Paper acknowledges positively D's approach to purity

[Robert Jacques]

On Mon, 15 Sep 2008 20:33:58 -0400, bearophile <bearophileHUGS at lycos.com>  
wrote:
> How have you used D with CUDA? I'd like to know a bit more technical  
> details.

Let's see. There's a couple of parts.
1) Created stub omf files to load the DLL correctly (CUDA doesn't support  
dynamic linking yet). The main issue here was there was a mangling change  
between what D expected and what NVIDIA provided.
2) Converted the CUDA device header over to D (done with htod with some  
manual clean up)
3) Made use of a template and mixin combo to simplify declaring all the  
methods of the CUDA types: i.e. float3, etc. I should refactor this to a  
template and aliases.
4) Made a high level api to simplify using CUDA. (Heavy use of templates  
again to type check, etc. everything). Here's an example:

  Device gpu = Device();                                         // Find  
the first CUDA compatible device
  writefln(gpu);                                                 // Print  
out its slot and type
  auto kernel = new Module("kernel.cubin");                      // Load a  
compiled CUDA module from a file
  float[] data = ...;                                            // Load  
some data
  auto image = new Texture!(float)(kernel,"image",640,480,data); // Place  
it into a texture
  auto ouput = new    cu1D!(float)(640*480);                     //  
Allocate an output array
  auto  func = new  Kernel!(float*)(kernel,"func",image);        // Get the  
function you want
  func[[40,30],[16,16]](output);                                 // Launch  
the function
  gpu.wait;                                                      // Wait  
for the GPU to finish
  write_output(output.dup);                                      // Get the  
results

I still write all the kernels in C and compile with nvcc though. Recently,  
PTX (aka CUDA assembly) support came to the driver, so writing D code that  
generates CUDA directly is theoretically possible.

> And do you know about BSGP (that I think is a step toward the future of  
> such things)?
> http://www.kunzhou.net/

I hadn't heard about BSGP, so thanks. The general concept of BSGP has been  
around for a few years now, Microsoft's Accelerator  
(research.microsoft.com/act/) comes to mind (And apparently Microsoft is  
the licensor for BSGP too), though the implementation looks more modern  
(at first glance).