search

lattice / quda

QUDA is a library for performing calculations in lattice QCD on GPUs.
https://lattice.github.io/quda
Other
286 stars 94 forks source link

cpuGaugeField constructor segfault #78

Closed jpfoley closed 12 years ago

jpfoley commented 12 years ago

Trying to fix issue 76. Compiles fine now, but when I call the cpuGaugeField constructor using QUDA_REFERENCE_FIELD_CREATE + QUDA_QDP_GAUGE_ORDER, I get a seg fault. The same constructor with QUDA_REFERENCE_FIELD_CREATE + QUDA_MILC_GAUGE_ORDER works fine. QUDA_ZERO_FIELD_CREATE + QUDA_QDP_GAUGE_ORDER is also okay.

maddyscientist commented 12 years ago

I guess the first question should be: what is your use case here? Are you creating a host field in QUDA_QDP_GAUGE_ORDER and passing the pointer to the cpuGaugeField? Since this use case is what all the dslash and solver tests use, I'm surprised that this one is seg faulting.

jpfoley commented 12 years ago

The seg fault arises in the constructor, which takes a reference to GaugeFieldParam as an argument. The pointer to the data is passed later using the setGauge method, but I have commented this out to make sure that the seg fault is in the constructor.

On 08/21/2012 05:19 PM, mikeaclark wrote:

I guess the first question should be: what is your use case here? Are you creating a host field in QUDA_QDP_GAUGE_ORDER and passing the pointer to the cpuGaugeField? Since this use case is what all the dslash and solver tests use, I'm surprised that this one is seg faulting.

— Reply to this email directly or view it on GitHub https://github.com/lattice/quda/issues/78#issuecomment-7919884.

jpfoley commented 12 years ago

I should also say that GaugeFieldParam::gauge is set to NULL.

On 08/21/2012 05:19 PM, mikeaclark wrote:

I guess the first question should be: what is your use case here? Are you creating a host field in QUDA_QDP_GAUGE_ORDER and passing the pointer to the cpuGaugeField? Since this use case is what all the dslash and solver tests use, I'm surprised that this one is seg faulting.

— Reply to this email directly or view it on GitHub https://github.com/lattice/quda/issues/78#issuecomment-7919884.

maddyscientist commented 12 years ago

That is why then. Creating a cpuGaugeField with QUDA_REFERENCE_CREATE doesn't allocate any memory, instead it just points the gauge pointer at the one specified by the param. This is to wrap QUDA metadata around a gauge field that has already been created by another application.

What exactly are you trying to do here?

jpfoley commented 12 years ago

Understood. I am trying to modify unitarize_link_test.cpp to work with the latest version of quda. Previously, I had used the computeFatLinkQuda function to generate a fattened gauge field. This takes as input and returns ordinary host arrays. I had then passed a pointer to the host array containing the fattened field to an instance of cpuGaugeField and unitarized that. I had used milc gauge-field ordering in the past, and that seemed to work (actually, the ordering was screwed up when I copied the data to the device, but it didn't affect my check). The packGhost function now explicitly forbids milc field ordering, so I switched to QDP ordering, which is when the segmentation fault appeared. I'll modify my code so the correct pointer is passed on instantiation, and add a check in the cpuGaugeField constructor. I will probably also modify computeFatLinkQuda to take and return instances of cpuGaugeField instead of simple host arrays, which will also simplify the test code significantly.

Thanks.

maddyscientist commented 12 years ago

I see you've closed this, did you resolve the problem you were having?

The packGhost always has forbidden MILC ordering that I am aware of, but this is something that can fixed fairly easily. I recently extented support for BQCD ordering, so this is fairly recently in my brain, so I can do this if you need it.

jpfoley commented 12 years ago

Well, I figured that it's not really a bug. I'll push the changes I've made today. It's true that packGhost has never supported the MILC ordering, but the check that excludes the MILC ordering scheme explicitly was only added recently. In unitarize_link_test.cpp, I had used MILC ordering previously, and the code didn't complain and the test passed, so there didn't seem to be a problem. However, this worked only because I was checking the unitarization, and not trying to match the fat links produced on the host and device. I.e., I just wanted to generate a field of non-su(3) links to pass to the unitarization function, and I did not care about the field ordering. It would be useful to be able to use packGhost on MILC-ordered fields, but it is not an urgent priority.

On 08/22/2012 10:53 AM, mikeaclark wrote:

I see you've closed this, did you resolve the problem you were having?

The packGhost always has forbidden MILC ordering that I am aware of, but this is something that can fixed fairly easily. I recently extented support for BQCD ordering, so this is fairly recently in my brain, so I can do this if you need it.

— Reply to this email directly or view it on GitHub https://github.com/lattice/quda/issues/78#issuecomment-7940986.