render faster

[wpalfi]

fyi you can speed up rendering by ~50%. Replace

large-steps-pytorch/scripts/render.py

Lines 210 to 213 in e03b40e

	# Sample incoming radiance
	pixel_normals = dr.interpolate(n[None, ...], rast, f)[0]
	# Sample envmap using the SH approximation
	light = self.sh.eval(pixel_normals)

by

vert_light = self.sh.eval(n).contiguous()
light = dr.interpolate(vert_light[None, ...], rast, f)[0]

[bathal1]

Nice, thanks! Do you have any insight as to why swapping those two lines speeds up computation so much?

[wpalfi]

Because the diffuse shading is view-independent and there are less vertices than pixels.

• The argument shapes of dr.interpolate() are the same in both cases.
• The argument size of self.sh.eval() is B*H*W*3 = 16*512**2*3 =~ 1e7 in the original code, but only V*3 =~ 1e5 in my version.

[bathal1]

Thanks! This has been updated now.