GPU Batch 10 #1193
GPU Batch 10 #1193
Conversation
Add SYCL testing of normal dist Add CUDA testing of normal dist Add NVRTC testing of normal dist NVRTC fixes Move headers for NVRTC support Add GPU support to inverse gaussian dist Add NVRTC testing of inverse Gaussian dist Add CUDA testing of inverse gaussian dist Add SYCL testing of inverse gaussian dist Add GPU support to lognormal dist Add SYCL testing of lognormal dist Add CUDA testing of lognormal dist Add nvrtc testing of lognormal dist Add GPU support to negative binomial dist Avoid float_prior on GPU platform Add NVRTC testing of negative binomial dist Fix ambiguous use of nextafter Add CUDA testing of negative binomial dist Fix float_prior workaround Add SYCL testing of negative binomial dist Add GPU support to non_central_beta dist Add SYCL testing of nc beta dist Add CUDA testing of nc beta dist Enable generic dist handling on GPU Add GPU support to brent_find_minima Add NVRTC testing of nc beta dist Add utility header Replace non-functional macro with new function Add GPU support to non central chi squared dist Add SYCL testing of non central chi squared dist Add missing macro definition Markup generic quantile finder Add CUDA testing of non central chi squared dist Add NVRTC testing of non central chi squared dist Add GPU support to the non-central f dist Add SYCL testing of ncf Add CUDA testing of ncf dist Add NVRTC testing of ncf dist Add GPU support to students_t dist Add SYCL testing of students_t dist Add CUDA testing of students_t Add NVRTC testing of students_t dist Workaround for header cycle Add GPU support to pareto dist Add SYCL testing of pareto dist Add CUDA testing of pareto dist Add NVRTC testing of pareto dist Add missing header Add GPU support to poisson dist Add SYCL testing of poisson dist Add CUDA testing of poisson dist Add NVRTC testing of poisson dist Add forward decl for NVRTC platform Add GPU support to rayleigh dist Add CUDA testing of rayleigh dist Add SYCL testing of rayleigh dist Add NVRTC testing of rayleigh dist Add GPU support to triangular dist Add SYCL testing of triangular dist Add NVRTC testing of triangular dist Add CUDA testing of triangular dist Add GPU support to the uniform dist Add CUDA testing of uniform dist Add SYCL testing of uniform dist Add NVRTC testing of uniform dist Fix missing header Add markers to docs
|
The use of quadrature in the non central T is a new addition to improve accuracy in the tails: IMO it would not be unreasonable to disable this on the GPU. Likewise there may be other functions which need slimming down for GPU usage, even though that may mean a lower quality of implementation. The trick is deciding where to draw the line ;) |
That makes sense. Now that we have more substantial testing there are some clear outliers appearing in testing run time: https://github.com/cppalliance/cuda-math/actions/runs/10741618717/job/29792448169?pr=21#step:9:262. Once things are correct we can take optimization steps. @NAThompson is also interested in the prospect of the quadrature algorithms being used by others on device. |
Nod, that's actually a good use case given that the algorithms are in principle massively parallel. In that case it probably makes more sense for the Integrand to be on the device, and the quadrature routine the glue that pulls the results together? |
I think so. I have to work through how to make a shared memory region on the device that won't be leaked. Right now nick uses a shared_ptr to vectors with all the data. Neither of those facilities are available on device. CUDA has a version if malloc and free which make it easy enough to implement vector, but a shared pointer is more involved. |
The focus of this PR was to round out support for the distribution now that many of the fundamental special functions have GPU support. Adds support to the following distributions: inverse gaussian, lognormal, non central beta, non central f, non central chi squared, negative binomial, normal, pareto, poisson, rayleigh, students t, triangular and uniform.
The three remaining unsupported distributions are the hyperexponential, hypergeometric, and non-central t. The first two need related special functions support. Non-central T needs GPU support for quadrature built out, specifically exp-sinh. My initial thoughts are restricting quadrature to NVCC and NVRTC based on available APIs for memory allocations, but we will see.
On device CI runs can be found: cppalliance/cuda-math#21.
CC: @dschmitz89, @izaid, @steppi.