在调用 fastLmPure 后,如何让 R 就地修改该矩阵?
我试图在循环中运行 fastLmPure
,避免复制大矩阵。
我已将矩阵预先分配到我需要的大小,这只是我需要更改值的最后一列。即它不会增长。
我将其归结为下面的最小案例,这些案例说明了我遇到的问题。
我希望矩阵的修改就地完成,但我得到了一个副本,因为引用计数器已通过调用 Rcpp 而增加。
Rcpp
不会修改 X,那么为什么会出现额外的引用计数,导致 R 在我下次修改时进行复制?
从 fastLm 返回执行后如何获得就地修改?
(我使用的是 R 4.1.2,并通过控制台运行以避免任何 RStudio 环境窗格引用问题。)
示例 1:
library(RcppArmadillo)
minimal_fastLm <- function() {
y <- c(1,2)
X <- matrix(data = c(1,1,2,3), nrow = 2, ncol = 2)
.Internal(inspect(X))
model <- .Call("_RcppArmadillo_fastLm_impl", X, y, PACKAGE = "RcppArmadillo")
.Internal(inspect(X))
X[, 2] <- c(3,4)
.Internal(inspect(X))
}
minimal_fastLm()
给出输出(为了清晰起见进行了编辑):
通过调用 fastLm 地址不变,但引用计数增加,然后在修改时进行复制。 (参见粗体)。 <代码> @0x0000000017401560 14 RELSXP g0c3 [REF(1),ATT] (len=4, tl=0) 1,1,2,3 ... @0x0000000017401560 14 RELSXP g0c3 [REF(2),ATT] (len=4, tl=0) 1,1,2,3 ... @0x00000000174ad568 14 RELSXP g0c3 [REF(1),ATT] (len=4, tl=0) 1,1,3,4 ...
示例 2: 甚至更小的(纯 Rcpp)
library(Rcpp)
cppFunction('int rawSEXP(SEXP X) { return(1); }')
cppFunction('int asNumMat(NumericMatrix X) { return(1); }')
pureRcpp <- function() {
X <- matrix(data = c(1,1,2,3), nrow = 2, ncol = 2)
.Internal(inspect(X))
X[, 2] <- c(4,5) # Initially modifies in place because X has REF(1).
.Internal(inspect(X))
rawSEXP(X) # Call Rcpp function with raw SEXP
.Internal(inspect(X))
X[, 2] <- c(6,7) # Still modifies in place because X has REF(1).
.Internal(inspect(X))
asNumMat(X) # Call Rcpp function that casts to NumericMatrix
.Internal(inspect(X))
X[, 2] <- c(8,9) # Causes a copy because X has REF(3).
.Internal(inspect(X))
}
pureRcpp()
X 也从未被 Rcpp 修改过。我也不希望这样。我只是想读取数据。
那么当 NumericMatrix 超出范围时,为什么引用计数器没有减回 1?
I'm trying to run fastLmPure
in a loop, avoiding copying a large matrix.
I've pre-allocated the matrix to the size I need, and it's only the last column I need to change values in. i.e. it doesn't grow.
I've boiled it down to the minimal cases below, which illustrate the problem I'm having.
I expect the modification of matrix to be done in-place, but instead I get a copy because ref counter has been incremented by the call out to Rcpp.
Rcpp
doesn't modify X, so why does an extra refcount hang around, which causes R to make a copy when I next modify?
How can I get a modify-in-place after execution returns from fastLm?
(I'm using R 4.1.2, and running via console to avoid any RStudio env pane reference issues.)
Example 1:
library(RcppArmadillo)
minimal_fastLm <- function() {
y <- c(1,2)
X <- matrix(data = c(1,1,2,3), nrow = 2, ncol = 2)
.Internal(inspect(X))
model <- .Call("_RcppArmadillo_fastLm_impl", X, y, PACKAGE = "RcppArmadillo")
.Internal(inspect(X))
X[, 2] <- c(3,4)
.Internal(inspect(X))
}
minimal_fastLm()
Gives output (edited for clarity):
Address unchanged through call to fastLm, but ref count increased, then copy made on modify. (See bold).
@0x0000000017401560 14 REALSXP g0c3 [REF(1),ATT] (len=4, tl=0) 1,1,2,3
...
@0x0000000017401560 14 REALSXP g0c3 [REF(2),ATT] (len=4, tl=0) 1,1,2,3
...
@0x00000000174ad568 14 REALSXP g0c3 [REF(1),ATT] (len=4, tl=0) 1,1,3,4
...
Example 2:
Even more minimal (pure Rcpp)
library(Rcpp)
cppFunction('int rawSEXP(SEXP X) { return(1); }')
cppFunction('int asNumMat(NumericMatrix X) { return(1); }')
pureRcpp <- function() {
X <- matrix(data = c(1,1,2,3), nrow = 2, ncol = 2)
.Internal(inspect(X))
X[, 2] <- c(4,5) # Initially modifies in place because X has REF(1).
.Internal(inspect(X))
rawSEXP(X) # Call Rcpp function with raw SEXP
.Internal(inspect(X))
X[, 2] <- c(6,7) # Still modifies in place because X has REF(1).
.Internal(inspect(X))
asNumMat(X) # Call Rcpp function that casts to NumericMatrix
.Internal(inspect(X))
X[, 2] <- c(8,9) # Causes a copy because X has REF(3).
.Internal(inspect(X))
}
pureRcpp()
X was never modified by Rcpp. Nor do I want it to be. I just want to read data.
So why was ref counter not decremented back to 1 when NumericMatrix went out of scope?
如果你对这篇内容有疑问,欢迎到本站社区发帖提问 参与讨论,获取更多帮助,或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱,如果其他用户或者作者回复了您的评论,将不能在第一时间通知您!
发布评论
评论(2)
R 使用写时复制,并且在来自 R 或 C(++) 代码的写入之间不会发生变化——更新的值只是更新的值。
因此,如果您想要可变的非复制数据,您可能必须自己在非 R 数据结构中管理数据。一个简单的犰狳向量或矩阵,你注意不要继承 R 数据就可以了。
分配一次,根据需要修改值,循环到您想要的内容。这应该符合你的要求,除非我误解了什么。
R uses copy-on-write and that doesn't change between the write coming from R or C(++) code -- an updated value is just an updated value.
So you if want mutable non-copied data you may have to manage the data yourself in a non R data structure. A simple Armadillo vector or matrix where you take care not to inherit from R data should do.
Allocate that once, modify values as you please, loop to your heart's content. That should suit your requirement, unless I misunderstood something.
经过一番挖掘,我发现 Rcpp 可以做一些清理工作,但实际上却没有。
作为错误提出,并确定了解决方案。
将在下一版本的 Rcpp 中修复。
https://github.com/RcppCore/Rcpp/pull/1205
所以,答案是:
After some digging, I found that there's some cleaning up that Rcpp could be doing, but isn't.
Raised as a bug, with solution identified.
Will be fixed in the next release of Rcpp.
https://github.com/RcppCore/Rcpp/pull/1205
So, the answer is: