, 64 or 128 bytes) 64 bytes = 4 vectors 4 vec_msum()’s + 4 vec_sums() + 3 mul()’s + 3 add()’s + 1 div() ≈ 20 cycles • AltiVec enabled • LPM and range search under 1000 cycles • The above estimate is based on other’s implementation on Pentium M • No restriction on key length, i.e. works for IPv4, IPv6, or very long keys. Performance is related to # of entries • On-the-fly table update • Benchmark information scheduled in July 2005 Co-operation with customers?
- 35. Freescale Semiconductor Confidential Proprietary. Freescale™ and the Freescale logo are trademarks of Freescale Semiconductor, Inc. All other product or service names are the property of their respective owners. © Freescale Semiconductor, Inc. 2004
Editor's Notes
- #31: Unfortunately libmotovec is currently a library without many “books.” However, we believe that the functions that we do have are important for applications that do a lot of data movement through the processor’s register files. The wider width and improved bandwidth to cache of the AltiVec register set allow functions like memcpy or memset in AltiVec to be faster than the same function using the general purpose register file.
- #32: Some data movement functions are very similar to memcpy but are called by a different name such as __copy_tofrom_user which is found in the Linux kernel and copies data from kernel to user space and vice versa. Our AltiVec-enabled memcpy can easily be modified to speed up these functions as well. More importantly, if there is other work to be done on the data being moved – like calculating a checksum – that work can largely be completely hidden under the memory latency of memcpy. Many applications already know this and provide functions like the Linux checksum calculation or checksum-while-copying function. Page_copy is a really trivial example of copying 4K byte pages of memory from one location to another.
- #33: The library has proven very easy to use. A customer’s existing object modules can be linked with the AltiVec-enabled library by inserting the library on the linker command line ahead of the compiler’s libc library. The customer’s object modules then link to the symbols in the AltiVec library instead of the compiler-supplied functions.