DPCtl Usage

DPCtl provides API to manage specific SYCL resources for SYCL-based Python packages. DPNP uses DPCtl as a global SYCL queue manager. Below code illustrates simple usage of DPNP in combination with dpCtl.

import dpctl
import dpnp as np

with dpctl.device_context("opencl:gpu"):
    x = np.array([1, 2, 3])
    s = np.sum(x)

For more information please refer to DPCtl’s documentation.

Example

import time

try:
    import dpnp
except ImportError:
    import os
    import sys

    sys.path.insert(0, os.path.abspath('.'))
    import dpnp

import numpy


def run(executor, size, test_type, repetition):
    x = executor.reshape(executor.arange(size * size, dtype=test_type), (size, size))

    times = []
    for _ in range(repetition):
        start_time = time.perf_counter()
        result = executor.sum(x)
        end_time = time.perf_counter()
        times.append(end_time - start_time)

    return numpy.median(times), result


def example():
    test_repetition = 5
    for test_type in [numpy.float64, numpy.float32, numpy.int64, numpy.int32]:
        type_name = numpy.dtype(test_type).name
        print(f"...Test data type is {type_name}, each test repetitions {test_repetition}")

        for size in [64, 128, 256, 512, 1024, 2048, 4096]:
            time_numpy, result_numpy = run(numpy, size, test_type, test_repetition)
            time_dpnp, result_dpnp = run(dpnp, size, test_type, test_repetition)

            if result_dpnp == result_numpy:
                verification = True
            else:
                verification = f"({result_dpnp} != {result_numpy})"

            msg = f"type:{type_name}:N:{size:4}:NumPy:{time_numpy:.3e}:SYCL:{time_dpnp:.3e}"
            msg += f":ratio:{time_numpy/time_dpnp:6.2f}:verification:{verification}"
            print(msg)


if __name__ == "__main__":
    try:
        import dpctl

        with dpctl.device_context("opencl:gpu") as gpu_queue:
            gpu_queue.get_sycl_device().dump_device_info()
            example()

    except ImportError:
        example()