write_block: Write array blocks

In DelayedArray: A unified framework for working transparently with on-disk and in-memory array-like datasets

Description

Use write_block to write a block of array data to a RealizationSink object. The function is typically used in the context of block processing of array-like objects (typically DelayedArray objects but not necessarily).

Usage

write_block(sink, viewport, block)

## Backend-agnostic RealizationSink constructor:
AutoRealizationSink(dim, dimnames=NULL, type="double", as.sparse=FALSE)

## Get/set the "automatic realization backend":
getAutoRealizationBackend()
setAutoRealizationBackend(BACKEND=NULL)
supportedRealizationBackends()

Arguments

sink	A **writable** array-like object, typically a RealizationSink derivative. Some important notes: DelayedArray objects are NEVER writable, even when they don't carry delayed operations (e.g. HDF5Array objects from the HDF5Array package), even when they don't carry delayed operations and have all their data in memory (e.g. RleArray objects). RealizationSink is a virtual class so sink must be an instance of a **concrete** RealizationSink subclass (e.g. an HDF5RealizationSink object from the HDF5Array package). RealizationSink derivatives are considered array-like objects i.e. they have dimensions and possibly dimnames. Although write_block() will typically be used on a RealizationSink derivative, it can also be used on an ordinary array or other writable in-memory array-like arrays like dgCMatrix objects from the Matrix package.
viewport	An ArrayViewport object compatible with sink, that is, such that refdim(viewport) is identical to dim(sink).
block	An ordinary (dense) array or SparseArraySeed object of the same dimensions as viewport.
dim	The dimensions (specified as an integer vector) of the RealizationSink object to create.
dimnames	The dimnames (specified as a list of character vectors or NULLs) of the RealizationSink object to create.
type	The type of the data that will be written to the RealizationSink object to create.
as.sparse	Whether the data should be written as sparse or not to the RealizationSink object to create. Not all realization backends support this.
BACKEND	NULL (the default), or a single string specifying the name of a realization backend e.g. "HDF5Array" or "RleArray" etc...

Details

*** The RealizationSink API ***

The DelayedArray package provides a simple API for writing blocks of array data to disk (or to memory): the "RealizationSink API". This API allows the developper to write code that is agnostic about the particular on-disk (or in-memory) format being used to store the data.

Here is how to use it:

1. Create a realization sink.

2. Write blocks of array data to the realization sink with one or several calls to write_block().

3. Close the realization sink with close().

4. Coerce the realization sink to DelayedArray.

A realization sink is formally represented by a RealizationSink object. Note that RealizationSink is a virtual class with various concrete subclasses like HDF5RealizationSink from the HDF5Array package, or RleRealizationSink. Each subclass implements the "RealizationSink API" for a specific realization backend.

To create a realization sink, use the specific constructor function. This function should be named as the class itself e.g. HDF5RealizationSink().

To create a realization sink in a backend-agnostic way, use AutoRealizationSink(). It will create a RealizationSink object for the current automatic realization backend (see below).

Once writing to the realization sink is completed, the RealizationSink object must be closed (with close(sink)), then coerced to DelayedArray (with as(sink, "DelayedArray"). What specific DelayedArray derivative this coercion will return depends on the specific class of the RealizationSink object. For example, if sink is an HDF5RealizationSink object from the HDF5Array package, then as(sink, "DelayedArray") will return an HDF5Array instance (the HDF5Array class is a DelayedArray subclass).

*** The automatic realization backend ***

The automatic realization backend is a user-controlled global setting that indicates what specific RealizationSink object AutoRealizationSink() should return. In the context of block processing of a DelayedArray object, this controls where/how realization happens e.g. as an ordinary array if not set (i.e. set to NULL), or as an HDF5Array object if set to "HDF5Array", or as an RleArray object if set to "RleArray", etc...

Use getAutoRealizationBackend() or setAutoRealizationBackend() to get or set the automatic realization backend.

Use supportedRealizationBackends() to get the list of realization backends that are currently supported.

*** Cross realization backend compatibility ***

Two important things to keep in mind for developers aiming at writing code that is compatible across realization backends:

• Realization backends don't necessarily support concurrent writing.

  More precisely: Even though it is safe to assume that any DelayedArray object will support concurrent read_block() calls, it is not so safe to assume that any RealizationSink derivative will support concurrent calls to write_block(). For example, at the moment, HDF5RealizationSink objects do not support concurrent writing.

  This means that in order to remain compatible across realization backends, code that contains calls to write_block() should NOT be parallelized.

• Some realization backends are "linear write only", that is, they don't support random write access, only linear write access.

  Such backends will provide a relization sink where the blocks of data must be written in linear order (i.e. by ascending rank). Furthermore, the geometry of the blocks must also be compatible with linear write access, that is, they must have a "first-dim-grows-first" shape. Concretely this means that the grid used to walk on the relization sink must be created with something like:

      colAutoGrid(sink)

  for a two-dimensional sink, or with something like:

      defaultAutoGrid(sink, block.shape="first-dim-grows-first")

  for a sink with an arbitrary number of dimensions.

  See ?defaultAutoGrid for more information.

  For obvious reasons, "linear write only" realization backends do not support concurrent writing.

Value

For write_block(), the modified array-like object sink.

For AutoRealizationSink(), a RealizationSink object for the current automatic realization backend.

For getAutoRealizationBackend, NULL (no backend set yet) or a single string specifying the name of the automatic realization backend currently in use.

For supportedRealizationBackends, a data frame with 1 row per supported realization backend.

See Also

• ArrayViewport objects.

• SparseArraySeed objects.

• read_block.

• blockApply and family for convenient block processing of an array-like object.

• defaultAutoGrid and family to generate automatic grids to use for block processing of array-like objects.

• HDF5RealizationSink objects in the HDF5Array package.

• HDF5-dump-management in the HDF5Array package to control the location and physical properties of automatically created HDF5 datasets.

• RleArray objects.

• DelayedArray objects.

• array objects in base R.

Examples

## ---------------------------------------------------------------------
## USING THE "RealizationSink API": EXAMPLE 1
## ---------------------------------------------------------------------

## -- STEP 1 --
## Create a realization sink. Note that instead of creating a
## realization sink by calling a backend-specific sink constructor
## (e.g. HDF5Array::HDF5RealizationSink), we set the "automatic
## realization backend" to "HDF5Array" and use backend-agnostic
## constructor AutoRealizationSink():
setAutoRealizationBackend("HDF5Array")
sink <- AutoRealizationSink(c(35L, 50L, 8L))
dim(sink)

## -- STEP 2 --
## Define the grid of viewports to walk on. Here we define a grid made
## of very small viewports on 'sink'. Note that, for real-world use cases,
## block processing will typically use grids made of much bigger
## viewports, usually obtained with defaultAutoGrid() or family.
## Also please note that this grid would not be compatible with "linear
## write only" realization backends. See "Cross realization backend
## compatibility" above in this man page for more information.
sink_grid <- RegularArrayGrid(dim(sink), spacings=c(20, 20, 4))

## -- STEP 3 --
## Walk on the grid, and, for each of viewport, write random data to it.
for (bid in seq_along(sink_grid)) {
    viewport <- sink_grid[[bid]]
    block <- array(runif(length(viewport)), dim=dim(viewport))
    sink <- write_block(sink, viewport, block)
}

## -- An alternative to STEP 3 --
FUN <- function(viewport, sink) {
    block <- array(runif(length(viewport)), dim=dim(viewport))
    write_block(sink, viewport, block)
}
sink <- viewportReduce(FUN, sink_grid, sink, verbose=TRUE)

## -- STEP 4 --
## Close the sink and turn it into a DelayedArray object:
close(sink)
A <- as(sink, "DelayedArray")
A

setAutoRealizationBackend()  # unset automatic realization backend

## ---------------------------------------------------------------------
## USING THE "RealizationSink API": EXAMPLE 2
## ---------------------------------------------------------------------

## Say we have a 3D array and want to collapse its 3rd dimension by
## summing the array elements that are stacked vertically, that is, we
## want to compute the matrix M obtained by doing sum(A[i, j, ]) for all
## valid i and j. This is very easy to do with an ordinary array:
collapse_3rd_dim <- function(a) apply(a, MARGIN=1:2, sum)

## or, in a slightly more efficient way:
collapse_3rd_dim <- function(a) {
    m <- matrix(0, nrow=nrow(a), ncol=ncol(a))
    for (z in seq_len(dim(a)[[3]]))
        m <- m + a[ , , z]
    m
}

## With a toy 3D array:
a <- array(runif(8000), dim=c(25, 40, 8))
dim(collapse_3rd_dim(a))
stopifnot(identical(sum(a), sum(collapse_3rd_dim(a))))  # sanity check

## Now say that A is so big that even M wouldn't fit in memory. This is
## a situation where we'd want to compute M block by block:

## -- STEP 1 --
## Create the 2D realization sink:
setAutoRealizationBackend("HDF5Array")
sink <- AutoRealizationSink(dim(a)[1:2])
dim(sink)

## -- STEP 2 --
## Define two grids: one for 'sink' and one for 'a'. Since we're going
## to walk on the two grids simultaneously, read a block from 'a' and
## write it to 'sink', we need to make sure that we define grids that
## are "aligned". More precisely the two grids must have the same number
## of viewports and the viewports in one must correspond to the viewports
## in the other one:
sink_grid <- colAutoGrid(sink, ncol=10)
a_spacings <- c(dim(sink_grid[[1L]]), dim(a)[[3]])
a_grid <- RegularArrayGrid(dim(a), spacings=a_spacings)
dims(sink_grid)  # dimensions of the individual viewports
dims(a_grid)     # dimensions of the individual viewports

## Here is how to check that the two grids are "aligned":
stopifnot(identical(length(sink_grid), length(a_grid)))
stopifnot(identical(dims(sink_grid), dims(a_grid)[ , -3]))

## -- STEP 3 --
## Walk on the two grids simultaneously:
for (bid in seq_along(sink_grid)) {
    ## Read block from 'a'.
    a_viewport <- a_grid[[bid]]
    block <- read_block(a, a_viewport)
    ## Collapse it.
    block <- collapse_3rd_dim(block)
    ## Write the collapsed block to 'sink'.
    sink_viewport <- sink_grid[[bid]]
    sink <- write_block(sink, sink_viewport, block)
}

## -- An alternative to STEP 3 --
FUN <- function(sink_viewport, sink) {
    ## Read block from 'a'.
    bid <- currentBlockId()
    a_viewport <- a_grid[[bid]]
    block <- read_block(a, a_viewport)
    ## Collapse it.
    block <- collapse_3rd_dim(block)
    ## Write the collapsed block to 'sink'.
    write_block(sink, sink_viewport, block)
}
sink <- viewportReduce(FUN, sink_grid, sink, verbose=TRUE)

## -- STEP 4 --
## Close the sink and turn it into a DelayedArray object:
close(sink)
M <- as(sink, "DelayedArray")
M

## Sanity check:
stopifnot(identical(collapse_3rd_dim(a), as.array(M)))

setAutoRealizationBackend()  # unset automatic realization backend

## ---------------------------------------------------------------------
## USING THE "RealizationSink API": AN ADVANCED EXAMPLE
## ---------------------------------------------------------------------

## Say we have 2 matrices with the same number of columns. Each column
## represents a biological sample:
library(HDF5Array)
R <- as(matrix(runif(75000), ncol=1000), "HDF5Array")   # 75 rows
G <- as(matrix(runif(250000), ncol=1000), "HDF5Array")  # 250 rows

## Say we want to compute the matrix U obtained by applying the same
## binary functions FUN() to all samples i.e. U is defined as:
##
##   U[ , j] <- FUN(R[ , j], G[ , j]) for 1 <= j <= 1000
##
## Note that FUN() should return a vector of constant length, say 200,
## so U will be a 200x1000 matrix. A naive implementation would be:
##
##   pFUN <- function(r, g) {
##       stopifnot(ncol(r) == ncol(g))  # sanity check
##       sapply(seq_len(ncol(r)), function(j) FUN(r[ , j], g[ , j]))
##   }
##
## But because U is going to be too big to fit in memory, we can't
## just do pFUN(R, G). So we want to compute U block by block and
## write the blocks to disk as we go. The blocks will be made of full
## columns. Also since we need to walk on 2 matrices at the same time
## (R and G), we can't use blockApply() or blockReduce() so we'll use
## a "for" loop.

## Before we get to the "for" loop, we need 4 things:

## 1. Two grids of blocks, one on R and one on G. The blocks in the
##    two grids must contain the same number of columns. We arbitrarily
##    choose to use blocks of 150 columns:
R_grid <- colAutoGrid(R, ncol=150)
G_grid <- colAutoGrid(G, ncol=150)

## 2. The function pFUN(). It will take 2 blocks as input, 1 from R
##    and 1 from G, apply FUN() to all the samples in the blocks,
##    and return a matrix with one columns per sample:
pFUN <- function(r, g) {
    stopifnot(ncol(r) == ncol(g))  # sanity check
    ## Return a matrix with 200 rows with random values. Completely
    ## artificial sorry. A realistic example would actually need to
    ## apply the same binary function to r[ ,j] and g[ , j] for
    ## 1 <= j <= ncol(r).
    matrix(runif(200 * ncol(r)), nrow=200)
}

## 3. A RealizationSink object where to write the matrices returned
##    by pFUN() as we go:
setAutoRealizationBackend("HDF5Array")
U_sink <- AutoRealizationSink(c(200L, 1000L))

## 4. Finally, we create a grid on U_sink with viewports that contain
##    the same number of columns as the corresponding blocks in R and G:
U_grid <- colAutoGrid(U_sink, ncol=150)

## Note that the three grids should have the same number of viewports:
stopifnot(length(U_grid) == length(R_grid))
stopifnot(length(U_grid) == length(G_grid))

## 5. Now we can proceed. We use a "for" loop to walk on R and G
##    simultaneously, block by block, apply pFUN(), and write the
##    output of pFUN() to U_sink:
for (bid in seq_along(U_grid)) {
    R_block <- read_block(R, R_grid[[bid]])
    G_block <- read_block(G, G_grid[[bid]])
    U_block <- pFUN(R_block, G_block)
    U_sink <- write_block(U_sink, U_grid[[bid]], U_block)
}

## An alternative to the "for" loop is to use viewportReduce():
FUN <- function(U_viewport, U_sink) {
    bid <- currentBlockId()
    R_block <- read_block(R, R_grid[[bid]])
    G_block <- read_block(G, G_grid[[bid]])
    U_block <- pFUN(R_block, G_block)
    write_block(U_sink, U_viewport, U_block)
}
U_sink <- viewportReduce(FUN, U_grid, U_sink, verbose=TRUE)

close(U_sink)
U <- as(U_sink, "DelayedArray")
U

setAutoRealizationBackend()  # unset automatic realization backend

## ---------------------------------------------------------------------
## VERY BASIC (BUT ALSO VERY ARTIFICIAL) USAGE OF THE
## read_block()/write_block() COMBO
## ---------------------------------------------------------------------

###### On an ordinary matrix ######
m1 <- matrix(1:30, ncol=5)

## Define a viewport on 'm1':
block1_dim <- c(4, 3)
viewport1 <- ArrayViewport(dim(m1), IRanges(c(3, 2), width=block1_dim))

## Read/tranform/write:
block1 <- read_block(m1, viewport1)
write_block(m1, viewport1, block1 + 1000L)

## Define another viewport on 'm1':
viewport1b <- ArrayViewport(dim(m1), IRanges(c(1, 3), width=block1_dim))

## Read/tranform/write:
write_block(m1, viewport1b, block1 + 1000L)

## No-op:
m <- write_block(m1, viewport1, read_block(m1, viewport1))
stopifnot(identical(m1, m))

########## On a 3D array ##########
a3 <- array(1:60, 5:3)

## Define a viewport on 'a3':
block3_dim <- c(2, 4, 1)
viewport3 <- ArrayViewport(dim(a3), IRanges(c(1, 1, 3), width=block3_dim))

## Read/tranform/write:
block3 <- read_block(a3, viewport3)
write_block(a3, viewport3, block3 + 1000L)

## Define another viewport on 'a3':
viewport3b <- ArrayViewport(dim(a3), IRanges(c(3, 1, 3), width=block3_dim))

## Read/tranform/write:
write_block(a3, viewport3b, block3 + 1000L)

## No-op:
a <- write_block(a3, viewport3, read_block(a3, viewport3))
stopifnot(identical(a3, a))

## ---------------------------------------------------------------------
## LESS BASIC (BUT STILL VERY ARTIFICIAL) USAGE OF THE
## read_block()/write_block() COMBO
## ---------------------------------------------------------------------

grid1 <- RegularArrayGrid(dim(m1), spacings=c(3L, 2L))
grid1
length(grid1)  # number of blocks defined by the grid
read_block(m1, grid1[[3L]])  # read 3rd block
read_block(m1, grid1[[1L, 3L]])

## Walk on the grid, colum by column:
m1a <- m1
for (bid in seq_along(grid1)) {
    viewport <- grid1[[bid]]
    block <- read_block(m1a, viewport)
    block <- bid * 1000L + block
    m1a <- write_block(m1a, viewport, block)
}
m1a

## Walk on the grid, row by row:
m1b <- m1
for (i in seq_len(dim(grid1)[[1]])) {
  for (j in seq_len(dim(grid1)[[2]])) {
    viewport <- grid1[[i, j]]
    block <- read_block(m1b, viewport)
    block <- (i * 10L + j) * 1000L + block
    m1b <- write_block(m1b, viewport, block)
  }
}
m1b

## ---------------------------------------------------------------------
## supportedRealizationBackends() AND FAMILY
## ---------------------------------------------------------------------

getAutoRealizationBackend()  # no backend set yet

supportedRealizationBackends()
setAutoRealizationBackend("HDF5Array")
getAutoRealizationBackend()  # backend is set to "HDF5Array"
supportedRealizationBackends()

getHDF5DumpChunkLength()
setHDF5DumpChunkLength(500L)
getHDF5DumpChunkShape()

sink <- AutoRealizationSink(c(120L, 50L))
class(sink)  # HDF5-specific realization sink
dim(sink)
chunkdim(sink)

grid <- defaultAutoGrid(sink, block.length=600)
for (bid in seq_along(grid)) {
    viewport <- grid[[bid]]
    block <- 101 * bid + runif(length(viewport))
    dim(block) <- dim(viewport)
    sink <- write_block(sink, viewport, block)
}

close(sink)
A <- as(sink, "DelayedArray")
A

setAutoRealizationBackend()  # unset automatic realization backend

DelayedArray documentation built on March 25, 2021, 6:01 p.m.