This is caused by kernel_volume_shadow, which incorrectly uses the FLT_MAX ray distance from the sun sampling when applying the absorption in the glass volume - and for a distance of ~1e38, it makes sense that even the slightest absorption would cause the color to be saturated.

I guess the transparent shadow calculation should clip the distance somehow, I'll look into it.

We can probably clip distance by doing something like

However, that would be tricky to get real max distance for scene with procedural world volume. Maybe that case we can simply disable from clipping for the time being.

Why is clipping by the world bounds needed? If there are no precision issues and the mesh is closed, the ray distance used for computing the transmittance should already be short.

I think the issue is that we are missing a kernel_volume_stack_enter_exit in shadow_blocked for direct lighting with a transmission BSDF. We need to take into account that we are lighting from the other side of the surface and update the volume stack accordingly, same as we do for indirect light in kernel_path_surface_bounce.

Was just slapping code together, thought it was indeed clamping issue. Looking into actual cause now, and it's much deeper.

@Brecht Van Lommel (brecht), we should indeed discard element from volume stack for shadow_blocked, otherwise sampling is still considered to be happening inside of the volume which is wrong. Here is an updated patch

The confusing part was that 2.77a was rendering this file correct, so i thought we made a mistake in volume stack somewhere. But appears it was never a case. The issue here is that after rB9b6ed3a we are not ignoring volume closures which weight is less than
CLOSURE_WEIGHT_CUTOFF (before that refactor absorption was removed). So maybe we should do something like this:

For P526, I think it should only do kernel_volume_stack_enter_exit for transmission rays? Otherwise looks good to me.

If we do P527 then we should remove the duplicate cutoff test in bsdf_alloc and bsdf_alloc_osl. I don't think we should be doing this cutoff for volume closures though, unless we also take into account the distance. The cutoff is 1e-5f currently which is actually not that low if multiply it by distance of 1e5f. For BSDFs these values are already scene scale independent so it's safer to have a cutoff.

Yeah, think explicit check for transmission rays would be good. Just initially thought enter_exit will not do anything if that wasn't a transmission ray because it wouldn't be reflected in the stack then. In any case, will add check and commit.

I will leave changes about closure cutoff for later, you've raised the valid point of cutoff not being small enough.maybe we can cutoff at FLT_EPS or so, but let's make it another story.

@Brecht Van Lommel (brecht), actually, either i'm doing something wrong, or there is something else going on here. What i did is, instead of enter_exit in original patch did it like this:

	if(state->flag & PATH_RAY_TRANSMIT) {
		kernel_volume_stack_enter_exit(kg, sd, ps->volume_stack);
	}

and this gives artifacts in this file:

Without check for transmission rays there are no artifacts. So still need to look into this..

I think the test should be for the direct lighting ray:

if(dot(sd->Ng, ray->D) < 0.0f) {
    kernel_volume_stack_enter_exit(kg, sd, ps->volume_stack);
}

@Brecht Van Lommel (brecht), indeed. it we need to check whether shadow ray is being transmitted or reflected back, not the previous bounce (previous bounce might have been reflection inside of the volume). Just hoped to save some multiplications.. Will run more tests now and commit.