JuliaSIMD / JuliaSIMD/LoopVectorization.jl
Support for RGB colors
- Lenguaje dominante
- Julia
- Estrellas
- 789
- Forks
- 73
- Métricas de merge de PR
- Sin PR fusionados en 30 d
Descripción
Aside from #66, I think one of the last obstacles to using this in ImageFiltering is the handling of color array elements, for example `RGB(0.8f0, 0.3f0, 0.5f0)` (which is just a struct of 3 `Float32`s). Through TiledIteration it's possible to convert the data into a suitable format, for example `RGB{N0f8}` can be converted to `RGB{Float32}` because [`TileBuffer` takes an element type as an argument](https://github.com/JuliaArrays/TiledIteration.jl#allocating-and-managing-temporary-storage). (I note that's really old syntax, with modern Julia I should change it to `TileBuffer{T}(undef, tileaxs)`.) Consequently it will be converted before it is used; because convolution/correlation uses each element multiple times, this conversion is well worth doing if it gets us better performance and allows us to use this package. (Another benefit of copying is that we can guarantee stridedness.)
So the question is, what kind of format should I convert it to? Should we define explict support for RGB, or would it be better to convert to an `SVector{3}` or similar? I kind of favor the latter since it would presumably be more generic.
I suspect this will alter the loop unrolling/SIMD packing, once I understand what representation you'd like here.
Guía de contribución
No hay ninguna guía de contribución indexada para este repositorio
Línea de trabajo
Start with the issue discussion and the linked TiledIteration TileBuffer documentation, then compare the proposed RGB and SVector representations for convolution and correlation. A decision on the supported representation and its effect on loop unrolling and SIMD packing is needed before implementation can be scoped.
Escrito por el modelo de indexación a partir del texto del issue.
Evaluación
- Stack tecnológico
- julia
- Área
- performance
- Tipo de issue
- Nueva funcionalidad
- Dificultad
- 5/5
- Tiempo estimado
- Más de una semana
- Estado de actividad
- Estancado
- Claridad
- Necesita aclaración
- Aptitud para principiantes
- 25/100