coarse.comp knows the maximum stack depth, and can pre-allocate scratch
space for kernel4.comp. Kernel4 no longer contains allocations nor
control barriers.
The invocation local blend stack is gone as well; it didn't seem to make
any difference in performance to always use global memory for pushing
and popping.
Signed-off-by: Elias Naur <mail@eliasnaur.com>
See http://ssp.impulsetrain.com/gamma-premult.html for a description of
the format.
Pre-multiplied alpha only matters for translucent objects; draw a few
such shapes in the test render.
Signed-off-by: Elias Naur <mail@eliasnaur.com>
Reclaims the space waste from splitting fill mode commands from fill
commands.
For example, a CmdStroke + CmdColor use an extra tag word compared to
the former combined CmdStroke. This change shaves off that one word.
In the future, we can pack several command tags into one tag word,
saving even more space.
Fixes#66
Signed-off-by: Elias Naur <mail@eliasnaur.com>
This change completes general support for stroked fills for clips and
images.
Annotated_size increases from 28 to 32, because of the linewidth field
added to AnnoImage. Stroked image fills are presumably rare, and if
memory pressure turns out to be a bottleneck, we could replace the
linewidth field with a separate AnnoLinewidth elements.
Updates #70
Signed-off-by: Elias Naur <mail@eliasnaur.com>
Before this change, every command (FillColor, FillImage, BeginClip)
had (or would need) stroke, (non-zero) fill and solid variants.
This change adds a command for each fill mode and their parameters,
reducing code duplication and adds support for stroked FillImage and
BeginClip as a side-effect.
The rest of the pipeline doesn't yet support Stroked FillImage and
BeginClip. That's a follow-up change.
Since each command includes a tag, this change adds an extra word for
each fill and stroke. That waste is also addressed in a follow-up.
Updates #70
Signed-off-by: Elias Naur <mail@eliasnaur.com>
Encode stroke vs fill as tag flags, thereby reducing the number of scene
elements. Encoding change only, no functional changes.
The previous Stroke and Fill commands are merged to one command,
FillColor. The encoding to annotated element is divergent, which is
fixed when annotated elements move to tag flags.
Updates #70
Signed-off-by: Elias Naur <mail@eliasnaur.com>
Commit 9afa9b86b6 added Rust support for
encoding flags into elements. This change adds support to shaders by
introducing variant tag structs:
struct VariantTag {
uint tag;
uint flags;
}
and returning them from Variant_tag functions.
It also adds a flags argument to write functions for enum variants that
include TagFlags.
No functionality changes.
Updates #70
Signed-off-by: Elias Naur <mail@eliasnaur.com>
FillImage is like Fill, except that it takes its color from one or
more image atlases.
kernel4 uses a single image for non-Vulkan hosts, and the dynamic sized array
of image descriptors on Vulkan.
A previous version of this commit used textures. I think images are a better
choice for piet-gpu, for several reasons:
- Texture sampling, in particular textureGrad, is slow on lower spec devices
such as Google Pixel. Texture sampling is particularly slow and difficult to
implement for CPU fallbacks.
- Texture sampling need more parameters, in particular the full u,v
transformation matrix, leading to a large increase in the command size. Since
all commands use the same size, that memory penalty is paid by all scenes, not
just scenes with textures.
- It is unlikely that piet-gpu will support every kind of fill for every
client, because each kind must be added to kernel4.
With FillImage, a client will prepare the image(s) in separate shader stages,
sampling and applying transformations and special effects as needed. Textures
that align with the output pixel grid can be used directly, without
pre-processing.
Note that the pre-processing step can run concurrently with the piet-gpu pipeline;
Only the last stage, kernel4, needs the images.
Pre-processing most likely uses fixed function vertex/fragment programs,
which on some GPUs may run in parallel with piet-gpu's compute programs.
While here, fix a few validation errors:
- Explicitly enable EXT_descriptor_indexing, KHR_maintenance3,
KHR_get_physical_device_properties2.
- Specify a vkDescriptorSetVariableDescriptorCountAllocateInfo for
vkAllocateDescriptorSets. Otherwise, variable image2D arrays won't work (but
sampler2D arrays do, at least on my setup).
Updates #38
Signed-off-by: Elias Naur <mail@eliasnaur.com>
As described in #62, the non-deterministic scene monoid may result in
slightly different transformations for path segments in an otherwise
closed path.
This change ensures consistent transformation across paths in three steps.
First, absolute transformations computed by the scene monoid is stored
along with path segments and annotated elements.
Second, elements.comp no longer transforms path segments. Instead, each
segment is stored untransformed along with a reference to its absolute
transformation.
Finally, path_coarse performs the transformation of path segments.
Because all segments in a path share a single transformation reference,
the inconsistency in #62 is avoided.
Fixes#62
Signed-off-by: Elias Naur <mail@eliasnaur.com>
The NVIDIA shader compiler bug that forced splitting of the state struct
into primitive types is now fixed.
Signed-off-by: Elias Naur <mail@eliasnaur.com>
Defining MEM_DEBUG in mem.h will add a size field to Alloc and enable
bounds and alignment checks for every memory read and write.
Notes:
- Deriving an Alloc from Path.tiles is unsound, but it's more trouble to
convert Path.tiles from TileRef to a variable sized Alloc.
- elements.comp note that "We should be able to use an array of structs but the
NV shader compiler doesn't seem to like it". If that's still relevant, does
the shared arrays of Allocs work?
Signed-off-by: Elias Naur <mail@eliasnaur.com>
The binning shader supports up to N_TILE bins. To efficiently cover wide or
tall viewports, convert the rigid N_TILE_X x N_TILE_Y bin layout to a variable
width_in_bins x height_in_bins layout.
Signed-off-by: Elias Naur <mail@eliasnaur.com>
If WIDTH_IN_TILES or HEIGHT_IN_TILES are not divisible by N_TILE_X or N_TILE_Y
respectively, the previously unconditional Cmd_End_write would write out of
bounds.
Signed-off-by: Elias Naur <mail@eliasnaur.com>
Merge all static and dynamic buffers to just one, "memory". Add a malloc
function for dynamic allocations.
Unify static allocation offsets into a "config" buffer containing scene setup
(number of paths, number of path segments), as well as the memory offsets of
the static allocations.
Finally, set an overflow flag when an allocation fail, and make sure to exit
shader execution as soon as that triggers. Add checks before beginning
execution in case the client wants to run two or more shaders before checking
the flag.
The "state" buffer is left alone because it needs zero'ing and because it is
accessed with the "volatile" keyword.
Fixes#40
Signed-off-by: Elias Naur <mail@eliasnaur.com>
The previous attempt to fix inconsistent intersections because of floating
point inaccuracy[0] missed two cases.
The first case is that for top intersections with the very first row would fail
the test
tag == PathSeg_FillCubic && y > y0 && xbackdrop < bbox.z
In particular, y is not larger than y0 when y0 has been clipped to 0.
Fix that by re-introducing the min(p0.y, p1.y) < tile_y0 check that does work
and is just as consistent. Add similar check, min(p0.x, p1.x) < tile_x0, for
deciding when to clip the segment to the left edge (but keep consistent xray check
for deciding left edge *intersections*).
The second case is that the tracking left intersections in the [xray, next_xray]
range of tiles may fail when next_xray is forced to last_xray, the final xray value.
Fix that case by computing next_xray explicitly, before looping over the
x tiles. The code is now much simpler.
Finally, ensure that xx0 and xx1 doesn't overflow the allocated number of tiles
by clamping them *after* setting them. Adjust xx0 to include xray, just as xx1
is adjusted; I haven't seen corruption without it, but it's not obvious xx0
always includes xray.
While here, replace a "+=" on a guaranteed zero value to just "=".
Updates #23
[0] 29cfb8b63e
Signed-off-by: Elias Naur <mail@eliasnaur.com>
The state header is only one word (flags), not two.
Move the partition atomic counter to a separate field instead of state[0],
simplifying state offset calculations.
Signed-off-by: Elias Naur <mail@eliasnaur.com>
The finite precision of floating point computations can lead the coarse
renderer into inconsistent tile intersections, which implies impossible line
segments such as lines with gaps or double intersections. The winding number
algorithm is sensitive to these errors which show up as incorrectly filled
paths.
This change forces all intersections to be consistent.
First, the floating point top edge intersection test is removed; top edge intersections are
completely determined by left edge intersections.
Then, left edge intersections are inserted from the tile with the last top edge
intersection. The next top edge is then fixed to be the last tile with a left
edge intersection.
More details in the patch comments.
Fixes#23
Signed-off-by: Elias Naur <mail@eliasnaur.com>
Eliminates the precision loss of the subtraction in the sign(end.x - start.x)
expression in kernel4. That's important for the next change that avoids
inconsistent line intersections in path_coarse.
Updates #23
Signed-off-by: Elias Naur <mail@eliasnaur.com>
Provide images to fine rasterization kernel as readonly textures with a
sampler, rather than storage images. That lets us use the GPU's hardware
for sampling, which should be considerably more efficient.
There are a bunch of parameters that are hardcoded, but it does seem to
work.
This patch passes a dynamically sized array of textures to the fine
rasterizer.
A bunch of the low level Vulkan stuff is done, but only enough of the
shaders and encoders to do minimal testing. We'll want to switch from
storage images to sampled images, track the actual array of textures
during encoding, use that to build the descriptor set (which will need
to be more dynamic), and of course run image elements through the
pipeline.
Progress towards #38
We keep a small window of the clip stack in registers in the fine
rasterization kernel, and when that window is exceeded, spill to global
memory, so the clip stack can be unbounded.
I realized there's a problem with encoding clip bboxes relative to the
current transform (see #36 for a more detailed explanation), so this is
changing it to absolute bboxes.
This more or less gets clips working. There are optimization
opportunities (all-clear and all-opaque mask tiles), and it doesn't deal
with overflow of the blend stack, but it seems to basically work.
Actually handle transforms in RenderCtx (was implemented in renderer but
not actually plumbed through). This also requires maintaining a state
stack, which will also be required for clipping.
This PR also starts work on encoding clipping, including tracking
bounding boxes.
WIP, none of this is tested yet.
The Vulkan and OpenGL specifications offer only weak forward progress guarantees, and
in practice several mobile devices fail to complete the decoupled lookback
spinloop without mitigation.
This patch implements Raph's suggestion from the "Forward Progress"
section from
https://raphlinus.github.io/gpu/2020/04/30/prefix-sum.html
Signed-off-by: Elias Naur <mail@eliasnaur.com>
Expand the the final kernel4 stage to maintain a per-pixel mask.
Introduce two new path elements, FillMask and FillMaskInv, to fill
the mask. FillMask acts like Fill, while FillMaskInv fills the area
outside the path.
SVG clipPaths is then representable by a FillMaskInv(0.0) for every nested
path, preceded by a FillMask(1.0) to clear the mask.
The bounding box for FillMaskInv elements is the entire screen; tightening of
the bounding box is left for future work. Note that a fullscreen bounding
box is not hopelessly inefficient because completely filling a tile with
a mask is just a single CmdSolidMask per tile.
Fixes#30
Signed-off-by: Elias Naur <mail@eliasnaur.com>
Both the Vulkan and OpenGL ES spec allow implementations to limit workgroups to
128 threads. Add a LG_WG_FACTOR setting for easy switching between 128 and 256
threads, with 256 being kept as the default setting.
Manually tested that LG_WG_FACTOR = 0 (128 threads) works as expected.
Signed-off-by: Elias Naur <mail@eliasnaur.com>
The transformation determinant is signed, but we're only interested in
the absolute scale for transforming linewidths.
Signed-off-by: Elias Naur <mail@eliasnaur.com>
This is a bit of a revert of the load-balanced ("more parallel") coarse
path rasterizer, but includes fills and also uses atomicExchange.
I'm doing it this way because it should be considerably easier to do
flattening in this structure, even though there will be some performance
regression.
Have a more-parallel read of the tile structures based on bbox coverage,
and only set the bit when the tile isn't empty.
This is a speedup, but there is some duplicated work and it is possible
to improve it further.
Path segments are unsorted, but other elements are using the same
sort-middle approach as before.
This is a checkpoint. At this point, there are unoptimized versions
of tile init and coarse path raster, but it isn't wired up into a
working pipeline. Also observing about a 3x performance regression in
element processing, which needs to be investigated.
In kernel 4, compute a chunk of pixels rather than just one per thread.
This is a dramatic speedup.
(This commit cherry-picked from another working branch)
Another speedup might be to special-case when the number of chunks in a
stroke or fill command is 1, then the segment header doesn't need
allocation and memory traffic is reduced. But right now we'll avoid the
complexity.
Coarse rasterization wasn't entirely taking line width into account.
Also fix swizzle in matrix (not yet used). And fix missing End command
in ptcl output (hasn't been a problem because buffer was cleared).
Trying to fit it into the fancy monad doesn't really work, so use a
more straightforward approach to compute it from the aggregate.
Also add yEdge logic (basically copying piet-metal). With a fix to
ELEMENT_BINNING_RATIO (which I had simply gotten wrong), the example
renders almost correctly, with small bounding box artifacts.
Write the right_edge to the binning output.
More work on encoding the fill/stroke distinction and plumbing that
through the pipeline. This is a bit unsatisfying because of the code
duplication; having an extra fill/stroke bool might be better, but I
want to avoid making the structs bigger (this could be solved by
better packing in the struct encoding).
Fills are plumbed through to the last stage. Backdrop is WIP.
This should get the "right_edge" value for each segment plumbed through
to the binning phase. It also needs to be plumbed to coarse raster and
wired up there.
Also considering WIP because none of this logic has been tested yet.
Compute tile coverage of segments using optimized algorithm. This
algorithm does a bit of setup, then uses an efficient formula to
compute the span per scan-line.
