`Undefined variable error` in kernel after update to `0.9.34`

[nHackel]

Hello, a kernel I wrote with KernalAbstractions.jl is now broken with version 0.9.34. I've reduced my kernel to this MWE:

 using KernelAbstractions

function foo(arr::AbstractMatrix{T}) where T
	backend = get_backend(arr)

	@kernel function foo_kernel(res, @Const(arr))
		row = @index(Group, Linear)
		grid_stride = prod(@groupsize)
		localIdx = @index(Local, Linear)
		N = size(arr, 2)
		
		shared = @localmem eltype(res) grid_stride
		shared[localIdx] = zero(eltype(res))
		
		tmp = zero(eltype(res))
		for i = localIdx:grid_stride:N
			tmp += arr[row, i]
		end
		shared[localIdx] = tmp
		@synchronize


		@private s = div(min(grid_stride, N), Int32(2))
		while s > Int32(0)
			if localIdx <= s
			shared[localIdx] = shared[localIdx] + shared[localIdx + s]
			end
			s >>= 1
			@synchronize
		end
	
		if localIdx == 1
			res[row] = shared[localIdx]
		end
	end
	
	res = similar(arr, size(arr, 1))
	kernel = foo_kernel(backend, 512)
	kernel(res, arr; ndrange = (512, size(arr, 1)))
	KernelAbstractions.synchronize(backend)
	return res
end


using CUDA
A = cu(rand(1024, 1024))
foo(A)

I'll try to reduce it even further.

I've run into this issue with both CUDA.jl and AMDGPU.jl with my custom kernel, however I could only reproduce the error with the MWE on CUDA.jl, since I don't have access to the AMD card at the moment.

With CUDA.jl I get the following error:

ERROR: a undefined variable error was thrown during kernel execution on thread (449, 1, 1) in block (40, 1, 1).
Stacktrace:                                                                                                              
[1] macro expansion at C:\...\.julia\packages\KernelAbstractions\sWSE0\src\KernelAbstractions.jl:242
[2] gpu_foo_kernel at C:\...\.julia\packages\KernelAbstractions\sWSE0\src\macros.jl:318
[3] gpu_foo_kernel at .\none:0                                                                                                                                                                                                                 

ERROR: LoadError: KernelException: exception thrown during kernel execution on device Quadro RTX 5000
Stacktrace:                                                                                                              
[1] check_exceptions()

(Note that on Windows I can't copy paste the MWE directly into the REPL and have to include it from a file, while on Linux it works. On both systems I can reproduce the error)

[vchuravy]

Thanks for opening the issue. Please see the changelog for the release here https://juliagpu.github.io/KernelAbstractions.jl/stable/#0.9.34

Your issue is likely solved by marking:

@uniform grid_stride = prod(@groupsize)

(Note that I would expect this to already error with the CPU backend on prior versions)

[nHackel]

Just with:

	@kernel function foo_kernel(res, @Const(arr))		
                grid_stride = 512
		shared = @localmem Float32 grid_strade
	end

the error also happens and with

	@kernel function foo_kernel(res, @Const(arr))		
                grid_stride = 512
		shared = @localmem Float32 512
	end

it works

[nHackel]

@vchuravy ah thanks for the quick reply! I'll try it out and get back to you

[nHackel]

Thank you, it works with both the MWE and the actual kernel!

The kernel is used as a linear operator for image reconstruction in magnetic particle imaging, a type of tomographic medical imaging. We never stumbled over the CPU error on previous versions, because for the CPU reconstructions we interact with the operator differently and bypassed the KernelAbstractions implementation