You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 7 Next »

Feb 28, 2017

mikhail:

  • looked at skbeam centering algorithm
  • sz compression

silke:

  • try SZ compression on some XPP/XCS datasets

sioan:

  • try SZ compression on some SXR/AMO dataset (works trivially with georgi's dataset)
  • spatial FPGA code working

clemens:

  • angular integrations: background is not symmetric (scattering from masks etc).  take difference of t0 (no signal) and later data (with signal) and that is symmetric and allows determination of beam center.  beam centers changes every half-hour so calibrations would have to be quick.
  • have results for CPU/KNL
  • found benchmark of EPICS common-mode correction, found performance problem in ndarray.
  • looked at intel vtune

chuck:

  • wavelet stuff takes 3ms on 250x250 image with pywt
  • visually physics results look OK

chris:

  • working with legion people on gpu-based two-threshold lag

Feb 21, 2017

sioan:

  • working on spatial.  good progress

clemens:

  • found slowdown for python bin count:  memory copy due to bad types
  • went through nersc presentation slides
  • need to use the fast local ram on knl
  • to do: how much does it cost to get 10GB/s through a node?
  • working with timbvd on beam-position centering
  • why is it hard to do in angular integrations in real-time?  detector corrections

mikhail:

  • working on sz from argonne
  • difficulty running examples, but fixed
  • only runs on floats
  • timing measurements for sz
  • automatic centering algorithm in skbeam/xicam/silke's huff algorithm

chuck:

  • away at LBL xtallography workshop:  pessimistic about drp
    • unit-cell not known
    • bad pixels, which change over time (good enough, but room for improvement in our dark runs)
      • compare bad-pixel maps from super-users to dark-run results
    • geometry
  • python scaling problem work
  • students look at pywavelets:  zero-pads for performance reasons

Feb 14, 2017

Chuck:

  • Tried SPI wavelet compression: 4x (on top of a 4x binned image)
  • quantifying the error at goal: 3Angstrom resolution (currently 9nm)

Mikhail:

  • looking at igor/matt compression
  • much faster than gzip, but limit is 2x compression
  • igor used multi-threading, matt did not
  • to do: enter igor stuff in spreadsheet, work on argonne package

silke:

  • work on pyfai
  • ran on  gpu102
  • compared with henrik bincount method
  • csr ocl is 4x as fast 1x henrik code, their numpy version is very slow (30x slower than henrik's)
  • turn it into a metric (standardize on 1024x1024 single-precision "bincount" method, if pyfai will do that)

clemens:

  • benchmarking phi
  • python is 8x worse than C
  • to do: test at SLAC and test with matrix multiply
  • to do: run all corrections and see if we can get 10GB/s

sioan:

  • running spatial builds

Feb 7, 2017

sioan:
possible users to contact:

  • nora berrah
  • dan higley
  • mankowsky
  • urs staub (elsa abreu)
  • marcus kubin (slow andor)
  • emmanuel jal
  • sam vinko
  • christoph bostedt

to do:

  • finishing populating list
  • will work on algorithms or data transfer
  • come up with examples where we need shot-by-shot position measurements with a reticle

silke:
cases where people rerun analysis:

  • some people reran roi's 3 months after
  • some people had a crappy detector and had to rerun analysis

possible users to contact:

  • timbvd
  • mike kozina
  • reis / mariano
  • henrik
  • ulvestad
  • marco

to do: will work on algorithms or data transfer

mikhail:

  • szip (proprietary)
  • lzf+shuffle (1.5 for calibrated, 1.7 for raw)
  • shuffle (5%-20%)
  • best: gzip with shuffle (2.2 for calibrated)
  • wait on SZ compressor results
  • look at igor algorithm
  • picking a feature-extraction

clemens:

  • comparing python/C++: still working on it
  • zmq/mpi/gasnet
  • gasnet: 3 year old python bindings
  • setup mtg with elliott/wonchan?
  • talk with DAQ about DRP re-configuration

Jan 27, 2017

mikhail:

  • henrik algorithm
  • filipe maia algorithm
  • running on gpus/knl
  • evaluate compression package from argonne that amedeo will send out

clemens:

  • why only 10GB/s with bincount?
  • why c++ 10x faster?
  • scaling plot with cores

silke:

  • understand required info for each expt (talk to timbvd about real-time angular integration difficulties)

sioan:

  • populate amo/sxr 2016 table like Silke

cpo:

  • evaluate pyfai
  • work on gpu peakfinding with legion with zhenglin
  • purchase knl/gpu/nvme boxes
  • No labels