some optimizations

[alex-knyaz]

removed markout_sobel_patches, faster on my computer

[alex-knyaz]

50% is in grad_to_dp, 50% in memmove

[alex-knyaz]

Wanted to try fixed-point type and hit bug in clang. Wild!

[sirynka]

./run.sh 
[INFO] directory `./build/` already exists
[INFO] ./build/stb_image.o is up to date
[INFO] ./build/stb_image_write.o is up to date
[INFO] CMD: cc -Wall -Wextra -ggdb -O3 -o ./build/main main.c ./build/stb_image.o ./build/stb_image_write.o -lm
[INFO] CMD: ./build/main ./images/Broadway_tower_edit.jpg ./images/Broadway_tower_squashed.jpg
OK: generated ./images/Broadway_tower_squashed.jpg
[INFO] Resizing took 3.995774secs

[INFO] directory `./build/` already exists
[INFO] ./build/stb_image.o is up to date
[INFO] ./build/stb_image_write.o is up to date
[INFO] CMD: cc -Wall -Wextra -ggdb -O3 -fopenmp-simd -fopenmp -o ./build/main main.c ./build/stb_image.o ./build/stb_image_write.o -lm
[INFO] CMD: ./build/main ./images/Broadway_tower_edit.jpg ./images/Broadway_tower_squashed.jpg
OK: generated ./images/Broadway_tower_squashed.jpg
[INFO] Resizing took 0.673004secs

Alex image is different

83.25% faster but produces slightly different image.
Does not produce any visible artifacts)

img

[alex-knyaz]

Ye. It is also non-deterministic. You can fix this by commenting out second '#pragma omp parallel for' (the one before dp recomputation loop).

[alex-knyaz]

It is deterministic, but images, are not the same. Interesting. Investigating.