oneMKL Summary Statistics Usage Model

Description

A typical algorithm for summary statistics is as follows:

1. Create and initialize an object for dataset.

2. Call the summary statistics routine to calculate the appropriate estimate.

The following example demonstrates how to calculate mean values for a 3-dimensional dataset filled with random numbers. For dataset creation, the make_dataset helper function is used.

Buffer-based example

#include <iostream>
#include <vector>

#include "CL/sycl.hpp"
#include "oneapi/mkl/stats.hpp"

int main() {
    sycl::queue queue;

    const size_t n_observations = 1000;
    const size_t n_dims = 3;
    std::vector<float> x(n_observations * n_dims);
    // fill x storage with random numbers
    for(int i = 0; i < n_dims, i++) {
      for(int j = 0; j < n_observations; j++) {
            x[j + i * n_observations] = float(std::rand()) / float(RAND_MAX);
        }
    }
    //create buffer for dataset
    sycl::buffer<float, 1> x_buf(x.data(), x.size());
    // create buffer for mean values
    sycl::buffer<float, 1> mean_buf(n_dims);
    // create oneapi::mkl::stats::dataset
    auto dataset = oneapi::mkl::stats::make_dataset<oneapi::mkl::stats::layout::row_major>(n_dims, n_observations, x_buf);


    oneapi::mkl::stats::mean(queue, dataset, mean_buf);


    // create host accessor for mean_buf to print results
    auto acc = mean_buf.template get_access<sycl::access::mode::read>();


    for(int i = 0; i < n_dims; i++) {
      std::cout << "Mean value for dimension " << i << ": " << acc[i] << std::endl;
    }
    return 0;
}

USM-based example

#include <iostream>
#include <vector>

#include "CL/sycl.hpp"
#include "oneapi/mkl/stats.hpp"

int main() {
    sycl::queue queue;

    const size_t n_observations = 1000;
    const size_t n_dims = 3;

    sycl::usm_allocator<float, sycl::usm::alloc::shared> allocator(queue);

    std::vector<float, decltype(allocator)> x(n_observations * n_dims, allocator);
    // fill x storage with random numbers
    for(int i = 0; i < n_dims, i++) {
      for(int j = 0; j < n_observations; j++) {
            x[j + i * n_observations] = float(std::rand()) / float(RAND_MAX);
        }
    }
    std::vector<float, decltype(allocator)> mean_buf(n_dims, allocator);
    // create oneapi::mkl::stats::dataset
    auto dataset = oneapi::mkl::stats::make_dataset<oneapi::mkl::stats::layout::row_major>(n_dims,  n_observations, x);

   sycl::event event = oneapi::mkl::stats::mean(queue, dataset, mean);
   event.wait();
   for(int i = 0; i < n_dims; i++) {
     std::cout << "Mean value for dimension " << i << ": " << mean[i] << std::endl;
   }
   return 0;
}

USM usage

You can also use USM with raw pointers by using the sycl::malloc_shared/malloc_device functions.

Parent topic: Summary Statistics