01-03-2014, 07:28 PM
mkay. i can't tell how that grayscale improves but well... and...
the fuq. that text description before is already boiled down version to your shader's sharpen function. but right it might cost a lil much. time and such. randomizing sample locations aren't predictable by the shader compiler/optimizer. doesn't schedule cache reads well. it's atomic values waiting in the gpu. it sux. but... i thought you do the optimizing. like... use fixed scale patterns (fixed and prediactable shader code) to enlarge the scan region. use half pixel locations with bilinear filter, you effectively sample 4 pixels with 1 shot - is faster then 4 sample reads. you accumulate the luma in your function anyway. one can do that 9-tap with 4 samples (memory loads). search for it. and do the math.
the fuq. that text description before is already boiled down version to your shader's sharpen function. but right it might cost a lil much. time and such. randomizing sample locations aren't predictable by the shader compiler/optimizer. doesn't schedule cache reads well. it's atomic values waiting in the gpu. it sux. but... i thought you do the optimizing. like... use fixed scale patterns (fixed and prediactable shader code) to enlarge the scan region. use half pixel locations with bilinear filter, you effectively sample 4 pixels with 1 shot - is faster then 4 sample reads. you accumulate the luma in your function anyway. one can do that 9-tap with 4 samples (memory loads). search for it. and do the math.