From dfba71dedef85993b46b5c5b406f91ba985c6639 Mon Sep 17 00:00:00 2001 From: Vito Caputo Date: Sun, 30 Jul 2023 23:23:44 -0700 Subject: modules/checkers: add blurb about context clones There's an outstanding issue surrounding the need for context clones, and I'd just like to write something down somewhere before it falls off my radar. Presently it's just checkers that exercises this need, so it makes sense to put it here for now, until I get around to actually taking action on the issue. --- src/modules/checkers/checkers.c | 40 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 40 insertions(+) (limited to 'src/modules/checkers') diff --git a/src/modules/checkers/checkers.c b/src/modules/checkers/checkers.c index f04ebf0..d6530eb 100644 --- a/src/modules/checkers/checkers.c +++ b/src/modules/checkers/checkers.c @@ -106,6 +106,46 @@ static til_module_context_t * checkers_create_context(const til_module_t *module /* since checkers is already threaded, create an n_cpus=1 context per-cpu */ if (til_module_create_contexts(module, stream, seed, ticks, 1, ctxt->setup->fill_module_setup, n_cpus, ctxt->fill_module_contexts) < 0) return til_module_context_free(&ctxt->til_module_context); + + /* XXX: you might be wondering: why not just use a single context, render into a tile-sized frame buffer in prepare_frame(), + * and just copy from that into the filled cells in threaded render_fragment()? it would avoid all this context clones + * sharing the same path complexity... and if the copy applied alpha/transparency, you'd still get overlay effects. + * What gives? + * + * The answer is: overlay effects go beyond simple transparencies. With fragment snapshotting, arbitrary sampling of + * anywhere within the incoming fragment's contents while writing the fragment is facilitated, so you get use cases like + * drizzle,style=map which treat the "puddle" as a sort of normal map and read in pixel colors from wherever the normal + * aims. And there's sampler overlays like voronoi and checkers, which fill arbitrary regions using the incoming color + * from an arbitrary pixel. For any of these, rendering once into an intermediate single tile-sized frame buffer would + * sample the blank tile-sized frame buffer, so they'd just be blank. It's more like what they'd need is to store the + * relative coordinates to read from per-pixel, instead of the final pixel data, in the tile-sized frame buffer, combined + * with encoding what operation to do on what was read. Instead of that complexity, I'm just cloning the contexts with + * identical seed+settings, and running them per-cpu. It may be interesting to experiment with storing source coordinates + * with color+operation information in a different "sampler" type of frame... but all these modules would need to then + * learn to store that instead of just always producing pixels. It'd be a substantial invasive change, and would constrain + * modules to work within what that sampler frame could express, whereas this context cloning lets the module authors just + * implement whatever they want for producing pixels. + * + * It might seem like a lot of trouble just for a silly checkers module, but ensuring this works properly for checkers + * removes the friction of doing the same kind of fill_module= instancing for concurrency in any future modules. + * + * TODO: a remaining problem to address surrounding context clones is the "ref" builtin, and til_stream_find_module_contexts() + * which it relies on. Imagine checkers uses fill_module=ref,path=/path/to/context/without/clones with things as-is. + * There would be n_cpus ref contexts in checkers, and each of those rendering concurrently will find the same single context + * at /path/to/context/without/clones on the stream. So they'd racily render while sharing a single context, which might + * crash - depending on what module that context belongs to. What needs to happen here is when the "ref" clones lookup + * the context @ path to use, they need a clone number to propagate down to til_stream_find_module_contexts(), which + * would have been assigned when the n_cpus ref clones were created for checkers. Then til_stream_find_module_contexts() + * can ensure the context returned has the same clone number within the found path. But there's still a need to actually + * create the clones when they don't exist. Keep in mind, the /path/to/context/without/clones will have just a + * single context, and we're going to want n_cpus of them from the same path via "ref". So the one that exists needs to be + * cloned on demand, in a race-free manner likely requiring some locking, and it will require the cloned context's module + * to perform the actual cloning in all but the simples of cases. So it looks like til_module_t.clone_context() is needed, + * and should be asserted as present when til_module_t.destroy_context() is present (and vice versa), as the presence of + * either strongly implies the context refers to external resources. And in the absence of a clone_context() method, we + * can have a default "dumb" clone that just duplicates the context with a malloc() and memcpy(). This would also require + * adding a size member to til_module_context_t, which is straightforward to add. + */ } return &ctxt->til_module_context; -- cgit v1.2.3