By using this site, you agree to our updated Privacy Policy and our Terms of Use. Manage your Cookies Settings.
455,130 Members | 1,365 Online
Bytes IT Community
+ Ask a Question
Need help? Post your question and get tips & solutions from a community of 455,130 IT Pros & Developers. It's quick & easy.

Scripting and block data manipulation- how to preserve performance

P: n/a
I have an application that manipulates large arrays of image data of
various types, all the usual arithmetic operations on the data objects
are supported. With careful design and tricks similar to those in
UBLAS (http://www.boost.org/libs/numeric/ublas/doc/) I have been able
to avoid temporaries where appropriate and carefully optimize code
(C++) that evaluates expressions, for example

Image A, B, C, D, E
A=((B+C)/D)+E

In this expression A is evaluated directly by looping through the
elements of B,C,D,E.

This all works fine and dandy if the image expressions are known at
compile time, but when it comes to scripting at run-time the
opimizations are lost. The only technique that preserves efficiency is
to generate object code for the image expression at run-time. In other
words build a simple expression compiler into the application and have
it generate native assembler as needed.

Is there precedent for this type of approach? Comments, ideas and
pointers to existing implementations would be most welcome.
Jul 23 '05 #1
Share this Question
Share on Google+
9 Replies


P: n/a
Code4u wrote:
UBLAS (http://www.boost.org/libs/numeric/ublas/doc/) I have been able
to avoid temporaries where appropriate and carefully optimize code
(C++) that evaluates expressions, for example

Image A, B, C, D, E
A=((B+C)/D)+E

In this expression A is evaluated directly by looping through the
elements of B,C,D,E.

This all works fine and dandy if the image expressions are known at
compile time, but when it comes to scripting at run-time the
opimizations are lost. The only technique that preserves efficiency
is to generate object code for the image expression at run-time.


Not the only. Your scripting can do the same that the C++ compiler does,
interpret adequately the expression and generate calls to the compose
operator functions. The difference in speed from this approach to generated
machine code will be very small, and the complexity and portability much
better.

--
Salu2
Jul 23 '05 #2

P: n/a
On Sun, 17 Jul 2005 23:57:11 +0200, Julián Albo <JU********@terra.es>
wrote:
Code4u wrote:
UBLAS (http://www.boost.org/libs/numeric/ublas/doc/) I have been able
to avoid temporaries where appropriate and carefully optimize code
(C++) that evaluates expressions, for example

Image A, B, C, D, E
A=((B+C)/D)+E

In this expression A is evaluated directly by looping through the
elements of B,C,D,E.

This all works fine and dandy if the image expressions are known at
compile time, but when it comes to scripting at run-time the
opimizations are lost. The only technique that preserves efficiency
is to generate object code for the image expression at run-time.


Not the only. Your scripting can do the same that the C++ compiler does,
interpret adequately the expression and generate calls to the compose
operator functions. The difference in speed from this approach to generated
machine code will be very small, and the complexity and portability much
better.


If I understand you correctly (perhaps not), I don't see how this
would generate efficient code. It's pretty easy to generate a series
of operator calls at run-time but this will not result in the loop
unrolling that maximises efficiency. In the above example the
following operators would be called in sequence:

operator+
operator/
operator+

In each operator call a loop iterates over the scalar values. The
problem is it is much less efficient than a combined operation:

for (size_t i=0; i<scalarCount; ++i)
{
a[i]=((b[i]+c[i])/d[i])+e[i];
}

In case the reason is not apparent- the above code will make much
better use of CPU cache because of the higher locality of reference
and uses a trivial amount of temporary storage. When tested with the
Visual Studio 2003 compiler I measure an average slowdown of x2 for
the version that uses composition of operators.

But how do we generate the equivalent object code at run-time? One
approach is to build an expression compiler into the application,
generate the required assembler, and call it.
Jul 23 '05 #3

P: n/a
Code4u wrote:
This all works fine and dandy if the image expressions are known at
compile time, but when it comes to scripting at run-time the
opimizations are lost. The only technique that preserves efficiency
is to generate object code for the image expression at run-time.
Not the only. Your scripting can do the same that the C++ compiler does,
interpret adequately the expression and generate calls to the compose
operator functions. The difference in speed from this approach to
generated machine code will be very small, and the complexity and
portability much better.


If I understand you correctly (perhaps not), I don't see how this
would generate efficient code. It's pretty easy to generate a series
of operator calls at run-time but this will not result in the loop
unrolling that maximises efficiency. In the above example the
following operators would be called in sequence:

operator+
operator/
operator+

In each operator call a loop iterates over the scalar values. The
problem is it is much less efficient than a combined operation:

for (size_t i=0; i<scalarCount; ++i)
{
a[i]=((b[i]+c[i])/d[i])+e[i];
}


And that is that I say. Your scripting engine can analyze the sequence and
call the combined operation instead of each individual operation in
sequence.

The loop unrolling probably can't be done in the scripting without great
effort. But if the operators are costly (and if you are working with image
data I suppose they are) the benefits of the loop unrolling will be
minimal.

And if you want the maximum efficience at all cost... well, you can write
the entire application in hand-optimized assembler.
But how do we generate the equivalent object code at run-time? One
approach is to build an expression compiler into the application,
generate the required assembler, and call it.


Don't know how your existing scripting engine actually works. If it
generates a stack machine, for example, a simple check os sequences of
operations in the machine can probably do the work. If it is a hand-coded
parser can be harder.

--
Salu2
Jul 23 '05 #4

P: n/a
Me
Code4u wrote:
I have an application that manipulates large arrays of image data of
various types, all the usual arithmetic operations on the data objects
are supported. With careful design and tricks similar to those in
UBLAS (http://www.boost.org/libs/numeric/ublas/doc/) I have been able
to avoid temporaries where appropriate and carefully optimize code
(C++) that evaluates expressions, for example

Image A, B, C, D, E
A=((B+C)/D)+E

In this expression A is evaluated directly by looping through the
elements of B,C,D,E.

This all works fine and dandy if the image expressions are known at
compile time, but when it comes to scripting at run-time the
opimizations are lost. The only technique that preserves efficiency is
to generate object code for the image expression at run-time. In other
words build a simple expression compiler into the application and have
it generate native assembler as needed.

Is there precedent for this type of approach? Comments, ideas and
pointers to existing implementations would be most welcome.


I'd profile this to see if it is really necessary (I'm assuming you're
doing this to eliminate temporaries and not to increase precision). For
example, instead of trying to convert that expression to:

for (size_t i = 0; i < len; ++i)
A[i]=((B[i]+C[i])/D[i])+E[i];

somehow. Try converting it to:

for (size_t i = 0; i < len; ++i)
A[i] = B[i] + C[i];

for (size_t i = 0; i < len; ++i)
A[i] = A[i] / D[i];

for (size_t i = 0; i < len; ++i)
A[i] = A[i] + E[i];

which doesn't involve any runtime code generation at all.

Jul 23 '05 #5

P: n/a
On Mon, 18 Jul 2005 08:34:53 +0200, Julián Albo <JU********@terra.es>
wrote:
And that is that I say.
Huh?
Your scripting engine can analyze the sequence and
call the combined operation instead of each individual operation in
sequence.

The loop unrolling probably can't be done in the scripting without great
effort. But if the operators are costly (and if you are working with image
data I suppose they are) the benefits of the loop unrolling will be
minimal.

Actually it's not, the average is a 2x speedup. That's huge.
And if you want the maximum efficience at all cost... well, you can write
the entire application in hand-optimized assembler.


I think there's a disconnect here. You don't seem to understand the
issue at hand.
Jul 23 '05 #6

P: n/a
Code4u wrote:
I think there's a disconnect here. You don't seem to understand the
issue at hand.


Maybe, but it's also possible thay you don't undesrtand what I propose.

--
Salu2
Jul 23 '05 #7

P: n/a
On Mon, 18 Jul 2005 16:15:16 +0200, Julián Albo <JU********@terra.es>
wrote:
Code4u wrote:
I think there's a disconnect here. You don't seem to understand the
issue at hand.


Maybe, but it's also possible thay you don't undesrtand what I propose.


You proposed composing the expression at run-time, yes? Such
composition is what I'm trying to avoid because it does not make
efficient use of cache. If I'm wrong, please restate and help me
understand what you propose, simple source code would help.

Thanks.
Jul 23 '05 #8

P: n/a
Code4u wrote:
You proposed composing the expression at run-time, yes? Such
composition is what I'm trying to avoid because it does not make
efficient use of cache.


Then I misunderstand you, I thinked you were worried about the use of
individual operations instead of the composites.

--
Salu2
Jul 23 '05 #9

P: n/a
Code4u wrote:
I have an application that manipulates large arrays of image data of
various types, all the usual arithmetic operations on the data objects
are supported. With careful design and tricks similar to those in
UBLAS (http://www.boost.org/libs/numeric/ublas/doc/) I have been able
to avoid temporaries where appropriate and carefully optimize code
(C++) that evaluates expressions, for example

Image A, B, C, D, E
A=((B+C)/D)+E

In this expression A is evaluated directly by looping through the
elements of B,C,D,E.

This all works fine and dandy if the image expressions are known at
compile time, but when it comes to scripting at run-time the
opimizations are lost. The only technique that preserves efficiency is
to generate object code for the image expression at run-time. In other
words build a simple expression compiler into the application and have
it generate native assembler as needed.

Is there precedent for this type of approach? Comments, ideas and
pointers to existing implementations would be most welcome.


I imagine that Matlab would be quite good at this sort of thing. So try
to find out how it does it.
Jul 23 '05 #10

This discussion thread is closed

Replies have been disabled for this discussion.