ACE april 9 Per, Miquel, Erik, Alexandra, Behrooz, sam brainstorming about where we can sacrifice accuracy fixed point calculations? what application areas? dsp? geophysical apps -- stencil applications? (could also relax sync) can we put together a set of "dwarves" for approx computing? identify different kinds of behaviors in applications -- could put together benchmark suite consisting both of applications and kernels distilled from them Chalmers trolls? search-based applications? Erik re: rasterization using fixed point Tom Forsyth? no Michael Abrash (the Larrabee team at RAD was Michael, Atman Binstock, Tom Forsyth, Mike Sartain) https://software.intel.com/en-us/articles/rasterization-on-larrabee software rasterization on Larrabee look for Dr. Dobbs articles can figure out how many bits needed based on number of pixels and precision required Miquel: can contact BSC people who work on algorithms (about kernels) Reverse Time Migration (they work w/ Repsol) Per: benchmark suite for GPGPUs? sam: what about software model? we start by instrumenting/annotating applications, but need to make sure there is a clear path to automatic compilation Miquel: autotuning for accuracy? sam: what about runtime system? Miquel: "correctness" is application dependent -- need API that programmer uses to make app interact w/ correctness-assessment system component? Erik: making programmers figure out exactly how many bits is too much, but there should be a way to look at precision of input and desired precision of output sam: put together a reading list Miquel: use Box for article repository Per: set of benchmarks for GPGPUs -- Tor Aamodt is one of the leaders Erik: low precision input, discrete cosign xform to frequency space to filter low frequency changes, then xlate back to pixels (same w/ wavelet compression -- lossy, but preserves edges and sharp changes). frequency space representation uses large floats, though. opportunity? sam: what about image analysis applications? (surveillance? face detection?) Per: problem w/ those apps is "accuracy" hard to make sense of intuitively sam: algorithm-based fault tolerance? Behrooz: automotive applications? brake-by-wire action items: *Christos will present at next meeting on April 24th, 4128 *Erik will look up articles on software rasterization/Larrabee *Sally will set up wiki, post Berkeley dwarves article *Alexandra will look up benchmarks for GPGPUs, and find out about compiling OpenCL apps for GPGPUsim (check out Tor Aamodt's group at Univ. of British Columbia) *Miquel will help sam w/ Box setup after wiki is in place and look for iterative solvers and ABFT Behrooz: will look up brake-by-wire stuff *ALL: look at more application areas mailing list so far: Per Stenström , Johan Karlsson , Ioannis Sourdis , Sally McKee , Erik Sintorn , Christos Dimitrakakis , Alexandra Angerd , Fatemeh Ayatolahi , Behrooz Sangchoolie , Miquel Perucas , Ulf Assarsson