Multi GPU

On multi-device clusters, let’s see how to select the card on which a KeOps operation will be performed.

Setup

Standard imports:

import numpy as np
import torch
from matplotlib import pyplot as plt

from pykeops.numpy import Genred

Define the list of gpu ids to be tested:

# By default we assume that there are two GPUs available with 0 and 1 labels:

gpuids = [0, 1] if torch.cuda.device_count() > 1 else [0]

KeOps Kernel

Define some arbitrary KeOps routine:

formula = "Square(p-a) * Exp(x+y)"
variables = ["x = Vi(3)", "y = Vj(3)", "a = Vj(1)", "p = Pm(1)"]

dtype = "float32"  # May be 'float32' or 'float64'

Tests with the NumPy API

my_routine = Genred(formula, variables, reduction_op="Sum", axis=1, dtype=dtype)

[pyKeOps] Warning : keyword argument dtype in Genred is deprecated ; argument is ignored.

Generate some data, stored on the CPU (host) memory:

M = 3000
N = 5000
x = np.random.randn(M, 3).astype(dtype)
y = np.random.randn(N, 3).astype(dtype)
a = np.random.randn(N, 1).astype(dtype)
p = np.random.randn(1).astype(dtype)

Launch our routine on the CPU, for reference:

c = my_routine(x, y, a, p, backend="CPU")

And on our GPUs, with copies between the Host and Device memories:

for gpuid in gpuids:
    d = my_routine(x, y, a, p, backend="GPU", device_id=gpuid)
    print(
        "Relative error on gpu {}: {:1.3e}".format(
            gpuid, float(np.mean(np.abs((c - d) / c)))
        )
    )

    # Plot the results next to each other:
    for i in range(3):
        plt.subplot(3, 1, i + 1)
        plt.plot(c[:40, i], "-", label="CPU")
        plt.plot(d[:40, i], "--", label="GPU {}".format(gpuid))
        plt.legend(loc="lower right")

    plt.tight_layout()
    plt.show()

Relative error on gpu 0: 1.214e-07

Tests with the PyTorch API

import torch
from pykeops.torch import Genred

my_routine = Genred(formula, variables, reduction_op="Sum", axis=1, dtype=dtype)

[pyKeOps] Warning : keyword argument dtype in Genred is deprecated ; argument is ignored.

First, we keep the data on the CPU (host) memory:

x = torch.from_numpy(x)
y = torch.from_numpy(y)
a = torch.from_numpy(a)
p = torch.from_numpy(p)
c = torch.from_numpy(c)

for gpuid in gpuids:
    d = my_routine(x, y, a, p, backend="GPU", device_id=gpuid)
    print(
        "Relative error on gpu {}: {:1.3e}".format(
            gpuid, float(torch.abs((c - d.cpu()) / c).mean())
        )
    )

    # Plot the results next to each other:
    for i in range(3):
        plt.subplot(3, 1, i + 1)
        plt.plot(c.cpu().numpy()[:40, i], "-", label="CPU")
        plt.plot(d.cpu().numpy()[:40, i], "--", label="GPU {}".format(gpuid))
        plt.legend(loc="lower right")

    plt.tight_layout()
    plt.show()

Relative error on gpu 0: 1.214e-07

Second, we load the data on the GPU (device) of our choice and let KeOps infer the device_id automatically:

for gpuid in gpuids:
    with torch.cuda.device(gpuid):
        # Transfer the data from Host to Device memory.
        # N.B.: The first call to ".cuda()" may take several seconds for each device.
        #       This is a known PyTorch issue.
        p, a, x, y = p.cuda(), a.cuda(), x.cuda(), y.cuda()

        # Call our KeOps routine:
        d = my_routine(x, y, a, p, backend="GPU")
        print(
            "Relative error on gpu {}: {:1.3e}".format(
                gpuid, float(torch.abs((c - d.cpu()) / c).mean())
            )
        )

        # Plot the results next to each other:
        for i in range(3):
            plt.subplot(3, 1, i + 1)
            plt.plot(c.cpu().numpy()[:40, i], "-", label="CPU")
            plt.plot(d.cpu().numpy()[:40, i], "--", label="GPU {}".format(gpuid))
            plt.legend(loc="lower right")

        plt.tight_layout()
        plt.show()

Relative error on gpu 0: 1.214e-07