JuliaSIMD / JuliaSIMD/LoopVectorization.jl

How to handle user defined functions?

オープン
#55 コメント 2 件 リアクション 0 件 担当者 0 名 GitHub で見る
主要言語
Julia
スター
789
フォーク
73
PR マージ指標
30日以内にマージされた PR はありません

説明

I get the following error
```julia
ERROR: MethodError: no method matching calc_A(::VectorizationBase.SVec{16,Int16})
```
where `calc_A` is a user defined function. Do I need to implement `calc_A(::VectorizationBase.SVec{16,Int16})` myself? Why isn't there an error for the other user defined functions? My first guess was because they are inlined, but `calc_A` should be inlined, too.

Here is the code for that:
```julia
@inline function calc_A(x::Int16)
p = 15; r = 1; A = Int16(23170); C = Int16(-425)
n = 7
a = 7
x² = (x * x) >> n
rounding = one(x) << (p - a - 1)
(A + rounding + (x² * C) >> r) >> (p - a)
end
@inline function calc_B(x::Int16)
p = 14; r = 3; B = Int16(-17790); D = Int16(351)
n = 7
a = 7
x² = (x * x) >> n
rounding = one(x) << (p - a - 1)
(rounding + x * (B + (x² * D) >> r) >> n) >> (p - a)
end
@inline function get_first_bit_sign(x)
n = 7
mysign(x << (sizeof(x) * 8 - n - 1))
end
@inline function get_second_bit_sign(x)
n = 7
mysign(x << (sizeof(x) * 8 - n - 2))
end
@inline function get_quarter_angle(x)
n = 7
x & (one(x) << n - one(x)) - one(x) << (n - 1)
end
@inline mysign(x) = 2 * signbit(x) - 1

function gen_sincos!(sins, coss, phases)
@avx for i = 1:2500
first_bit_sign = get_first_bit_sign(phases[i])
second_bit_sign = get_second_bit_sign(phases[i])
quarter_angle = get_quarter_angle(phases[i])
A = calc_A(quarter_angle)
B = calc_B(quarter_angle)

coss[i] = second_bit_sign * (first_bit_sign * A + B)
sins[i] = second_bit_sign * (A - first_bit_sign * B)
end
end
sins = Vector{Int16}(undef, 2500)
coss = Vector{Int16}(undef, 2500)
phases = Vector{Int16}(undef, 2500)
gen_sincos!(sins, coss, phases)
```
**EDIT:** Alright I found the solution myself: If I write `function calc_A(x)` instead of `function calc_A(x::Int16)`, it will work. So should I omit the declaration of `x`? What if I have a different function for `x::Int32`?

コントリビューションガイド

このリポジトリのコントリビューションガイドは索引されていません

調査の方向性

Start with gen_sincos! and the @avx loop, then compare how calc_A is defined with the other user-defined functions in the example. Check how the vectorized loop dispatches calc_A for quarter_angle and verify the behavior for Int16 and Int32 inputs. Done means documenting the required method definitions or confirming the supported dispatch pattern.

索引モデルが issue の本文から書いたものです。

評価

技術スタック
julia
領域
performance
issue の種類
バグ
難易度
3/5
見積もり時間
1〜2日
活発さ
停滞
明瞭さ
おおむね明確
初心者へのやさしさ
35/100

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。