Computing with GPUs

There is web site all about using GPUs for general purpose computing. One particularly useful looking article goes into some of the options and issues of coding for GPUs.

problem with extractor5.09

In the course of Suzaku pipeline testing Lorella spotted a problem I introduced during the attempted speed-up of the fixwmp routine. Pixels with non-zero counts show up even within the part of the region which is supposed to be set to -1.

The particular case of Suzaku demonstrated a conceptual issue here. Those pixels showing up within the supposed excluded region are real events. What is happening is that the region is defined in sky coordinates but the WMAP is written in detector coordinates. I calculate a conversion to determine which WMAP pixels are within the image region however it is a time-averaged solution. In the pre-5.09 version the satellite attitude wobble results in some events being incorrectly excluded from the WMAP. It was these events that the 5.09 version actually included however they are potentially confusing to users.

The fix is checked in as v5.10.

xselect and Chandra grating data

Jane Turner reported that the current version of xselect fails to read Chandra HETG evt2 data. This is due to a bug in xsel_mdb.f which I think only matters for Chandra grating files.

XSPEC units for setplot wave

In v12.5 we switched the y-axis units under setplot wave to per Hz instead of per Angstrom. Since this has not been universally popular we have added an initialization option to decide between the two cases. This is available as patch 12.5.0t. That and other recent patches can be obtained from the usual place and are summarized below

12.5.0r	Fix needed for all etable model calculations. Our thanks to Roderick Johnstone for pointing this out. Report added on Feb 12, 2009.
12.5.0s	When the both the absori and pexriv models are loaded during the same XSPEC session, whichever model is loaded second will not work properly. Note: this bug originated in version 12.5.0. Report added on Feb 13, 2009.
12.5.0t	This adds an option for choosing Y-axis units when plotting in setplot wave mode. The selection is made by the WAVE_PLOT_UNITS entry, which has been added to the new version of the Xspec.init file. The units may now be specified in Hz (the default for version 12.5.0), or angstroms as they were prior to 12.5.0. This applies to plots of data, counts, and all variants of model and ufspec plots. Report added on Feb 19, 2009.
12.5.0u	Minor code changes mostly to clean up the messages reported when running XSPEC under the valgrind software tool. Report added on Feb 19, 2009.
12.5.0v	Multiplicative model components which are created with the mdefine command should not be multiplied by the energy bin widths. Our thanks to Roderick Johnstone for pointing this out. Report added on Feb 19, 2009.

non-integer COUNTS column

Steve Snowden ran into a silent problem because he made spectrum files with non-integer COUNTS columns. This is in violation of the standard and many ftools will simply silently truncate the float value to the nearest integer. I've added checks in the low-level library routines to write warnings in this case.

Update on 5/7/09. It turns out that the RXTE s/w writes PHA files with a real COUNTS column although the values are integers. So, I've improved the checking so that no warnings are given if the values are integers even if the column is defined as real. If the values really are non-integer then a warning is written and the values truncated to integers.

Gelman on statistics

I'm going to start linking to interesting blog posts by Andrew Gelman because they often have important points to remember. Today he talks about a couple of reviews of "The Black Swan" and includes the following note on Bayesian inference :

From a philosophical point of view, I think the most important point of confusion about Bayesian inference is the idea that it's about computing the probability that a model is true. In all the areas I've ever worked on, the model is never true. But what you can do is find out that certain important aspects of the data are highly unlikely to be captured by the fitted model, which can facilitate a "model shift" moment. This sort of falsification is why I believe Popper's philosophy of science to be a good fit to Bayesian data analysis.

Recent XSPEC bug fixes

The following are all available from the usual place.

12.5.0l	This slightly modifies the state of newly untied parameters. When a parameter is untied, its frozen/unfrozen setting should retain the setting of the parameter(s) to which it was linked, rather than be restored to its original value. Our thanks to Phil Evans and Delphine Porquet for pointing this out. Report added on Jan 14, 2009.
12.5.0m	When steppar is executed, it should not automatically run a fit at the end to place the parameters in their best-fit state. Instead it should restore the parameters to their pre-steppar values, regardless of the fit state. Our thanks to Tim Kallman for pointing this out. Report added on Jan 14, 2009.
12.5.0n	Bug fix for a C-to-Fortran boolean conversion error, affecting the default setting for the thermal broadening flag of the bapec and bvapec models. Unless overridden by the APECTHERMAL setting, thermal broadening should be ON by default for bapec and bvapec. This bug has only been observed on Solaris platforms. Report added on Jan 15, 2009.
12.5.0o	When a delcomp operation is performed on the unnamed (default) model and no spectra are loaded, the model calculation is not updated. Our thanks to Laura Brenneman for pointing this out. Report added on Jan 21, 2009.
12.5.0p	This relaxes the OGIP requirement that spectral data files must contain a RESPFILE keyword, primarily to allow the loading of certain RXTE/PCA files. Our thanks to David Smith for pointing this out. Report added on Jan 30, 2009.
12.5.0q	A new option has been added to the tclout command: tclout gain. See the online tclout manual entry for proper syntax and usage. Prior to this, it was not possible to retrieve the values of gain parameters from tclout. Our thanks to Frank Haberl for pointing this out. Report added on Feb 03, 2009.

xselect and region filters

I've added clarification to the on-line help for xselect about region filters. It is dangerous to make region filters in Image coordinates - the correct options are WCS or Physical. If the xselect image binning is not 1 then regions in Image coordinates will not be handled correctly.

Chandra CY11 webspec updated

Updated webspec with the final ARFs for Chandra Cycle 11.

Word-cloud of the 171 Astro2010 NOI titles

recent xspec patches

12.5.0j	Parameter links both to and from a model which is not associated with any spectra (shown as "Active/Off" in the model display), should not be removed when the model's state gets updated (such as when using the dummyrsp command). Our thanks to Laura Brenneman for pointing this out. Report added on Dec 24, 2008.
12.5.0k	The dummyrsp command may cause a crash when entering a very large number of energy bins. Our thanks to Joel Frioriksson for pointing this out. Report added on Dec 31, 2008.

webspec for Chandra

I've updated webspec to use the Chandra CY11 rmfs and arfs released by the CUC. Note that in mid January they will be releasing new arfs with corrected contamination corrections and effective areas.

xspec patch for energies extend

12.5.0i The energies extend option does not work properly for the case of a data group containing multiple responses whose energy bins differ. Our thanks to Joel Frioriksson for pointing this out. Report added on Dec 15, 2008.

webspec updated to include pile-up

At the request of Andrea Prestwich at the CUC I've added the pile-up model to WebSpec. In the simple interface there is a check box to added it while in the advanced interface the user is expected to select pileup as the option for the first model component.

One tweak could be to set the default pile-up parameters appropriately for each instrument.

cfitsio region filtering bug

The recent cfitsio version region filtering fails when the first region is an exclude. This was an error I introduced when extending to multiple components as allowed in the FITS REGION extension. The fix should be available next week. Meanwhile a work-around is to include an initial region covering the entire field.

Update 12/15/08: There is also a memory-handling error in the code which reads a FITS REGION. This bug is only triggered if the extension contains a large number (>100) of exclude regions.

extractor bug for detector coordinate regions

Dave Henley reported problems with the new extractor when using detector coordinates for the region. This is because I wasn't correctly handling the case when the WCS type is blank. This is corrected in v5.09 which should be released as part of HEAsoft patch update next week.

xspec 12.5 patches

Another set of recent xspec patches. Two of these are minor mistakes in the 12.5 release, the others cover relatively obscure problems.

12.5.0e	In the FNINIT function (for users linking the XSPEC models library into their own programs), the default model data directory name needs to be updated from "modelIonData" to "modelData" to coincide with v12.5.0 restructuring. Our thanks to John Houck for pointing this out. Report added on Dec 05, 2008.
12.5.0f	Patch to improve the behavior of an exceptional case of Levenberg-Marquardt fitting, where a parameter becomes pegged due to a zero second-derivative matrix diagonal element AND the user chooses to exit before convergence. This prevents an attempt at a covariance calculation on the parameter, which otherwise might lead to a segmentation fault. Our thanks to Stefano Bianchi and Jeremy Sanders for pointing this out. Report added on Dec 05, 2008.
12.5.0g	Fix to the swind1 model function. Now that it is included as part of XSPEC's built-in models library, it should no longer look to the $LMODDIR symbol for finding the model data directory. Our thanks to Delphine Porquet for pointing this out. Report added on Dec 09, 2008.
12.5.0h	When doing a 2-panel plot of plot counts/lcounts chisq and the current fit statistic is cstat, the Y-axis scaling of the chisq panel is wrong. Our thanks to Fill Humphrey for pointing this out. Report added on Dec 09, 2008.

extractor bug for annular regions

Phil Evans spotted a problem in the new release when using an annulus region type. The region filtering is performed correctly however the WMAP created is too small in the case when the WMAP and image coordinates are the same. This is actually an error I introduced in the cfitsio code (which is however not used internally in cfitsio so doesn't change its standard use). A work-around is to use the elliptannulus region type instead of annulus. I also have a fix within the extractor code to get around this.

This bug is important for Swift analysis.

xselect and ASCA

The switch to wild cards in xselect.mdb was a little too enthusiastic in the case of the ASCA SIS - the catcol list still needs to be specified for individual detectors. The current version causes an error when running make obscat for ASCA SIS. I have checked in a fixed version.

script to split *.rsp files into *.rmf and *.arf

I've written a little perl script split_rsp_to_rmf_arf which takes an input response matrix file and splits it into a response matrix file with unit total response at each energy and an arf which gives the total response at each energy. This can be used to mock up in v12 xspec the behaviour of v11 xspec with /b models. An example is given on the wiki section on backgrounds.