Skip to contents

Changelog

Source: NEWS.md

anvl (development version)

Breaking changes

  • transform_gradient() is renamed to graph_gradient().
  • prim_rng_bit_generator() is replaced by prim_sample_bits().
  • The RNG state is now handled by {anvl} itself and RNG states no longer have to be passed around. This comes with functions such as nv_set_seed(). See the random numbers article for more details.

Features

  • AnvlArrays can now be serialized.
  • New nv_bool(), nv_int{8,16,32,64}(), nv_uint{8,16,32,64}(), nv_float32() and nv_float64() create an array of the data type in their name.
  • New functions for the binomial distribution: nv_{d,p,q}binom().
  • New functions for the exponential distribution: nv_{d,p,q}exp().
  • A side effects mechanism was added to anvl, which drives the RNG handling and provides a basis for more side effects in the future. See the side effects section of the internals article for more details.
  • Registering custom classes that can be (un-)flattened is now supported via pjrt::register_tree_node().
  • Bugfix: nv_cumsum(), nv_cumprod(), nv_cummax(), nv_cummin(), nv_extract_diag(), nv_trace() and nv_quantile() now accept an array with a size-0 axis; nv_quantile() reducing over one is an informative error.
  • New prim_custom_gradient() computes a value with one function and differentiates it with another. A rule_reverse() with a forward pass can take the place of the latter, so that the backward pass reads what the forward pass computed.
  • prim_scatter()’s update_fn can be given by name: "replace" (the new default, replacing NULL), "add", "sub", "mul", "max" or "min". The gradient of a scatter is implemented for each of them, not only for replacement; an update_fn that is a function is not differentiated.
  • nv_subset_assign() and [<- take update_fn too, so nv_subset_assign(x, i, value = v, update_fn = "add") adds v to the subset, every value of an index that is repeated in i.

Bug fixes

  • nv_rnorm() returns NaN for a NaN mean or a non-finite or negative sd, like rnorm().
  • An unsigned dynamic index of 0 (in [, [<-, prim_dynamic_slice(), prim_dynamic_update_slice()) no longer wraps around to the last position.
  • prim_convolution() and nv_conv1d() / nv_conv2d() / nv_conv3d() are now differentiable.
  • AnvlArrays can now be saved with saveRDS() and friends; they load onto the default device, not the one they were saved from. Arrays of the pjrt backend no longer cache their device in a $device field.
  • A jitted function called without array inputs whose program is pinned to a device by an array it closes over no longer recompiles when the default device changes.
  • A classed R value, such as a bit64::integer64 array, is an error under jit() instead of being read through its storage type.
  • nv_rnorm() returns NaN for a NaN mean or a non-finite or negative sd, like rnorm().
  • nv_rbinom() accepts size = 0, like rbinom(), and nv_sample() gives a clear error for an empty population.
  • nv_seq() counts down at an unsigned data type instead of erroring.
  • The gradient of an empty strided prim_static_slice() has the shape of its input.
  • The gradient of x^y is 0 instead of NaN in y at base 0 and in x at exponent 0.
  • The gradient of nv_pad() / prim_pad() works for negative edge padding.
  • The gradient of prim_broadcast_in_axes() is no longer permuted for broadcast axes that are not in increasing order.
  • An NA in an array() subset index now gives an informative error.
  • The common data type of arrays and R values no longer depends on the order of the arguments; the R values take the data type the arrays agree on.
  • nv_squeeze(x, axes = integer(0)) no longer drops every axis.
  • nv_subset() accepts an R value as x, like nv_subset_assign() does.
  • nv_rnorm() no longer returns Inf / -Inf when a uniform draw is exactly 0.
  • nv_mean(), nv_var() and nv_sd() no longer overflow on an integer input, which is now
  • The gradient of prim_dot_general() with respect to lhs is right when
  • nv_squeeze(x, axes = integer(0)) no longer drops every axis.
  • nv_subset() accepts an R value as x, like nv_subset_assign() does.
  • nv_rnorm() no longer returns Inf / -Inf when a uniform draw is exactly 0.
  • nv_mean(),nv_var()andnv_sd()` no longer overflow on an integer input, which is now summed at the default float.
  • New nv_bool(), nv_int{8,16,32,64}(), nv_uint{8,16,32,64}(), nv_float32() and nv_float64() create an array of the data type in their name.
  • New functions for the Poisson distribution: nv_dpois(), nv_ppois(), and nv_qpois().
  • New functions for the binomial distribution: nv_dbinom(), nv_pbinom() and nv_qbinom().
  • New functions for the exponential distribution: nv_dexp(), nv_pexp(), and nv_qexp().
  • nv_chol() / prim_chol() now lower to a LAPACK (CPU) / cuSOLVER (CUDA) potrf custom call, which is about ten times faster on CPU.
  • nv_triangular_solve() / prim_triangular_solve(), and with them nv_solve(), now lower to a BLAS trsm custom call on CPU, which runs several times faster from about 100 rows on and compiles faster.
  • Bugfix: the gradient of prim_dot_general() with respect to lhs is right when lhs has batching axes but no free axes.
  • nv_broadcast_arrays() broadcasts a size-1 axis that meets a size-0 axis to size 0.
  • An update that prim_scatter() drops as out of bounds no longer gets a gradient when unique_indices = TRUE.

Performance

  • Converting AnvlArrays to R is now faster.
  • Cholesky decomposition is now much faster.
  • Linear algebra primitives in f32 are now faster on windows.

anvl 0.5.1

  • Bugfix: Add r-xla r-universe repo to description again.

anvl 0.5.0

This release contained many breaking changes. From now on, the API will be more stable. Because of the amount of changes in this release, we only list some highlights and top-level overviews here.

Breaking changes

  • Many internal functions and classes were renamed, some no longer needed ones removed.
  • The main API functions and primitives, as well as their parameters were renamed to be more consistent with base R.
  • The package now uses col-major semantics to be more consistent with base R. This is e.g. noticeable in prim_flatten, the broadcasting functions, prim_bitcast_convert, and the prim_cum* primitives.
  • The @jit roxygen tag was removed; wrap functions in jit() at the definition instead.
  • The type system of {anvl} was changed to avoid the problems reported in issue #373. Specifically, the ambiguity system was replaced with the RData system and a new system of rules for type promotions. With it, also the promotion behavior of various primitives and API functions was improved.
  • When creating an AnvlArray from R data, the inputs are now always checked for NAs, as well as out-of-range issues.
  • as_array() and the as.double() / as.integer() / bit64::as.integer64() / as.logical() methods take check = "warn", "err" or FALSE instead of a flag, following {pjrt}, and warn by default about a value R’s type cannot hold. Write check = "err" where you wrote check = TRUE, and check = FALSE to materialize silently.
  • common_dtype() now errors for ui64 and a signed integer instead of returning i64, which could not hold every ui64 value. Convert one of them with nv_convert().
  • There is now exactly one backend used at a time and it is configured via the anvl.backend option. With this change the device_arg parameter was removed from jit() as it is no longer needed.
  • The internal graph representation is now pure. Only the top-level graph can now have constants.

Features

  • Many new API functions and extensions to existing ones.
  • Some new primitives (including prim_scan).
  • Added support for configuring the default device.
  • Added support for configuring the default data types.
  • as_anvl_array() gained a .promote argument, naming the data type the input is brought to, as as_anvl_arrays() already had.
  • Error messages are improved across the package.
  • nv_array() now accepts a raw() vector.
  • as.vector() now returns a bit64::integer64 for integer data types that do not fit into R’s 32 bit integers.
  • Added support for more base R generics.
  • Improved documentation of API functions and primitives.

Performance

  • nv_quantile() and nv_median() select the needed order statistics with top_k instead of a full sort.
  • prim_top_k() gained indices; without them the CUDA lowering uses an unstable sort of the values and a slice instead of the CHLO op, which costs no more than a full sort there. nv_top_k(indices = FALSE) and the quantile fast path use it.

Bug fixes

  • Many bugs in the API functions and primitives have been fixed.
  • gradient() now handles values that prim_if() / prim_scan() sub-graphs close over, and treats values the differentiated function closes over as constants, as in JAX. Several individual gradient rules were fixed, too.
  • jit() no longer evaluates its arguments a second time.
  • Arrays created inside a trace or loaded from disk are now placed on the right device.
  • Type promotion was fixed in several API functions, and floating-point functions now refuse boolean inputs.
  • Input validation was tightened across the API and primitives, so invalid arguments now give a clear error instead of failing inside the backend, aborting the R session, or silently returning a wrong result.
  • Edge cases of subsetting, such as empty selections and ranges that count down, now work.
  • Improved the numerics of some functions, e.g. nv_mod() and nv_qnorm().

Tests

  • The environment variables that configure only the test suite are now spelled with an ANVL_TEST prefix: ANVL_TEST_SKIP_QUICKR. ANVL_TEST itself is unchanged.
  • The suite can be run with ANVL_DEFAULT_DEVICE=cpu:1, which makes anything allocating on the first CPU device rather than following the trace land on a device of its own instead of agreeing with everything else by accident. The default-device workflow runs it that way on the full-test label.
  • Moved some of pjrt’s dispatcher tests into anvl.

anvl 0.4.0

Breaking changes

  • Renamed dim/dims to axis/axes throughout the package (an axis is an index, a dimension is a size); ndims() is now naxes().
  • Renamed the primary array argument of prim_* / nv_* functions from operand to x.
  • Renamed the "xla" backend to "pjrt".
  • xla() has been removed; use jit() instead.
  • nv_rdunif() has been renamed to nv_sample_int(), mirroring R’s sample.int().
  • nv_runif()’s lower/upper arguments are now min/max, nv_rnorm()’s mu/sigma are now mean/sd, and nv_rbinom()’s n is now size, matching the corresponding R functions.

Bug fixes

  • nv_sample_int() (formerly nv_rdunif()) was off by one: the first integer was drawn twice as often as it should have been, and the last integer was never drawn at all.

Features

  • New nv_sample() samples from an arbitrary population.
  • New nv_lower_tri() and nv_upper_tri() (with nv_lower_tri_like() / nv_upper_tri_like()) return a boolean triangular mask for a given shape, mirroring base R’s lower.tri() / upper.tri().
  • New functions for the normal distribution: nv_dnorm(), nv_qnorm(), and nv_pnorm() thanks to Louis Aslett. They are implemented to be accurate far into either tail.
  • nv_rnorm()’s mean and sd now accept arrayish inputs.
  • Dimension arguments (dim, dims, dimension, permutation) now accept negative values that count from the end, so -1 refers to the last dimension.
  • Reshaping functions accept a single -1 in shape indicating a dimension to be inferred.
  • Added support for 1-3 dimensional convolutions, thanks to Troy Hernandez.
  • AnvlArray constructors and converters have gained a check argument that opts into scanning for NA values, see the “Gotchas” vignette for more information.
  • nv_var() and nv_sd() now default to dims = NULL, which reduces over all dimensions and returns a scalar, consistent with the other reductions.
  • trace_fn() gained an optimize argument controlling which graph optimization passes run on the traced graph. TRUE runs all passes, FALSE (default) runs none, and a character vector (e.g. c("inline_scalars", "remove_unused_constants")) selects a subset. jit() always traces with all passes enabled.
  • Improved the installation vignette

Performance

  • Most nv_*() API functions are now JIT-compiled internally (via a new @jit roxygen roclet), speeding up eager-mode execution.
  • Tracing (trace_fn()) performance has been improved.
  • StableHLO lowering has been sped up.
  • Calling jit()ted functions is now significantly faster.

Bug fixes

  • Reductions now reject dimensions that are out of range for the operand instead of silently ignoring them.
  • NULL is now treated as an empty node when flattening and unflattening trees.
  • nv_argmax() / nv_argmin() and nv_cummax() / nv_cummin() now break ties order-independently, so they return the same result on GPU as on CPU (#368). nv_argmax() / nv_argmin() prefer the smallest index; nv_cummax() / nv_cummin() prefer the last occurrence.
  • nv_diag() now errors on non-1-D input instead of silently producing an incorrect result.
  • jit() now rejects static arguments with reference semantics.
  • Error messages now speak of arrays instead of tensors.

anvl 0.3.0

Breaking Changes

New Features

  • On CPU, jitted XLA functions now back every non-aliased output with an R-owned RAWSXP. anvl appends a phantom donated input per unaliased output during lowering, allocates pjrt::pjrt_empty() buffers at execute time, and pjrt migrates the keepalive onto the output XPtr. The output’s host bytes are then managed by R’s GC.
  • Renamed user-facing API functions to match base R names: nv_sine() -> nv_sin(), nv_cosine() -> nv_cos(), nv_ceil() -> nv_ceiling(), nv_cholesky() -> nv_chol(). The corresponding primitives were renamed in step: prim_sine() -> prim_sin(), prim_cosine() -> prim_cos(), prim_cholesky() -> prim_chol().
  • nv_reduce_mean() was renamed to nv_mean().
  • nv_solve() no longer requires a to be symmetric positive-definite as it uses LU instead of Cholesky decomposition. Because of this, it is no longer differentiable, as the reverse rule for LU is not implemented yet.
  • nv_chol() / prim_chol() now default to lower = FALSE (upper-triangular factor), matching base R’s chol(). Previously defaulted to lower = TRUE.

New Features

Linear algebra

Element-wise math

  • New unary primitives and corresponding nv_*() functions: acos, acosh, asin, asinh, atan, atanh, cosh, sinh, digamma, lgamma, polygamma, erf, erf_inv, erfc.
  • New API functions nv_mod() (flooring remainder) and nv_trunc() (truncation toward zero).

Cumulative reductions

  • New primitives and corresponding nv_*() functions: cumsum, cumprod, cummax, cummin. prim_cumprod() does not yet have a reverse rule.

Sorting and searching

Array construction / shape

  • nv_array() gained a byrow argument that fills the array from an R object in row-major order, mirroring matrix(byrow = TRUE) (#165).
  • New nv_matrix(data, nrow, ncol, ...) which works like R’s matrix().
  • New API functions nv_rbind() and nv_cbind() and corresponding rbind() / cbind() generics.
  • New API function nv_flatten() for flattening to 1-D.

Misc

Other

  • nv_reduce_sum(), nv_reduce_prod(), nv_reduce_max(), nv_reduce_min(), nv_reduce_any(), nv_reduce_all() and nv_mean() now default dims = NULL, which reduces over all dimensions and returns a scalar. Previously, dims was required.

Bug Fixes

  • The overloaded %% operator now calls the new nv_mod() to be consistent with base R.
  • The reverse rule for prim_reduce_prod() no longer produces NaN / Inf gradients when the input contains zeros.
  • The CI now actually runs the torch-comparison tests.
  • nv_runif() not properly respects the lower argument.

anvl 0.2.0

Breaking Changes

  • The package was renamed from anvil to anvl to avoid a conflict with the Bioconductor package AnVIL.
  • AnvilTensor/nv_tensor were renamed to AnvlArray and nv_array to be more in line with R’s array(). Also, nv_aten() was renamed to nv_aval().
  • Subsetting with list() (e.g. x[list(1, 3)]) is no longer supported. Use array() to wrap the indices instead, e.g. x[array(c(1L, 3L))]. This mirrors the input convention used everywhere else in the package.
  • Removed debug mode.
  • Remove NSE support for nvl_if. It now requires passing 0-argument closures as true and false arguments.
  • Primitives renamed from nvl_* to prim_*. The underlying primitive object containing the rules and metadata is now part of the JitPrimitive function via the primitive attribute.

New Features

  • Better composability: jit()ted functions can now be used in other jit()-calls. This is the mechanism underlying the new eager mode.
  • Eager mode was added: This means, you can now do nv_add(1, nv_array(1:2)) and it will actually perform the computation and not only do type inference.
  • An experimental {quickr} backend was added It only runs on CPU for now and supports a subset of available operations. You can enable it via the backend argument in jit() and nv_array() or via the anvl.default_backend option.
  • New primitives:
    • nvl_cholesky() to compute the Cholesky decomposition of a matrix.
    • nvl_triangular_solve() to solve a system of linear equations with a triangular matrix.
  • New API functions (+ corresponding R generic implementations):
  • jit() now accepts integer positions for the static argument.
  • New S3 methods dim(), nrow(), ncol(), and length() for anvl arrays.
  • Printing tensors via nv_print() now also works on GPUs.
  • R vectors of length 1 and arrays are now auto-converted when being passed to jitted functions.
  • Improved device handling in jit()

Performance

  • Many operations are now done asynchronously, which improves performance, especially on GPUs.

Bug Fixes

  • +-Inf/NaN are correctly created for f64 when inlined into the XLA exectuable (#182). This caused wrong results with e.g. nv_reduce_max() when working with f64.
  • Corrected argument checks in nv_iota().
  • Fix check that wrt arguments in gradient() must be floats.
  • nv_subset() and nv_subset_assign() now error on trailing-comma subscripts (#273).

Documentation

  • New vignette on implementing Gaussian Processes.
  • New vignette on implementing Metropolis-Hastings sampling.

Platform support and installation

  • An installation guide was added.
  • Linux on ARM is now supported (CPU only).
  • To use the CUDA backend, it is now possible to install the cuda12.8 package (see installation guide), which only requires a compatible CUDA driver.

anvl 0.1.0

Initial release