www.digitalmars.com         C & C++   DMDScript  

D - Scientific, Graphics, and Game programming performance

reply "Craig Black" <cblack ara.com> writes:
I would like to toss around some ideas that would benefit scientific,
graphics, and game programming.  Consider the code below.

vector3 a(0, 1, 2), b(3, 4, 5), c(6, 7, 8), d;
d = a + b - c * a;

C++ uses expression templates for vector operations with arbitrary size.
C++ expression templates could transform the second line of code into:

for (int i = 0; i < 3; i++) d[i] = a[i] + b[i] - c[i] * a[i];

This is good because it eliminates temporary variables. However, for small
vectors it would be ideal to unroll the for loop to get:

d.x = a.x + b.x - c.x * a.x;
d.y = a.y + b.y - c.y * a.y;
d.z = a.z + b.z - c.z * a.z;

How could this kind of thing be done?

One way to solve this problem is with generics like C++ expression
templates.  Another solution may be to have built-in vector, and matrix
types, like with NVidia Cg.  Perhaps Cg's solution is better because you can
also make use of the graphics card instructions, boosting performance even
further.
Jun 25 2002
parent reply "Sean L. Palmer" <seanpalmer earthlink.net> writes:
D already has everything we need except operator overloading.

In D you'd just write:

alias float[3] vector3;

vector3 a = {0,1,2}, b = {3,4,5}, c = {6,7,8}, d;
d = a + b - c * a;

And the idea is that the compiler will already try to do exactly what you
want.

However this isn't implemented in the current D compiler.  (and the array
initialization syntax will crash the compiler.  ;)  But Walter has mentioned
that he will do it like this.

What I want is to be able to make Dot and Cross product operators, and
matrix multiply, inner product, and outer product should be part of the
language, they're so common.  Very fundamental vector math.  And quaternion
is just an extension of Complex which is already in the D language.  Why
stop there?  Tensor math is basically what computer graphicists use already.
Walter won't go for geometric algebra support but tensor math is essentially
inner product, outer product, and matrix multiply and matrix by vector
transforms.  And it uses square matrices which is what we want in graphics
most of the time.  Just about everything else (invert, transpose, etc) we
could write as functions.

I don't think Cg types have a place in standard D.  Graphics is too
specialized.  D programs run on the main CPU, not on the GPU.  Cg programs
run on the GPU only.

Instead of Cg types we need the compiler to recognize the special cases of
float[4], double[2], byte[16], short[8], int[4], etc and generate SIMD (SSE
or 3DNOW!) code.  Heck it could generate SIMD code anyway if it wanted.

Sean

"Craig Black" <cblack ara.com> wrote in message
news:afa36a$10nt$1 digitaldaemon.com...

 I would like to toss around some ideas that would benefit scientific,
 graphics, and game programming.  Consider the code below.

 vector3 a(0, 1, 2), b(3, 4, 5), c(6, 7, 8), d;
 d = a + b - c * a;

 C++ uses expression templates for vector operations with arbitrary size.
 C++ expression templates could transform the second line of code into:

 for (int i = 0; i < 3; i++) d[i] = a[i] + b[i] - c[i] * a[i];

 This is good because it eliminates temporary variables. However, for small
 vectors it would be ideal to unroll the for loop to get:

 d.x = a.x + b.x - c.x * a.x;
 d.y = a.y + b.y - c.y * a.y;
 d.z = a.z + b.z - c.z * a.z;

 How could this kind of thing be done?

 One way to solve this problem is with generics like C++ expression
 templates.  Another solution may be to have built-in vector, and matrix
 types, like with NVidia Cg.  Perhaps Cg's solution is better because you


can

 also make use of the graphics card instructions, boosting performance even
 further.
Jun 25 2002
parent reply "Craig Black" <cblack ara.com> writes:
 What I want is to be able to make Dot and Cross product operators, and
 matrix multiply, inner product, and outer product should be part of the
 language, they're so common.  Very fundamental vector math.  And


quaternion

 is just an extension of Complex which is already in the D language.  Why
 stop there?  Tensor math is basically what computer graphicists use


already.

 Walter won't go for geometric algebra support but tensor math is


essentially

 inner product, outer product, and matrix multiply and matrix by vector
 transforms.  And it uses square matrices which is what we want in graphics
 most of the time.  Just about everything else (invert, transpose, etc) we
 could write as functions.


Sounds good, but how do you plan on writing invert, transpose, etc for
arbitrary matrix dimensions?  Would you pass the dimensions of the matrix as
parameters?  Isn't this kinda messy?  Or could we use a template function
like you can in C++.

template<int a, int b>
float [b][a] Inverse(float [a][b] matrix) { ... }

Or you could describe it this way:

float [b][a] Inverse<int a, int b>(float [a][b] matrix) { ... }

This breed of template is nice but I presume would be difficult to
implement?


 I don't think Cg types have a place in standard D.  Graphics is too
 specialized.  D programs run on the main CPU, not on the GPU.  Cg programs
 run on the GPU only.


Yes Cg is based solely on NVidia chipset.


 Instead of Cg types we need the compiler to recognize the special cases of
 float[4], double[2], byte[16], short[8], int[4], etc and generate SIMD


(SSE

 or 3DNOW!) code.  Heck it could generate SIMD code anyway if it wanted.


Sounds good.

--Craig
Jun 25 2002
parent "Sean L. Palmer" <seanpalmer earthlink.net> writes:
I don't know about the scientific end of things but in graphics we mainly
need either 3x3, 3x4, or 4x4 matrices.  D already supports function
overloading so we can just overload those three cases:

void Invert(inout float[3][3]);
void Invert(inout float[4][3]);
void Invert(inout float[4][4]);

I haven't given that much thought to generics syntax.  I use C++ templates
all the time so that syntax is tolerable to me, but I know it could be
improved.  I just haven't sat down and experimented enough to know which way
would be best.  Admittedly that's a very personal yardstick.   ;)

Sean

"Craig Black" <cblack ara.com> wrote in message
news:afad7l$1t22$1 digitaldaemon.com...

 Sounds good, but how do you plan on writing invert, transpose, etc for
 arbitrary matrix dimensions?  Would you pass the dimensions of the matrix


as

 parameters?  Isn't this kinda messy?  Or could we use a template function
 like you can in C++.

 template<int a, int b>
 float [b][a] Inverse(float [a][b] matrix) { ... }

 Or you could describe it this way:

 float [b][a] Inverse<int a, int b>(float [a][b] matrix) { ... }

 This breed of template is nice but I presume would be difficult to
 implement?
Jun 26 2002