Re: LLVM-pipe: most powerful instructions set ;)
I see the point. I was thinking about using a small, simplified, reliable fixed set of basic GPU functions which are always supported, and call them from time to time to speed up some mass operations such as texture scaling (instead of MMXing big textures on CPU, call existing OpenGL driver and free the CPU for different operations while GPU scales them). But it DOES involve lots of data exchange, at least unless embedded GPU core can process them in the common memory space). So I see the point, the data exchange can "eat" all this profit. On 22.11.2022 14:06, Jose Fonseca wrote: Trying to combine two OpenGL implementations it's technically difficult. Furthermore one might thing combining two implementations gives the best of both, but easily gives the worst of both worlds, because there's overhead moving data between them. Image the worst case scenario: draw on driver A, copy data to driver B, draw on driver B, copy data to driver A. One might devise a heuristic to avoid switching so often, but the end result would end being the bulk of the rendering beind done on the most capable drvier (llvmpipe in this case), so not much better than always using llvmpipe. This is why technologies like NVIDIA SLI which divide the work across multiple identical GPUs tend to split the work in a granular fashion, like whole frames (GPU 1 draws frame 2*N, GPU 2 draws frame 2*N + 1). Furthermore, mixing different implementations easily leads to artifacts, like depth fighting, overlapped geometry or gaps between, because 3D APIs specification have somewhat lax rules about rasterization, and float conversion. Honestly, it's more effective to buy a more modern device. Jose *From:* mesa-dev on behalf of t...@profit-grand.ru *Sent:* Tuesday, November 15, 2022 12:48 *To:* mesa-dev@lists.freedesktop.org *Subject:* LLVM-pipe: most powerful instructions set ;) !! External Email First, thanks for saving my day. Software renderer works way faster than no renderer at all :) But there are an "instruction set" which is way faster than the MMX or SSE sets: it's an existing OpenGL 2.0 driver which is not sufficient in most cases, because at least 3.0 is required nowadays. Of course, it's not literally an instruction set, but a source of additional computing power for basic operations (assuming advanced operations are not supported by the GPU, making LLVM-pipe necessary for such systems). So, the actual question is: Can LLVM-pipe somehow support (today or in future versions) a "bridge mode", relying on a different OpenGL driver for basic operations (supported by some ancient GPU) and closing a gap between that driver and the required OpenGL version it's usual way, by the state-of-art software renderer? It's also somewhat related to my OGLOED driver idea (see my Github), the difference is "bridge mode" is local, OGLOED is not. But there are some diagrams which can help understanding my question :) !! External Email: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the sender.
Re: LLVM-pipe: most powerful instructions set ;)
Trying to combine two OpenGL implementations it's technically difficult. Furthermore one might thing combining two implementations gives the best of both, but easily gives the worst of both worlds, because there's overhead moving data between them. Image the worst case scenario: draw on driver A, copy data to driver B, draw on driver B, copy data to driver A. One might devise a heuristic to avoid switching so often, but the end result would end being the bulk of the rendering beind done on the most capable drvier (llvmpipe in this case), so not much better than always using llvmpipe. This is why technologies like NVIDIA SLI which divide the work across multiple identical GPUs tend to split the work in a granular fashion, like whole frames (GPU 1 draws frame 2*N, GPU 2 draws frame 2*N + 1). Furthermore, mixing different implementations easily leads to artifacts, like depth fighting, overlapped geometry or gaps between, because 3D APIs specification have somewhat lax rules about rasterization, and float conversion. Honestly, it's more effective to buy a more modern device. Jose From: mesa-dev on behalf of t...@profit-grand.ru Sent: Tuesday, November 15, 2022 12:48 To: mesa-dev@lists.freedesktop.org Subject: LLVM-pipe: most powerful instructions set ;) !! External Email First, thanks for saving my day. Software renderer works way faster than no renderer at all :) But there are an "instruction set" which is way faster than the MMX or SSE sets: it's an existing OpenGL 2.0 driver which is not sufficient in most cases, because at least 3.0 is required nowadays. Of course, it's not literally an instruction set, but a source of additional computing power for basic operations (assuming advanced operations are not supported by the GPU, making LLVM-pipe necessary for such systems). So, the actual question is: Can LLVM-pipe somehow support (today or in future versions) a "bridge mode", relying on a different OpenGL driver for basic operations (supported by some ancient GPU) and closing a gap between that driver and the required OpenGL version it's usual way, by the state-of-art software renderer? It's also somewhat related to my OGLOED driver idea (see my Github), the difference is "bridge mode" is local, OGLOED is not. But there are some diagrams which can help understanding my question :) !! External Email: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the sender.
LLVM-pipe: most powerful instructions set ;)
First, thanks for saving my day. Software renderer works way faster than no renderer at all :) But there are an "instruction set" which is way faster than the MMX or SSE sets: it's an existing OpenGL 2.0 driver which is not sufficient in most cases, because at least 3.0 is required nowadays. Of course, it's not literally an instruction set, but a source of additional computing power for basic operations (assuming advanced operations are not supported by the GPU, making LLVM-pipe necessary for such systems). So, the actual question is: Can LLVM-pipe somehow support (today or in future versions) a "bridge mode", relying on a different OpenGL driver for basic operations (supported by some ancient GPU) and closing a gap between that driver and the required OpenGL version it's usual way, by the state-of-art software renderer? It's also somewhat related to my OGLOED driver idea (see my Github), the difference is "bridge mode" is local, OGLOED is not. But there are some diagrams which can help understanding my question :)