Cloo - Compute Language, Object Oriented

Thanks. No rush, though. I can get busy elsewhere in the code. As for you, I guess the priority ATM is OpenTK 1.0.

You can use any struct that has exactly 4 float fields:

struct Vector4f
{
    float x;
    float y;
    float z;
    float w;
 
    // methods here ....
}

If you don't have such a structure in your project you can use OpenTK.Vector4 instead.

I've never tried setting a local argument. Chapter 3.3 of the OpenCL specs:

Local Memory: A memory region local to a work-group. This memory region can be
used to allocate variables that are shared by all work-items in that work-group. It may be
implemented as dedicated regions of memory on the OpenCL device. Alternatively, the
local memory region may be mapped onto sections of the global memory.

Table 3.1 states that you cannot access (read or write) such arguments. You can only allocate them. Try removing the MemFlags.MemReadWrite flag when you create c and see what happens.

Edit: You can also create buffers without specifying an array:
ComputeBuffer<float> c = new ComputeBuffer<float>( context, flags, count );

No news about the "local" problem ?

Thx

Thank you very much for new release (0.3.1) and for a new test (KernelArgsTester).

In my environment (Intel CPU) compilation of the kernel fails with following error messages:

------------------| Start Kernel args test |------------------
C:\Users\xxx\AppData\Local\Temp\OCL9414.tmp.cl(4): error: kernel pointer
          arguments must point to addrSpace global, local, or constant
      kernel void k03(          image3d_t img ) {}
                  ^
 
C:\Users\xxx\AppData\Local\Temp\OCL9414.tmp.cl(14): error: a parameter
          cannot be allocated in a named address space
      kernel void k11( global   image3d_t img ) {}
                       ^
 
C:\Users\xxx\AppData\Local\Temp\OCL9414.tmp.cl(14): error: kernel pointer
          arguments must point to addrSpace global, local, or constant
      kernel void k11( global   image3d_t img ) {}
                  ^
 
3 errors detected in the compilation of "C:\Users\xxx\AppData\Local\Temp\OCL94
14.tmp.cl".
    image3d_t im
-------------------| End Kernel args test |-------------------

That's ok, but just for your info...

Are you going to implement some kind of automatic .NET to OpenCL kernel translation? There is a Brahma project http://brahma.ananthonline.net/ with some steps toward similar goal, but the project is completely stalled now. =((( It was an attempt to translate general computational LINQ expression to its parallel equivalent and to execute it on GPU using some deprecated DirectX GP GPU set of libraries.

I do not feel myself comfortable enough with writing parallel expressions on LINQ, but it is better to LINQ then to OpenCL. =) I mean it is not my first dream to study new deeply graphic oriented C-like dialect.

Have a fast code!
Anton.
http://kyta.spb.ru

No problem. There are chances that this is problems on my side: you provide kernels for nVidia implementation and I try to compile it using ATI's 2.0-beta4 driver.

I do not advertise LINQ (just mentioned Brahma project as reference): a) MS announced PLinq already; b) it is hard to write general computations in this syntax. We all like conditions and loops and all other procedural benefits C-like language gifts us. =)

Hmm!.. I wonder to start with System.Expression-to-Kernel conversion. System.Expression trees are very general way to represent general program tree with all its conditional branching and loops. LINQ is nomore then just a short way to write some complicated Expression tree...

But at the end of the game I dream to have some stand-alone .NET class written completely in C# [probably] without any external dependencies, [probably] completely covered with usual unit-tests. This class performs just one CPU intensive task and it _IS_ able to do the job. But it is too slow. Then I dream this class to be able to automatically analyze its own IL and to emit corresponding OpenCL kernel. After that (thanks to Cloo) it is just a few seconds to get 10-100 times speedup with OpenCL-on-CPU or even 100-1000 times speedup with OpenCL-on-GPU! Does anybody have robust IL-to-OpenCL translator? =DDD

It's a serious point. =|

Even a more serious point. But you are ready to show them the excellence, aren't you? ;-)

Have a fast code!
Anton.
http://kyta.spb.ru

How do you store your polygons in main memory?

Any way you like. If you declare a float2* on the kernel you can as well pass a float[]. The kernel will then read the items in pairs. Just don't forget that float2 is half the size (in items) so watch your for loops :)

Edit: Or you can use any struct that has exactly two floats:

struct MyFloat2
{
   float f1, f2;
}

You can use this struct in both ComputeBuffer or inside your OpenCL code:

struct MyFloat2
{
   float f1, f2;
}
 
kernel void K( global MyFloat2* floatArray )
{
...
}

I don't have much experience with SQL and I've never seen WKB before. I don't even know what you're trying to achieve (rendering?) therefore I'm not sure what to suggest.
Instead I'm posting the signature of a kernel that does some sort of hybrid raytracing. You probably don't need all this so take it as purely inspirational code:

kernel void render( 
           read_only  int     width,  // frame width
           read_only  int     height, // frame height
    global write_only float4* renderbuffer,  // the rendered image
    global            float*  depthbuffer,   // depth buffer information
    global read_only  float*  caminfo,  // camera info: location, view direction, up vector, fov, clipping distances
    global read_only  float*  inversematrix,  //  the inverse of the ransformation matrix for this mesh. it is applied to the ray before ray-mesh intersection test occurs
           read_only  float4  boundvol, // the bounding sphere around this mesh
    global read_only  float*  vertices, // vertex positions (v1.x, v1.y, v1.z, v2.x, v2.y, v2.z, etc. )
    global read_only  uint*   indices,   // indices that point to the vertices array. specifies how to create triangles (v3, v6, v2)
           read_only  uint    prim_count ) // number of triangles in indices

As you see you can pass your polys any way you like. Except for some limits that OpenCL specifies that is.
Passing a string would look like this.
OpenCL code:

kernel void Test( global char* str )
{...}

Cloo code:

ComputeBuffer<char> str = new ... ;
kernel.SetMemoryArgument( 0, str );

Last but not least: watch out for char encoding differences between .NET and OpenCL C.

This is not possible. OpenCL forbids pointer to pointer kernel arguments. Therefore you cannot pass array of structs that contain an inner array. As a result Cloo issues a warning (or crashes) before you attempt such a thing.
If your strings are always char[30] you can pack them all inside a ComputeBuffer of length 30*5000 (or whatever). You can then read/write different pieces of the buffer through ComputeCommandQueue.Read/Write methods.

Since last message I changed the code: now the job is done in just one kernel call. But now there another issue: my computer hangs during computations (on GPU). The only way to "awake it" is to press reset button. I would like to pay special attention: on CPU exactly the same kernel works fine.

Is there any work around for my case? I try to execute kernel for work item with dimensions 256x256x256. Such dimension is supported by GPU according to platform info. Also it has 9 computing units (CPU has only 2 computing units).

After the kernel is started, my video freezes: no mouse movements, no cursor blinks. For shorter tasks it awakes after a while, but THIS long-running kernel completely kills the PC. Does anybody else experience similar problem? What's a solution?

Thank you in advance,
Anton.
http://kyta.spb.ru

PS OS Ubuntu 10.04 64bit, ATI Stream SDK 2.1, Mono 2.6.4, Cloo 0.7.2, ATI Radeon HD 5750.