Re: Matrix multiplication performance

palik imre <[email protected]>
Newsgroups gmane.comp.lib.boost.ublas
Message-ID <[email protected]>
Hi Michael,

I had a look on your AVX microkernel on my old AMD box.  Congratulations,  it is twice as fast as what I managed to get out of my optimised C kernel.

I wonder if I can catch up with that using newer gcc + SIMD arrays or intrinsics.

Cheers,

Imre

--------------------------------------------
On Fri, 22/1/16, Michael Lehn <[email protected]> wrote:

 Subject: Re: [ublas] Matrix multiplication performance
 To: "palik imre" <[email protected]>
 Cc: "ublas mailing list" <[email protected]>
 Date: Friday, 22 January, 2016, 21:27
 
 Wow Imre!
 
 Ok, that is actually a
 significant difference :-)
 
 I have just added a new version to my site. 
 Unfortunately the computer I used for creating
 the page does not have posix_memalign().  So I
 had to hack my own function for that.  But
 I think for a proof of concept it will do.
 
 But most of all I also added
 micro-kernel that are (fairly) optimised for AVX and FMA. 
 The
 micro-kernel for AVX require MR=4,
 NR=8.  For FMA it requires MR=4, NR=16.  Otherwise
 the reference implementation gets selected. 
 For all parameters default values now can be
 overwritten when compiling, e.g.
 
     g++ -O3 -Wall
 -std=c++11 -DHAVE_AVX  -DBS_D_MC=512  matprod.cc
 
 The optimised micro kernels
 are only included when compiled with -DHAVE_AVX or
 -DHAVE_FMA
 
 I put all this
 stuff here
 
     http://www.mathematik.uni-ulm.de/~lehn/test_ublas/session2/page01.html
 
 also the tar-ball 
 
     http://www.mathematik.uni-ulm.de/~lehn/test_ublas/session2.tgz
 
 contains all required
 files:
 
    
 session2/avx.hpp
     session2/fma.hpp
     session2/gemm.hpp
    
 session2/matprod.cc
 
 Of
 course I would be interested how all this performs on other
 platforms.
 
 
 Cheers,
 
 Michael
 
 
 On 22
 Jan 2016, at 16:20, palik imre <[email protected]>
 wrote:
 
 > Hi All,
 > 
 > In the meantime I
 enabled avx2 ...
 > 
 >
 Theoretical CPU performance maximum: clock * vector-size *
 ALUs/Core * 2 ~ 69.6GFLOPS
 > So it is at
 ~25%
 > 
 > Compiler is
 gcc 4.8.3
 > 
 > vanilla
 run:
 > $ g++ -Wall -W -std=gnu++11 -Ofast
 -march=core-avx2 -mtune=core-avx2 -g -DNDEBUG -I gemm -o
 matprod matprod.cc
 > $ ./matprod
 > #   m     n 
    k  uBLAS:   t1   
    MFLOPS   Blocked:   t2 
     MFLOPS        Diff nrm1
 >   100   100   100 
 0.000350402      5707.73      0.00116104     
 1722.59               0
 >   200   200   200   0.00445094 
     3594.74      0.00819996      1951.23       
        0
 >   300   300   300 
   0.0138515      3898.49       0.0266515 
     2026.15     1.06987e-06
 >   400   400   400 
   0.0266447      4803.96       0.0613506 
     2086.37     4.01475e-06
 >   500   500   500 
   0.0424372      5891.06        0.119345     
 2094.77     8.99605e-06
 >   600   600   600 
   0.0724648      5961.51        0.203187     
 2126.12     1.63618e-05
 >   700   700   700 
    0.115464      5941.25        0.325834 
     2105.36     2.69547e-05
 >   800   800   800 
    0.173003      5918.98        0.480655 
     2130.43     4.09449e-05
 >   900   900   900 
    0.248077       5877.2       
 0.689972      2113.13     5.87376e-05
 >  1000  1000  1000      0.33781   
   5920.49        0.930591      2149.17 
    8.16264e-05
 >  1100 
 1100  1100       0.5149      5169.93   
      1.25507         2121 
    0.000108883
 >  1200 
 1200  1200     0.670628      5153.38     
    1.62732      2123.74 
    0.000141876
 >  1300 
 1300  1300     0.852692      5153.09     
    2.06708      2125.71 
    0.000180926
 >  1400 
 1400  1400      1.06695      5143.65     
    2.56183      2142.22 
    0.000225975
 >  1500 
 1500  1500       1.3874   
    4865.2         3.16532     
 2132.49     0.000278553
 > 
 1600  1600  1600      1.77623      4612.03     
     3.8137      2148.05     0.000338106
 >  1700  1700  1700   
    2.3773      4133.26     
    4.56665      2151.69 
    0.000404458
 >  1800 
 1800  1800      3.06381      3807.03     
    5.40317      2158.73      0.00048119
 >  1900  1900  1900   
    3.9039      3513.92     
    6.37295      2152.53 
    0.000564692
 >  2000 
 2000  2000      4.79166      3339.13     
    7.43399      2152.28 
    0.000659714
 >  2100 
 2100  2100      6.04946      3061.76     
    8.62429      2147.65 
    0.000762223
 >  2200 
 2200  2200      7.39085       2881.4   
      9.86237      2159.32 
    0.000875624
 >  2300 
 2300  2300      9.00453      2702.42     
    11.2513      2162.78      0.00100184
 >  2400  2400  2400      10.3952   
   2659.68         12.7491      2168.62 
     0.00113563
 >  2500  2500  2500 
     12.2283      2555.55     
    14.4615      2160.92      0.00128336
 >  2600  2600  2600      13.8912   
   2530.51         16.1965      2170.34 
     0.00144304
 >  2700  2700  2700 
      15.391      2557.72     
    18.1998      2162.99      0.00161411
 >  2800  2800  2800      17.5673   
   2499.19         20.2171      2171.63 
     0.00180035
 >  2900  2900  2900 
     19.4621      2506.31     
    22.5482      2163.28      0.00199765
 >  3000  3000  3000      21.4506   
   2517.42         24.9477      2164.53 
     0.00221028
 >  3100  3100  3100 
       23.71      2512.95     
    27.5144      2165.48      0.00243877
 >  3200  3200  3200      25.9051   
   2529.85         30.2816      2164.22 
     0.00267766
 >  3300  3300  3300 
     28.1949      2549.18          33.176     
 2166.45      0.00293379
 >  3400 
 3400  3400      30.7235      2558.56     
    36.0156      2182.61   
    0.0032087
 >  3500  3500 
 3500      34.0419      2518.95     
    39.3929      2176.79      0.00349827
 >  3600  3600  3600      37.0562   
   2518.12         42.7524      2182.62 
     0.00380447
 >  3700  3700  3700 
     39.7885      2546.11     
    46.4748      2179.81      0.00412621
 >  3800  3800  3800      43.6607   
   2513.56         50.2119      2185.62 
      0.0044694
 >  3900 
 3900  3900      46.5104      2550.78     
    54.4822      2177.56      0.00482355
 >  4000  4000  4000      50.6098   
   2529.15         58.7686      2178.03 
     0.00520289
 > 
 >
 tuned run:
 > 
 > $ g++
 -Wall -W -std=gnu++11 -Ofast -march=core-avx2
 -mtune=core-avx2 -g -DNDEBUG -I gemm -o matprod2
 matprod2.cc
 > $ ./matprod2
 > #   m     n 
    k  uBLAS:   t1   
    MFLOPS   Blocked:   t2 
     MFLOPS        Diff nrm1
 >   100   100   100 
 0.000351671      5687.13     0.000316612   
   6316.88               0
 >   200   200   200   0.00419531 
     3813.78      0.00159044      10060.1       
        0
 >   300   300   300 
   0.0141153      3825.62      0.00421113     
 12823.2     1.07645e-06
 >   400   400   400 
   0.0291599      4389.59      0.00858138       
 14916     4.00614e-06
 >   500   500   500 
   0.0483492      5170.72       0.0166519 
     15013.3     8.96808e-06
 >   600   600   600 
   0.0725783      5952.19       0.0279634 
     15448.7     1.63386e-05
 >   700   700   700 
    0.113891      6023.29        0.043077 
       15925     2.69191e-05
 >   800   800   800 
    0.171416      5973.79   
    0.0627796        16311 
    4.09782e-05
 >   900   900   900 
    0.243677      5983.32   
    0.0922766      15800.3 
    5.88092e-05
 >  1000 
 1000  1000     0.335158      5967.33     
   0.123339      16215.5     8.15988e-05
 >  1100  1100  1100 
    0.515776      5161.15     
    0.16578      16057.5 
    0.000108991
 >  1200 
 1200  1200     0.662706      5214.98     
   0.205989      16777.6     0.000141824
 >  1300  1300  1300 
    0.845952      5194.15     
    0.27637        15899      0.00018111
 >  1400  1400  1400      1.06712   
   5142.82        0.332118      16524.2 
    0.000225958
 >  1500 
 1500  1500      1.38147      4886.11       
 0.409224      16494.6     0.000278265
 >  1600  1600  1600      1.72238   
   4756.21        0.492314      16639.8 
    0.000338095
 >  1700 
 1700  1700      2.38508      4119.77       
 0.603566      16279.9     0.000404362
 >  1800  1800  1800      3.12034   
   3738.05        0.717409      16258.5 
    0.000481575
 >  1900 
 1900  1900      3.93668      3484.66       
 0.824933      16629.2     0.000564727
 >  2000  2000  2000      4.76038   
   3361.07        0.941643      16991.6     
 0.00065862
 >  2100  2100  2100     
 5.90627      3135.99         1.12226   
   16504.2     0.000762307
 >  2200  2200  2200      7.26419   
   2931.64         1.28213      16609.9 
    0.000876699
 >  2300 
 2300  2300      8.88171      2739.79     
    1.45247      16753.5      0.00100222
 >  2400  2400  2400      10.4956   
   2634.26         1.62705      16992.7 
     0.00113566
 >  2500  2500  2500 
      11.913      2623.18     
    1.87499      16666.7      0.00128371
 >  2600  2600  2600      13.7057   
   2564.77          2.1156      16615.6     
 0.00144259
 >  2700  2700  2700     
 15.5959      2524.13         2.33957   
   16826.1      0.00161501
 >  2800 
 2800  2800      17.1121      2565.67     
    2.57445      17053.8      0.00179901
 >  2900  2900  2900      19.4167   
   2512.16         2.92445      16679.4 
     0.00199764
 >  3000  3000  3000 
     21.3239      2532.37     
    3.18891      16933.7      0.00220999
 >  3100  3100  3100      23.5049   
   2534.88          3.5305      16876.4     
 0.00243845
 >  3200  3200  3200     
 25.7362      2546.45         3.81708   
   17169.1      0.00267581
 >  3300 
 3300  3300      28.4467      2526.62     
    4.25869        16877      0.00293513
 >  3400  3400  3400      30.4607   
   2580.63         4.67999      16796.6 
     0.00320688
 >  3500  3500  3500 
     33.7737      2538.96     
    5.04289      17004.1      0.00349667
 >  3600  3600  3600      36.9633   
   2524.45           5.414      17235.3 
     0.00380237
 >  3700  3700  3700 
     39.5153      2563.71     
    6.04875      16748.2      0.00412583
 >  3800  3800  3800      42.9412   
   2555.68         6.48985      16910.1 
     0.00446785
 >  3900  3900  3900 
     46.5282      2549.81     
    7.05844        16808      0.00482701
 >  4000  4000  4000      50.2218   
   2548.69         7.42442      17240.4 
     0.00520272
 > 
 >
 As the generated machine code is completely different, I
 guess gcc notices the aligned alloc, and uses the alignment
 information for optimisation.
 > 
 > Cheers,
 > 
 > Imre
 > 
 > 
 > On Friday, 22
 January 2016, 15:09, Michael Lehn <[email protected]>
 wrote:
 > 
 > 
 > Hi Imre,
 > 
 > thanks for running the benchmarks.  Of
 course you are right that using aligned memory for the
 buffers improves
 > performance.  I also
 did not really put any effort in optimising the parameters
 MC, NC, KC, MR and NR.  I will
 > compare
 different variants and report them on the website
 > 
 >     http://www.mathematik.uni-ulm.de/~lehn/test_ublas/index.html
 > 
 > I modified my
 benchmark program such that it also computes the FLOPS as
 > 
 >     FLOPS =
 2*m*n*k/time_elpased
 > 
 > See
 > 
 >     http://www.mathematik.uni-ulm.de/~lehn/test_ublas/download/session1/matprod.cc
 > 
 > Could you re-run
 your benchmarks and post the different MFLOPS you get? That
 is important for actually tuning thing.
 >
 On my machine my code only reaches 20% of the peak
 performance (about 5 GFLOPS instead of 25.6
 GFLOPS).   So
 > a speedup of
 2.5 would be impressive but still far from peak
 performance.
 > 
 >
 Cheers,
 > 
 >
 Michael
 > 
 > 
 > On 22 Jan 2016, at 11:03, palik imre
 <[email protected]>
 wrote:
 > 
 >> Sorry
 for posting twice more or less the same thing.  I got
 confused with javascript interfaces.
 >> 
 >> It seems I
 also forgot to enable avx for my last measurements.  With
 that + my blocking and alignment changes, performance
 according to my tests is something like 250% higher than
 running Michael's original code (with avx).
 >> 
 >> Cheers,
 >> 
 >> Imre
 >> 
 >> 
 >> On Friday, 22 January 2016, 10:33,
 palik imre <[email protected]>
 wrote:
 >> 
 >>
 
 >> Hi Michael,
 >> 
 >> your
 blocksizes are far from optimal.  MR & NR should be
 multiples of the L1 cache line size (i.e. 16 for double on
 Intel).  Also, the blocks should be allocated aligned to L1
 cache lines (e.g., via posix_memalign()).
 >> 
 >> This alone
 brought something like 50% speedup for my square matrix
 test.
 >> 
 >> I
 will have a look at the other parameters + the whole thing
 via perf during the weekend.
 >> 
 >> Cheers,
 >> 
 >> Imre
 >> 
 >> 
 >> 
 >> On Friday, 22 January 2016, 0:28,
 "[email protected]"
 <[email protected]>
 wrote:
 >> 
 >>
 
 >> Subject: Re: [ublas] Matrix
 multiplication performance
 >>
 Message-ID: <[email protected]>
 >> Content-Type: text/plain;
 charset="windows-1252"
 >>
 
 >> Hi Nasos,
 >>
 
 >> first of all I don?t want to take
 wrong credits and want to point out that this is not my
 algorithm.  It is based on
 >> 
 >>     http://www.cs.utexas.edu/users/flame/pubs/blis2_toms_rev3.pdf
 >> 
 >> 
    https://github.com/flame/blis
 >> 
 >> For a few
 cores (4-8) it can easily made multithreaded.  For
 many-cores like Intel Xeon Phi this is a bit more
 >> sophisticated but still not too
 hard.  The demo I posted does not use micro kernels that
 exploit SSE, AVX or
 >> FMA
 instructions.  With that the matrix product is on par with
 Intel MKL.  Just like BLIS. For my platforms I wrote
 >> my own micro-kernels but the interface
 of function ugemm is compatible to BLIS.
 >> 
 >> Maybe you
 could help me to integrate your code in the benchmark
 example I posted above.
 >> 
 >> About Blaze:  Do they have their own
 implementation of a matrix-matrix product?  It seems to
 require a
 >> tuned BLAS implementation
 (?Otherwise you get only poor performance?) for the
 matrix-matrix product.
 >> IMHO they
 only have tuned the ?easy? stuff like BLAS Level1 and
 Level2.  In that case it makes more
 >> sense to compare the performance with
 the actual underlying GEMM implementation.  But if I am
 wrong,
 >> let me know.
 >> 
 >> About the
 block size: In my experience you get better performance if
 you chose them dynamically at runtime
 >> depending on the problem size. 
 Depending on the architecture you can just specify ranges
 like 256 - 384 for
 >> blocking factor
 MC.  In my code it then also needs to satisfy the
 restriction that it can be divided by factor MR.
 >> I know that doing things at compile
 time is a C++ fetish.  But the runtime overhead is
 negligible and having
 >> blocks of
 similar sizes easily pays of.
 >> 
 >> Cheers,
 >> 
 >> Michael
 >>
 *************************************
 >> 
 >> 
 >> 
 >> 
 >>
 _______________________________________________
 >> ublas mailing list
 >> [email protected]
 >> http://lists.boost.org/mailman/listinfo.cgi/ublas
 >> Sent to: [email protected]
 > 
 > 
 > 
_______________________________________________
ublas mailing list
[email protected]
http://lists.boost.org/mailman/listinfo.cgi/ublas
Sent to: [email protected]
lmpx.com only provides a reader for public news (NNTP) servers. It is not affiliated with the servers or forums shown here and is not responsible for the content of articles, which is written by their respective authors.