Handling different implementations of the same algorithm

L

luca

Hi,

correct my analysis if i am wrong, please. I am writing a software in
the computer vision field. Now, i have several algorithms that can be
regarded as low-level and could be viewed as a single routine. Is it a
wrong attitude to write a simple class with a single method that do
the job? Should i use a static method of a namespace or a class
containing static low-level routines? Consider that sometimes it could
be useful to create an object, because memory allocation could be done
only at construction (or better, initialization) time (and i could
reuse the object untile its destruction). The burden of handling such
memory would be demanded to the object itself, instead of having to
manage memory to pass to a static method.

Another question i have is the following: suppose i have three
different implementation of the same algorithm. One is written in
plain C++, one is written in OpenCL (for parallel processing, would
result in different C++ code, with special function to be called,
kernel programs that run on the device to be compiled and so on), one
is written in playin vertex/program shaders for GPU (call it GPU
implementation, it would use OpenGL for example). The algorithm do the
same job (for example, processing an image and returnings the a list
of coordinates of special pixels), but the method to call it and the
way it returns data can be very different. The problem is the
following:

1.
I will drop C++ implementation of the low-level algorithms in future,
but for now, OpenCL is not so widespread, and i am forced to implement
things in different ways.

2.
C++ plain implementation results in code to be executed on the CPU.
Memory is passed/returned as usual (through pointers).

3.
OpenCL/GPU implementation: although they are different under every
aspect, share a common thing: memory must be transferred from main
system
to device memory for input parameters and the contrary for output
parameters.

So, 2 and 3 call for some sort of interface for input/output
parameters. Even the format of this parameters is important. For
example, a graph implemented using pointers would be not suited to be
passed to the device memory. But this aspect would be decided in
advance (whatever is the implementation of the low-level algorithm,
the format of the input parameters must be such that no pointers are
valid and so on...), so no problem.

But, one thing to keep in mind and which is important is the
following: low-level algorithms usually are called one after another,
that is, one run A1, followed by A2, followed by A3 and so on. Now, it
can happen that A_{i+1} uses the results provided by A_i and it would
be a waste of time to transfer memory from device to main system at
the end of A_i: just tell A_{i+1} that the parameter needed is already
on the device memory and has handle XXX.

Ok, it is clear that some form of interface for input/output
parameters is needed. This object would have a method Create( void
*memory, etc...) and would handle all of this.

So, i have an interface for memory object to be used as input/output
parameters for low-level algorithsm. Ok, but how would you create this
algorithms?
I mean, mixing CPU with GPU implementation would be a stupid thing to
do. OpenCL uses the concept of Context. A natural way would be to
create an interface IContext and having diffrerent context for
different implementation (CPU, OPENCL, OPENGL for example). Then i
would implement a factory for low-level algorithms. This factory
would create IAlgorithm objects (another interface, but could be a
void *) given:

1. the name of the algorithm (for example: ALGO_EDGEDETECTOR_SOBEL)
2. the actual context

Finally, who construct the context? The client of the software will
use a static method, for example:

IContext *ctx = ContextFactory::CreateContext( CTX_OPENCL );

So, to sum up:

.. a context concept, IContext, created by a context factory
.. a way of handling input/output for low-level algorithm
(IMemoryObject could be the basic interafce)
.. a factory for classes implementing low-level algorithms

Is this general desing a mess? What do you think? How would you handle
this situation?

Thank you very much!
Luca

PS.
I have posted this msg already to C++.moderated, but my original
intent was to post this here, because it is faster! Sorry for cross-
posting.
 
N

Nick Keighley

I think you need to read up on some software design patterns.

correct my analysis if i am wrong, please. I am writing a software in
the computer vision field. Now, i have several algorithms that can be
regarded as low-level and could be viewed as a single routine. Is it a
wrong attitude to write a simple class with a single method that do
the job?

this is similar to the Strategy pattern. Hide the actual concrete
implementation behind an abstract interface. the rest of the program
need never know what did the calculation.

Should i use a static method of a namespace or a class
containing static low-level routines? Consider that sometimes it could
be useful to create an object, because memory allocation could be done
only at construction (or better, initialization) time (and i could
reuse the object untile its destruction). The burden of handling such
memory would be demanded to the object itself, instead of having to
manage memory to pass to a static method.

yes, this seems like a good reason /not/ to use a static method
Another question i have is the following: suppose i have three
different implementation of the same algorithm. One is written in
plain C++, one is written in OpenCL (for parallel processing, would
result in different C++ code, with special function to be called,
kernel programs that run on the device to be compiled and so on), one
is written in playin vertex/program shaders for GPU (call it GPU
implementation, it would use OpenGL for example). The algorithm do the
same job (for example, processing an image and returnings the a list
of coordinates of special pixels), but the method to call it and the
way it returns data can be very different. The problem is the
following:

1.
I will drop C++ implementation of the low-level algorithms in future,
but for now, OpenCL is not so widespread, and i am forced to implement
things in different ways.

2.
C++ plain implementation results in code to be executed on the CPU.
Memory is passed/returned as usual (through pointers).

3.
OpenCL/GPU implementation: although they are different under every
aspect, share a common thing: memory must be transferred from main
system
to device memory for input parameters and the contrary for output
parameters.

So, 2 and 3 call for some sort of interface for input/output
parameters. Even the format of this parameters is important. For
example, a graph implemented using pointers would be not suited to be
passed to the device memory. But this aspect would be decided in
advance (whatever is the implementation of the low-level algorithm,
the format of the input parameters must be such that no pointers are
valid and so on...), so no problem.

But, one thing to keep in mind and which is important is the
following: low-level algorithms usually are called one after another,
that is, one run A1, followed by A2, followed by A3 and so on. Now, it
can happen that A_{i+1} uses the results provided by A_i and it would
be a waste of time to transfer memory from device to main system at
the end of A_i: just tell A_{i+1} that the parameter needed is already
on the device memory and has handle XXX.

Ok, it is clear that some form of interface for input/output
parameters is needed. This object would have a method Create( void
*memory, etc...) and would handle all of this.

So, i have an interface for memory object to be used as input/output
parameters for low-level algorithsm. Ok, but how would you create this
algorithms?

I don't think I know enough about your problem. Could the A1..An
functions have some sort of context passed to them. The context would
handle the details of memory management
I mean, mixing CPU with GPU implementation would be a stupid thing to
do. OpenCL uses the concept of Context.

ah! :)
A natural way would be to
create an interface IContext and having diffrerent context for
different implementation (CPU, OPENCL, OPENGL for example). Then i
would implement a factory for low-level algorithms. This factory
would create IAlgorithm objects (another interface, but could be a
void *) given:

I'm not sure why void* is needed.
1. the name of the algorithm (for example: ALGO_EDGEDETECTOR_SOBEL)
2. the actual context

Finally, who construct the context? The client of the software will
use a static method, for example:

IContext *ctx = ContextFactory::CreateContext( CTX_OPENCL );

So, to sum up:

. a context concept, IContext, created by a context factory
. a way of handling input/output for low-level algorithm
(IMemoryObject could be the basic interafce)
. a factory for classes implementing low-level algorithms

Is this general desing a mess? What do you think? How would you handle
this situation?

well it *sounds* sort of reasonable
 

Ask a Question

Want to reply to this thread or ask your own question?

You'll need to choose a username for the site, which only take a couple of moments. After that, you can post your question and our members will help you out.

Ask a Question

Members online

Forum statistics

Threads
473,744
Messages
2,569,484
Members
44,903
Latest member
orderPeak8CBDGummies

Latest Threads

Top