Performance work driven by perf profiles; all solver
outputs are checked against frozen outputs of version 0.2.0 (relative
differences below 1e-11) and the Julia-derived reference solutions in
the tests.
wdl()/wig() on the CPU now predict the
document barycenters with the same batched Gibbs-kernel iteration the
training uses, on a whole batch of documents at once. Before, inference
solved one barycenter per document and, with the default
method = "auto", did so in the log domain, which cost about
20 times the whole training (2000 NYT headlines, 1 epoch: 33 minutes,
now 33 seconds). method, threshold and
n_threads of the barycenter control are documented as
unused by WDL.sinkhorn() and
barycenter() exponentiate with a vectorised rational
approximation (1-2 ulp of exp()) instead of one libm call
per matrix entry, and log sinkhorn() no longer computes a
column soft-min per iteration whose only use, the column term of the
error, is identically zero. 1000 x 1000 log Sinkhorn, 1000 iterations:
19 s to 8 s on one thread.n_threads) spin briefly between the
short parallel sections of an iteration instead of sleeping; 12 threads
now give about 5x on the forward passes where they gave 3x.sinkhorn()
and barycenter() (with_grad = TRUE) no longer
materialize the M x N softmax matrices: the forward pass keeps its row
and column soft-mins (M + N numbers per iteration), from which each
adjoint product is one fused sweep over the cost matrix, split over the
threads by column blocks. 1000 x 1000 log Sinkhorn with gradient, 1000
iterations: 15.6 s to 13.2 s on one thread, 3.8 s to 2.1 s on 12; 800 x
800 log barycenter with gradients: 37.9 s to 33.9 s and 10.9 s to 5.2
s.barycenter()) use one dsymv/dgemv
per column rather than a level-3 call that repacks the N x N kernel
every time: 800 x 800 barycenter with 4 sources, 0.79 s to 0.34 s.wdl(): the per-document cublasDgemv
launches of the weight adjoint are one strided-batched GEMM (2000
headlines: 5.6 s to 4.5 s)..Call with registered
routines) and calls the BLAS/LAPACK libraries shipped with R directly
through a small internal matrix layer, which also makes the installed
package much smaller. The package now has no compiled-code dependencies
at all.sinkhorn() and
barycenter(): the soft-min kernels no longer materialize
the M x N matrix C - f 1' - 1 g' on every iteration, and
worker threads (n_threads) are created once per call
instead of once per iteration.wdl() on CPU: the softmax Jacobian is applied in
O(N) instead of O(N^2) per column, an unused matrix product per backward
step was removed, and the per-batch history buffers use about half the
memory.wdl() now draws its random initialization from R’s RNG
(rnorm()), so set.seed() reproduces the
documented R sequence. Fits with the same seed therefore differ from
version 0.1.0, which used Armadillo’s own generator.wdl() returned weights,
docs_pred and docs_dist in the shuffled
training order (the default shuffle = TRUE) next to
docs in the input order, so wig() summed
document scores into the wrong periods. Per-document outputs are now put
back into the input order.wig() failed when wig_control was
given without group_unit.wdl_specs()/wig_specs() now honor the
values passed in barycenter_control (method,
max_iter, …) instead of silently overriding them; the
wdl_control seed entry is named seed (it was
documented as rng_seed but read as seed).
verbose defaults to 0 when omitted from a partial
sinkhorn_control/barycenter_control.rwig no longer
sets the BLAS thread count to 1 for the whole session when attached. The
startup message now explains the recommendation and, if RhpcBLASctl is
installed (it is now only suggested), reports the current thread count
and the call to change it.group_unit is now
passed to cut() (?cut.Date), so it accepts
“day”, “week”, “month”, “quarter”, “year” and multiples such as “2
months”. Weeks start on Monday.n_threads > 0) log barycenter
crashed when the cost matrix had more columns than rows.nvcc (recent CUDA toolkits default to sm_52,
which lacks double-precision atomicAdd). Set
RWIG_CUDA_ARCH to override the detected flag, or
RWIG_NO_CUDA=1 to build without GPU support even when CUDA
is installed.