gpu.js/example.js

"use strict";

// Dependency:
var fs = require("fs");
var path = require("path");
var cwd = __dirname;
var nooocl = require('nooocl');
var CLHost = nooocl.CLHost;
var CLContext = nooocl.CLContext;
var CLBuffer = nooocl.CLBuffer;
var CLCommandQueue = nooocl.CLCommandQueue;
var NDRange = nooocl.NDRange;
var CLError = nooocl.CLError;
var fastcall = require("fastcall");
var ref = fastcall.ref;
var double = ref.types.double;

// Initialize OpenCL then we get host, device, context, and a queue
var host = CLHost.createV11();
var defs = host.cl.defs;

var platforms = host.getPlatforms();
var device;
function searchForDevice(hardware) {
  platforms.forEach(function (p) {
    var devices = hardware === "gpu" ? p.gpuDevices() : p.cpuDevices();
    devices = devices.filter(function (d) {
      // Is double precision supported?
      // See: https://www.khronos.org/registry/cl/sdk/1.1/docs/man/xhtml/clGetDeviceInfo.html
      return true;
      return d.doubleFpConfig &
        (defs.CL_FP_FMA | defs.CL_FP_ROUND_TO_NEAREST | defs.CL_FP_ROUND_TO_ZERO | defs.CL_FP_ROUND_TO_INF | defs.CL_FP_INF_NAN | defs.CL_FP_DENORM);
    });
    if (devices.length) {
      device = devices[0];
    }
    if (device) {
      return false;
    }
  });
}

searchForDevice("gpu");
if (!device) {
  console.warn("No GPU device has been found, searching for a CPU fallback.");
  searchForDevice("cpu");
}

if (!device) {
  throw new Error("No capable OpenCL 1.1 device has been found.");
}
else {
  console.log("Running on device: " + device.name + " - " + device.platform.name);
}

var context = new CLContext(device);
var queue = new CLCommandQueue(context, device);

// Initialize data on the host side:
var n = 1000;
var bytes = n * double.size;

var h_a = new Buffer(n * double.size);
var h_b = new Buffer(n * double.size);
var h_c = new Buffer(n * double.size);

// Initialize vectors on host
for (var i = 0; i < n; i++) {
  var offset = i * double.size;
  double.set(h_a, offset, 0.1 + 0.2);
  double.set(h_b, offset, 0);
}

// Create device memory buffers
var d_a = new CLBuffer(context, defs.CL_MEM_READ_ONLY, bytes);
var d_b = new CLBuffer(context, defs.CL_MEM_READ_ONLY, bytes);
var d_c = new CLBuffer(context, defs.CL_MEM_WRITE_ONLY, bytes);

// Copy memory buffers
// Notice: the is no synchronous operations in NOOOCL,
// so there is no blocking_write parameter there.
// All writes and reads are asynchronous.
queue.enqueueWriteBuffer(d_a, 0, bytes, h_a);
queue.enqueueWriteBuffer(d_b, 0, bytes, h_b);

// It's time to build the program.
var kernelSourceCode = `
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
__kernel void vecAdd(  __global double *a,
                       __global double *b,
                       __global double *c,
                       const unsigned int n)
{
    //Get our global thread ID
    int id = get_global_id(0);

    //Make sure we do not go out of bounds
    if (id < n)
        c[id] = a[id] + b[id];
}
`;
var program = context.createProgram(kernelSourceCode);

console.log("Building ...");
// Building is always asynchronous in NOOOCL!
nooocl.scope(function () {
  return program.build("-cl-fast-relaxed-math")
    .then(function () {
      var buildStatus = program.getBuildStatus(device);
      var buildLog = program.getBuildLog(device);
      console.log(buildLog);
      if (buildStatus < 0) {
        throw new CLError(buildStatus, "Build failed.");
      }
      console.log("Build completed.");

      // Kernel stuff:
      var kernel = program.createKernel("vecAdd");

      kernel.setArg(0, d_a);
      kernel.setArg(1, d_b);
      kernel.setArg(2, d_c);
      // Notice: in NOOOCL you have specify type of value arguments,
      // because there is no C compatible type system exists in JavaScript.
      kernel.setArg(3, n, "uint");

      // Ranges:
      // Number of work items in each local work group
      var localSize = new NDRange(64);
      // Number of total work items - localSize must be devisor
      var globalSize = new NDRange(Math.ceil(n / 64) * 64);

      console.log("Launching the kernel.");

      // Enqueue the kernel asynchronously
      queue.enqueueNDRangeKernel(kernel, globalSize, localSize);

      // Then copy back the result from the device to the host asynchronously,
      // when the queue ends.
      // We should query a waitable queue which returns an event for each enqueue operations,
      // and the event's promise can be used for continuation of the control flow on the host side.
      console.log("Waiting for result.");
      return queue.waitable().enqueueReadBuffer(d_c, 0, bytes, h_c).promise
        .then(function() {
          // Data gets back to host, we're done:

          var sum = 0;
          for (var i = 0; i < n; i++) {
            var offset = i * double.size;
            sum += double.get(h_c, offset);
          }

          console.log("Final result: " + sum / n);
        });
    });
});

console.log("(Everything after this point is asynchronous.)");