www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.learn - Vector operations optimization.

reply "Comrad" <comrad.karlovich googlemail.com> writes:
I'd like to try d in computational physics. One of the most 
appealing features of the d is implementation of arrays, but to 
be really usable this has to work FAST.
So here http://dlang.org/arrays.html it is stated, that:

"Im­ple­men­ta­tion note: many of the more com­mon vec­tor 
op­er­a­tions are ex­pected to take ad­van­tage of any 
vec­tor math in­struc­tions avail­able on the tar­get 
com­puter."

What is the status at the moment? What compiler and with which 
compiler flags I should use to achieve maximum performance?
Mar 21 2012
parent reply Trass3r <un known.com> writes:
 What is the status at the moment? What compiler and with which compiler  
 flags I should use to achieve maximum performance?

In general gdc or ldc. Not sure how good vectorization is though, esp. auto-vectorization. On the other hand the so called vector operations like a[] = b[] + c[]; are lowered to hand-written SSE assembly even in dmd.
Mar 22 2012
next sibling parent reply "Comrad" <comrad.karlovich googlemail.com> writes:
On Thursday, 22 March 2012 at 10:43:35 UTC, Trass3r wrote:
 What is the status at the moment? What compiler and with which 
 compiler flags I should use to achieve maximum performance?

In general gdc or ldc. Not sure how good vectorization is though, esp. auto-vectorization. On the other hand the so called vector operations like a[] = b[] + c[]; are lowered to hand-written SSE assembly even in dmd.

I had such a snippet to test: 1 import std.stdio; 2 void main() 3 { 4 double[2] a=[1.,0.]; 5 double[2] a1=[1.,0.]; 6 double[2] a2=[1.,0.]; 7 double[2] a3=[0.,0.]; 8 foreach(i;0..1000000000) 9 a3[]+=a[]+a1[]*a2[]; 10 writeln(a3); 11 } And I compared with the following d code: 1 import std.stdio; 2 void main() 3 { 4 double[2] a=[1.,0.]; 5 double[2] a1=[1.,0.]; 6 double[2] a2=[1.,0.]; 7 double[2] a3=[0.,0.]; 8 foreach(i;0..1000000000) 9 { 10 a3[0]+=a[0]+a1[0]*a2[0]; 11 a3[1]+=a[1]+a1[1]*a2[1]; 12 } 13 writeln(a3); 14 } And with the following c code: 1 #include <stdio.h> 2 int main() 3 { 4 double a[2]={1.,0.}; 5 double a1[2]={1.,0.}; 6 double a2[2]={1.,0.}; 7 double a3[2]; 8 unsigned i; 9 for(i=0;i<1000000000;++i) 10 { 11 a3[0]+=a[0]+a1[0]*a2[0]; 12 a3[1]+=a[1]+a1[1]*a2[1]; 13 } 14 printf("%f %f\n",a3[0],a3[1]); 15 return 0; 16 } The last one I compiled with gcc two previous with dmd and ldc. C code with -O2 was the fastest and as fast as d without slicing compiled with ldc. d code with slicing was 3 times slower (ldc compiler). I tried to compile with different optimization flags, that didn't help. Maybe I used the wrong ones. Can someone comment on this?
Mar 22 2012
parent Dmitry Olshansky <dmitry.olsh gmail.com> writes:
On 23.03.2012 9:57, Comrad wrote:
 On Thursday, 22 March 2012 at 10:43:35 UTC, Trass3r wrote:
 What is the status at the moment? What compiler and with which
 compiler flags I should use to achieve maximum performance?

In general gdc or ldc. Not sure how good vectorization is though, esp. auto-vectorization. On the other hand the so called vector operations like a[] = b[] + c[]; are lowered to hand-written SSE assembly even in dmd.

I had such a snippet to test: 1 import std.stdio; 2 void main() 3 { 4 double[2] a=[1.,0.]; 5 double[2] a1=[1.,0.]; 6 double[2] a2=[1.,0.]; 7 double[2] a3=[0.,0.];

Here is a culprit, the array ops [] are tuned for arbitrary long(!) arrays, they are not plain 1 simd SEE op. They are handcrafted loops(!) on SSE ops, cool and fast for arrays in general, not fixed pairs/trios/etc. I believe it might change in future, if compiler is able to deduce that size is fixed, and use more optimal code for small sizes.
 8 foreach(i;0..1000000000)
 9 a3[]+=a[]+a1[]*a2[];
 10 writeln(a3);
 11 }

 And I compared with the following d code:

 1 import std.stdio;
 2 void main()
 3 {
 4 double[2] a=[1.,0.];
 5 double[2] a1=[1.,0.];
 6 double[2] a2=[1.,0.];
 7 double[2] a3=[0.,0.];
 8 foreach(i;0..1000000000)
 9 {
 10 a3[0]+=a[0]+a1[0]*a2[0];
 11 a3[1]+=a[1]+a1[1]*a2[1];
 12 }
 13 writeln(a3);
 14 }

 And with the following c code:

 1 #include <stdio.h>
 2 int main()
 3 {
 4 double a[2]={1.,0.};
 5 double a1[2]={1.,0.};
 6 double a2[2]={1.,0.};
 7 double a3[2];
 8 unsigned i;
 9 for(i=0;i<1000000000;++i)
 10 {
 11 a3[0]+=a[0]+a1[0]*a2[0];
 12 a3[1]+=a[1]+a1[1]*a2[1];
 13 }
 14 printf("%f %f\n",a3[0],a3[1]);
 15 return 0;
 16 }

 The last one I compiled with gcc two previous with dmd and ldc. C code
 with -O2
 was the fastest and as fast as d without slicing compiled with ldc. d
 code with slicing was 3 times slower (ldc compiler). I tried to compile
 with different optimization flags, that didn't help. Maybe I used the
 wrong ones. Can someone comment on this?

-- Dmitry Olshansky
Mar 23 2012
prev sibling next sibling parent James Miller <james aatch.net> writes:
On 23 March 2012 18:57, Comrad <comrad.karlovich googlemail.com> wrote:
 On Thursday, 22 March 2012 at 10:43:35 UTC, Trass3r wrote:
 What is the status at the moment? What compiler and with which compiler
 flags I should use to achieve maximum performance?

In general gdc or ldc. Not sure how good vectorization is though, esp. auto-vectorization. On the other hand the so called vector operations like a[] =3D b[] + c[]=


 are lowered to hand-written SSE assembly even in dmd.

I had such a snippet to test: =C2=A01 import std.stdio; =C2=A02 void main() =C2=A03 { =C2=A04 =C2=A0 double[2] a=3D[1.,0.]; =C2=A05 =C2=A0 double[2] a1=3D[1.,0.]; =C2=A06 =C2=A0 double[2] a2=3D[1.,0.]; =C2=A07 =C2=A0 double[2] a3=3D[0.,0.]; =C2=A08 =C2=A0 foreach(i;0..1000000000) =C2=A09 =C2=A0 =C2=A0 a3[]+=3Da[]+a1[]*a2[]; =C2=A010 =C2=A0 writeln(a3); =C2=A011 } And I compared with the following d code: =C2=A01 import std.stdio; =C2=A02 void main() =C2=A03 { =C2=A04 =C2=A0 double[2] a=3D[1.,0.]; =C2=A05 =C2=A0 double[2] a1=3D[1.,0.]; =C2=A06 =C2=A0 double[2] a2=3D[1.,0.]; =C2=A07 =C2=A0 double[2] a3=3D[0.,0.]; =C2=A08 =C2=A0 foreach(i;0..1000000000) =C2=A09 =C2=A0 { =C2=A010 =C2=A0 =C2=A0 a3[0]+=3Da[0]+a1[0]*a2[0]; =C2=A011 =C2=A0 =C2=A0 a3[1]+=3Da[1]+a1[1]*a2[1]; =C2=A012 =C2=A0 } =C2=A013 =C2=A0 writeln(a3); =C2=A014 } And with the following c code: =C2=A01 #include =C2=A0<stdio.h> =C2=A02 int main() =C2=A03 { =C2=A04 =C2=A0 double a[2]=3D{1.,0.}; =C2=A05 =C2=A0 double a1[2]=3D{1.,0.}; =C2=A06 =C2=A0 double a2[2]=3D{1.,0.}; =C2=A07 =C2=A0 double a3[2]; =C2=A08 =C2=A0 unsigned i; =C2=A09 =C2=A0 for(i=3D0;i<1000000000;++i) =C2=A010 =C2=A0 { =C2=A011 =C2=A0 =C2=A0 a3[0]+=3Da[0]+a1[0]*a2[0]; =C2=A012 =C2=A0 =C2=A0 a3[1]+=3Da[1]+a1[1]*a2[1]; =C2=A013 =C2=A0 } =C2=A014 =C2=A0 printf("%f %f\n",a3[0],a3[1]); =C2=A015 =C2=A0 return 0; =C2=A016 } The last one I compiled with gcc two previous with dmd and ldc. C code wi=

 -O2
 was the fastest and as fast as d without slicing compiled with ldc. d cod=

 with slicing was 3 times slower (ldc compiler). I tried to compile with
 different optimization flags, that didn't help. Maybe I used the wrong on=

 Can someone comment on this?

The flags you want are -O2, -inline -release. If you don't have those, then that might explain some of the slow down on slicing, since -release drops a ton of runtime checks. Otherwise, I'm not sure why its so much slower, the druntime array ops are written using SIMD instructions where available, so it should be fast. -- James Miller
Mar 22 2012
prev sibling next sibling parent Trass3r <un known.com> writes:
 The flags you want are -O, -inline -release.

 If you don't have those, then that might explain some of the slow down
 on slicing, since -release drops a ton of runtime checks.

-noboundscheck option can also speed up things.
Mar 23 2012
prev sibling next sibling parent "Comrad" <comrad.karlovich gmail.com> writes:
On Friday, 23 March 2012 at 10:48:55 UTC, Dmitry Olshansky wrote:
 On 23.03.2012 9:57, Comrad wrote:
 On Thursday, 22 March 2012 at 10:43:35 UTC, Trass3r wrote:
 What is the status at the moment? What compiler and with 
 which
 compiler flags I should use to achieve maximum performance?

In general gdc or ldc. Not sure how good vectorization is though, esp. auto-vectorization. On the other hand the so called vector operations like a[] = b[] + c[]; are lowered to hand-written SSE assembly even in dmd.

I had such a snippet to test: 1 import std.stdio; 2 void main() 3 { 4 double[2] a=[1.,0.]; 5 double[2] a1=[1.,0.]; 6 double[2] a2=[1.,0.]; 7 double[2] a3=[0.,0.];

Here is a culprit, the array ops [] are tuned for arbitrary long(!) arrays, they are not plain 1 simd SEE op. They are handcrafted loops(!) on SSE ops, cool and fast for arrays in general, not fixed pairs/trios/etc. I believe it might change in future, if compiler is able to deduce that size is fixed, and use more optimal code for small sizes.

So currently there is no such an optimization exists for any d compiler?
Mar 23 2012
prev sibling parent "Comrad" <comrad.karlovich gmail.com> writes:
On Friday, 23 March 2012 at 11:20:59 UTC, Trass3r wrote:
 The flags you want are -O, -inline -release.

 If you don't have those, then that might explain some of the 
 slow down
 on slicing, since -release drops a ton of runtime checks.

-noboundscheck option can also speed up things.

dmd is anyway very slow, ldc2 was better, but still not fast enough.
Mar 23 2012