alex/vello
1
0
Fork 0
mirror of https://github.com/italicsjenga/vello.git synced 2024-10-18 07:21:30 +11:00
Code Issues Packages Projects Releases Wiki Activity
vello/piet-gpu/shader/coarse.comp

467 lines
19 KiB
GLSL
Raw Normal View History

all: add SPDX license headers Fixes #53 Signed-off-by: Elias Naur <mail@eliasnaur.com>
2020-12-12 01:01:48 +11:00
// SPDX-License-Identifier: Apache-2.0 OR MIT OR Unlicense
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
// The coarse rasterizer stage of the pipeline.
Shader documentation and a slight cleanup
2020-06-28 23:37:27 +10:00
//
// As input we have the ordered partitions of paths from the binning phase and
// the annotated tile list of segments and backdrop per path.
//
// Each workgroup operating on one bin by stream compacting
// the elements corresponding to the bin.
//
shaders: Add .clang-format and reformat Helps keeping the code tidy. Style is chosen to minimize diff, but contains a slight bit of personal taste.
2022-01-30 18:23:18 +11:00
// As output we have an ordered command stream per tile. Every tile from a path (backdrop + segment list) will be
// encoded.
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
#version 450
#extension GL_GOOGLE_include_directive : enable
unify GPU memory management 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>
2020-12-12 04:30:20 +11:00
#include "mem.h"
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
#include "setup.h"
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
layout(local_size_x = N_TILE, local_size_y = 1) in;
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
layout(binding = 1) readonly buffer ConfigBuf {
unify GPU memory management 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>
2020-12-12 04:30:20 +11:00
Config conf;
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
};
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
layout(binding = 2) readonly buffer SceneBuf {
uint[] scene;
};
#include "drawtag.h"
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
#include "bins.h"
Continue building out pipeline Plumbs the new tiling scheme to k4. This works (stroke only) but still has some performance issues.
2020-06-04 02:28:43 +10:00
#include "tile.h"
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
#include "ptcl.h"
Fix missing blend/clip logic 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).
2022-03-17 08:40:58 +11:00
#include "blend.h"
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
shader: introduce a crude setting for adjusting the maximum workgroup size 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>
2020-09-13 20:58:47 +10:00
#define LG_N_PART_READ (7 + LG_WG_FACTOR)
Parallel merge The fancy stuff :)
2020-05-31 08:37:34 +10:00
#define N_PART_READ (1 << LG_N_PART_READ)
Simplify ringbuf We don't really need a ring buffer, as we only read what we're actually going to process.
2020-05-31 14:12:55 +10:00
shared uint sh_elements[N_TILE];
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
Parallel merge The fancy stuff :)
2020-05-31 08:37:34 +10:00
// Number of elements in the partition; prefix sum.
shared uint sh_part_count[N_PART_READ];
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
shared Alloc sh_part_elements[N_PART_READ];
Parallel merge The fancy stuff :)
2020-05-31 08:37:34 +10:00
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
shared uint sh_bitmaps[N_SLICE][N_TILE];
Start work on parallel segment output Output of segments is in parallel. Getting closer, some problems with chaining but mostly correct.
2020-05-23 07:18:39 +10:00
Filter sparse tiles 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.
2020-06-04 10:55:42 +10:00
shared uint sh_tile_count[N_TILE];
// The width of the tile rect for the element, intersected with this bin
shared uint sh_tile_width[N_TILE];
shared uint sh_tile_x0[N_TILE];
shared uint sh_tile_y0[N_TILE];
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
Faster coarse raster Store a lot more tile context in shared memory and do the work from that.
2020-06-05 03:39:08 +10:00
// These are set up so base + tile_y * stride + tile_x points to a Tile.
shared uint sh_tile_base[N_TILE];
shared uint sh_tile_stride[N_TILE];
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
#ifdef MEM_DEBUG
// Store allocs only when MEM_DEBUG to save shared memory traffic.
shared Alloc sh_tile_alloc[N_TILE];
void write_tile_alloc(uint el_ix, Alloc a) {
sh_tile_alloc[el_ix] = a;
}
avoid non-uniform barrier control flow when exhausting memory 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>
2021-04-12 22:41:03 +10:00
Alloc read_tile_alloc(uint el_ix, bool mem_ok) {
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
return sh_tile_alloc[el_ix];
}
#else
void write_tile_alloc(uint el_ix, Alloc a) {
// No-op
}
avoid non-uniform barrier control flow when exhausting memory 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>
2021-04-12 22:41:03 +10:00
Alloc read_tile_alloc(uint el_ix, bool mem_ok) {
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
// All memory.
shaders: Add .clang-format and reformat Helps keeping the code tidy. Style is chosen to minimize diff, but contains a slight bit of personal taste.
2022-01-30 18:23:18 +11:00
return new_alloc(0, memory.length() * 4, mem_ok);
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
}
#endif
coarse,kernel4: make stroke, (non-zero) fill, solid separate commands 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>
2021-03-18 22:47:14 +11:00
// The maximum number of commands per annotated element.
#define ANNO_COMMANDS 2
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
// Perhaps cmd_alloc should be a global? This is a style question.
bool alloc_cmd(inout Alloc cmd_alloc, inout CmdRef cmd_ref, inout uint cmd_limit) {
unify GPU memory management 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>
2020-12-12 04:30:20 +11:00
if (cmd_ref.offset < cmd_limit) {
return true;
Start writing tiles This is the first checkpoint where it actually runs a pipeline end to end, though it's far from accurate.
2020-05-16 05:28:29 +10:00
}
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
MallocResult new_cmd = malloc(PTCL_INITIAL_ALLOC);
unify GPU memory management 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>
2020-12-12 04:30:20 +11:00
if (new_cmd.failed) {
return false;
}
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
CmdJump jump = CmdJump(new_cmd.alloc.offset);
Cmd_Jump_write(cmd_alloc, cmd_ref, jump);
cmd_alloc = new_cmd.alloc;
cmd_ref = CmdRef(cmd_alloc.offset);
coarse,kernel4: make stroke, (non-zero) fill, solid separate commands 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>
2021-03-18 22:47:14 +11:00
// Reserve space for the maximum number of commands and a potential jump.
cmd_limit = cmd_alloc.offset + PTCL_INITIAL_ALLOC - (ANNO_COMMANDS + 1) * Cmd_size;
unify GPU memory management 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>
2020-12-12 04:30:20 +11:00
return true;
Start writing tiles This is the first checkpoint where it actually runs a pipeline end to end, though it's far from accurate.
2020-05-16 05:28:29 +10:00
}
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
void write_fill(Alloc alloc, inout CmdRef cmd_ref, Tile tile, float linewidth) {
if (linewidth < 0.0) {
coarse.comp: extract area commands into function No functional changes. Signed-off-by: Elias Naur <mail@eliasnaur.com>
2021-03-31 04:56:09 +11:00
if (tile.tile.offset != 0) {
CmdFill cmd_fill = CmdFill(tile.tile.offset, tile.backdrop);
Cmd_Fill_write(alloc, cmd_ref, cmd_fill);
cmd_ref.offset += 4 + CmdFill_size;
} else {
Cmd_Solid_write(alloc, cmd_ref);
cmd_ref.offset += 4;
}
} else {
CmdStroke cmd_stroke = CmdStroke(tile.tile.offset, 0.5 * linewidth);
Cmd_Stroke_write(alloc, cmd_ref, cmd_stroke);
cmd_ref.offset += 4 + CmdStroke_size;
}
}
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
void main() {
// Could use either linear or 2d layouts for both dispatch and
// invocations within the workgroup. We'll use variables to abstract.
shaders: Add .clang-format and reformat Helps keeping the code tidy. Style is chosen to minimize diff, but contains a slight bit of personal taste.
2022-01-30 18:23:18 +11:00
uint width_in_bins = (conf.width_in_tiles + N_TILE_X - 1) / N_TILE_X;
coarse,binning: organize bins into width_in_bins x height_in_bins 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>
2020-12-24 21:53:17 +11:00
uint bin_ix = width_in_bins * gl_WorkGroupID.y + gl_WorkGroupID.x;
Merge one segment at a time No parallelism yet, but seems to improve performance.
2020-05-31 01:35:26 +10:00
uint partition_ix = 0;
unify GPU memory management 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>
2020-12-12 04:30:20 +11:00
uint n_partitions = (conf.n_elements + N_TILE - 1) / N_TILE;
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
uint th_ix = gl_LocalInvocationID.x;
Start writing tiles This is the first checkpoint where it actually runs a pipeline end to end, though it's far from accurate.
2020-05-16 05:28:29 +10:00
Faster coarse raster Store a lot more tile context in shared memory and do the work from that.
2020-06-05 03:39:08 +10:00
// Coordinates of top left of bin, in tiles.
uint bin_tile_x = N_TILE_X * gl_WorkGroupID.x;
uint bin_tile_y = N_TILE_Y * gl_WorkGroupID.y;
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
// Per-tile state
Faster coarse raster Store a lot more tile context in shared memory and do the work from that.
2020-06-05 03:39:08 +10:00
uint tile_x = gl_LocalInvocationID.x % N_TILE_X;
uint tile_y = gl_LocalInvocationID.x / N_TILE_X;
implement variable output sizing Signed-off-by: Elias Naur <mail@eliasnaur.com>
2020-12-18 10:55:21 +11:00
uint this_tile_ix = (bin_tile_y + tile_y) * conf.width_in_tiles + bin_tile_x + tile_x;
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
Alloc cmd_alloc = slice_mem(conf.ptcl_alloc, this_tile_ix * PTCL_INITIAL_ALLOC, PTCL_INITIAL_ALLOC);
CmdRef cmd_ref = CmdRef(cmd_alloc.offset);
coarse,kernel4: make stroke, (non-zero) fill, solid separate commands 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>
2021-03-18 22:47:14 +11:00
// Reserve space for the maximum number of commands and a potential jump.
uint cmd_limit = cmd_ref.offset + PTCL_INITIAL_ALLOC - (ANNO_COMMANDS + 1) * Cmd_size;
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
// The nesting depth of the clip stack
uint clip_depth = 0;
// State for the "clip zero" optimization. If it's nonzero, then we are
// currently in a clip for which the entire tile has an alpha of zero, and
// the value is the depth after the "begin clip" of that element.
uint clip_zero_depth = 0;
Start writing tiles This is the first checkpoint where it actually runs a pipeline end to end, though it's far from accurate.
2020-05-16 05:28:29 +10:00
Parallel merge The fancy stuff :)
2020-05-31 08:37:34 +10:00
// I'm sure we can figure out how to do this with at least one fewer register...
// Items up to rd_ix have been read from sh_elements
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
uint rd_ix = 0;
Parallel merge The fancy stuff :)
2020-05-31 08:37:34 +10:00
// Items up to wr_ix have been written into sh_elements
uint wr_ix = 0;
// Items between part_start_ix and ready_ix are ready to be transferred from sh_part_elements
uint part_start_ix = 0;
uint ready_ix = 0;
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
First cut at split blend stack 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
2022-05-17 04:01:03 +10:00
cmd_ref.offset += 4;
// Accounting for allocation of blend memory
uint render_blend_depth = 0;
uint max_blend_depth = 0;
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
uint drawmonoid_start = conf.drawmonoid_alloc.offset >> 2;
uint drawtag_start = conf.drawtag_offset >> 2;
uint drawdata_start = conf.drawdata_offset >> 2;
uint drawinfo_start = conf.drawinfo_alloc.offset >> 2;
avoid non-uniform barrier control flow when exhausting memory 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>
2021-04-12 22:41:03 +10:00
bool mem_ok = mem_error == NO_ERROR;
Correct rendering (on Intel) Handle multiple passes in coarse raster. Doesn't work on NV, WIP to find out why.
2020-05-16 13:57:07 +10:00
while (true) {
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
for (uint i = 0; i < N_SLICE; i++) {
sh_bitmaps[i][th_ix] = 0;
}
Parallel merge The fancy stuff :)
2020-05-31 08:37:34 +10:00
// parallel read of input partitions
do {
if (ready_ix == wr_ix && partition_ix < n_partitions) {
part_start_ix = ready_ix;
uint count = 0;
if (th_ix < N_PART_READ && partition_ix + th_ix < n_partitions) {
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
uint in_ix = (conf.bin_alloc.offset >> 2) + ((partition_ix + th_ix) * N_TILE + bin_ix) * 2;
count = read_mem(conf.bin_alloc, in_ix);
uint offset = read_mem(conf.bin_alloc, in_ix + 1);
shaders: Add .clang-format and reformat Helps keeping the code tidy. Style is chosen to minimize diff, but contains a slight bit of personal taste.
2022-01-30 18:23:18 +11:00
sh_part_elements[th_ix] = new_alloc(offset, count * BinInstance_size, mem_ok);
Parallel merge The fancy stuff :)
2020-05-31 08:37:34 +10:00
}
// prefix sum of counts
for (uint i = 0; i < LG_N_PART_READ; i++) {
if (th_ix < N_PART_READ) {
sh_part_count[th_ix] = count;
}
barrier();
if (th_ix < N_PART_READ) {
Make shaders cross-platform Translate all piet-gpu shaders into DXIL and MSL; move generated files into the shader/gen directory.
2021-12-04 10:49:58 +11:00
if (th_ix >= (1u << i)) {
count += sh_part_count[th_ix - (1u << i)];
Parallel merge The fancy stuff :)
2020-05-31 08:37:34 +10:00
}
}
barrier();
}
if (th_ix < N_PART_READ) {
sh_part_count[th_ix] = part_start_ix + count;
}
barrier();
ready_ix = sh_part_count[N_PART_READ - 1];
partition_ix += N_PART_READ;
}
// use binary search to find element to read
uint ix = rd_ix + th_ix;
avoid non-uniform barrier control flow when exhausting memory 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>
2021-04-12 22:41:03 +10:00
if (ix >= wr_ix && ix < ready_ix && mem_ok) {
Parallel merge The fancy stuff :)
2020-05-31 08:37:34 +10:00
uint part_ix = 0;
for (uint i = 0; i < LG_N_PART_READ; i++) {
Make shaders cross-platform Translate all piet-gpu shaders into DXIL and MSL; move generated files into the shader/gen directory.
2021-12-04 10:49:58 +11:00
uint probe = part_ix + (uint(N_PART_READ / 2) >> i);
Parallel merge The fancy stuff :)
2020-05-31 08:37:34 +10:00
if (ix >= sh_part_count[probe - 1]) {
part_ix = probe;
}
}
ix -= part_ix > 0 ? sh_part_count[part_ix - 1] : part_start_ix;
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
Alloc bin_alloc = sh_part_elements[part_ix];
BinInstanceRef inst_ref = BinInstanceRef(bin_alloc.offset);
BinInstance inst = BinInstance_read(bin_alloc, BinInstance_index(inst_ref, ix));
Simplify ringbuf We don't really need a ring buffer, as we only read what we're actually going to process.
2020-05-31 14:12:55 +10:00
sh_elements[th_ix] = inst.element_ix;
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
}
Parallel merge The fancy stuff :)
2020-05-31 08:37:34 +10:00
barrier();
wr_ix = min(rd_ix + N_TILE, ready_ix);
} while (wr_ix - rd_ix < N_TILE && (wr_ix < ready_ix || partition_ix < n_partitions));
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
// We've done the merge and filled the buffer.
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
// Read one element, compute coverage.
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
uint tag = Drawtag_Nop;
Faster coarse raster Store a lot more tile context in shared memory and do the work from that.
2020-06-05 03:39:08 +10:00
uint element_ix;
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
if (th_ix + rd_ix < wr_ix) {
Faster coarse raster Store a lot more tile context in shared memory and do the work from that.
2020-06-05 03:39:08 +10:00
element_ix = sh_elements[th_ix];
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
tag = scene[drawtag_start + element_ix];
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
}
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
Smarter line segment coverage 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.
2020-05-20 01:20:45 +10:00
// Bounding box of element in pixel coordinates.
Faster coarse raster Store a lot more tile context in shared memory and do the work from that.
2020-06-05 03:39:08 +10:00
uint tile_count;
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
switch (tag) {
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
case Drawtag_FillColor:
case Drawtag_FillImage:
case Drawtag_FillLinGradient:
Radial gradients 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.
2022-03-18 09:00:08 +11:00
case Drawtag_FillRadGradient:
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
case Drawtag_BeginClip:
case Drawtag_EndClip:
uint drawmonoid_base = drawmonoid_start + 4 * element_ix;
New clip implementation 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
2022-02-18 11:25:41 +11:00
uint path_ix = memory[drawmonoid_base];
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
Path path = Path_read(conf.tile_alloc, PathRef(conf.tile_alloc.offset + path_ix * Path_size));
Faster coarse raster Store a lot more tile context in shared memory and do the work from that.
2020-06-05 03:39:08 +10:00
uint stride = path.bbox.z - path.bbox.x;
sh_tile_stride[th_ix] = stride;
int dx = int(path.bbox.x) - int(bin_tile_x);
int dy = int(path.bbox.y) - int(bin_tile_y);
int x0 = clamp(dx, 0, N_TILE_X);
int y0 = clamp(dy, 0, N_TILE_Y);
int x1 = clamp(int(path.bbox.z) - int(bin_tile_x), 0, N_TILE_X);
int y1 = clamp(int(path.bbox.w) - int(bin_tile_y), 0, N_TILE_Y);
sh_tile_width[th_ix] = uint(x1 - x0);
sh_tile_x0[th_ix] = x0;
sh_tile_y0[th_ix] = y0;
tile_count = uint(x1 - x0) * uint(y1 - y0);
// base relative to bin
uint base = path.tiles.offset - uint(dy * stride + dx) * Tile_size;
sh_tile_base[th_ix] = base;
shaders: Add .clang-format and reformat Helps keeping the code tidy. Style is chosen to minimize diff, but contains a slight bit of personal taste.
2022-01-30 18:23:18 +11:00
Alloc path_alloc = new_alloc(path.tiles.offset,
(path.bbox.z - path.bbox.x) * (path.bbox.w - path.bbox.y) * Tile_size, mem_ok);
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
write_tile_alloc(th_ix, path_alloc);
Smarter line segment coverage 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.
2020-05-20 01:20:45 +10:00
break;
default:
Faster coarse raster Store a lot more tile context in shared memory and do the work from that.
2020-06-05 03:39:08 +10:00
tile_count = 0;
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
break;
}
Smarter line segment coverage 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.
2020-05-20 01:20:45 +10:00
Filter sparse tiles 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.
2020-06-04 10:55:42 +10:00
// Prefix sum of sh_tile_count
sh_tile_count[th_ix] = tile_count;
for (uint i = 0; i < LG_N_TILE; i++) {
barrier();
Make shaders cross-platform Translate all piet-gpu shaders into DXIL and MSL; move generated files into the shader/gen directory.
2021-12-04 10:49:58 +11:00
if (th_ix >= (1u << i)) {
tile_count += sh_tile_count[th_ix - (1u << i)];
Filter sparse tiles 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.
2020-06-04 10:55:42 +10:00
}
barrier();
sh_tile_count[th_ix] = tile_count;
}
barrier();
uint total_tile_count = sh_tile_count[N_TILE - 1];
for (uint ix = th_ix; ix < total_tile_count; ix += N_TILE) {
// Binary search to find element
uint el_ix = 0;
for (uint i = 0; i < LG_N_TILE; i++) {
Make shaders cross-platform Translate all piet-gpu shaders into DXIL and MSL; move generated files into the shader/gen directory.
2021-12-04 10:49:58 +11:00
uint probe = el_ix + (uint(N_TILE / 2) >> i);
Filter sparse tiles 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.
2020-06-04 10:55:42 +10:00
if (ix >= sh_tile_count[probe - 1]) {
el_ix = probe;
}
}
Fix missing blend/clip logic 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).
2022-03-17 08:40:58 +11:00
uint element_ix = sh_elements[el_ix];
uint tag = scene[drawtag_start + element_ix];
Filter sparse tiles 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.
2020-06-04 10:55:42 +10:00
uint seq_ix = ix - (el_ix > 0 ? sh_tile_count[el_ix - 1] : 0);
uint width = sh_tile_width[el_ix];
uint x = sh_tile_x0[el_ix] + seq_ix % width;
uint y = sh_tile_y0[el_ix] + seq_ix / width;
avoid non-uniform barrier control flow when exhausting memory 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>
2021-04-12 22:41:03 +10:00
bool include_tile = false;
New clip implementation 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
2022-02-18 11:25:41 +11:00
if (mem_ok) {
shaders: Add .clang-format and reformat Helps keeping the code tidy. Style is chosen to minimize diff, but contains a slight bit of personal taste.
2022-01-30 18:23:18 +11:00
Tile tile = Tile_read(read_tile_alloc(el_ix, mem_ok),
TileRef(sh_tile_base[el_ix] + (sh_tile_stride[el_ix] * y + x) * Tile_size));
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
bool is_clip = (tag & 1) != 0;
New clip implementation 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
2022-02-18 11:25:41 +11:00
// Always include the tile if it contains a path segment.
// For draws, include the tile if it is solid.
// For clips, include the tile if it is empty - this way, logic
// below will suppress the drawing of inner elements.
Initial implementation of blend modes * 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
2022-03-01 04:38:14 +11:00
// For blends, include the tile if
// (blend_mode, composition_mode) != (Normal, SrcOver)
Fix missing blend/clip logic 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).
2022-03-17 08:40:58 +11:00
bool is_blend = false;
if (is_clip) {
uint drawmonoid_base = drawmonoid_start + 4 * element_ix;
uint scene_offset = memory[drawmonoid_base + 2];
uint dd = drawdata_start + (scene_offset >> 2);
uint blend = scene[dd];
More blend mode fixes 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.
2022-05-18 07:34:14 +10:00
is_blend = (blend != BlendComp_clip);
Fix missing blend/clip logic 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).
2022-03-17 08:40:58 +11:00
}
Initial implementation of blend modes * 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
2022-03-01 04:38:14 +11:00
include_tile = tile.tile.offset != 0 || (tile.backdrop == 0) == is_clip
Fix is_clip nit Not necessary to test is_clip when is_blend is set. Also recompiles shaders on Windows machine.
2022-03-31 01:27:29 +11:00
|| is_blend;
Continuing work on clips 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.
2020-11-21 04:26:02 +11:00
}
if (include_tile) {
Faster coarse raster Store a lot more tile context in shared memory and do the work from that.
2020-06-05 03:39:08 +10:00
uint el_slice = el_ix / 32;
Make shaders cross-platform Translate all piet-gpu shaders into DXIL and MSL; move generated files into the shader/gen directory.
2021-12-04 10:49:58 +11:00
uint el_mask = 1u << (el_ix & 31);
Faster coarse raster Store a lot more tile context in shared memory and do the work from that.
2020-06-05 03:39:08 +10:00
atomicOr(sh_bitmaps[el_slice][y * N_TILE_X + x], el_mask);
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
}
}
Filter sparse tiles 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.
2020-06-04 10:55:42 +10:00
Start writing tiles This is the first checkpoint where it actually runs a pipeline end to end, though it's far from accurate.
2020-05-16 05:28:29 +10:00
barrier();
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
// Output draw objects for this tile. The thread does a sequential walk
// through the draw objects.
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
uint slice_ix = 0;
uint bitmap = sh_bitmaps[0][th_ix];
avoid non-uniform barrier control flow when exhausting memory 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>
2021-04-12 22:41:03 +10:00
while (mem_ok) {
Continue building out pipeline Plumbs the new tiling scheme to k4. This works (stroke only) but still has some performance issues.
2020-06-04 02:28:43 +10:00
if (bitmap == 0) {
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
slice_ix++;
if (slice_ix == N_SLICE) {
break;
}
bitmap = sh_bitmaps[slice_ix][th_ix];
Continue building out pipeline Plumbs the new tiling scheme to k4. This works (stroke only) but still has some performance issues.
2020-06-04 02:28:43 +10:00
if (bitmap == 0) {
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
continue;
}
}
Continue building out pipeline Plumbs the new tiling scheme to k4. This works (stroke only) but still has some performance issues.
2020-06-04 02:28:43 +10:00
uint element_ref_ix = slice_ix * 32 + findLSB(bitmap);
Simplify ringbuf We don't really need a ring buffer, as we only read what we're actually going to process.
2020-05-31 14:12:55 +10:00
uint element_ix = sh_elements[element_ref_ix];
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
Continue building out pipeline Plumbs the new tiling scheme to k4. This works (stroke only) but still has some performance issues.
2020-06-04 02:28:43 +10:00
// Clear LSB
bitmap &= bitmap - 1;
Adding backdrop logic Calculation of backdrops kinda works but with issues, so WIP.
2020-05-21 04:48:05 +10:00
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
uint drawtag = scene[drawtag_start + element_ix];
Start writing tiles This is the first checkpoint where it actually runs a pipeline end to end, though it's far from accurate.
2020-05-16 05:28:29 +10:00
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
if (clip_zero_depth == 0) {
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
Tile tile = Tile_read(read_tile_alloc(element_ref_ix, mem_ok),
TileRef(sh_tile_base[element_ref_ix] +
(sh_tile_stride[element_ref_ix] * tile_y + tile_x) * Tile_size));
uint drawmonoid_base = drawmonoid_start + 4 * element_ix;
uint scene_offset = memory[drawmonoid_base + 2];
uint info_offset = memory[drawmonoid_base + 3];
uint dd = drawdata_start + (scene_offset >> 2);
uint di = drawinfo_start + (info_offset >> 2);
switch (drawtag) {
case Drawtag_FillColor:
float linewidth = uintBitsToFloat(memory[di]);
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
unify GPU memory management 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>
2020-12-12 04:30:20 +11:00
break;
}
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
write_fill(cmd_alloc, cmd_ref, tile, linewidth);
uint rgba = scene[dd];
Cmd_Color_write(cmd_alloc, cmd_ref, CmdColor(rgba));
tightly pack fine rasterizer commands 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>
2021-03-19 06:17:04 +11:00
cmd_ref.offset += 4 + CmdColor_size;
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
break;
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
case Drawtag_FillLinGradient:
Start work on gradients 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.
2021-06-24 04:50:51 +10:00
if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
break;
}
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
linewidth = uintBitsToFloat(memory[di]);
write_fill(cmd_alloc, cmd_ref, tile, linewidth);
Start work on gradients 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.
2021-06-24 04:50:51 +10:00
CmdLinGrad cmd_lin;
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
cmd_lin.index = scene[dd];
cmd_lin.line_x = uintBitsToFloat(memory[di + 1]);
cmd_lin.line_y = uintBitsToFloat(memory[di + 2]);
cmd_lin.line_c = uintBitsToFloat(memory[di + 3]);
Start work on gradients 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.
2021-06-24 04:50:51 +10:00
Cmd_LinGrad_write(cmd_alloc, cmd_ref, cmd_lin);
cmd_ref.offset += 4 + CmdLinGrad_size;
break;
Radial gradients 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.
2022-03-18 09:00:08 +11:00
case Drawtag_FillRadGradient:
if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
break;
}
linewidth = uintBitsToFloat(memory[di]);
write_fill(cmd_alloc, cmd_ref, tile, linewidth);
CmdRadGrad cmd_rad;
cmd_rad.index = scene[dd];
// Given that this is basically a memcpy, we might consider
// letting the fine raster read the info itself.
cmd_rad.mat = uintBitsToFloat(uvec4(memory[di + 1], memory[di + 2],
memory[di + 3], memory[di + 4]));
cmd_rad.xlat = uintBitsToFloat(uvec2(memory[di + 5], memory[di + 6]));
cmd_rad.c1 = uintBitsToFloat(uvec2(memory[di + 7], memory[di + 8]));
cmd_rad.ra = uintBitsToFloat(memory[di + 9]);
cmd_rad.roff = uintBitsToFloat(memory[di + 10]);
Cmd_RadGrad_write(cmd_alloc, cmd_ref, cmd_rad);
cmd_ref.offset += 4 + CmdRadGrad_size;
break;
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
case Drawtag_FillImage:
linewidth = uintBitsToFloat(memory[di]);
implement FillImage command and sRGB support 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>
2020-12-29 08:02:39 +11:00
if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
break;
}
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
write_fill(cmd_alloc, cmd_ref, tile, linewidth);
uint index = scene[dd];
uint raw1 = scene[dd + 1];
ivec2 offset = ivec2(int(raw1 << 16) >> 16, int(raw1) >> 16);
Cmd_Image_write(cmd_alloc, cmd_ref, CmdImage(index, offset));
tightly pack fine rasterizer commands 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>
2021-03-19 06:17:04 +11:00
cmd_ref.offset += 4 + CmdImage_size;
implement FillImage command and sRGB support 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>
2020-12-29 08:02:39 +11:00
break;
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
case Drawtag_BeginClip:
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
if (tile.tile.offset == 0 && tile.backdrop == 0) {
clip_zero_depth = clip_depth + 1;
} else {
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
unify GPU memory management 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>
2020-12-12 04:30:20 +11:00
break;
}
coarse,kernel4: make stroke, (non-zero) fill, solid separate commands 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>
2021-03-18 22:47:14 +11:00
Cmd_BeginClip_write(cmd_alloc, cmd_ref);
tightly pack fine rasterizer commands 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>
2021-03-19 06:17:04 +11:00
cmd_ref.offset += 4;
First cut at split blend stack 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
2022-05-17 04:01:03 +10:00
render_blend_depth++;
max_blend_depth = max(max_blend_depth, render_blend_depth);
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
}
clip_depth++;
break;
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
case Drawtag_EndClip:
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
clip_depth--;
New clip implementation 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
2022-02-18 11:25:41 +11:00
if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
break;
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
}
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
write_fill(cmd_alloc, cmd_ref, tile, -1.0);
uint blend = scene[dd];
Cmd_EndClip_write(cmd_alloc, cmd_ref, CmdEndClip(blend));
Initial implementation of blend modes * 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
2022-03-01 04:38:14 +11:00
cmd_ref.offset += 4 + CmdEndClip_size;
First cut at split blend stack 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
2022-05-17 04:01:03 +10:00
render_blend_depth--;
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
break;
}
} else {
// In "clip zero" state, suppress all drawing
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
switch (drawtag) {
case Drawtag_BeginClip:
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
clip_depth++;
break;
Variable size encoding of draw objects 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
2022-03-03 09:44:03 +11:00
case Drawtag_EndClip:
Optimize clips Optimize tiles with clip masks that are all-zero or all-one. Part of #36
2020-11-28 03:42:21 +11:00
if (clip_depth == clip_zero_depth) {
clip_zero_depth = 0;
}
clip_depth--;
break;
Restore FillMask logic Per discussion, don't remove FillMask until we get unbounded clip stacks.
2020-11-22 02:00:03 +11:00
}
Start writing tiles This is the first checkpoint where it actually runs a pipeline end to end, though it's far from accurate.
2020-05-16 05:28:29 +10:00
}
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
}
Correct rendering (on Intel) Handle multiple passes in coarse raster. Doesn't work on NV, WIP to find out why.
2020-05-16 13:57:07 +10:00
barrier();
Start output stage in coarse pass Still very much WIP but it's progress.
2020-05-15 10:06:45 +10:00
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
rd_ix += N_TILE;
shaders: Add .clang-format and reformat Helps keeping the code tidy. Style is chosen to minimize diff, but contains a slight bit of personal taste.
2022-01-30 18:23:18 +11:00
if (rd_ix >= ready_ix && partition_ix >= n_partitions)
break;
Correct rendering (on Intel) Handle multiple passes in coarse raster. Doesn't work on NV, WIP to find out why.
2020-05-16 13:57:07 +10:00
}
implement variable output sizing Signed-off-by: Elias Naur <mail@eliasnaur.com>
2020-12-18 10:55:21 +11:00
if (bin_tile_x + tile_x < conf.width_in_tiles && bin_tile_y + tile_y < conf.height_in_tiles) {
all: add optional memory checks 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>
2020-12-24 22:00:53 +11:00
Cmd_End_write(cmd_alloc, cmd_ref);
First cut at split blend stack 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
2022-05-17 04:01:03 +10:00
if (max_blend_depth > BLEND_STACK_SPLIT) {
// TODO: allocate blend memory and write result
}
coarse.comp: don't write Cmd_End to tiles out of bounds 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>
2020-12-18 10:29:25 +11:00
}
Starting coarse rasterizer Working down the pipeline. WIP
2020-05-14 08:35:19 +10:00
}
Reference in a new issue Copy permalink