This is the core logic for robust dynamic memory. There are changes to both shaders and the driver logic.
On the shader side, failure information is more useful and fine grained. In particular, it now reports which stage failed and how much memory would have been required to make that stage succeed.
On the driver side, there is a new RenderDriver abstraction which owns command buffers (and associated query pools) and runs the logic to retry and reallocate buffers when necessary. There's also a fairly significant rework of the logic to produce the config block, as that overlaps the robust memory.
The RenderDriver abstraction may not stay. It was done this way to minimize code disruption, but arguably it should just be combined with Renderer.
Another change: the GLSL length() method on a buffer requires additional infrastructure (at least on Metal, where it needs a binding of its own), so we now pass that in as a field in the config.
This also moves blend memory to its own buffer. This worked out well because coarse rasterization can simply report the size of the blend buffer and it can be reallocated without needing to rerun the pipeline. In the previous state, blend allocations and ptcl writes were interleaved in coarse rasterization, so a failure of the former would require rerunning coarse. This should fix#83 (finally!)
There are a few loose ends. The binaries haven't (yet) been updated (I've been testing using a hand-written test program). Gradients weren't touched so still have a fixed size allocation. And the logic to calculate the new buffer size on allocation failure could be smarter.
Closes#175
Adds a test to visualize the blend modes. Fixes a dumb bug in blend.h and also a more subtle issue where default blending is not the same as clipping, as the former needs to always push a blend group (to cause isolation) and the latter does not. This might be something we need to get back to.
This should fix the rendering, so it fairly closely resembles the Mozilla reference image. There's also a compile-time switch to disable sRGB conversion, which is (sadly) needed for compatible rendering.
Split the blend stack into register and memory segments. Do blending in registers up to that size, then spill to memory if needed.
This version may regress performance on Pixel 4, as it uses common memory for the blend stack, rather than keeping that memory read-only in fine rasterization, and using a separate buffer for blend stack. This needs investigation. It's possible we'll want to have single common memory as a config option, as it pools allocations and decreases the probability of failure.
Also a flaw in this version: there is no checking of memory overflow.
For understanding code history: this commit largely reverts #77, but there were some intervening changes to blending, and this commit also implements the split so some of the stack is in registers.
Closes#156
The blending math had two errors: first, colors were not separated for the purpose of blending (blending was wrongly applied to premultiplied values), and second, alpha was applied over-aggressively to the alpha channel.
This PR does *not* address the issue of gamma correctness. That is a complex issue and should probably be handled in the short term by disabling sRGB conversions and doing the internal math in sRGB color space rather than linear. This will degrade the quality of antialiasing but on the other hand give spec-compliant results for compositing.
We remove the plus-darker mode as its specification does not appear to be valid. The plus-lighter mode remains as it is quite useful for cross-fading effects.
Also the generated shaders were compiled on mac so the DXIL is unsigned. Those should be compiled on Windows before this PR is merged. (and we should figure out a better strategy for all that)
This patch adds radial gradients, including both the piet API and some
new methods specifically to support COLRv1, including the ability to
transform the gradient separately from the path.
We always do BeginClip/EndClip if it's a solid tile and the blend mode
is not default.
Also fix missing entry in pipeline layout (affects Vulkan but not Metal).
This patch switches to a variable size encoding of draw objects.
In addition to the CPU-side scene encoding, it changes the representation of intermediate per draw object state from the `Annotated` struct to a variable "info" encoding. In addition, the bounding boxes are moved to a separate array (for a more "structure of "arrays" approach). Data that's unchanged from the scene encoding is not copied. Rather, downstream stages can access the data from the scene buffer (reducing allocation and copying).
Prefix sums, computed in `DrawMonoid` track the offset of both scene and intermediate data. The tags for the CPU-side encoding have been split into their own stream (again a change from AoS to SoA style).
This is not necessarily the final form. There's some stuff (including at least one piet-gpu-derive type) that can be deleted. In addition, the linewidth field should probably move from the info to path-specific. Also, the 1:1 correspondence between draw object and path has not yet been broken.
Closes#152
This just runs ninja on the piet-gpu/shaders on a Windows machine, so
translated shaders match the existing pipeline.
At some point, we'll rework this to reduce friction.
* Add blend and composition mode enums to API
* Mirror these in the shaders
* Add new public blend function to PietGpuRenderContext that mirrors clip
* Plumb the modes through the pipeline from scene to kernel4
This PR reworks the clip implementation. The highlight is that clip bounding box accounting is now done on GPU rather than CPU. The clip mask is also rasterized on EndClip rather than BeginClip, which decreases memory traffic needed for the clip stack.
This is a pretty good working state, but not all cleanup has been applied. An important next step is to remove the CPU clip accounting (it is computed and encoded, but that result is not used). Another step is to remove the Annotated structure entirely.
Fixes#88. Also relevant to #119
This exposes interfaces to render glyphs into a texture atlas. The main changes are:
* Methods to plumb raw Metal GPU resources (device, texture, etc) into piet-gpu-hal objects.
* A new glyph_render API specialized to rendering glyphs. This is basically the same as just painting to a canvas, but will allow better caching (and has more direct access to fonts, bypassing the Piet font type which is underdeveloped).
* Ability to render to A8 target in addition to RGBA.
WIP, there are some rough edges, not least of which is that the image format changes are only on mac and cause compile errors elsewhere.
Make max workgroup size 256 and respect LG_WG_FACTOR.
Because the monoid scans only support a height of 2, this will reduce
the maximum scene complexity we can render. But it also increases
compatibility. Supporting larger scans is a TODO.
This is one of the stages in the new element pipeline. It's a simple
one, just a prefix sum of a couple counts, and some of it will probably
get merged with a downstream stage, but we'll do it separately for now
for convenience.
This patch also contains an update to Vulkan tools 1.2.198, which
accounts for the large diff of translated shaders.
This patch contains the core of the path stream processing, though some
integration bits are missing. The core logic is tested, though
combinations of path types, transforms, and line widths are not (yet).
Progress towards #119
There's a bit of reorganizing as well. Shader stages are made available
from piet-gpu to the test rig, config is now a proper structure
(marshaled with bytemuck).
This commit just has the transform stage, which is a simple monoid scan
of affine transforms.
Progress toward #119
The flag read needs acquire semantics. There are a number of ways that
could be expressed, but a generally portable way is to have a barrier
after. However, in the translation to Metal, that barrier needs to be in
uniform control flow. This patch does some workarounds to ensure that.
memoryBarrierBuffer is mapped to the threadgroup_barrier function in
Metal, which is a control barrier that must be executed by all threads
(or none). This change establishes that property for the two memory
barriers we have.
While here, remove ENABLE_IMAGE_INDICES completely; it was disabled in
an earlier change.
Signed-off-by: Elias Naur <mail@eliasnaur.com>
Separate out render context upload from renderer creation. Upload ramps
to GPU buffer. Encode gradients to scene description. Fix a number of
bugs in uploading and processing.
This renders gradients in a test image, but has some shortcomings. For
one, staging buffers need to be applied for a couple things (they're
just host mapped for now). Also, the interaction between sRGB and
premultiplied alpha isn't quite right. The size of the gradient ramp
buffer is fixed and should be dynamic.
And of course there's always more optimization to be done, including
making the upload of gradient ramps more incremental, and probably
hashing of the stops instead of the processed ramps.
Don't recompute the parameters from quadratic subdivision, but rather
retain them across the two phases (summing the subdivision estimate, and
generating the subdivisions). The motivation for this is that the values
were subtly different (differing by 1 or 2 least signficant bits) across
the two phases. It *might* also be faster depending on ALU/memory
relative performance.
Fixes#107
WIP. Most of the GPU-side work should be done (though it's not tested
end-to-end and it's certainly possible I missed something), but still
needs work on encoding side.
The compute shaders have a check for the succesful completion of their
preceding stage. However, consider a shader execution path like the
following:
void main()
if (mem_error != NO_ERROR) {
return;
}
...
malloc(...);
...
barrier();
...
}
and shader execution that fails to allocate memory, thereby setting
mem_error to ERR_MALLOC_FAILED in malloc before reaching the barrier. If
another shader execution then begins execution, its mem_eror check will
make it return early and not reach the barrier.
All GPU APIs require (dynamically) uniform control flow for barriers,
and the above case may lead to GPU hangs in practice.
Fix this issue by replacing the early exits with careful checks that
don't interrupt barrier control flow.
Unfortunately, it's harder to prove the soundness of the new checks, so
this change also clears dynamic memory ranges in MEM_DEBUG mode when
memory is exhausted. The result is that accessing memory after
exhaustion triggers an error.
Signed-off-by: Elias Naur <mail@eliasnaur.com>
