Implement relaxed SIMD dot product instructions #4586
Conversation
| template<size_t Lanes, | ||
| size_t Factor, | ||
| LaneArray<Lanes * Factor> (Literal::*IntoLanes)() const> | ||
| static Literal dot(const Literal& left, const Literal& right) { |
There was a problem hiding this comment.
Adding a brief comment about what this does can be helpful, like, "lane-wise multiply signed n-bit integers in the two input vectors and add adjacent pairs of the full (n*2)-bit results". I wasn't aware that adjacent pairs are added in the result before checking the instruction description myself.
Apparently it ignores the high bit on one of the inputs. Actually, now that I write that out, I realize that this is not what I implemented in the interpreter. Will fix that. |
|
Actually, the text on the proposal issue is confusing, so I think the code is already correct. The output of the instruction is deterministic if the input is up to 7 bits, but otherwise it may have one of two behaviors. The code in this PR implements the same behavior as ARM, so I don't think it needs to change. |
As proposed in WebAssembly/relaxed-simd#52.