the list). The tradeoffs are platform dependent. The "nobreak" version

seen slower for big vectors).

It may be hard to get a universally better version. Still, a version

suggests we should be checking for Inf anyway.

individual arguments of the sum, it the sum is found to have NaN).

> Hi Tomas,

>

> Can you share the full code for your benchmark, compiler options, and performance results so that I can try to reproduce them? There are a lot of variables that can affect the results. Private email is fine if it is too much for the mailing list.

>

> I am measuring on Knight's Landing (KNL) that was released in November. KNL is not a co-processor so no offload is necessary. R executes directly on the Phi, which looks like a multi-core machine with 64 cores.

>

> Robert

>

> -----Original Message-----

> From: Tomas Kalibera [mailto:

[hidden email]]

> Sent: Monday, January 16, 2017 12:00 PM

> To: Cohn, Robert S <

[hidden email]>

> Cc:

[hidden email]
> Subject: Re: [Rd] accelerating matrix multiply

>

>

> Hi Robert,

>

> thanks for the report and your suggestions how to make the NaN checks faster.

>

> Based on my experiments it seems that the "break" in the loop actually can have positive impact on performance even in the common case when we don't have NaNs. With gcc on linux (corei7), where isnan is inlined, the "break" version uses a conditional jump while the "nobreak" version uses a conditional move. The conditional jump is faster because it takes advantage of the branch prediction. Neither of the two versions is vectorized (only scalar SSE instructions used).

>

> How do you run R on Xeon Phi? Do you offload the NaN checks to the Phi coprocessor? So far I tried without offloading to Phi, icc could vectorize the "nobreak" version, but the performance of it was the same as "break".

>

> For my experiments I extracted NaN checks into a function. This was the "break" version (same performance as the current code):

>

> static __attribute__ ((noinline)) Rboolean hasNA(double *x, int n) {

> for (R_xlen_t i = 0; i < n; i++)

> if (ISNAN(x[i])) return TRUE;

> return FALSE;

> }

>

> And this was the "nobreak" version:

>

> static __attribute__ ((noinline)) Rboolean hasNA(double *x, int n) {

> Rboolean has = FALSE;

> for (R_xlen_t i = 0; i < n; i++)

> if (ISNAN(x[i])) has=TRUE;

> return has;

> }

>

> Thanks,

> Tomas

>

> On 01/11/2017 02:28 PM, Cohn, Robert S wrote:

>>> Do you have R code (including set.seed(.) if relevant) to show on how

>>> to generate the large square matrices you've mentioned in the

>>> beginning? So we get to some reproducible benchmarks?

>> Hi Martin,

>>

>> Here is the program I used. I only generate 2 random numbers and reuse them to make the benchmark run faster. Let me know if there is something I can do to help--alternate benchmarks, tests, experiments with compilers other than icc.

>>

>> MKL LAPACK behavior is undefined for NaN's so I left the check in, just made it more efficient on a CPU with SIMD. Thanks for looking at this.

>>

>> set.seed (1)

>> m <- 30000

>> n <- 30000

>> A <- matrix (runif(2),nrow=m,ncol=n)

>> B <- matrix (runif(2),nrow=m,ncol=n)

>> print(typeof(A[1,2]))

>> print(A[1,2])

>>

>> # Matrix multiply

>> system.time (C <- B %*% A)

>> system.time (C <- B %*% A)

>> system.time (C <- B %*% A)

>>

>> -----Original Message-----

>> From: Martin Maechler [mailto:

[hidden email]]

>> Sent: Tuesday, January 10, 2017 8:59 AM

>> To: Cohn, Robert S <

[hidden email]>

>> Cc:

[hidden email]
>> Subject: Re: [Rd] accelerating matrix multiply

>>

>>>>>>> Cohn, Robert S <

[hidden email]>

>>>>>>> on Sat, 7 Jan 2017 16:41:42 +0000 writes:

>>> I am using R to multiply some large (30k x 30k double) matrices on a

>>> 64 core machine (xeon phi). I added some timers to src/main/array.c

>>> to see where the time is going. All of the time is being spent in the

>>> matprod function, most of that time is spent in dgemm. 15 seconds is

>>> in matprod in some code that is checking if there are NaNs.

>>>> system.time (C <- B %*% A)

>>> nancheck: wall time 15.240282s

>>> dgemm: wall time 43.111064s

>>> matprod: wall time 58.351572s

>>> user system elapsed

>>> 2710.154 20.999 58.398

>>>

>>> The NaN checking code is not being vectorized because of the early

>>> exit when NaN is detected:

>>>

>>> /* Don't trust the BLAS to handle NA/NaNs correctly: PR#4582

>>> * The test is only O(n) here.

>>> */

>>> for (R_xlen_t i = 0; i < NRX*ncx; i++)

>>> if (ISNAN(x[i])) {have_na = TRUE; break;}

>>> if (!have_na)

>>> for (R_xlen_t i = 0; i < NRY*ncy; i++)

>>> if (ISNAN(y[i])) {have_na = TRUE; break;}

>>>

>>> I tried deleting the 'break'. By inspecting the asm code, I verified

>>> that the loop was not being vectorized before, but now is vectorized.

>>> Total time goes down:

>>>

>>> system.time (C <- B %*% A)

>>> nancheck: wall time 1.898667s

>>> dgemm: wall time 43.913621s

>>> matprod: wall time 45.812468s

>>> user system elapsed

>>> 2727.877 20.723 45.859

>>>

>>> The break accelerates the case when there is a NaN, at the expense of

>>> the much more common case when there isn't a NaN. If a NaN is

>>> detected, it doesn't call dgemm and calls its own matrix multiply,

>>> which makes the NaN check time insignificant so I doubt the early

>>> exit provides any benefit.

>>>

>>> I was a little surprised that the O(n) NaN check is costly compared

>>> to the O(n**2) dgemm that follows. I think the reason is that nan

>>> check is single thread and not vectorized, and my machine can do 2048

>>> floating point ops/cycle when you consider the cores/dual issue/8 way

>>> SIMD/muladd, and the constant factor will be significant for even

>>> large matrices.

>>>

>>> Would you consider deleting the breaks? I can submit a patch if that

>>> will help. Thanks.

>>>

>>> Robert

>> Thank you Robert for bringing the issue up ("again", possibly).

>> Within R core, some have seen somewhat similar timing on some platforms (gcc) .. but much less dramatical differences e.g. on macOS with clang.

>>

>> As seen in the source code you cite above, the current implementation was triggered by a nasty BLAS bug .. actually also showing up only on some platforms, possibly depending on runtime libraries in addition to the compilers used.

>>

>> Do you have R code (including set.seed(.) if relevant) to show on how to generate the large square matrices you've mentioned in the beginning? So we get to some reproducible benchmarks?

>>

>> With best regards,

>> Martin Maechler

>>

>> ______________________________________________

>>

[hidden email] mailing list

>>

https://stat.ethz.ch/mailman/listinfo/r-devel>