when I work with online examples I get somewhat confused and since I went down the wrong path once, and might be on the wrong track again, I'd need some clarification:
Which matrix and vector objects are used with which backend (OpenMP, CUDA, OpenCL)?
for either of {CUDA,OpenCL,OpenMP} you need to use the ViennaCL-datatypes, i.e. viennacl::vector<>, viennacl::matrix<>, etc.
ViennaCL used to be OpenCL-only in the early days, so there was a strict distinction between CPU types (std::vector, ublas::vector) necessary. With the addition of the OpenMP backend to ViennaCL, a strict CPU vs. GPU distinction was no longer necessary, but it still helped with the standard use-case of ViennaCL for GPUs.
If you have any concrete suggestions on how to improve the documentation, we are more than happy to take your input :-)
Best regards,
Karli
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
About the documentation, an introduction that lowers the entry hurdle to a small step would help a lot.
Scope: Iterative solvers
Rules: Stick to best practices, the high speed options and restrict it to ViennaCL
Audience: "KGVP" (Kinder, Greise, Vorstände und Partner) or "How to Explain your Grandma How to use ViennaCL iterative solvers?"
Acceptance & Marketing effect: If people master this first step with ease, acceptance will go through the roof and the word will spread
One page could be enough .. here we go...:
Prerequisites: Install c++ (includes OpenMP to use multiple CPU cores), CUDA (optional for use of Nvidia GPUs), OpenCL (optional for use of Nvidia or AMD GPUs) on your computer
Get ready!
step 1: Download ViennaCL from the internet and extract it
step 2: Create a project folder "myproject"
step 3: Copy the folder "viennacl" which is inside the extracted directory from your download to "myproject", that's it, no "installation" required
done
Playtime!
1: Copy the following code to "myproject"
2: Compile it with the compiler flag for OpenMP and run it on the CPU first: ... - fopenmp, before you run it make sure you tell the computer how many cores or threads to use, on Linux that's done by executing the command 'export OMP_NUM_THREADS=8' ,here to use 8 threads, before you execute your programm
3: Play with the different solvers, preconditioners and relevant types of matrices
4: Compile it for GPUs and experience the difference
Have fun.
Code, we need to work out in detail: (for now just a proposed structure)
// A) Define which "backend", OpenMP or CUDA or OpenCL you want to use, default is OpenMP#define VIENNACL_WITH_OPENMP // #define VIENNACL_WITH_CUDA // #define VIENNACL_WITH_OPENCL // B) Add c++ includes if requiredCodegoeshere...// C) Add ViennaCL includes// For matrix, vectors and solvers#include"../viennacl/vector.hpp"#include"../viennacl/compressed_matrix.hpp"#include"../viennacl/linalg/gmres.hpp"#include"../viennacl/linalg/bicstab.hpp"#include"../viennacl/linalg/cg.hpp"#include"../viennacl/linalg/mixed_precision_cg.hpp"// For preconditioners#include"../viennacl/linalg/detail/ilu/block_ilu.hpp"#include"../viennacl/linalg/amg.hpp"#include"../viennacl/linalg/ichol.hpp"...note:allpreconditionersshouldbeincluded// D) Create the matrix and vector objects needed; we want to solve A x x = b// matrixviennacl::compressed_matrix<ScalarType>A(n,m,nnz);// sparse matrix A with size nxm and nnz nonzero values// vector(s) for x (soution) and right-hand-side often called rhs or bviennacl::vector<ScalarType>x(n);// size nviennacl::vector<ScalarType>b(n);// size n// E1) Code to load sample matrix from matrix marketCodegoeshere...// E2) Code to load or generate b/rhsCodegoeshere...// F) Complete sample setup of all preconditioners// Preconditioner 1: name -----------------------------------------Codegoeshere...// Preconditioner 2: name ------------------------------------------Codegoeshere......// Preconditioner N: AMG -------------------------------------------Codegoeshere...// G) Create an (optional) initial guess for xCodegoeshere...// H) Complete sample setup of all four solvers incl. a preconditioner that works best for the selected matrix// Solver 1: GMRES ----------------------------------------------Codegoeshere...// Solver 2: bicstab --------------------------------------------Codegoeshere...// Solver 3: cg -------------------------------------------------Codegoeshere...// Solver 4: mixed_precision_cg ---------------------------------Codegoeshere...// I) Output that documents the performanceCodegoeshere...
note: The default settings should run out of the box, changes should be limited to copy/paste and commenting, uncommenting sections and there should be a youtube video showing a rookie doing the exercise. It should be like made of lego building blocks to play with.
What do you think?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I have a couple of questions:
1.) In what way is the current "Installation" in the online manual insufficient? http://viennacl.sourceforge.net/doc/manual-installation.html It describes just the steps you mentioned above?
2.) The suggested code is actually not too different from what is in examples/benchmark/solver.cpp - apparently it's not prominently placed. Did you find that example at all?
3.) Would you watch a youtube video for a library installation? I think that written documentation is much more efficient in providing that information. But I'd like to get feedback, since my own views are not necessarily relevant ;-)
Best regards,
Karli
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I am sure the information is somewhere on the website in the examples... sometimes easy to find and digest, at times "incomplete" at first sight with a reference to the "missing" bids...
From my experience a quick, complete, step-by-step and modular introduction can lower the entry barrier for a newcomer a lot.
Regarding your qustions:
1: Yes, it's absolutely fine but it should be part of any introduction, it's step 1
(What's not so good regarding examples is e.g., in the download package, there's a build directory but cmake works in the example directory so the build directory is of no use and confusing)
2: Yes, there are examples, I liked the table showing what they are about and what's the related backend but what I'd like better is a toolbox with complete blocks of code and a note about the relevant header files. Some examples or code snippets on the website are followed by a note refering to another e.g. preconditioner regarding configuration details or configuration options are not fully exploited, covering 8 of 10 options, (where) would people search for more options if they are not mentioned in relevant examples?. It may be good programing style to avoid the same code in two places but I found it cumbersome to copy the bits and pieces from different locations and to get the syntax right. Also important I think, there's little information about how to copy matrices and vectors from/to ViennaCL, that's a gerenral problem of linear algebra libraries, usually (talking about libraries in general) examples show how to create and load some random numbers with a loop which I think is not really relevant. Big issues I encountered are the data formats COO, CSR ... more precisely, what's the exact data structure (usually a set of arrays[] for the matrix) that needs to be prepared and how to load/copy data into the ViennaCL objects. I have seen examples where people tried to develop extensions for their applications to be able to use external libraries/solvers and failed with 90% of the solution ready because they didn't get the data conversion right.
3: I wouldn't watch a youtube video just about the installation (which is only copy/paste), written documentation is fine but is there an up-to-date version? The one I have is from version 1.5 which is fine for me so far.
Does that help?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello,
when I work with online examples I get somewhat confused and since I went down the wrong path once, and might be on the wrong track again, I'd need some clarification:
Which matrix and vector objects are used with which backend (OpenMP, CUDA, OpenCL)?
There are ViennaCL, std and ublas objects and e.g. example http://viennacl.sourceforge.net/viennacl-examples-vector.html uses
while example http://viennacl.sourceforge.net/viennacl-examples-iterative.html works with
~~~
// Set up some ublas objects (CPU):
ublas::vector<ScalarType> ublas_rhs;
// Set up some ViennaCL objects (GPU):
viennacl::vector<ScalarType> vcl_rhs;
~~~
What should I use to work with the OpenMP backend on the CPU utilising multiple CPU cores?
matrix object?
vector object?
What should I use to work with the CUDA or OpenCL backend on the GPU using multiple GPUs?
matrix object?
vector object?
T.
Hi Tario,
for either of {CUDA,OpenCL,OpenMP} you need to use the ViennaCL-datatypes, i.e.
viennacl::vector<>,viennacl::matrix<>, etc.ViennaCL used to be OpenCL-only in the early days, so there was a strict distinction between CPU types (std::vector, ublas::vector) necessary. With the addition of the OpenMP backend to ViennaCL, a strict CPU vs. GPU distinction was no longer necessary, but it still helped with the standard use-case of ViennaCL for GPUs.
If you have any concrete suggestions on how to improve the documentation, we are more than happy to take your input :-)
Best regards,
Karli
About the documentation, an introduction that lowers the entry hurdle to a small step would help a lot.
Scope: Iterative solvers
Rules: Stick to best practices, the high speed options and restrict it to ViennaCL
Audience: "KGVP" (Kinder, Greise, Vorstände und Partner) or "How to Explain your Grandma How to use ViennaCL iterative solvers?"
Acceptance & Marketing effect: If people master this first step with ease, acceptance will go through the roof and the word will spread
One page could be enough .. here we go...:
Prerequisites: Install c++ (includes OpenMP to use multiple CPU cores), CUDA (optional for use of Nvidia GPUs), OpenCL (optional for use of Nvidia or AMD GPUs) on your computer
Get ready!
step 1: Download ViennaCL from the internet and extract it
step 2: Create a project folder "myproject"
step 3: Copy the folder "viennacl" which is inside the extracted directory from your download to "myproject", that's it, no "installation" required
done
Playtime!
1: Copy the following code to "myproject"
2: Compile it with the compiler flag for OpenMP and run it on the CPU first: ... - fopenmp, before you run it make sure you tell the computer how many cores or threads to use, on Linux that's done by executing the command 'export OMP_NUM_THREADS=8' ,here to use 8 threads, before you execute your programm
3: Play with the different solvers, preconditioners and relevant types of matrices
4: Compile it for GPUs and experience the difference
Have fun.
Code, we need to work out in detail: (for now just a proposed structure)
note: The default settings should run out of the box, changes should be limited to copy/paste and commenting, uncommenting sections and there should be a youtube video showing a rookie doing the exercise. It should be like made of lego building blocks to play with.
What do you think?
Thanks, tario, for the suggestion.
I have a couple of questions:
1.) In what way is the current "Installation" in the online manual insufficient? http://viennacl.sourceforge.net/doc/manual-installation.html It describes just the steps you mentioned above?
2.) The suggested code is actually not too different from what is in
examples/benchmark/solver.cpp- apparently it's not prominently placed. Did you find that example at all?3.) Would you watch a youtube video for a library installation? I think that written documentation is much more efficient in providing that information. But I'd like to get feedback, since my own views are not necessarily relevant ;-)
Best regards,
Karli
I am sure the information is somewhere on the website in the examples... sometimes easy to find and digest, at times "incomplete" at first sight with a reference to the "missing" bids...
From my experience a quick, complete, step-by-step and modular introduction can lower the entry barrier for a newcomer a lot.
Regarding your qustions:
1: Yes, it's absolutely fine but it should be part of any introduction, it's step 1
(What's not so good regarding examples is e.g., in the download package, there's a build directory but cmake works in the example directory so the build directory is of no use and confusing)
2: Yes, there are examples, I liked the table showing what they are about and what's the related backend but what I'd like better is a toolbox with complete blocks of code and a note about the relevant header files. Some examples or code snippets on the website are followed by a note refering to another e.g. preconditioner regarding configuration details or configuration options are not fully exploited, covering 8 of 10 options, (where) would people search for more options if they are not mentioned in relevant examples?. It may be good programing style to avoid the same code in two places but I found it cumbersome to copy the bits and pieces from different locations and to get the syntax right. Also important I think, there's little information about how to copy matrices and vectors from/to ViennaCL, that's a gerenral problem of linear algebra libraries, usually (talking about libraries in general) examples show how to create and load some random numbers with a loop which I think is not really relevant. Big issues I encountered are the data formats COO, CSR ... more precisely, what's the exact data structure (usually a set of arrays[] for the matrix) that needs to be prepared and how to load/copy data into the ViennaCL objects. I have seen examples where people tried to develop extensions for their applications to be able to use external libraries/solvers and failed with 90% of the solution ready because they didn't get the data conversion right.
3: I wouldn't watch a youtube video just about the installation (which is only copy/paste), written documentation is fine but is there an up-to-date version? The one I have is from version 1.5 which is fine for me so far.
Does that help?