We need to set the size of the patch to be sufficiently large to include the covariant sources, while understanding how separated they need to be. A patch will have bordering objects called fixed galaxies. I think we want that a fixed galaxy in one patch not be an active galaxy in another patch running at the same time.
Clearly this depends on the size of objects. E.g., we're fitting out to 5-6 sigma in the outermost Gaussian mixture radius.
We are limited by the GPU shared memory to a particular number of active sources, likely fewer than 30.
We also need to be watchful that it presently appears that "non-interacting" pixels still take a noticeable amount of time to compute, like 25% of the interacting pixels. So we can't just set the patch size to be huge in the hopes that any single object interacts with only a small portion of the pixels.
So I think that we want the patches to be modest in size. But they have to handle the covariances, so probably that means that we want to get all of the active sources in a ~5" region and then extend that out to the radius of those sources. The result may be a 10" patch!
My intuition is that if two such patches abut, then the fixed galaxies in one are not the active galaxies in another. So this might be a simple way to dice up the dispatch problem, i.e., define a grid of 10" patches (which means that (5/10)^2 = 25% of objects are active), run them all for a few hundred iterations, then pick a new registration of the 10" grid. At least for the UDF application, where the data is already on a fixed pixel grid, that could be simple.
We need to set the size of the patch to be sufficiently large to include the covariant sources, while understanding how separated they need to be. A patch will have bordering objects called fixed galaxies. I think we want that a fixed galaxy in one patch not be an active galaxy in another patch running at the same time.
Clearly this depends on the size of objects. E.g., we're fitting out to 5-6 sigma in the outermost Gaussian mixture radius.
We are limited by the GPU shared memory to a particular number of active sources, likely fewer than 30.
We also need to be watchful that it presently appears that "non-interacting" pixels still take a noticeable amount of time to compute, like 25% of the interacting pixels. So we can't just set the patch size to be huge in the hopes that any single object interacts with only a small portion of the pixels.
So I think that we want the patches to be modest in size. But they have to handle the covariances, so probably that means that we want to get all of the active sources in a ~5" region and then extend that out to the radius of those sources. The result may be a 10" patch!
My intuition is that if two such patches abut, then the fixed galaxies in one are not the active galaxies in another. So this might be a simple way to dice up the dispatch problem, i.e., define a grid of 10" patches (which means that (5/10)^2 = 25% of objects are active), run them all for a few hundred iterations, then pick a new registration of the 10" grid. At least for the UDF application, where the data is already on a fixed pixel grid, that could be simple.