thanks for the detailed answer :) I am definitely excited about your project. I am sure however far you will go with it , it will definitely benefit Pharo . Good luck and keep us informed :)
On Thu, May 8, 2014 at 11:55 PM, Ronie Salgado <ronies...@gmail.com> wrote: > Blender <http://www.blender.org/> is actually porting OpenCL for its >> rendering engine cycles. Cycles already uses CUDA. They say that OpenCL is >> quite buggy yet and not well supported. Whats your experience with it ? > > > This is actually a problem with the AMD implementation of OpenCL, and yes > it is buggy, but being improved over the time, specially after Valve is > going for Linux and the new generation of consoles is based in AMD hardware. > > Until AMD made the GCN architecture(HD7xxx and upward), they did not have > hardware support for function calls unlike NVIDIA, so they have to inline > everything. The Blender Cycles rendering engine is huge, and specially > their noise functions used for procedural textures. Because everything was > being in-lined, they ended running out of memory easily when compiling > their kernels for the first time. In my desktop machine with 3 GB of RAM I > end without of memory when trying Cycles with OpenCL, but in my laptop with > 8 GB of RAM it works, but not perfect because they haven't done(at least > the last year when I last tried ) too much testing with OpenCL. > > With GCN AMD should have fixed the problem, but they keep using the same > intermediate representation that they used for their previouse generation, > so they keep inlining everything. In some time AMD should fix the problem, > specially when most of the new generation of consoles uses GCN and Valve > coming to Linux. > > My recommendation is to keep OpenCL kernels as simple as possible, which > should also be done for performance. > > > Are you going to expose also an API for non graphic stuff ? >> > > I am exposing some parts of the OpenCL API one by one. OpenCL is a small > API, but it is designed for C, so I am trying to make it a bit more > Smalltalk friendlier. But just a bit more. > > For testing the bindings, I have one test that performs a simple big > vector addition. So, it can be used for non graphics stuffs. > Graphics are cooler to show. > > do you think that we could have a openCL back-end for athens? >> > > Yes, it should be possible. A scanline renderer in OpenCL should be really > interesting of implementing. > > But, I think that a better option is to use OpenGL with the Stencil Buffer > and manual evaluation of the curves. I have done it before for vector > graphics, and I need an OpenGL based Athens backend for Roassal 3D and my > next video-game graphics engine. > I also tried some time ago with the paper present here: > http://ramisantina.com/blog/2011/09/resolution-independent-curves-at-graphicon-2011/. > I am interested in extending his technique from quadratic NURBS into > cubic NURBS. > > They are lot of different ways for making 2D vector graphics in the GPU, > and I want to try some of them for profiling. > > > > do you think that we could avoid having openCL text inside Pharo but have >> a kind of DSL? >> > > I hate MATLAB, so I would love a DSL for working with matrices, vectors > and use Roassal/GraphET for displaying the data and plotting. This should > be very easy to implement. > > For some domains, making a DSL should be easy. > > In general, a DSL for the GPGPU is not going to be enough. Some algorithms > have to be adapted explicitly for the GPU, taking in account the presence > of different memory models, such as the local memory. If we want a generic > model, we need to make a different Pharo-ish programming language. > > The work with UFFI/VirtualCPU should make that very long term task easy. > But for now, lets just use some DSLs. > > Greetings, > Ronie > > > 2014-05-08 11:06 GMT-04:00 stepharo <steph...@free.fr>: > > Hi ronie >> >> Some questions: >> do you think that we could have a openCL back-end for athens? >> >> do you think that we could avoid having openCL text inside Pharo but >> have a kind of DSL? >> >> Stef >> >> >> On 8/5/14 08:32, Ronie Salgado wrote: >> >>> Hello, >>> >>> I started working in making some basic bindings for OpenCL using >>> NativeBoost. >>> >>> Currently I can run basics OpenCL OpenCL kernels. I tried to generate >>> quickly some procedural images which was a success. >>> >>> I am going to be showing some more cool stuffs when mixing >>> OpenCL+OpenGL+SDL2. >>> >>> Currently uploading my work into http://smalltalkhub.com/#!/~ >>> ronsaldo/OpenCL <http://smalltalkhub.com/#%21/%7Eronsaldo/OpenCL> . No >>> configuration or stable release yet. >>> >>> Greetings, >>> Ronie >>> >> >> >> >
