Link issues

doc/vision.md

@@ -36,7 +36,7 @@ One challenge facing piet-gpu is the lack of adequate infrastructure for portabl
 
 I strongly believe that Vulkan is emerging as a viable low-level platform for utilizing GPU compute resources. I’m also not the only one thinking along these lines. The [VkFFT] project is an impressive demonstration that a Vulkan deployment of one math-intensive algorithm can be just as performant as the CUDA version. In addition, there are early steps toward running machine learning workloads on Vulkan, particularly TensorFlow Lite.
 
-Of course, while it’s possible to run Vulkan on a pretty wide range of hardware, it doesn’t solve all portability problems. “Runs Vulkan” is not a binary, but rather a portal to a vast matrix of optional features and limits from the various combinations of hardware, drivers, and compatibility shims ([vulkan.gpuinfo.org] is an excellent resource). In particular, Apple forces the use of Metal. In theory, MoltenVk — or, more generally, the [Vulkan Portability Extension] — lets you run Vulkan code on Apple hardware, but in practice it doesn’t quite work (see #42), and there are compatibility and integration advantages to DX12 over Vulkan on Windows; older CPU generations such as Haswell and Broadwell don’t support Vulkan at all. To this end, I’ve started a portability layer (piet-gpu-hal) which should be able to run natively on these other API’s.
+Of course, while it’s possible to run Vulkan on a pretty wide range of hardware, it doesn’t solve all portability problems. “Runs Vulkan” is not a binary, but rather a portal to a vast matrix of optional features and limits from the various combinations of hardware, drivers, and compatibility shims ([vulkan.gpuinfo.org] is an excellent resource). In particular, Apple forces the use of Metal. In theory, MoltenVk — or, more generally, the [Vulkan Portability Extension] — lets you run Vulkan code on Apple hardware, but in practice it doesn’t quite work (see [#42]), and there are compatibility and integration advantages to DX12 over Vulkan on Windows; older CPU generations such as Haswell and Broadwell don’t support Vulkan at all. To this end, I’ve started a portability layer (piet-gpu-hal) which should be able to run natively on these other API’s.
 
 ### Why not wgpu?
 
@@ -100,12 +100,11 @@ Even so, I believe it’s possible to do even better. A CPU-bound renderer has b
 
 The details of what I have in mind could be a blog post in and of itself, but I’ll sketch out the highlights.
 
-Perhaps the most important quality problem is that of so-called “conflation artifacts,” the seams that happen when compositing antialiased elements (see #49). Most of the academic literature on 2D rendering on GPU addresses this question. I think it’s practical to do in the piet-gpu architecture, basically by swapping out soft-alpha compositing in the fine rasterizer with one based on supersampling. Some of the academic literature also takes the opportunity at that stage in the pipeline to apply a reconstruction filter more sophisticated than a box filter, but I am not yet convinced that the improvement is worth it, especially as physical display resolution increases.
+Perhaps the most important quality problem is that of so-called “conflation artifacts,” the seams that happen when compositing antialiased elements (see [#49]). Most of the academic literature on 2D rendering on GPU addresses this question. I think it’s practical to do in the piet-gpu architecture, basically by swapping out soft-alpha compositing in the fine rasterizer with one based on supersampling. Some of the academic literature also takes the opportunity at that stage in the pipeline to apply a reconstruction filter more sophisticated than a box filter, but I am not yet convinced that the improvement is worth it, especially as physical display resolution increases.
 
 The next major area of potential quality improvement is getting gamma right. This is a surprisingly tricky area, as a theoretically “correct” approach to gamma often yields text and hairline strokes that appear weak and spindly. Another concern is document compatibility; simply changing the gamma of the colorspace in which alpha blending happens will change the color of the result. Likely, a perfect solution to this problem will require cooperation with the application driving the renderer; if it is designed with gamma-perfect rendering in mind, there is no real problem, but otherwise it’s likely that various heuristics will need to be applied to get good results. (Note that [stem darkening] is one approach used specifically for text rendering, and among other things is a source of considerable variation between platforms.)
 
-When driving low-dpi displays (which still exist), one opportunity to improve quality is more sophisticated RGB [subpixel rendering]
-. Currently, that’s basically text-only, but could be applied to vector rendering as well, and often doesn’t survive sophisticated compositing, as an RGBA texture with a transparent background cannot represent RGB subpixel text. One solution is to do compositing with per-channel alpha, which can be done very efficiently when compositing in a compute shader, but would be a serious performance problem if intermediate texture buffers needed to be written out to global memory.
+When driving low-dpi displays (which still exist), one opportunity to improve quality is more sophisticated RGB [subpixel rendering]. Currently, that’s basically text-only, but could be applied to vector rendering as well, and often doesn’t survive sophisticated compositing, as an RGBA texture with a transparent background cannot represent RGB subpixel text. One solution is to do compositing with per-channel alpha, which can be done very efficiently when compositing in a compute shader, but would be a serious performance problem if intermediate texture buffers needed to be written out to global memory.
 
 These potential quality improvements may well provide the answer to the question, “why move to a new rendering architecture instead of incrementally improving what we’ve got now?”
 
@@ -174,3 +173,5 @@ I'm really excited to see where this goes. I think there's the potential to buil
 [blurred rounded rectangle]: https://raphlinus.github.io/graphics/2020/04/21/blurred-rounded-rects.html
 [gio]: https://gioui.org/
 [Minikin]: https://android.googlesource.com/platform/frameworks/minikin/
+[#42]: https://github.com/linebender/piet-gpu/issues/42
+[#49]: https://github.com/linebender/piet-gpu/issues/49