Compile time unrolled @parameter for loop?

[paugier]

It's possible to write

    check_bench_1_nelts[2](original_time)
    check_bench_1_nelts[4](original_time)
    check_bench_1_nelts[8](original_time)
    check_bench_1_nelts[16](original_time)
    check_bench_1_nelts[32](original_time)

Given that there are @parameter if and @parameter fn, wouldn't it make sense to be able to write something like:

@parameter
for nelts in [2, 4, 8, 16, 32]:
    check_bench_1_nelts[nelts](original_time)

[lattner]

I think this is supported if your for loop is unrolled but I'm not entirely sure the phase ordering here: does loop unroll happen before elaboration @weiweichen ?

[weiweichen]

loop unrolling currently happens after elaboration because we don't know the trip count till elaboration happens. Currently we only support for with range, not for i in [...] yet, and to get the loop bounds and step, we need elaboration to get those values out of range. I think a @parameter for makes sense if we can get the loop info before elaboration, maybe when we have support for

for nelts in [2, 4, 8, 16, 32]:

or we can add something like @parameter for.

[mikowals]

There is now VariadicList.__iter__() to allow for width in VariadicList[Int](2, 4, 8, 16, 32) but width will be a dynamic value. All the values are available at compile time, it is just that the workaround using unroll is still necessary:

alias simd_widths = VariadicList[Int](2, 4, 8, 16, 32)
@parameter
fn simd_loop[width_index: Int]():
    alias width = simd_widths[width_index] # VariadicList.__getitem__() produces a compile time value
    check_bench_1_nelts[width](original_time)

# VariadicList.__len__() produces a compile time value
unroll[len(simd_widths), simd_loop]()

Should there be a feature request to either implement @parameter on for loops or to make for width in VariadicList[Int](2, 4, 8, 16, 32) produce a compile time value without needing a decorator?

[weiweichen]

I think @parameter will still be needed like it is for @parameter if for the compiler to know that the loop iteration is a constant value, regardless whether VariadicList.__iter__() is a constant value or not. In other words, IMO, the key feature is to add @parameter for where the loop iterator can be used as a compile time constant/"parameter" in the loop body, and then VariadicList can leverage this new feature.