0:00:15a all
0:00:16would have to know
0:00:17my name is the one who's on i'm phones or nation and universe
0:00:24you know
0:00:25signal processing and williams the a team and uh
0:00:29much computation you yeah i especially
0:00:32indeed you these are the
0:00:34a polite detail signal processing i
0:00:37detection of processing
0:00:39that it would in terms of two
0:00:41or high sampling applications
0:00:48you you gaze
0:00:50yeah actually so we
0:00:52used to use uh
0:00:55we we as i is the circus about
0:00:58you know what you had a architecture based the sheep use that you know it was
0:01:02also to
0:01:03and place up to the of a stopwatch
0:01:07here you have this to some problems
0:01:12four is the problem is so that she
0:01:14the other sam she can share everybody
0:01:17so that you can
0:01:19how do you have a lot
0:01:21word or
0:01:23i guess was always signal processing algorithms are high high hell of it
0:01:29one cup to what your voice you can so i would assume so uh that that based on feedback
0:01:35so you could have an adaptive filter is uh
0:01:40one audible discontinuity
0:01:42and the
0:01:44another one is uh architecture space people were not or don't have outermost in so i
0:01:51E twenty four we
0:01:53a pit i had already would you
0:01:55jones there are some
0:01:57well what have the
0:01:58but we do all word uh sometimes at the dollar alignment or word awhile soon corning journal where the white
0:02:04in top artists communication
0:02:06or what
0:02:10the algorithm that
0:02:13i'm going to project here at least uh all the two but that is a filter
0:02:20yeah so you know what are the two that is a filter is uh use the for channel equalization noise
0:02:26beamforming and many applications
0:02:30and all the team of filter
0:02:32so uh components available mood you up with that
0:02:35so why not forties we use uh
0:02:38this kind of able more generality
0:02:41so that is awful wants to have the
0:02:46but the video over panel
0:02:47panel of them that you can get we do
0:02:50this kind of will
0:02:51so show that it had on that you can be so high say
0:02:55this days ten so the filter
0:02:58then don't the video will have that you is just
0:03:00ten so
0:03:02for a parallel or yeah we needed to see
0:03:05are the kind of a of that so there's of your point for the the thought of it of it
0:03:12a not that question is so
0:03:14why they all that to elect is if it does all that table
0:03:17tenth to filter is um
0:03:19wall frequent to used
0:03:22well the to of the or the T what had the delay line filters
0:03:26uh that you've got the to had a light that had a
0:03:30that that can be paralysed
0:03:32but the what where there is a bit is do so
0:03:35we each to pet lights all that to that use a it
0:03:39and the use of those two gonna a system performance such as uh
0:03:43what that patients with the
0:03:45yeah but with a kind of battle for that you will that is
0:03:55well we try to implement the disagreed the sin the and the mote acquired this dog
0:04:03first uh
0:04:05i'm going to
0:04:08i'm going to
0:04:10pretty pretty discuss a sin then white choir architectures
0:04:15a i'm like the and the pad a computational of all that two would like to the filters
0:04:22i'm with the discuss a or of the and take out of a parallel port
0:04:28and that top home with this the work so
0:04:30have a a have a red id H a noble the in you can all see creations
0:04:35the you you can see K john
0:04:37so i use the for all the two that two filters
0:04:40then i'm going to sure
0:04:42experiment to reject to and
0:04:51in space space if you
0:04:54seven that had other processing feature are you
0:04:59i don't use those who
0:05:02but also pa skylark take your
0:05:05we use a a what that you call a compile a couple a C
0:05:10does second
0:05:12parallel architectures is seemed that picture
0:05:14that is a single are used stuck what but data
0:05:17for the case of which and architecture
0:05:21we don't one instruction
0:05:23you can come down to
0:05:25say in in this a four addition
0:05:29at one time
0:05:32this is a very efficient
0:05:35or processing approach
0:05:39this so
0:05:40architecture or shows the
0:05:42some or word of then
0:05:44date i an a line this say
0:05:48you put be
0:05:51one to at the age get a to be you on a you want to P tool
0:05:54then and
0:05:55uh i there should be
0:05:58well well i want to step and then a novel
0:06:09and always so and architecture i
0:06:11use a problem too
0:06:14then what do but to to
0:06:19and the doors
0:06:22on not the architect had our a processor architecture that are used the shape you is a a to clock
0:06:31you know in in the you this to be here
0:06:34well do you are off
0:06:35for the claw up a lot in in soul do on global "'cause" side
0:06:40in you we're most there we here because a
0:06:44on more and more out so is uh
0:06:47implemented on the chip
0:06:50so we here form what quark architecture we need a to and play by to pursue of the so this
0:06:56is all point for the set of that the have of the region
0:06:59and of force so that that have that have a really do we need to consider either the synchronization model
0:07:04where the peak or do do you do pause or not
0:07:07working in no
0:07:11uh so so all
0:07:13for ten so speed that can be
0:07:17the me to the bias in coney need additional way
0:07:28that that this of the whole
0:07:31in shall
0:07:35you can and priori
0:07:37they turn that there a and so that that that of the to
0:07:40and uh for the seemed architecture that is a date other of it and the V
0:07:45you to can see that date all alignment or what
0:07:48and of for the case of but entire as as the you for O D video of a are you
0:07:53miss of the media the for
0:07:55and of for the case on work well
0:07:58but T deal will have other things a proportional out to do a mobile courses so
0:08:03the or number bubble pause so
0:08:05you know that
0:08:13this is the gradient all that to will
0:08:16that is to filter
0:08:19and that this a fit closest of will
0:08:22yeah a uh several steps
0:08:24but don't operation and each is that
0:08:27is a i didn't colour
0:08:29and the one important thing is that
0:08:32do fit a coefficient to a a or
0:08:35or that thilo by
0:08:36low cut and a signal
0:08:39or all
0:08:41we can all that to dull
0:08:43phil top quite efficient to using the global
0:08:47that our signal
0:08:49in my to process the implementation
0:08:52it is a very important to use a local feed the X so
0:08:56oh one all the T job a parallel and i guess some for this a filter
0:09:01use of coming in from the fact to their uh you can use a look card field of F for
0:09:07for all that station
0:09:08yeah so
0:09:11as a of filter shows
0:09:13filter coefficient the K
0:09:15and the filter coefficient to a each she our or or that to the by a local fusion
0:09:28as a i
0:09:29mentioned a or you'll
0:09:31one are be also a is so
0:09:35pat a parallel i do each is that
0:09:39okay so this so the this is open code the crunch sort of a parallel region
0:09:44here are to do so
0:09:46we pull type dining
0:09:48that used is
0:09:49at each is stage you close that
0:09:54compared this to figure your
0:09:58and this to the guillaume E
0:10:01we put pipelining registers at each class it
0:10:08but yeah i didn't type or but you turn
0:10:11you couldn't do
0:10:12computational what this so these the D just that
0:10:15independent to
0:10:18but computational ball
0:10:21or it this just that have and this is step can not be
0:10:24yeah does same it there because of type dining they "'cause" so each is that
0:10:30oh one step delay does so
0:10:33then you we compute at time and for this is that
0:10:36then we need to to compute a and plus eight
0:10:40and plus a one for this is that so
0:10:42uh yeah this a
0:10:44well if you're the
0:10:45yeah this uh
0:10:47oh it is it shows times that
0:10:49and that is so it shows others
0:10:51yeah stage a number
0:10:53it the we need to top i some skew the process
0:10:59and you a
0:11:00what we do know is sort the pipelining used to this
0:11:04yeah the that in use these of high more because the but normal pipelining registers is just them
0:11:15this so the band to your boat this top low is the total number of able don't do really of
0:11:20a parallel a jim is a the emitted the bike to two
0:11:26and uh so on a approach is to use a T turn have that have a the computation take that
0:11:32a back a our computation
0:11:34means that
0:11:35we we try to compute to board with and press that the time so you want for of four is
0:11:41the uh what that to for the
0:11:43you can a like take out of that that the um
0:11:46and the you C O A you can use the ten thousand
0:11:52for eight are that it had a yeah
0:11:54since the V
0:11:56yeah fast fast racing many many samples of but they are in signal processing
0:12:00don't think a pair of a jim is not be to the by filter told us so use some well
0:12:05i didn't
0:12:06so is to more idea
0:12:08but i use a more complex
0:12:11here L
0:12:12and so the state G
0:12:16we norm value of the key and minus a lot of F and minus than
0:12:20and do we need the to compute to D
0:12:24and and a computation computation
0:12:26couldn't be transformed or to store the
0:12:30the you know the currency creation so that
0:12:33you shown use of high complex about
0:12:36do you can be a compute either
0:12:38we that depend noise do not depend on it means the
0:12:41a a three the back
0:12:44but i are he and my so one
0:12:47depend you and yeah and minus one and minus one
0:12:56yeah so
0:12:58well this can be translated to to yeah and minus so one and
0:13:03you participate bit i E N minus one M mine some X the N so
0:13:08this it becomes up
0:13:09for told a constant coefficient
0:13:12the in you know the can creation
0:13:24yeah is
0:13:26there are several you creations here
0:13:32and that
0:13:33or not the case a used computational able to yeah and
0:13:37in this case
0:13:40can always still be come but to do to
0:13:42time be quite efficient to the you know the colours see cajun
0:13:47so okay K M and comes at eight K end times K N and minus one
0:13:52plus X K N
0:13:53for the case of what time they are the in you you can seek shown
0:13:57oh that's all the to be more complex so then though
0:14:00parallel order like duration of will
0:14:02tossed the coefficient case
0:14:10i and like to put a forty
0:14:12well you spain
0:14:13how be compared a like you the in you know the common seek kitchen
0:14:17you know
0:14:19the in you the currency cajun and use a the back so then
0:14:24that makes the dependence problem because also
0:14:27to compute to Y N we need the why and my so on to Y M minus want and
0:14:33oh so you since a very difficult to each is seems not a large got to compute the why and
0:14:39a Y and minus and Y and plus the simultaneous it but it is
0:14:53why i though we so say you at the times a why my some plus sixty at all
0:15:00oh why why mine a one is so
0:15:04initial condition
0:15:07since that this is a
0:15:09the in you know
0:15:11yeah different C creation
0:15:13you couldn't you could
0:15:14do calm or just leading to two part
0:15:17one is not time you to a solution
0:15:19and the yeah is a particular solution
0:15:22so you for be you go lot the initial condition you can compute to
0:15:27particular solution
0:15:28and that you if we can see though one you'd a particular solution
0:15:32a particular solution
0:15:34can be computed that you had other people's so
0:15:37yeah i i even lot thing in each can be shy and that that breaks the dependence problem
0:15:43then to problems are how we can
0:15:46well compensate for that depend problem
0:15:50you you these days so
0:15:51the in your the currency K john
0:15:57we can a polite blue i'd the mess that to compensate for the
0:16:02well effect all good initial condition
0:16:05so by computing
0:16:07had you can also lose and you a solution
0:16:10oh a to D
0:16:12no become paralysed a computational able you know the can seek age
0:16:21this so at the forces that
0:16:23it do we compute the particular solution
0:16:26and that does step two
0:16:28we prop a okay to the initial condition
0:16:31and that the propagation is kind to the
0:16:33by i in particular solution and the initial condition
0:16:37and the finally we compute the companies solution
0:16:41so do you do a of computation
0:16:44you the some or where the pattern or idea a symbol for the would for the case of O
0:16:49for although though the in you know the because uh
0:16:51though or why there is uh not more than twice
0:16:54though number of will she of computation
0:16:58so that instead
0:16:59you full we use of four processors
0:17:02in C D you can guess say a a on the tools that the pose and top was speed the
0:17:08and the for the constant the coefficient the case of tar speed or can be a how you
0:17:14okay this is the payment L reader
0:17:17a a did X meant to be the
0:17:19well you inter
0:17:21oh for the quasi C pure
0:17:23and the this so has a
0:17:26for a she in the
0:17:34and that this this shows of though
0:17:36or reject to this is the fiddle the
0:17:39and it's of for a a was sixty in
0:17:42and the
0:17:43a a this is it you can sell which are to then
0:17:47processing L samples
0:17:49and that this is thus in the parallel the
0:17:52uh a processing and sample
0:17:54so by a plan seemed that
0:17:57oh be you get the speed P T in
0:17:59a about two hundred all
0:18:01fifty percent the tool
0:18:03two seven used to represent
0:18:06and that this is so these are two for data that the parallel region
0:18:12we sin the so
0:18:14this to mean that using in the instructions so
0:18:17you i
0:18:19no data level parallel to a and a be the that we get
0:18:23or or or are they are higher speed up
0:18:25well don't well but the or over a to lead them is of for whistle four
0:18:32a for the case to an to quell
0:18:34one one show that had L later
0:18:37we can hard but you get any game because some to quite cheap you
0:18:42well need the some C duration mobile but
0:18:47yeah a a a i'm so we do we do some sort of a parallel region with the multiple acquired
0:18:52B can that
0:18:53a a up
0:18:57or sort is it talk case or but data that that have a and we the not to quell and
0:19:01the you couldn't get a bar one hundred the pose in both speed up
0:19:06and the by combining portal
0:19:08data to better parallel do using by to core and C and you get a ball
0:19:13for hundred the nist
0:19:16so to pose and top was speed up
0:19:24this so all experiment experiment a are two shows that
0:19:29oh i
0:19:34for pat computation to read the took why cpus though effect of well
0:19:40in a mean you cajun and that's and corny duration will where the P comes of thirty problem
0:19:45so we need a on this or that
0:19:47or tech and
0:19:48or a but this kind of a single you a channel with that we do
0:19:53i'm sure about had out of a job it it a very difficult to to up so will
0:19:57this kind of a sink one you H job would that be do
0:20:01by applying data never parallel and you can
0:20:05oh well or would this kind of a or where the and the still get some
0:20:14this is the company no but this do you have we implement T the to but tight this fit on
0:20:20but to in the
0:20:22for type the void no bought that will filter is
0:20:26and uh we get
0:20:29fairly close speed uh with a small right toys C then and police in the in
0:20:36for data that have are processing and
0:20:38we couldn't guess
0:20:39speed uh
0:20:40for some more to Y and dark picture
0:20:44and uh
0:20:46we believe but that this top to can be a provide the to are don't or the T that two
0:20:51and those was a sure D
0:20:54yeah or that to beamforming are you would
0:21:07or questions from you
0:21:16uh this is a
0:22:49okay there are no other questions