Assert failure: Unhandled instruction

[KeKsBoTer]

The following code sample results in this error:
(0): error 99999: Slang compilation aborted due to an exception of N5Slang13InternalErrorE: assert failure: Unhandled instruction
slang shader:

[AutoPyBindCUDA]
[CUDAKernel]
void gaussian_grad_bwd()
{
    var xz = diffPair(1.);
    bwd_diff(foo_gradient)(xz, float2(1.0));
}

[Differentiable]
float foo(float x)
{
    return exp(x);
}

[Differentiable]
float2 foo_gradient(float x)
{
    var x_diff = diffPair(x);

    bwd_diff(foo)(
        x_diff,
        1.0);

    return x_diff.d;
}

compiled with slangtorch:

import slangtorch
slangtorch.loadModule("error.slang")

The issue is the statement bwd_diff(foo)(x_diff,1.0);

slangtorch version: 1.2.6

[bmillsNV]

@saipraveenb25 can you help to take a look?

[saipraveenb25]

@KeKsBoTer:
This code looks like it is invoking bwd_diff multiple times on the same code. A double backwards pass is not currently well-supported.
Does your use-case require two reverse-mode derivative passes? The resulting code is usually very inefficient (atleast on the GPU).

For higher-order derivatives, it is (usually) much more efficient to use a bwd_diff over one or more fwd_diff (forward-mode derivative) calls to sweep the Hessian matrix row-by-row.

If you only need one reverse-mode pass, then there is no need to call bwd_diff again in gaussian_grad_bwd

[KeKsBoTer]

Thanks for the quick reply!

I do use the forward mode now and it works for my case.

I just wanted to report this bug.

[bmillsNV]

@saipraveenb25 can you please update the autodiff user guide to ensure we include this restriction, then please close this issue.