gtc_fall_cupy_v6.py

# Copyright (c) 2019-2020, NVIDIA CORPORATION.
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import cupy as cp
import numpy as np
import sys

from cupy import prof
from scipy import signal


# CuPy: Version 6
# Implementations a user level cache from version 2.
# Seperates 32 bit and 64 bit versions to 
# reduce register pressure from version 3.
# Allows --use_fast_math flag to kernel compile from version 4.
# Move CUDA to cu file and build a fatbin to load with CuPy from version 5.
# Add launch bounds to CUDA kernel to assist compiler.


_kernel_cache = {}


def _lombscargle(x, y, freqs, pgram, y_dot):

    if (str(pgram.dtype)) in _kernel_cache:
        kernel = _kernel_cache[(str(pgram.dtype))]
    else:
        module = cp.RawModule(path="./_lombscargle_lb.fatbin")
        kernel = _kernel_cache[(str(pgram.dtype))] = module.get_function("_cupy_lombscargle_" + str(pgram.dtype))
        print("Registers", kernel.num_regs)

    device_id = cp.cuda.Device()
    numSM = device_id.attributes["MultiProcessorCount"]
    threadsperblock = (128, )
    blockspergrid = (numSM * 20,)    

    kernel_args = (
            x.shape[0],
            freqs.shape[0],
            x,
            y,
            freqs,
            pgram,
            y_dot,
        )

    kernel(blockspergrid, threadsperblock, kernel_args)

    cp.cuda.runtime.deviceSynchronize()


def lombscargle(
    x,
    y,
    freqs,
    precenter=False,
    normalize=False,
):

    x = cp.asarray(x)
    y = cp.asarray(y)
    freqs = cp.asarray(freqs)
    pgram = cp.empty(freqs.shape[0], dtype=freqs.dtype)

    assert x.ndim == 1
    assert y.ndim == 1
    assert freqs.ndim == 1

    # Check input sizes
    if x.shape[0] != y.shape[0]:
        raise ValueError("Input arrays do not have the same size.")

    y_dot = cp.zeros(1, dtype=y.dtype)
    if normalize:
        cp.dot(y, y, out=y_dot)

    if precenter:
        y_in = y - y.mean()
    else:
        y_in = y

    _lombscargle(x, y_in, freqs, pgram, y_dot)

    return pgram


if __name__ == "__main__":

    dtype = sys.argv[1]
    loops = int(sys.argv[2])

    A = 2.0
    w = 1.0
    phi = 0.5 * np.pi
    frac_points = 0.9  # Fraction of points to select

    in_samps = 2 ** 10
    out_samps = 2 ** 20

    np.random.seed(1234)
    r = np.random.rand(in_samps)
    x = np.linspace(0.01, 10 * np.pi, in_samps)
    x = x[r >= frac_points]
    y = A * np.cos(w * x + phi)
    f = np.linspace(0.01, 10, out_samps)

    # Use float32 if b32 passed
    if dtype == 'float32':
        x = x.astype(np.float32)
        y = y.astype(np.float32)
        f = f.astype(np.float32)

    d_x = cp.array(x)
    d_y = cp.array(y)
    d_f = cp.array(f)

    # Run baseline with scipy.signal.lombscargle
    with prof.time_range("scipy_lombscargle", 0):
        cpu_lombscargle = signal.lombscargle(x, y, f)

    # Run Numba version
    with prof.time_range("cupy_lombscargle", 1):
        gpu_lombscargle = lombscargle(d_x, d_y, d_f)

    # Copy result to host
    gpu_lombscargle = cp.asnumpy(gpu_lombscargle)

    # Compare results
    np.testing.assert_allclose(cpu_lombscargle, gpu_lombscargle, 1e-3)    

    # Run multiple passes to get average
    for _ in range(loops):
        with prof.time_range("cupy_lombscargle_loop", 2):
            gpu_lombscargle = lombscargle(d_x, d_y, d_f)