How many addiction memory will multisample needs?

I don't exactly know how the multisampling works, but a basic way would be for 4X to actually have a buffer twice the width and height of the screen, 9X would have a buffer 3 times as high and wide. In these cases you would take up 4 and 9 times as much memory as the original screen buffer.

for 1024x1024x32 that's 4Mb to start with - meaning you'd be up to 16/36Mb with 4X/9X AA.

Can somebody confirm/refute my reasoning?

Quote:

Original post by d000hg I don't exactly know how the multisampling works, but a basic way would be for 4X to actually have a buffer twice the width and height of the screen, 9X would have a buffer 3 times as high and wide. In these cases you would take up 4 and 9 times as much memory as the original screen buffer. for 1024x1024x32 that's 4Mb to start with - meaning you'd be up to 16/36Mb with 4X/9X AA. Can somebody confirm/refute my reasoning?

Multisampling is a filtering technique, therefore no additional memory will be required.
The rendering performance will go down but the memory requirements stay the same.
The buffer will be 4Mib big.

Seems correct. Additional memory could be used for padding, if the surface dimensions would be non-powers of 2.

EDIT: True supersampling stores all the sub-pixels separately, so you can write them separately, thus taking the more memory the more samples there is per pixel.

Quote:

Original post by d000hg I don't exactly know how the multisampling works, but a basic way would be for 4X to actually have a buffer twice the width and height of the screen, 9X would have a buffer 3 times as high and wide. In these cases you would take up 4 and 9 times as much memory as the original screen buffer. for 1024x1024x32 that's 4Mb to start with - meaning you'd be up to 16/36Mb with 4X/9X AA. Can somebody confirm/refute my reasoning?

AFAIK what you describe is referred to as 'super-sampling'.
Indeed with supersampling, the buffer size depends on the multiplier. Thus with 4x supersampling you will end up with a 4096x4096 buffer for a 1024x1024 target.

Quote:

Original post by Anonymous Poster Quote: Original post by darookie Quote: Original post by d000hg I don't exactly know how the multisampling works, but a basic way would be for 4X to actually have a buffer twice the width and height of the screen, 9X would have a buffer 3 times as high and wide. In these cases you would take up 4 and 9 times as much memory as the original screen buffer. for 1024x1024x32 that's 4Mb to start with - meaning you'd be up to 16/36Mb with 4X/9X AA. Can somebody confirm/refute my reasoning? Multisampling is a filtering technique, therefore no additional memory will be required. The rendering performance will go down but the memory requirements stay the same. The buffer will be 4Mib big. I think multisample should take more memory. I have read the gdc2002_multisample document(which is opengl, btw). It says that: It(Multisample) doesn¡¯t happen for free More video memory is required: Vid_mem = sizeof(Front_buffer) + sizeof(Back_buffer) + num_samples * (sizeof(Front_buffer) +sizeof(ZS_buffer)) I don't know what does sizeof(Front_buffer) means. If this means the whole front buffer size, then multisample will need A LOT OF MEMORY!! I don't think this is reasonable so I wonder how multisample works in real hardware?

You are right about that! Even though hardware performs some trick to reduce the memory foot-print (e.g. num_samples / x * (sizeof(front_buffer) + sizeof(ZS_buffer)), the formula you provided seems correct. I supposed the hardware does that on a per-pixel basis, but I was wrong.
So the main difference between super-sampling and multi-sampling is that the former calculates a factor * buffer_size image, while the latter renders factor * buffer_size images (and applies some tricks to reduce memory bandwidth).

My apologies for the misinformation I provided.

Pat.