Yes, the C++ code has single indexing for vectors while D code has doouble indexing got matrixes
You may want to declare vectors in the begining of the outer loop
like
auto uxv = ux[i];
and operate with this vectors in the internal loop
D (and probably C/C++) can not vectorise double indexing like Fortran
Keep us in touch, I think it is a good example of porting and you can write a short blog post after (this would be very helpful for others)
Also, you may want to use https://github.com/libmir/mir-random . It implements C++ RNG standrd and more
@9il It would help a lot if you can extract a minimal example that shows that things are not vectorized/optimized well. There is so much going on in the current example that it's hard to analyze why things don't optimize well. Part of the problem could be that slices are used which don't optimize so well yet (it's a work-in-progress).
ldmd2 -inline -O -enable-cross-module-inlining -release -boundscheck=off -I mir-algorithm/source/ -output-s matrix_copy.d
@JohanEngelen Both variants are very slow
This is surprising because it is probably regression
I can not confirm it because ndslice is not compatable anymore with older versions
ldc2 --version
LDC - the LLVM D compiler (1.3.0git-a969bcf):
based on DMD v2.073.2 and LLVM 4.0.0
built with LDC - the LLVM D compiler (0.17.5git-64a274a)
Default target: x86_64-apple-darwin16.5.0
Host CPU: haswell
http://dlang.org - http://wiki.dlang.org/LDC
Morning. I posted that code as I went to sleep, now I finally had a chance to look at the assembly. As you correctly said, the
opIndex calls don't get inlined, which is expensive by itself and also prohibits all further optimization via alias analysis and ultimately vectorization.
Moving over to
a[i][j] style indexing improved the performance by about 2x, but it's still ~720ms vs 30ms.
I think it can be fixed during 48 hours
Anyhow, I'll be around and very happy to test anything or give any feedback I can.
It will be in DUB registry after less then a hour.
Do not forget to remove
dub.selections.jsonand update mir-algorithm version upto ~>0.6.2 if you use dub.
Also, you may want to add
"dflags-ldc":["-mcpu=native"]
Please let me know your benchmarks results for C++ code and both D variants. Thanks!
@9il Okay, the version with
a[i][j] style indexing now runs in 89 ms, compared to 31 ms from C++, so it's within 3x now.
Will go through the assembly now, see what's going on.
Also please try to transform the code like
foreach (i; 0..m)
{
foreach (j; 0..n)
{
// use matrix1[i, j], matrix2[i, j], matrix3[i, j]
}
}
to
foreach (i; 0..m)
{
auto v1 = matrix1[i];
auto v2 = matrix2[i];
auto v3 = matrix3[i];
foreach (j; 0..n)
{
// use v1[j], v2[j], v3[j]
}
}
The same for tensors
Okay, there was an aliasing issue similar to what I recently encountered in Julia, which I fixed by explicitly operating on temporary variables
ux0, etc., and only storing the results in the matrices at the end. This enabled some vectorization and brought the timing down to 55 ms. It still doesn't unroll the loop as extensively as clang does and the vectorization isn't quite complete, but we're now within 2x.
If I understand this correctly, what I did was a more extreme version of what you just suggested with hoisting out the rows?