Dealing with lack of floating point precision in OpenCL particle system -

-


Dealing with lack of floating point precision in OpenCL particle system -

i'm writing opencl based particle scheme speed visualizations of big scale networks. in essence, 2 phase problem phase 1 applies negative gravity each particle (typical n-bodies problem) repel , phase 2 attracts particles based on edges (or springs) between particles.

during each iteration of gravity algorithm each particle's location, represented pair of floats, impacted distance each other particle (classical physics model, no drag, keeping simple).

in situation 1 has spaced out square array of particles application of gravity should result in symmetry across both x , y axes. true @ origin of gravity application, on time lack of precision inherent in adding lots of floating point numbers results in little non-uniform deviations. this, in turn propagates through entire n-body scheme , loss of symmetry occurs.

one simple way avoid utilize double precision numbers, geforce 9600m gt on macbook pro not back upwards double precision numbers. so, what's nice way deal such problems in opencl? i've thought truncating floating point numbers i'm adding few decimals avoid problem, seems bit hokey.

this pretty mutual problem; on cpus, want avoid utilize of double precisions avoid factor of 2 in memory overhead/bandwidth.

a number of molecular dynamics , n-body codes written gpus utilize "mixed-precision" arithmetic; store particle positions , velocities single precisions, utilize double precision few key operations -- typically store position differences, , accumulate accelerations. (googling "mixed precision" "molecular dynamics" or "n-body" gives tonnes of results).

so reduces number of double precision calculations, not zero. implement higher precision arithmetic hardware natively supports, can software emulation, emulating double 2 floats. there venerable fortran library dsfun90 implemented this, , in this nvidia forum implemented similar in cuda (based on operations in nvidia's mandelbrot example). don't know of opencl implementation offhand copying on cuda should pretty straightforward. it's not fast native doubles, if it's few key operations it's not bad.

floating-point opencl precision

Comments

Post a Comment

Popular posts from this blog

How do I check if an insert was successful with MySQLdb in Python? -

-
How do I check if an insert was successful with MySQLdb in Python? - i have code: cursor = conn.cursor() cursor.execute(("insert new_files (videos_id, filename, " "is_processing) values (%s,%s,1)"), (id, filename)) logging.warn("%d", cursor.rowcount) if (cursor.rowcount == 1): logging.info("inserted values %d, %s", id, filename) else: logging.warn("failed insert values %d, %s", id, filename) cursor.close() fun is, cursor.rowcount always one, though updated database create videos_id unique key. is, insert fails because in tests same videos_id going appear (and when check database, nil inserted). whatever reason, rowcount 1 - logging.warn have spits out rowcount of 1. so, question: can utilize rowcount work out if insert went fine? if so, (presumably) doing wrong? otherwise, how check if insert went fine? your code not commit after modifications (your modifications rolled back). should...
Read more

delphi - blogger via idHTTP : error 400 bad request -

-
delphi - blogger via idHTTP : error 400 bad request - i'm trying post blogger using idhttp component, however, i'm getting "http/1.0 400 bad request" error. procedure tform1.button1click(sender: tobject); var request,response,req : tstringlist; auth,blogid : string; begin blogid := '00000000000000000000000'; request := tstringlist.create; response := tstringlist.create; req := tstringlist.create; idhttp1.request.connection := 'keep-alive'; idhttp1.request.contenttype := 'application/x-www-form-urlencoded'; idssliohandlersocket1.ssloptions.method := sslvsslv23; request.clear(); request.values['accounttype'] := 'google'; request.values['email'] := 'xxx@gmail.com'; request.values['passwd'] := 'yyy'; request.values['service'] := 'blogger'; response.text :=idhttp1.post('https://www.google.com/accounts/clientlogin',request); auth := resp...
Read more

postgresql - ERROR: operator is not unique: unknown + unknown -

-
postgresql - ERROR: operator is not unique: unknown + unknown - i have postgis database , have compute new values rows in new column. these values should average of values of several columns. query: insert bdps (da_m) values (avg('da_1'+'da_2'+'da_3'+'da_4'+'da_5'+'da_6'+'da_7')); in query bdps database, da_m new column , da_1 da_7 existing columns have double precision type. da_m created using alter table bdps add together column da_m double precision; i next error: error: operator not unique: unknown + unknown line 2: values (avg('da_1'+'da_2'+'da_3'+'da_4'+'da_5'+'da_6'+'da_7... ^ hint: not take best candidate operator. might need add together explicit type casts. ********** error ********** error: operator not unique: unknown + unknown sql state: 42725 hint: not take best candidate operator. might need add togethe...
Read more