Is it any faster to read array slices than just elements of an array?

[Marco Leise via Digitalmars-d-learn]

Am Wed, 08 Apr 2015 17:01:43 +0000
schrieb "ZILtoid1991" <ziltoidtheomnicent at gmail.com>:

> While I technically finished the 0.2 version of my graphics 
> engine which has a reasonable speed at low internal resolutions 
> and with only a couple of sprites, but it still gets bottlenecked 
> a lot. First I'll throw out the "top-down determination 
> algorhythm" as it requires constant memory paging (alrought it 
> made much more sense when the engine was full OO and even slower).
> 
> Instead I'll use a overwriting ("bottom-up") method. It still 
> needs constant updates and I have to remove the per sprite 
> transparency key and use a per layer key, however it requires 
> much less paging, and still have the ability of unbound layer 
> numbers and sprite count with unbound sizes.
> 
> I also came up with the idea of reading slices out from the 
> graphical elements to potentially speed up the process a bit, 
> especially as the custom bitmaps it uses are 16bit for palette 
> operations, so per pixel read operations would waste a portion of 
> memory bus. So should I write a method for the bitmap class which 
> gets a line from it? (an array slice as it contains the data in a 
> single 1D array to avoid jagged arrays on a future expansion for 
> a scaler) And can I write an array slice at a position of an 
> array? (to reduce writeback calls)

I don't get the whole picture, but an array slice is just a
pointer and a length, so it doesn't access the pixels at all.
If you know you will be going over the whole scanline, by all
means get an array slice (or a C style raw pointer) and
process it in as big chunks as possible. For example
copying with 128-bit SSE registers will be much faster than
16-bit ushorts. (SSE4 also offers a table lookup function.)
Beyond that - since it is a gfx engine - why not use OpenGL
and drop the palette lookup altogether ?

-- 
Marco