Handle MTLCommandBuffer Error Logs

[tgymnich]

Access logs to check for errors.

• MTLCommandBufferError
• WWDC Session: Debug GPU-side errors in Metal

Relevant TODO:

Metal.jl/src/compiler/execution.jl

Lines 303 to 304 in 6a760a6

	# TODO: access logs here to check for errors
	# https://developer.apple.com/videos/play/wwdc2020/10616/

[vovw]

taking a shot at this, do i need to handle all the error cases?

• Timeout errors (kernel took too long)
• Page faults
• Out of memory conditions
• Invalid resource access
• Stack overflows
• Access revocation
• Other unspecified errors

also each check should be a simple

    if error.code == MTL.Timeout
                @error "GPU kernel timed out" kernel_name=nameof(kernel.f)

like so?

[maleadt]

Maybe it's sufficient to just render the NSError with @error "GPU kernel failure" exception=(the_nserror, a_backtrace)? You probably can't throw from within that context, but some information will be needed to identify the specific launch that failed, so it'll need some experimentation to make this useful.