# \`pack\_bits\` inside of vectorize fails constraint check **URL:** \ **Category:** Mojo\ **Tags:** 25\_3\ **Created:** [May 6, 2025, 6:37pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384 "2025-05-06T18:37:44Z")\ **Posts on this page:** 12\ **Page:** 1
**Author:** ![duck\_tape](https://sea1.discourse-cdn.com/flex001/user_avatar/forum.modular.com/duck_tape/32/432_2.png) [@duck\_tape](https://forum.modular.com/u/duck_tape)\ **Post date:** [May 6, 2025, 6:37pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/1 "2025-05-06T18:37:44Z")
```mojo import math from algorithm import vectorize from bit import pop_count from memory import pack_bits from sys import simdwidthof alias U8_SIMD_WIDTH = simdwidthof[DType.uint8]() fn count_nuc_content[ simd_width: Int, nuc: UInt8 ](sequence: Span[UInt8]) -> UInt: var count = 0 var ptr = sequence.unsafe_ptr() # This works # alias nuc_vector = SIMD[DType.uint8, simd_width](nuc) # var aligned_end = math.align_down(len(sequence), simd_width) # for offset in range(0, aligned_end, simd_width): # var vector = ptr.offset(offset).load[width=simd_width]() # var mask = vector == nuc_vector # var packed = pack_bits(mask) # count += Int(pop_count(packed)) # for offset in range(aligned_end, len(sequence)): # count += 1 if sequence[offset] == nuc else 0 # This does not @parameter fn count_nucs[width: Int](offset: Int): alias nuc_vector = SIMD[DType.uint8, width](nuc) var vector = ptr.offset(offset).load[width=width]() var mask = vector == nuc_vector var packed = pack_bits(mask) # ^ constraint failed: the width of the bool vector must be the same as the bitwidth of the target type count += Int(pop_count(packed)) vectorize[count_nucs, simd_width](len(sequence)) return count def main(): var seq = "ACTGACTGACGCCCCCCCCCCCCTTTTTTTTT".as_bytes() var count = count_nuc_content[U8_SIMD_WIDTH, ord("C")](seq) print("C Content:", count) ``` The commented out “manual” SIMD works. But the vectorize application does not. `pack_bits` says: ```txt open-source/max/mojo/stdlib/stdlib/builtin/constrained.mojo:58:6: note: constraint failed: the width of the bool vector must be the same as the bitwidth of the target type ``` Gist is the slightly more fleshed out example: [count\_nucleotide.mojo · GitHub](https://gist.github.com/sstadick/1e3a12d17718cc338e253970512dc9bc) ---
**Author:** ![sora](https://avatars.discourse-cdn.com/v4/letter/s/ad7895/32.png) [@sora](https://forum.modular.com/u/sora)\ **Post date:** [May 6, 2025, 7:01pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/2 "2025-05-06T19:01:15Z")
I implemented `pack_bits`, but I’m a bit conflicted about whether we should support the `Scalar[bool]` case. Maybe we should just return a `Scalar[bool]` when given one, that would be slightly more consistent. That said, you’re still blocked by `pop_count`, which only accepts an integral `dtype`. That said, since `pack_bits` is supposed to be a very thin wrapper around `pop.bitcast`, the fact that the following doesn’t compile seems to argue against that. ```mojo fn main(): val = Scalar(True) res = __mlir_op.`pop.bitcast`[_type = Scalar[DType.bool]._mlir_type](val.value) # error: MLIR verification error: # 'pop.bitcast' op operand type # '!pop.scalar' # and result type # '!pop.scalar' # are cast incompatible ``` ---
**Author:** ![duck\_tape](https://sea1.discourse-cdn.com/flex001/user_avatar/forum.modular.com/duck_tape/32/432_2.png) [@duck\_tape](https://forum.modular.com/u/duck_tape)\ **Post date:** [May 6, 2025, 7:13pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/3 "2025-05-06T19:13:35Z")
Ah, it’s failing because it will get called with width=1. That at least makes sense now. Would it be too egregious to have `pack_bits` return a `UInt8` when given a `Bool`? But I can at least fix it on my end now: ```mojo @parameter if width == 1: count += 1 if mask else 0 else: var packed = pack_bits(mask) count += Int(pop_count(packed)) ``` ---
**Author:** ![duck\_tape](https://sea1.discourse-cdn.com/flex001/user_avatar/forum.modular.com/duck_tape/32/432_2.png) [@duck\_tape](https://forum.modular.com/u/duck_tape)\ **Post date:** [May 6, 2025, 7:15pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/4 "2025-05-06T19:15:44Z")
Maybe the best fix is just an improved error message for the specific case where `width == 1`. Once you pointed that out things made a lot more sense, I was just missing that that was even a possibility. ---
**Author:** ![sora](https://avatars.discourse-cdn.com/v4/letter/s/ad7895/32.png) [@sora](https://forum.modular.com/u/sora)\ **Post date:** [May 6, 2025, 7:30pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/5 "2025-05-06T19:30:22Z")
Two small issues: - Should we go with `scalar -> scalar` or `scalar -> scalar`? - Either way, we lose the nice invariant: `width == bitwidthof[Scalar[new_type]]()`. That said, I’m not opposed to either option if we can find a good justification. We could even add a `pack_bits` function to the `bit` module that implements one of these more specific semantics (PR welcome). P.S. You could also write `count += Int(mask)`. ---
**Author:** ![sora](https://avatars.discourse-cdn.com/v4/letter/s/ad7895/32.png) [@sora](https://forum.modular.com/u/sora)\ **Post date:** [May 6, 2025, 7:32pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/6 "2025-05-06T19:32:40Z")
Indeed we can probably do better with `constrained` (`*extra: StaticString` for error messages) these days. Are you willing to make a PR? ---
**Author:** ![duck\_tape](https://sea1.discourse-cdn.com/flex001/user_avatar/forum.modular.com/duck_tape/32/432_2.png) [@duck\_tape](https://forum.modular.com/u/duck_tape)\ **Post date:** [May 6, 2025, 7:52pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/7 "2025-05-06T19:52:49Z")
Totally up for making that PR! Added to my Mojo stdlib todo list, might be a few days. I agree that the `width == bitwidthof` invariant is nice to maintain… but I can see wanting everything to work with scalar too. I’ll make a PR for the improved error message since I think that really covers the biggest need, and pushes the handling of width=1 to the caller at least as a starting point. Good point on `Int(mask)` 👍 ---
**Author:** ![adakkak](https://avatars.discourse-cdn.com/v4/letter/a/b19c9b/32.png) [@adakkak](https://forum.modular.com/u/adakkak)\ **Post date:** [May 7, 2025, 3:08am UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/8 "2025-05-07T03:08:13Z")
We recently added the bitset type [BitSet | Modular](https://docs.modular.com/mojo/stdlib/collections/bitset/BitSet# __init__ ) so you could write the code as `len(Bitset(some_lhs == some_rhs))` You can see the implementation at [modular/mojo/stdlib/src/collections/bitset.mojo at main · modular/modular · GitHub](https://github.com/modular/modular/blob/main/mojo/stdlib/src/collections/bitset.mojo#L118) ---
**Author:** ![duck\_tape](https://sea1.discourse-cdn.com/flex001/user_avatar/forum.modular.com/duck_tape/32/432_2.png) [@duck\_tape](https://forum.modular.com/u/duck_tape)\ **Post date:** [May 7, 2025, 12:23pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/9 "2025-05-07T12:23:12Z")
Interesting, I’ll give that a try too and maybe bench the two against each other. ---
**Author:** ![sora](https://avatars.discourse-cdn.com/v4/letter/s/ad7895/32.png) [@sora](https://forum.modular.com/u/sora)\ **Post date:** [May 7, 2025, 2:22pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/10 "2025-05-07T14:22:54Z")
Maybe we should use `pack_bits` in the `BitSet` from `SIMD[bool, _]`constructor. Also, you might want to use `_words.unsafe_get(i)` throughout the codebase, since `InlineArray. __getitem__ ` has bounds checking and assertions enabled by default. ---
**Author:** ![duck\_tape](https://sea1.discourse-cdn.com/flex001/user_avatar/forum.modular.com/duck_tape/32/432_2.png) [@duck\_tape](https://forum.modular.com/u/duck_tape)\ **Post date:** [May 7, 2025, 2:51pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/11 "2025-05-07T14:51:32Z")
Some benchmarks with BitSet. [bench\_bitset.mojo · GitHub](https://gist.github.com/sstadick/09b4cbfeb42c0248d89bbe399f2cb454?permalink_comment_id=5569601#gistcomment-5569601) I was already working on some benchmarks for different methods so I included BitSet in the matrix. It’s just counting the GC content (in a simplistic way) in the whole genome. See comment on `main` for prepping the input data. * * * | name | met (ms) | iters | GB/s | | --- | --- | --- | --- | | Manual SIMD, width 32 | 171.47048266666664 | 6 | 18.71625981970759 | | Vectorized, width 32 | 169.0004497142857 | 7 | 18.989808076994233 | | BitSet, width 32 | 46201.248465 | 2 | 0.06946319010039766 | | Manual SIMD, width 64 | 159.27863557142857 | 7 | 20.148879939147864 | | Vectorized, width 64 | 158.84849485714287 | 7 | 20.203440441070633 | | BitSet, width 64 | 47842.169988 | 2 | 0.06708069692083299 | | Naive | 5174.215648 | 2 | 0.6202459122940677 | * * * ---
**Author:** ![system](https://us1.discourse-cdn.com/flex001/uploads/modular/original/1X/2751e0fbdc595a99718b216730957e9db4448cfd.jpeg) [@system](https://forum.modular.com/u/system)\ **Post date:** [November 3, 2025, 2:51pm UTC](https://forum.modular.com/t/pack-bits-inside-of-vectorize-fails-constraint-check/1384/12 "2025-11-03T14:51:42Z")
This topic was automatically closed 180 days after the last reply. New replies are no longer allowed.