Repository navigation
Issue with performance using dpnp #1363
Description
The code on the dpnp using gpu or cpu devices is slower that using numpy library. But should be faster.
dpnp - Device(level_zero:gpu:0)
540 ms ± 32.8 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
dpnp - Device(opencl:gpu:0)
731 ms ± 77.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
dpnp - Device(opencl:cpu:0)
660 ms ± 36.3 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
numpy
290 ms ± 16.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
The code on the dpnp:
DISPLAY_W, DISPLAY_H = 800, 600
OFFSET_X = 1.4*DISPLAY_W//2
OFFSET_Y = DISPLAY_H//2
ZOOM = 2.5/DISPLAY_H
MAX_ITER = 30
import os
os.environ["SYCL_DEVICE_FILTER"] = "level_zero:gpu"
#os.environ["SYCL_DEVICE_FILTER"] = "opencl:gpu"
#os.environ["SYCL_DEVICE_FILTER"] = "cpu"
import dpnp as np
def color_by_intensity(intensity):
c1 = np.asarray([0.0, 0.0, 0.2])
c2 = np.asarray([1.0, 0.7, 0.9])
c3 = np.asarray([0.6, 1.0, 0.2])
intensity = np.broadcast_to(intensity[:, :, np.newaxis], intensity.shape + (3,))
return np.where(intensity < 0.5, c3*intensity + c2*(1.0-intensity), c1*intensity + c2*(1.0-intensity))
def mandelbrot(x, y, zoom):
xx = (x - OFFSET_X) * zoom
yy = (y - OFFSET_Y) * zoom
c = xx + 1j * yy[:, np.newaxis]
n_iter = np.full(c.shape, 0, dtype=np.int32) # 2d array
z = np.zeros(c.shape, dtype=np.complex64) # 2d array too
mask = (n_iter < MAX_ITER) # Initialize with True
for i in range(MAX_ITER):
z[mask] = z[mask]**2 + c[mask]
mask = mask & (np.abs(z) <= 2.0)
n_iter[mask] = i
intensity = n_iter.T / MAX_ITER
values = (color_by_intensity(intensity)*255).astype(np.int32)
return values
import matplotlib.pyplot as plt
%matplotlib inline
class Fractal:
def __init__(self, zoom):
self.values = np.full((DISPLAY_W, DISPLAY_H, 3), 0, dtype=np.int32)
self.zoom = zoom
self.x = np.linspace(0, DISPLAY_W, num=DISPLAY_W, dtype=np.float32)
self.y = np.linspace(0, DISPLAY_H, num=DISPLAY_H, dtype=np.float32)
print (self.x.device)
def calculate(self):
self.values = mandelbrot(self.x, self.y, self.zoom)
def draw(self):
cpu_values = np.asnumpy(self.values)
plt.imshow(cpu_values)
def main():
fractal = Fractal(ZOOM)
%timeit fractal.calculate()
fractal.draw()
if __name__ == "__main__":
main()
The same code using numpy:
DISPLAY_W, DISPLAY_H = 800, 600
OFFSET_X = 1.4*DISPLAY_W//2
OFFSET_Y = DISPLAY_H//2
ZOOM = 2.5/DISPLAY_H
MAX_ITER = 30
import numpy as np
def color_by_intensity(intensity):
c1 = np.asarray([0.0, 0.0, 0.2])
c2 = np.asarray([1.0, 0.7, 0.9])
c3 = np.asarray([0.6, 1.0, 0.2])
intensity = np.broadcast_to(intensity[:, :, np.newaxis], intensity.shape + (3,))
return np.where(intensity < 0.5, c3*intensity + c2*(1.0-intensity), c1*intensity + c2*(1.0-intensity))
def mandelbrot(x, y, zoom):
xx = (x - OFFSET_X) * zoom
yy = (y - OFFSET_Y) * zoom
c = xx + 1j * yy[:, np.newaxis]
n_iter = np.full(c.shape, 0) # 2d array
z = np.empty(c.shape, np.csingle) # 2d array too
mask = (n_iter < MAX_ITER) # Initialize with True
for i in range(MAX_ITER):
z[mask] = z[mask]**2 + c[mask]
mask = mask & (np.abs(z) <= 2.0)
n_iter[mask] = i
intensity = n_iter.T / MAX_ITER
values = (color_by_intensity(intensity)*255).astype(np.int32)
return values
import matplotlib.pyplot as plt
%matplotlib inline
class Fractal:
def __init__(self, zoom):
self.values = np.full((DISPLAY_W, DISPLAY_H, 3), 0, dtype=np.int32)
self.zoom = zoom
#self.need_recalculate = True
self.x = np.linspace(0, DISPLAY_W, num=DISPLAY_W, dtype=np.float32)
self.y = np.linspace(0, DISPLAY_H, num=DISPLAY_H, dtype=np.float32)
def calculate(self):
self.values = mandelbrot(self.x, self.y, self.zoom)
def draw(self):
plt.imshow(self.values)
def main():
fractal = Fractal(ZOOM)
%timeit fractal.calculate()
fractal.draw()
if __name__ == "__main__":
main()
Activity