opencl experiment

exp_opencl.ml

#require "owl_opencl";;

let prog_s = "
  __kernel void hello_kernel(__global const float *a,
                             __global const float *b,
                             __global float *result)
  {
    int gid = get_global_id(0);
    result[gid] = a[gid] + b[gid];
  }
";;

let l = Owl_opencl.Platform.get_platforms ();;
let p = Owl_opencl.Platform.get_info l.(0);;
let m = Owl_opencl.Device.get_devices l.(0);;
let d = Owl_opencl.Device.get_info m.(0);;
let ctx = Owl_opencl.Context.create m;;
let ctx_info = Owl_opencl.Context.get_info ctx;;
let program = Owl_opencl.Program.create_with_source ctx [|prog_s|];;
Owl_opencl.Program.build program m;;
let program_info = Owl_opencl.Program.get_info program;;
let kernel = Owl_opencl.Kernel.create program "hello_kernel";;
Owl_opencl.Kernel.get_info kernel;;

let cmdq = Owl_opencl.CommandQueue.create ctx m.(0);;
Owl_opencl.CommandQueue.get_info cmdq;;

let x = Arr.uniform [|10;10;10|];;
let buf = Owl_opencl.Buffer.create ~flags:[|Owl_opencl_generated.cl_MEM_USE_HOST_PTR|] ctx x;;

hello_world.ml

#require "owl_opencl";;

let prog_s = "
  __kernel void hello_kernel(__global const float *a,
                             __global const float *b,
                             __global float *result)
  {
    int gid = get_global_id(0);
    result[gid] = a[gid] + b[gid];
  }
";;

let l = Owl_opencl.Platform.get_platforms ();;
let m = Owl_opencl.Device.get_devices l.(0);;
let gpu = [|m.(1)|];;
let ctx = Owl_opencl.Context.create gpu;;
let cmdq = Owl_opencl.CommandQueue.create ctx gpu;;
let program = Owl_opencl.Program.create_with_source ctx [|prog_s|];;
Owl_opencl.Program.build program gpu;;
let kernel = Owl_opencl.Kernel.create program "hello_kernel";;

let a = Dense.Ndarray.S.uniform [|1024|];;
let b = Dense.Ndarray.S.uniform [|1024|];;
let a' = Owl_opencl.Buffer.create ~flags:[|Owl_opencl_generated.cl_MEM_USE_HOST_PTR|] ctx a ;;
let b' = Owl_opencl.Buffer.create ~flags:[|Owl_opencl_generated.cl_MEM_USE_HOST_PTR|] ctx b ;;
Kernel.set_arg