I'm using a fairly up-to-date Manjaro Linux. I've noticed that R's torch does not seem utilize my CPU (AMD Ryzen 7, series 5000) to its full extent. Further, using torch_set_num_threads does not seem to take any effect, as the code takes roughly the same amount of time. An equivalent Python code does not have these issues.
I'm attaching a reproducible example, modelled after one available at documentation of Python torch library. Let me know whether the setup is correct and whether You observe a similar effect on Your side.
size <- 1024
set.seed(2024)
X <- matrix(runif(size^2), size, size)
Y <- matrix(runif(size^2), size, size)
for(n_threads in c(1,2,3,4,5,6,7,8)){
torch::torch_set_num_threads(n_threads)
library(torch)
message <- paste0("Number of threads: ", torch_get_num_threads(), "\n")
cat(message)
t1 <- torch_tensor(X)
t2 <- torch_tensor(Y)
time_start <- Sys.time()
out <- microbenchmark::microbenchmark(torch_matmul(t1,t2), times = 7)
time_stop <- Sys.time()
print((time_stop - time_start) / 7)
print(out)
detach(package:torch)
}
print(sessionInfo())
Output :
Number of threads: 1
Time difference of 0.4125318 secs
Unit: milliseconds
expr min lq mean median uq max neval
torch_matmul(t1, t2) 385.0982 386.6031 392.8908 387.9618 390.5209 422.9279 7
Number of threads: 2
Time difference of 0.4029093 secs
Unit: milliseconds
expr min lq mean median uq max neval
torch_matmul(t1, t2) 383.9286 385.3504 387.0242 387.8626 388.6715 389.3341 7
Number of threads: 3
Time difference of 0.412262 secs
Unit: milliseconds
expr min lq mean median uq max neval
torch_matmul(t1, t2) 390.5602 393.3822 395.6617 394.6437 397.4655 402.7326 7
Number of threads: 4
Time difference of 0.4133101 secs
Unit: milliseconds
expr min lq mean median uq max neval
torch_matmul(t1, t2) 389.1392 389.7427 398.4571 390.4203 390.6904 448.7739 7
Number of threads: 5
Time difference of 0.4015262 secs
Unit: milliseconds
expr min lq mean median uq max neval
torch_matmul(t1, t2) 383.1957 385.1875 386.9126 387.8565 388.3859 390.1893 7
Number of threads: 6
Time difference of 0.4085842 secs
Unit: milliseconds
expr min lq mean median uq max neval
torch_matmul(t1, t2) 387.7209 390.1149 391.9693 391.7547 393.8607 396.3586 7
Number of threads: 7
Time difference of 0.4061831 secs
Unit: milliseconds
expr min lq mean median uq max neval
torch_matmul(t1, t2) 389.0678 389.5376 391.5841 390.2344 392.4632 397.7849 7
Number of threads: 8
Time difference of 0.4225501 secs
Unit: milliseconds
expr min lq mean median uq max neval
torch_matmul(t1, t2) 396.3549 398.5 406.618 400.3949 404.1815 444.2131 7
R version 4.4.0 (2024-04-24)
Platform: x86_64-pc-linux-gnu
Running under: Manjaro Linux
Matrix products: default
BLAS: /usr/lib/libblas.so.3.12.0
LAPACK: /usr/lib/liblapack.so.3.12.0
locale:
[1] LC_CTYPE=en_GB.UTF-8 LC_NUMERIC=C LC_TIME=pl_PL.UTF-8
[4] LC_COLLATE=en_GB.UTF-8 LC_MONETARY=pl_PL.UTF-8 LC_MESSAGES=en_GB.UTF-8
[7] LC_PAPER=pl_PL.UTF-8 LC_NAME=C LC_ADDRESS=C
[10] LC_TELEPHONE=C LC_MEASUREMENT=pl_PL.UTF-8 LC_IDENTIFICATION=C
time zone: Europe/Warsaw
tzcode source: system (glibc)
attached base packages:
[1] stats graphics grDevices utils datasets methods base
loaded via a namespace (and not attached):
[1] microbenchmark_1.4.10 processx_3.8.4 bit_4.0.5 compiler_4.4.0
[5] magrittr_2.0.3 cli_3.6.2 tools_4.4.0 rstudioapi_0.16.0
[9] torch_0.13.0 Rcpp_1.0.12 bit64_4.0.5 coro_1.0.4
[13] callr_3.7.6 ps_1.7.6 rlang_1.1.3
Hi,
I'm using a fairly up-to-date Manjaro Linux. I've noticed that R's torch does not seem utilize my CPU (AMD Ryzen 7, series 5000) to its full extent. Further, using
torch_set_num_threadsdoes not seem to take any effect, as the code takes roughly the same amount of time. An equivalent Python code does not have these issues.I'm attaching a reproducible example, modelled after one available at documentation of Python torch library. Let me know whether the setup is correct and whether You observe a similar effect on Your side.
Output :