BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20191206T142451Z
LOCATION:Analog 1\, 2
DTSTART;TZID=Europe/Stockholm:20180625T130900
DTEND;TZID=Europe/Stockholm:20180625T131300
UID:isc_hpc_ISC High Performance 2018_sess103_phd129@linklings.com
SUMMARY:(PhD02) Mixed precision multi-level approach for FFTs
DESCRIPTION:PhD Forum\n\n(PhD02) Mixed precision multi-level approach for 
 FFTs\n\nBrunn\n\nFast Fourier Transform (FFT) is a common tool for spectra
 l analysis of\nPDEs. Applications such as medical image registration in br
 ain tumor\nresearch involve many large FFTs dominating the overall runtime
 . For\nthis research, efficiently solving the forward and backward Poisson
 \nequation and computing derivatives in the spectral space on\nheterogeneo
 us hardware platforms with native FP16 support is the main\nobjective.\n\n
 For large three dimensional FFTs on distributed architectures the\nruntime
  is dominated by the communication, since the data need to be\ntransposed 
 several times. To decrease the communication volume and\nimprove the compu
 tational throughput low precision floating-point\nnumbers can be used. How
 ever, restricting the accuracy not only\ninfluences stability, but also ra
 ises numerical problems due to the\nlimited dynamic range of floating-poin
 t numbers.\n\nTo overcome these issues, our aim is to develop a mixed prec
 ision\nmulti-level approach for FFTs. The basic idea is to split a given h
 igh\nprecision data on a fine grid into a coarser grid of same accuracy an
 d\na fine grid with lower accuracy. The coarse grid has less\ncomputationa
 l volume and the remaining low precision data on the fine\ngrid stays with
 in the dynamic range, as high frequency modes of\ntypical application data
  decay fast. One of the core issues of this\nstudy is the error analysis o
 f this mixed precision splitting approach.\n\nPasses: Conference Pass, Exh
 ibition Pass, HPC Accelerators, Heterogeneous Systems, Parallel Algorithms
 , Parallel Applications\n\nTag: Conference Pass, Exhibition Pass, HPC Acce
 lerators, Heterogeneous Systems, Parallel Algorithms, Parallel Application
 s
URL:https://2018.isc-program.com/presentation/?id=phd129&sess=sess103
END:VEVENT
END:VCALENDAR

