performance - Matlab + CUDA slow in solving matrix-vector equation A*x=B -



performance - Matlab + CUDA slow in solving matrix-vector equation A*x=B -

i calculating equation a*x=b, matrix , b vector, x reply (unknown) vector.

hardware specs: intel i7 3630qm (4 cores), nvidia geforce gt 640m (384 cuda cores)

here's example:

>> a=rand(5000); >> b=rand(5000,1); >> agpu=gpuarray(a); >> bgpu=gpuarray(b); >> tic;a\b;toc; elapsed time 1.382281 seconds. >> tic;agpu\bgpu;toc; elapsed time 4.775395 seconds.

somehow gpu much slower... why? slower in fft, inv, lu calculations, should related matrix division.

however, gpu much faster in matrix multiplication (the same data):

>> tic;a*b;toc; elapsed time 0.014700 seconds. >> tic;agpu*bgpu;toc; elapsed time 0.000505 seconds.

the main question why gpu a\b (mldivide) slow comparing cpu?

updated

here more results when a, b (on cpu), aa, bb (on gpu) rand(5000):

>> tic;fft(a);toc; elapsed time *0.117189 *seconds. >> tic;fft(aa);toc; elapsed time 1.062969 seconds. >> tic;fft(aa);toc; elapsed time 0.542242 seconds. >> tic;fft(aa);toc; elapsed time *0.229773* seconds. >> tic;fft(aa);toc;

bold times stable times. gpu twice slower. way, why gpu more slower on first 2 attempts? compiled twice firstly?

in addition:

>> tic;sin(a);toc; elapsed time *0.121008* seconds. >> tic;sin(aa);toc; elapsed time 0.020448 seconds. >> tic;sin(aa);toc; elapsed time 0.157209 seconds. >> tic;sin(aa);toc; elapsed time *0.000419 *seconds

after 2 calculations gpu incredibly faster in sin calculations.

so, still, why gpu slow in matrix division, fft , similar calculations, though fast in matrix multiplication , trigonometry? question should not that... gpu should faster in these calculations because matlab has released overlapped functions (mldivide, fft) gpu.

could help me solve these issues, please? :)

please read how matlab calculates solutions. help understand why gpu slower.

i'll seek in few words.

a*x=b becomes l*(u*x=y)=b, l*u=a

so matlab makes l*u (this process cannot done parallel far know instead steps can done parallel, due nature) then matlab solves l*y=b , finds y. (this process cannot done parallel each step requires info previous) then matlab solves u*x=y , finds x. (this process cannot done parallel each step requires info previous)

so gpu clock slower cpu, , since processes cannot done parallel, cpu faster. , no, unless come improve method (good luck!) gpu slower except in specific cases.

performance matlab matrix cuda linear-algebra

Comments

Post a Comment

Popular posts from this blog

web services - java.lang.NoClassDefFoundError: Could not initialize class net.sf.cglib.proxy.Enhancer -

web services - java.lang.NoClassDefFoundError: Could not initialize class net.sf.cglib.proxy.Enhancer - when trying request rest service @ time of response getting exception.. feb 21, 2013 2:34:49 pm com.sun.jersey.spi.container.containerresponse mapmappablecontainerexception severe: exception contained within mappablecontainerexception not mapped response, re-throwing http container java.lang.noclassdeffounderror: not initialize class net.sf.cglib.proxy.enhancer @ org.hibernate.proxy.pojo.cglib.cgliblazyinitializer.getproxyfactory(cgliblazyinitializer.java:117) @ org.hibernate.proxy.pojo.cglib.cglibproxyfactory.postinstantiate(cglibproxyfactory.java:43) @ org.hibernate.tuple.entity.pojoentitytuplizer.buildproxyfactory(pojoentitytuplizer.java:162) @ org.hibernate.tuple.entity.abstractentitytuplizer.<init>(abstractentitytuplizer.java:135) @ org.hibernate.tuple.entity.pojoentitytuplizer.<init>(pojoentitytuplizer.java:55) @ org.hibernate.tuple.e...
Read more

Accessing MATLAB's unicode strings from C -

Accessing MATLAB's unicode strings from C - how can access underlying unicode info of matlab strings through matlab engine or mex c interfaces? here's example. let's set unicode characters in utf-8 encoded file test.txt, read as fid=fopen('test.txt','r','l','utf-8'); s=fscanf(fid, '%s') in matlab. now if first feature('defaultcharacterset', 'utf-8') , c engevalstring(ep, "s") , output text file utf-8. proves matlab stores unicode internally. if mxarraytostring(enggetvariable(ep, "s")) , unicode2native(s, 'latin-1') give me in matlab: non-latin-1 characters replaced character code 26. need getting access underlying unicode info c string in unicode format (utf-8, utf-16, etc.), , preserving non-latin-1 characters. is possible? my platform os x, matlab r2012b. addendum: documentation explicitly states "[mxarraytostring()] supports multibyte encoded characters...
Read more

javascript - mongodb won't find my schema method in nested container -

javascript - mongodb won't find my schema method in nested container - i trying access method of schema stored within mixed container. here situation : have cases model can many different things, have schema each of these things stored in "casecontent" mixed property. var caseschema = mongoose.schema({ casecontent : {}, object : {type:string, default : "null"}, collision : {type : boolean, default : false} }); the casecontent property filled model of 1 of schemas, 1 exemple : var treeschema = new mongoose.schema({ applecount : {type : number, default : 3} }); treeschema.methods.dostuff = function (data) { console.log('hey, listen'); homecoming true; }; then, want utilize method of schema original container : caseschema.methods.dostuff = function (data) { if (this.casecontent.dostuff !== undefined) { this.casecontent.dostuff(); console.log('it worked'); } else { ...
Read more