Noise in textures back
Board:
Home
Board index
Raytracing
General Development
(L) [2014/09/24] [ziu] [Noise in textures] Wayback!Hello,
I am using OpenCL which does not support mipmapped textures. So I get all this wonderful weird noise in detailed textures when viewed at some distance. I am also using horrendous bandwith to handle gigantic textures even if they are shrunk to the size of a pixel on the screen. Any suggestions? Sampling filters to use? Implementing mipmapping by myself on OpenCL 1.1? Is CUDA any better at this?
I am also considering to use texture atlases to improve texturing performance. OpenCL really needs better texturing.
Questions, questions...
(L) [2014/09/25] [Dade] [Noise in textures] Wayback!>> ziu wrote:I am using OpenCL which does not support mipmapped textures. So I get all this wonderful weird noise in detailed textures when viewed at some distance. I am also using horrendous bandwith to handle gigantic textures even if they are shrunk to the size of a pixel on the screen. Any suggestions? Sampling filters to use? Implementing mipmapping by myself on OpenCL 1.1? Is CUDA any better at this?
But what kind of rendering are you doing ? I mean, in a normal path tracer with a proper image reconstruction filter (and many samples per pixel in order to get a noise free image) I never felt the need of texture mapping filtering  [SMILEY :?:]
 >> ziu wrote:I am also considering to use texture atlases to improve texturing performance. OpenCL really needs better texturing.
In my opinion OpenCL image support is just prehistoric and useless: it looks like something done to expose a 3Dfx Voodoo graphic feature  [SMILEY :lol:]
(L) [2014/09/25] [ziu] [Noise in textures] Wayback!>> Dade wrote:But what kind of rendering are you doing ? I mean, in a normal path tracer with a proper image reconstruction filter (and many samples per pixel in order to get a noise free image) I never felt the need of texture mapping filtering   
You are correct of course. Supersampling with a filter would solve these issues. I am just doing primary rays + ambient occlusion with one primary ray per pixel. It is just that mipmapping seems a lot cheaper computationally and then there is always the question of the texture bandwidth. You are trashing the cache a lot with these large textures, especially if they are far away and a mere stamp on screen. Mipmapping would reduce the required memory bandwidth to render the scene.
(L) [2014/09/25] [ingenious] [Noise in textures] Wayback!>> Dade wrote:...with a proper image reconstruction filter (and many samples per pixel in order to get a noise free image) I never felt the need of texture mapping filtering   
Tell this to a production rendering TD and he'll have a good laugh [SMILEY :)] And it's not as much about the noise as it is about managing the frikkin memory/network bandwidth  [SMILEY :o] Try rendering a scene with 100GB of textures.
(L) [2014/09/25] [Dade] [Noise in textures] Wayback!>> ingenious wrote:Dade wrote:...with a proper image reconstruction filter (and many samples per pixel in order to get a noise free image) I never felt the need of texture mapping filtering   
Tell this to a production rendering TD and he'll have a good laugh  And it's not as much about the noise as it is about managing the frikkin memory/network bandwidth   Try rendering a scene with 100GB of textures.
GPUs have nowadays only 8 or 16GB, something like 100GB of texture maps is against the law of physics. Another clear example of the superiority of GPUs over CPUs  [SMILEY :lol:]
P.s. out of core rendering is considered black magic and forbidden too.
(L) [2014/09/25] [ingenious] [Noise in textures] Wayback!>> Dade wrote:something like 100GB of texture maps is against the law of physics.
Forget physics, it's against common sense [SMILEY :)] And yet 100GB-texture scenes are not considered too large nowadays by production standards. So yes, mip-mapping is pretty much crucial for performance.
back