Re: [PATCH 3/3] boost::ublas increasing the range of BLAS level 3 benchmarks
palik imre <[email protected]>
| Newsgroups | gmane.comp.lib.boost.ublas |
|---|---|
| Message-ID | <[email protected]> |
Sent request to https://github.com/uBLAS/ublas/tree/feature/ublas00004_simd_gemm Why doesn't it gets sent to the mailing list like the one about the week equality check? -------------------------------------------- On Mon, 14/3/16, Nasos Iliopoulos <[email protected]> wrote: Subject: Re: [ublas] [PATCH 3/3] boost::ublas increasing the range of BLAS level 3 benchmarks To: [email protected] Date: Monday, 14 March, 2016, 16:12 Only in exceptional cases we make pull requests or changes in the master . Master only merges off develop ( that in turn merges off feature/bug branches). So the https://github.com/uBLAS/ublas/tree/feature/ublas00004_simd_gemm is the correct branch to request a pull. Pull requests go to https://github.com/uBLAS/ublas and NOT https://github.com/boostorg/ublas). I see the pull request in the boostorg repo, so please perform it in the ublas repo. I need to clarify this in the wiki because probably it is not very obvious. -Nasos On 03/13/2016 02:26 PM, palik imre wrote: A bit of confusion here. I created a fork of the feature branch you sent, as I didn't have the rights to push there. Then I sent a pull request for that. Should I fork the master instead? Thanks, Imre On Sunday, 13 March 2016, 19:03, palik imre <[email protected]> wrote: Results for low dimmension. More data would exceed mailing list limits: # m original: t1 MFLOPS original: t1 MFLOPS Diff nrm3 gemm: t2 MFLOPS Diff nrm4 mixed: t2 MFLOPS Diff nrm5 1 2.1263e-07 9.40601 1.32802e-07 15.06 0 1.36006e-07 14.7052 0 6.31318e-07 3.16798 0 2 2.28189e-07 70.1173 1.37767e-07 116.138 0 1.59801e-07 100.125 0 6.653e-07 24.0493 0 3 2.649e-07 203.851 1.5541e-07 347.468 0 1.54267e-07 350.042 0 6.98766e-07 77.2791 0 4 3.35269e-07 381.783 2.4183e-07 529.297 0 2.12891e-07 601.247 0 6.65688e-07 192.282 0 5 3.53868e-07 706.478 2.30977e-07 1082.36 0 2.53215e-07 987.303 0 7.07933e-07 353.141 0 6 4.2987e-07 1004.95 2.59713e-07 1663.37 0 2.54867e-07 1695 0 8.17448e-07 528.474 0 7 5.39621e-07 1271.26 4.51043e-07 1520.92 7.98975e-09 5.12948e-07 1337.37 7.98975e-09 8.76363e-07 782.781 0 8 6.18993e-07 1654.3 6.38988e-07 1602.53 4.17673e-09 6.37931e-07 1605.19 4.17673e-09 8.92556e-07 1147.27 0 9 7.73683e-07 1884.49 7.26336e-07 2007.34 3.30697e-09 8.00656e-07 1821.01 3.30697e-09 1.09762e-06 1328.33 0 10 9.27569e-07 2156.17 8.31827e-07 2404.35 1.94317e-09 8.72131e-07 2293.23 1.94317e-09 1.16572e-06 1715.68 0 11 1.13882e-06 2337.5 1.03275e-06 2577.58 1.27501e-09 1.08775e-06 2447.25 1.27501e-09 1.16439e-06 2286.17 0 12 1.26427e-06 2733.59 1.40013e-06 2468.34 8.50076e-10 1.39562e-06 2476.32 8.50076e-10 1.01202e-06 3414.97 0 13 1.5751e-06 2789.66 1.64811e-06 2666.09 5.39864e-10 1.66862e-06 2633.32 5.39864e-10 1.61517e-06 2720.45 0 14 1.79595e-06 3055.77 1.89937e-06 2889.37 4.08632e-10 1.6485e-06 3329.09 0 1.65016e-06 3325.73 0 15 2.14056e-06 3153.37 2.24248e-06 3010.06 2.73316e-10 1.6875e-06 3999.99 0 1.80164e-06 3746.59 0 16 2.38996e-06 3427.67 2.63386e-06 3110.27 2.30152e-10 1.74627e-06 4691.14 0 1.91648e-06 4274.49 0 17 2.93315e-06 3349.98 3.08031e-06 3189.94 1.85538e-10 2.17697e-06 4513.62 0 2.13505e-06 4602.23 0 18 3.3771e-06 3453.85 3.23863e-06 3601.52 1.20251e-10 2.23225e-06 5225.23 0 2.36877e-06 4924.07 0 19 4.19699e-06 3268.53 4.02621e-06 3407.17 1.07796e-10 2.29651e-06 5973.4 0 2.44714e-06 5605.72 0 20 4.27777e-06 3740.27 4.86115e-06 3291.4 8.37665e-11 2.26798e-06 7054.74 0 2.44016e-06 6556.96 0 21 5.58038e-06 3319.13 5.51606e-06 3357.83 5.93714e-11 2.61705e-06 7077.43 0 2.90197e-06 6382.56 0 22 5.46208e-06 3898.88 5.50258e-06 3870.19 5.76987e-11 2.85448e-06 7460.56 0 3.09923e-06 6871.39 0 23 7.26813e-06 3348.04 6.48407e-06 3752.89 4.47169e-11 3.03986e-06 8004.98 0 3.16566e-06 7686.86 0 24 6.56421e-06 4211.93 7.20581e-06 3836.9 3.61275e-11 2.84288e-06 9725.35 0 2.81577e-06 9818.99 0 25 7.97135e-06 3920.29 7.80654e-06 4003.06 3.02957e-11 4.04575e-06 7724.16 0 4.15001e-06 7530.11 0 26 8.59272e-06 4090.9 8.46934e-06 4150.5 2.53217e-11 4.1795e-06 8410.58 0 4.36958e-06 8044.71 0 27 1.05527e-05 3730.41 9.66865e-06 4071.51 1.97479e-11 4.24268e-06 9278.57 0 4.64476e-06 8475.37 0 28 9.77679e-06 4490.63 1.0918e-05 4021.26 1.71505e-11 4.41728e-06 9939.14 0 4.55165e-06 9645.73 0 29 1.23574e-05 3947.28 1.15308e-05 4230.22 1.54399e-11 4.96383e-06 9826.69 0 5.27042e-06 9255.05 0 30 1.25312e-05 4309.24 1.23192e-05 4383.4 1.38837e-11 5.36616e-06 10063.1 0 5.57707e-06 9682.51 0 31 1.41019e-05 4225.11 1.41554e-05 4209.15 1.12822e-11 5.56749e-06 10701.8 0 5.87983e-06 10133.3 0 32 1.44935e-05 4521.76 1.74419e-05 3757.38 9.5502e-12 5.91291e-06 11083.5 0 6.07622e-06 10785.7 0 33 1.68922e-05 4254.86 1.62224e-05 4430.55 8.00562e-12 6.51645e-06 11029.6 0 6.62821e-06 10843.7 0 34 1.73001e-05 4543.8 1.68924e-05 4653.46 7.54927e-12 6.83433e-06 11501.9 0 6.95343e-06 11304.9 0 35 2.07166e-05 4139.2 2.15962e-05 3970.61 6.52939e-12 7.06462e-06 12137.9 0 7.53811e-06 11375.5 0 36 1.98326e-05 4704.97 2.13473e-05 4371.14 5.68874e-12 6.703e-06 13920.9 0 6.99365e-06 13342.4 0 37 2.3838e-05 4249.78 2.23655e-05 4529.56 5.11318e-12 8.87253e-06 11417.9 0 9.13862e-06 11085.5 0 38 2.35903e-05 4652.09 2.48122e-05 4422.99 4.71306e-12 9.24238e-06 11874 0 9.27922e-06 11826.9 0 39 2.79913e-05 4238.39 2.64576e-05 4484.09 4.20714e-12 9.68511e-06 12249.5 0 9.95689e-06 11915.2 0 40 2.60131e-05 4920.6 2.9098e-05 4398.93 3.42002e-12 9.80308e-06 13057.1 0 1.04198e-05 12284.3 0 41 3.13419e-05 4398.01 3.03942e-05 4535.14 3.13757e-12 1.07587e-05 12812.2 0 1.10016e-05 12529.3 0 42 3.10015e-05 4779.64 3.20343e-05 4625.54 2.91245e-12 1.09989e-05 13471.9 0 1.16031e-05 12770.4 0 43 3.6527e-05 4353.33 3.49908e-05 4544.46 2.71446e-12 1.13164e-05 14051.7 0 1.20516e-05 13194.4 0 44 3.36654e-05 5060.62 3.86435e-05 4408.71 2.49076e-12 1.16151e-05 14667.8 0 1.21377e-05 14036.2 0 45 3.95282e-05 4610.63 3.98562e-05 4572.69 2.12037e-12 1.26784e-05 14374.8 0 1.32723e-05 13731.6 0 46 3.96351e-05 4911.6 4.17105e-05 4667.22 1.96734e-12 1.27302e-05 15292.2 0 1.34346e-05 14490.3 0 47 4.63424e-05 4480.69 4.50811e-05 4606.05 1.77515e-12 1.33133e-05 15596.9 0 1.39354e-05 14900.7 0 48 4.31748e-05 5122.99 5.0325e-05 4395.11 1.75073e-12 1.32491e-05 16694.3 0 1.3501e-05 16382.8 0 49 4.93001e-05 4772.77 5.11402e-05 4601.03 1.48788e-12 1.6222e-05 14504.9 0 1.72531e-05 13638 0 First group is legacy axpy_prod(), second group is legacy prod(), third group is legacy prod for low dimensions, and gemm() for high dimmension. fourth group is gemm(). As the legacy version is expression template based, it can possibly provide some further advantages when the operations are chained. I put some defines in place, that would make possible to force the legacy version as the default, as opposed to the runtime switched version. Imre On Friday, 11 March 2016, 14:21, Nasos Iliopoulos <[email protected]> wrote: Regardless, these are great figures. Can you please run them comparing the simple uBlas implementation for matrices from 2 to 100 with the gemm based one with a signle thread? I wonder when the control statement starts to play a role. What do you think should be the plan to switch from multi-core to to single-threaded so as to not get all the communication hit for smaller matrices? - Nasos _______________________________________________ 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] -----Inline Attachment Follows----- _______________________________________________ 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]