0:00:15after after everyone
0:00:19but they meant hobbies the pair implementation for multi
0:00:22that mission we them in simple improve
0:00:25cool mode decomposition
0:00:28i first that were talking about the big one for the uh
0:00:31much what the measuring simple in three promoting something we call we
0:00:37and what our what the why we wanna palace that and lay i with the
0:00:41show you how we per rest of that and then show the result
0:00:45in cruise
0:00:49and we I D's that fine
0:00:51frequency and it's for the multi
0:00:54um at at mission data
0:00:57and as you can be price that time to the spatial uh to this
0:01:00spatial domain data
0:01:04it's hits be under uh multi data
0:01:07different uh and that a he's P the block is uh uh empirical mode decomposition emd
0:01:13and i in emd D
0:01:16only can work on the one D
0:01:18and then we use then
0:01:20is then that use a uh is that a
0:01:24use then uh
0:01:26a a bad know
0:01:28i didn't know it so we have a in simple
0:01:31empirical mode decomposition because record E M D but i E N D self
0:01:35still working on the one D data
0:01:39E E E and D provide the possibility to extend to the much but i'm use that because of making
0:01:44a cake
0:01:45have a a uh
0:01:46some probabilities of copper that to uh
0:01:51a a to to date a uh along the sent mission but uh two
0:01:55post cruel uh neighbouring sliced data
0:02:00so a i sure that uh
0:02:03the part a diagram for E and D
0:02:05so in the you know the you you are around the multi but i have a procedure to extract that
0:02:09interesting model
0:02:11uh frequency
0:02:12i we live
0:02:13and the each i have for see joe you have a several shift in procedure
0:02:17in uh
0:02:19in the several system procedure in that
0:02:21it just different procedure
0:02:23we best your try to find a a a a envelope and use use that
0:02:27that difference of the envelope
0:02:29and uh all the of of envelope a and all real signal
0:02:33to to the and have a
0:02:36and but to finesse in moderation
0:02:40in order to to lot we have to uh we have to to find the the we have to find
0:02:45the each trampling of the the stick no local each point of the second
0:02:49and the use a lap uh this lost if from a points to your uh to or the envelope
0:02:55and you in this
0:02:58more most come we use a spline to to build that
0:03:02the envelope
0:03:04we also know the the
0:03:06this my of issues
0:03:08uh is quite uh is
0:03:09hi close to do so
0:03:12and based on the fighting
0:03:14we with discover the strip the post us if civil inverse joe uh
0:03:19i a nineteen a percent of the computation
0:03:21in the emd
0:03:23uh as the or the several be a uh several or be of the block of the shift shifting
0:03:29and the sum uh some a shell from this process
0:03:32not not that i think that percent of the exclusion
0:03:37so and
0:03:45if we movie seem uh we use uh uh
0:03:48uh if what when you uh if he and the will come on the one C can a one T
0:03:52signal and if he's in sacks
0:03:55i he's out voice S C one to C K
0:03:58E and the
0:03:59can be that show ask the the for uh this speaker so for that each D don't we add the
0:04:05different with to generate
0:04:07uh a a a symbol set of a signal
0:04:09and they for the H T going we passed through order
0:04:12uh have the reading in the uh independent and the which are right it's
0:04:17it's own output and the we have a to the output
0:04:20a a and the
0:04:21we defined the final output
0:04:24uh as a as a uh as the ever to so
0:04:29so a
0:04:31so i so we can see the but you in the E do you have a a a lot of
0:04:35a uh a a lot of U M the you we need to S Q
0:04:39in a
0:04:40in the and me and D
0:04:41that's and we use a E and the S appeared building block
0:04:44so if we uh i i i use a two D uh two D image just example
0:04:49so we first uh a uh from the emd a the one dimension
0:04:55so which generates a
0:04:57some partial sell
0:04:58and the we perform a be uh would have a E and D
0:05:01and to they are not a technician
0:05:04so we have a a two D uh we have a two D result
0:05:07and we basis
0:05:08would best to uh
0:05:10oh oh
0:05:12which uh use the this as a coefficient uh and uh
0:05:16to generate the final result
0:05:19and the put in the uh
0:05:20the close in probably is to because so i'm only show in the in the pit
0:05:25and in a E i E and E we uh for the one i i i M emd of the
0:05:31well fight of by five have image
0:05:34oh we need to exit you have to the
0:05:36a three uh thirty meaning and
0:05:41and we can see that's a computation in uh come tension tense if and that network
0:05:46now use series
0:05:47the basic a were late and the need a is that uh
0:05:53it's a a a real
0:05:54so our go is try to the paradise that i was a set or a the and M D
0:05:59the people i told before i talk about uh
0:06:02how how we pay the i sure that the some X uh and some resolve the and even D
0:06:08so that's a the
0:06:10that's a a uh a you mean you of the uh
0:06:12a two D image of the
0:06:16i have put for from the T and the
0:06:19the or or don't you meet you can be T cost all of the uh
0:06:23because to the fight fight image
0:06:28that differ uh that if and that different or a more a frequency
0:06:34you can think of the the their first speakers is that high frequency of the
0:06:39uh a high i have to have a high was because the of the original so should
0:06:44and that that the last wise of the lowest frequency so that scrolls the the but high pass and low
0:06:53so i'll go is high to paradise the any and the uh use a general part uh general purpose processor
0:07:01so and
0:07:02base a general process forces uh process support the power a ration
0:07:06like uh uh a was P you
0:07:10or the
0:07:12uh and uh also the the many code you Q
0:07:15a graphic are a a a a
0:07:17graph uh graphic
0:07:18processing you need angle go graphic are
0:07:21so why we P could general part i job as processes
0:07:23yes system more common
0:07:25so if it comes to the almost every time has a
0:07:28if even cell phone has a
0:07:29but we are not talking on the set of
0:07:33also he's
0:07:33he uh he's dislike cycles ace
0:07:36compare the has to use used this is a work
0:07:39and you have some i for by the more uh flexibility so sometimes has you need to my the uh
0:07:44you need to to the function of your program
0:07:48a a so
0:07:49so that's the a some which was so uh
0:07:51uh general purpose
0:07:55so and i was lose the the that uh several the a pair of a model
0:08:01like a a uh old maybe for if use magic was if you and the could a and be that
0:08:06could have for the cheap used
0:08:07i also for uh
0:08:09that's that's some and then to open C or is target on there
0:08:13a a a lot of about picture
0:08:15like us if you to if U S P F P G A event i feature
0:08:19but we listen that's of future work
0:08:21so you nice talk we only talk about the open C O and the
0:08:24could the
0:08:25is uh for the see if you entropy was
0:08:30basic a all is the appropriate model for the multi close if you
0:08:33and he
0:08:35he use uh us that annotation to to label the loop
0:08:40use the have to label the loose when self
0:08:43but well is annotation
0:08:45the compatible to understand that which should double can be powerless
0:08:50and for the H uh for they should power loop
0:08:52that you to ration a large group can be to your schedule to this stem rows
0:08:57see use the rest
0:08:59and the rest
0:09:00the rest part the open all that still run time manager
0:09:06could i an a uh can model for the be to you
0:09:11could apple right the the two layer uh a to layer at a picture
0:09:15uh which are that at the red a plot to control to people program
0:09:20and the right
0:09:21is something that the scalar execution of the a uh
0:09:25i time did mod and there it is some fact that the scalar as a fusion of the middle of
0:09:32uh a tell uh
0:09:34which to uh
0:09:37major two should
0:09:38and that third rock face of the
0:09:40something like a vector speech itself
0:09:44the in in that there that each there were performed the send
0:09:47send reach an expression your in your codes
0:09:50and the
0:09:51in a just the record
0:09:54it can't if is the map a many need there it
0:09:57and the
0:10:00we have a how a a ha on shape on scratch the memory to come you to
0:10:05to help to come kate
0:10:07a that the red
0:10:09and in this that sent there to brock
0:10:11and that is that there about half true
0:10:14have to be uh independent
0:10:20yeah i sure the the out six to to the uh
0:10:23all opaque at imputation
0:10:25of the uh of a i mean E D
0:10:27so basically basic are signed the one third of uh wise if you there to handle the one em M
0:10:33so why would do would itself read
0:10:35a the the
0:10:37one stiff used there to handle the one
0:10:41because in the D we have the we share the an input data
0:10:46so we
0:10:46in a so this cage doing can put up of fight that they capture in in in the case
0:10:57so a all sure the X one
0:11:00if we would do the office uh the outer scheduling joint late words they would not have a ship they
0:11:04have sure so that was all on performance
0:11:11as as if you uh as
0:11:13you also provide that the vector or to um bit bidding you next Q shouldn't i say C
0:11:18so the in this case we only we we around the the i I C in table uh C compiler
0:11:24to help of us to to find a vector um vector iteration
0:11:30and the for high dimensional data
0:11:32and night
0:11:33the a two do you have a
0:11:35the you it process S Q a E and the in the
0:11:38one the and U S as the D and then not that mission but had
0:11:42for this i mission high dimensional data
0:11:45because you S is paid to use a different value of that
0:11:49you are the
0:11:50the members storage
0:11:51that direction
0:11:52so in this case is you was uh
0:11:56the performance of a cash will be slowed down
0:11:58so would base it to that in split as
0:12:01express a transport of its per the pull
0:12:06of the data
0:12:07so make a like
0:12:08um uh clay have
0:12:10um a like a sheet
0:12:11so that will provide the better duration of cash
0:12:16and the in that you for uh in the
0:12:18could i for station we have a different strategies for for different strategy
0:12:22we basic a signed the one if you there
0:12:25to the white and these that emd big how that
0:12:29it tells the the
0:12:31the because
0:12:33G you have a most to right
0:12:35so he's a he have to be the more find a more fine of grand uh full matt final in
0:12:40the program
0:12:42this cajun will five provide the mess of data power
0:12:46but that was a cost another not problem
0:12:49so we basis cage order
0:12:52sketch to what they're to the what one white
0:12:55is self for one
0:12:56the whole why wise
0:13:02so we
0:13:03so the
0:13:03but if you
0:13:04they tie in the uh the sent uh it in the same thread block
0:13:09may have to uh bridge direction
0:13:12uh because
0:13:13uh because we had as noise
0:13:17so would basic a or as this thing uh
0:13:20schedule wife this send X
0:13:22to the same that block
0:13:24uh a to the sent their brought to avoid the a range that richard
0:13:28that's all they all
0:13:29come from this send X
0:13:31they have a higher as us an energy
0:13:36so we basically scale use something that is
0:13:38is that often this
0:13:41the rest how we just by up there that one cache
0:13:48and in in the put i'm measure what didn't do that the the ah a
0:13:51the uh had them uh exposition transport
0:13:57but we that have a high dimensional man chorus in a that has is proper
0:14:02so we
0:14:03best again low that data how to the function of sketch pen memory
0:14:07so we will low the something like this
0:14:09to the question had a memory
0:14:12in disk
0:14:13and in the switch pay very with to this set the this to has S
0:14:18and i i'm because
0:14:19the pen we for scratchpad memories really hide and they since is and really low so
0:14:24we still can benefit
0:14:28so and an outer is some we didn't do the the uh it's is especially transpose
0:14:33even you to the transpose you still need to store data also this
0:14:37this this local transport
0:14:40she's is what ten remove the the read this memory S as
0:14:45and that's an an you should do you have a limited them a memory
0:14:49so compare was strip you
0:14:52how how C if you
0:14:56so in in your man you base you probably uh a basic can has uh i to the sits for
0:15:01key up lights
0:15:03but for the one to if you only can have a a for the three yeah up by man
0:15:08but in this
0:15:09in this
0:15:10and me and D
0:15:12uh we have to sub for the up to the uh a tough the up i is for fight to
0:15:16have a five tough
0:15:18to get the flow pair
0:15:21the nice
0:15:22uh in feasible so
0:15:24so would
0:15:24we basically
0:15:26to the pitch exclusion so with tie out that they uh how the data
0:15:30to several slice and it just slides can fit in
0:15:32that you memory
0:15:34and you this then says size
0:15:38lay can as the execute
0:15:40in the power or in that difference that size two
0:15:43in the sequential
0:15:46a list the hours
0:15:48uh all
0:15:49oh is yeah uh
0:15:51is very environment
0:15:52uh is we run on the
0:15:55and me idea and of
0:15:56full uh full stock it's
0:15:58a call
0:15:59uh to get a entails
0:16:02a X um X seven five fifty with this six will be out by memory
0:16:06but only the first
0:16:08first sixteen in are use
0:16:11an a if you run on the ready it uh i the yeah for me test uh C twenty fifty
0:16:16two you
0:16:17they only provide a a a a a very gap by memory
0:16:19and i would you could there for has us C to fifty
0:16:22because for for me
0:16:24from yeah at a trace their first at a first
0:16:28really top yeah under that a a of procedure for point is
0:16:32and i'll whole result is
0:16:35uh is
0:16:35is double precision
0:16:38and we'll on that's three and three uh
0:16:42test case
0:16:43and this is their size
0:16:45and that's zero exclusion we based
0:16:47we best the only schedule the one call
0:16:50on this motion
0:16:51so and that that
0:16:54then you is second
0:16:56so and the in the could ice occlusion
0:16:59uh in in the could i from and in implementation
0:17:03we can see we got a have for the uh a a wrong around uh a of fifty
0:17:07X P compare of a sequential budget
0:17:11and in all then
0:17:12and he implementation
0:17:14we gather up to that you avon
0:17:16uh speed up compare of a sequential
0:17:21so we also yeah evaluate the the scalability of the
0:17:25uh a T
0:17:28by changing the how many to read that we schedule
0:17:32we base
0:17:33basing for as a base to provide almost noon near
0:17:37no more almost a P scalability for prime T
0:17:40or if you only schedule no then
0:17:43a called
0:17:45but after lot
0:17:47he was like to right
0:17:49so the members some we believe that's the the
0:17:52the memory bound
0:17:54can i cannot for uh is limited
0:17:57and also lost
0:17:58probably no my fact will also
0:18:00cost not
0:18:04we also plot provide
0:18:06for the
0:18:07for the uh the present you just keep a of the exclusion
0:18:11uh uh the
0:18:13uh a the but uh a among the block
0:18:18oh my is basically provide an this then a simple find has a sequential
0:18:23and the in the
0:18:24a a good i'm from station the
0:18:27the profile is a the to the different
0:18:30the input in uh in separation pot is wrong slightly faster
0:18:34but and the though each mark tape to price size
0:18:38and the man is we pretty that in the present prize
0:18:42uh is
0:18:43i suitable for the message powers
0:18:46because they did not have a a um there's something in uh there's on the as
0:18:51but should at takes so you have a uh a a a that the sorry
0:18:55each are how you passed
0:18:57because uh
0:18:59because you cannot not fit the
0:19:01uh a the the where
0:19:04where they extra our point where happen
0:19:07so last that they'd had a and then you regular T
0:19:12to some rice
0:19:13and this one provide tools to there right uh to their label power or or i me and the from
0:19:18station four
0:19:20for cheap you C if you
0:19:23our program model is screwed i and with zero
0:19:26i we got the pretty good speed
0:19:28and just get up
0:19:32this what
0:19:34the call will be released and didn't in this way say
0:19:37and also this work this
0:19:41in this library
0:19:43uh power em D N E and the also all support
0:19:46i also support
0:19:49uh is not
0:19:51how uh how we to light is not show in the this uh
0:19:54uh this paper we published to
0:19:56another that france
0:19:59for the future work we will provide uh
0:20:02still wrote true you can support a a a a and D if you and and the ad dot texture
0:20:07and a
0:20:09you also my the hard to pin of you you can only
0:20:12use a one still scroll to long on the uh to on and you have to take
0:20:18i put that's all
0:20:19that's all i have
0:20:21and question
0:20:26a questions only
0:20:49oh we include
0:20:53and in the paper we showed the
0:20:55the uh chance uh
0:20:57that actions run times the make uh less than one person
0:21:02on percent of the total i'm person of the took place cushion
0:21:16like the speaker
0:21:17and that's C and uh