(In reply to Cameron Cawley from comment #0)
> however I haven't done
> extensive amounts of testing for all of the new paths.

If you can do some more testing, I'd wait, say, a week or
something.

Hey, if you look at the bug 4503
There is a testcase, an image, and a script to compare.

the testcase load a multi-color image, convert to all format and print some CRC for each step.
if you run it before the patch to have a reference and after: the compare script will tell if it breaks anything and the speed improvements.

(I think there was a few fix to add like ... 
https://hg.libsdl.org/SDL/rev/233b2a61cad1
https://hg.libsdl.org/SDL/rev/dff36de37426 )

maybe the format of SDL 1.2 are different and testcase should be adapted

Has the patch received more testing?

Created attachment 4515 [details]
test-case for SDL-1.2

Hello, I tried the test-case from the previous ticket, and there are much more differences that what I remembered of SDL-1.2.

So I ported it to SDL-1.2.
At least, it compiles and runs.

gcc fast_bench.c -Ibuild/include/SDL/ -Lbuild/lib -lSDL -lz -lcrypto


There are TODOs open:

- Need to use IMG_Load (instead of loadBMP) for a .png file, so that the image has an alpha channel for testing.

- Is there blending in SDL-1.2 ?

- Currently it has lots of pixel_format:
  It takes all SDL2 pixel formats and copies them as SDL1 pixel formats.
  Not sure if that are all relevant. ( BPP1 and BPP4 commented out )

It needs to be tested with the SDL-1.2 + patch

Created attachment 4516 [details]
compare script

The compare script hasn't changed much. I just updated to python3
(if needed:
- keys = db2.keys()
+ keys = list(db2.keys()) #python3
)

Usage: 

./testcase.out test.png  > ref1.txt
apply the patch
./testcase.out test.png  > ref2.txt

python compare_blitlog.py ref1.txt ref2.txt

The image must be relevant. big enough to mesure the image processing time.
Multi color + alpha so that we have all corner cases.
various size odd/even, pitch padding.

After fixing a crash: https://hg.libsdl.org/SDL/rev/0bbdbc7b3ff2 

I've just tested the patch, and all pixel format ARGB2101010 are failing. not sure if this is supported, but at there are not "unsupported()" neither ?

Otherwise, there are improvements as expected: 
like BGRX8888 ->       RGB24 with goes from 15 ms to 3 ms. 
and many others.

Created attachment 4517 [details]
test-case for SDL-1.2

Update the test-case to remove ARGB2101010 and use IMG_Load from SDL_image

Created attachment 4518 [details]
compare script

Update the script to print the "faster/slower" indication

Created attachment 4519 [details]
results file

This my result file if you want to look at.
It shows where there are blit improvements.

(NB: there are a few irrelevant suspicious/slower conversion because blits at two fast in those cases.)

Created attachment 4521 [details]
test-case for SDL-1.2

Update the test-case to re-add the SDL_GetPerformanceCounter + some padding/memcheck.

Created attachment 4522 [details]
results file

This the results file which show the the speed improvements on 50 conversions.
using the image  "IMG_2476_ColorKey_very_small.png" from bug 4503

The results file looks ok to me, 
So I have committed the patch https://hg.libsdl.org/SDL/rev/e901fc96cf25

Sylvain: Thanks for marvelous work testing this!
And thanks Cameron Cawley for the backport patch!

Closing this as fixed.

Closed as fixed.

Thanks a lot for sorting this out!