alex/vello
1
0
Fork 0
mirror of https://github.com/italicsjenga/vello.git synced 2024-10-17 23:11:30 +11:00
Code Issues Packages Projects Releases Wiki Activity
vello/piet-gpu/shader/elements.comp
Elias Naur d9d518b248 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>
2021-04-20 10:15:29 +02:00

411 lines
16 KiB
Plaintext
Raw Blame History

// SPDX-License-Identifier: Apache-2.0 OR MIT OR Unlicense
// The element processing stage, first in the pipeline.
//
// This stage is primarily about applying transforms and computing bounding
// boxes. It is organized as a scan over the input elements, producing
// annotated output elements.
#version 450
#extension GL_GOOGLE_include_directive : enable
#include "mem.h"
#include "setup.h"
#define N_ROWS 4
#define WG_SIZE 32
#define LG_WG_SIZE 5
#define PARTITION_SIZE (WG_SIZE * N_ROWS)
layout(local_size_x = WG_SIZE, local_size_y = 1) in;
layout(set = 0, binding = 1) readonly buffer ConfigBuf {
Config conf;
};
layout(set = 0, binding = 2) readonly buffer SceneBuf {
uint[] scene;
};
// It would be better to use the Vulkan memory model than
// "volatile" but shooting for compatibility here rather
// than doing things right.
layout(set = 0, binding = 3) volatile buffer StateBuf {
uint part_counter;
uint[] state;
};
#include "scene.h"
#include "state.h"
#include "annotated.h"
#include "pathseg.h"
#include "tile.h"
#define StateBuf_stride (4 + 2 * State_size)
StateRef state_aggregate_ref(uint partition_ix) {
return StateRef(4 + partition_ix * StateBuf_stride);
}
StateRef state_prefix_ref(uint partition_ix) {
return StateRef(4 + partition_ix * StateBuf_stride + State_size);
}
uint state_flag_index(uint partition_ix) {
return partition_ix * (StateBuf_stride / 4);
}
// These correspond to X, A, P respectively in the prefix sum paper.
#define FLAG_NOT_READY 0
#define FLAG_AGGREGATE_READY 1
#define FLAG_PREFIX_READY 2
#define FLAG_SET_LINEWIDTH 1
#define FLAG_SET_BBOX 2
#define FLAG_RESET_BBOX 4
#define FLAG_SET_FILL_MODE 8
// Fill modes take up the next bit. Non-zero fill is 0, stroke is 1.
#define LG_FILL_MODE 4
#define FILL_MODE_BITS 1
#define FILL_MODE_MASK (FILL_MODE_BITS << LG_FILL_MODE)
// This is almost like a monoid (the interaction between transformation and
// bounding boxes is approximate)
State combine_state(State a, State b) {
State c;
c.bbox.x = min(a.mat.x * b.bbox.x, a.mat.x * b.bbox.z) + min(a.mat.z * b.bbox.y, a.mat.z * b.bbox.w) + a.translate.x;
c.bbox.y = min(a.mat.y * b.bbox.x, a.mat.y * b.bbox.z) + min(a.mat.w * b.bbox.y, a.mat.w * b.bbox.w) + a.translate.y;
c.bbox.z = max(a.mat.x * b.bbox.x, a.mat.x * b.bbox.z) + max(a.mat.z * b.bbox.y, a.mat.z * b.bbox.w) + a.translate.x;
c.bbox.w = max(a.mat.y * b.bbox.x, a.mat.y * b.bbox.z) + max(a.mat.w * b.bbox.y, a.mat.w * b.bbox.w) + a.translate.y;
if ((a.flags & FLAG_RESET_BBOX) == 0 && b.bbox.z <= b.bbox.x && b.bbox.w <= b.bbox.y) {
c.bbox = a.bbox;
} else if ((a.flags & FLAG_RESET_BBOX) == 0 && (b.flags & FLAG_SET_BBOX) == 0 &&
(a.bbox.z > a.bbox.x || a.bbox.w > a.bbox.y))
{
c.bbox.xy = min(a.bbox.xy, c.bbox.xy);
c.bbox.zw = max(a.bbox.zw, c.bbox.zw);
}
// It would be more concise to cast to matrix types; ah well.
c.mat.x = a.mat.x * b.mat.x + a.mat.z * b.mat.y;
c.mat.y = a.mat.y * b.mat.x + a.mat.w * b.mat.y;
c.mat.z = a.mat.x * b.mat.z + a.mat.z * b.mat.w;
c.mat.w = a.mat.y * b.mat.z + a.mat.w * b.mat.w;
c.translate.x = a.mat.x * b.translate.x + a.mat.z * b.translate.y + a.translate.x;
c.translate.y = a.mat.y * b.translate.x + a.mat.w * b.translate.y + a.translate.y;
c.linewidth = (b.flags & FLAG_SET_LINEWIDTH) == 0 ? a.linewidth : b.linewidth;
c.flags = (a.flags & (FLAG_SET_LINEWIDTH | FLAG_SET_BBOX | FLAG_SET_FILL_MODE)) | b.flags;
c.flags |= (a.flags & FLAG_RESET_BBOX) >> 1;
uint fill_mode = (b.flags & FLAG_SET_FILL_MODE) == 0 ? a.flags : b.flags;
fill_mode &= FILL_MODE_MASK;
c.flags = (c.flags & ~FILL_MODE_MASK) | fill_mode;
c.path_count = a.path_count + b.path_count;
c.pathseg_count = a.pathseg_count + b.pathseg_count;
c.trans_count = a.trans_count + b.trans_count;
return c;
}
State map_element(ElementRef ref) {
// TODO: it would *probably* be more efficient to make the memory read patterns less
// divergent, though it would be more wasted memory.
uint tag = Element_tag(ref).tag;
State c;
c.bbox = vec4(0.0, 0.0, 0.0, 0.0);
c.mat = vec4(1.0, 0.0, 0.0, 1.0);
c.translate = vec2(0.0, 0.0);
c.linewidth = 1.0; // TODO should be 0.0
c.flags = 0;
c.path_count = 0;
c.pathseg_count = 0;
c.trans_count = 0;
switch (tag) {
case Element_Line:
LineSeg line = Element_Line_read(ref);
c.bbox.xy = min(line.p0, line.p1);
c.bbox.zw = max(line.p0, line.p1);
c.pathseg_count = 1;
break;
case Element_Quad:
QuadSeg quad = Element_Quad_read(ref);
c.bbox.xy = min(min(quad.p0, quad.p1), quad.p2);
c.bbox.zw = max(max(quad.p0, quad.p1), quad.p2);
c.pathseg_count = 1;
break;
case Element_Cubic:
CubicSeg cubic = Element_Cubic_read(ref);
c.bbox.xy = min(min(cubic.p0, cubic.p1), min(cubic.p2, cubic.p3));
c.bbox.zw = max(max(cubic.p0, cubic.p1), max(cubic.p2, cubic.p3));
c.pathseg_count = 1;
break;
case Element_FillColor:
case Element_FillImage:
case Element_BeginClip:
c.flags = FLAG_RESET_BBOX;
c.path_count = 1;
break;
case Element_EndClip:
c.path_count = 1;
break;
case Element_SetLineWidth:
SetLineWidth lw = Element_SetLineWidth_read(ref);
c.linewidth = lw.width;
c.flags = FLAG_SET_LINEWIDTH;
break;
case Element_Transform:
Transform t = Element_Transform_read(ref);
c.mat = t.mat;
c.translate = t.translate;
c.trans_count = 1;
break;
case Element_SetFillMode:
SetFillMode fm = Element_SetFillMode_read(ref);
c.flags = FLAG_SET_FILL_MODE | (fm.fill_mode << LG_FILL_MODE);
break;
}
return c;
}
// Get the bounding box of a circle transformed by the matrix into an ellipse.
vec2 get_linewidth(State st) {
// See https://www.iquilezles.org/www/articles/ellipses/ellipses.htm
return 0.5 * st.linewidth * vec2(length(st.mat.xz), length(st.mat.yw));
}
shared State sh_state[WG_SIZE];
shared uint sh_part_ix;
shared State sh_prefix;
void main() {
State th_state[N_ROWS];
// Determine partition to process by atomic counter (described in Section
// 4.4 of prefix sum paper).
if (gl_LocalInvocationID.x == 0) {
sh_part_ix = atomicAdd(part_counter, 1);
}
barrier();
uint part_ix = sh_part_ix;
uint ix = part_ix * PARTITION_SIZE + gl_LocalInvocationID.x * N_ROWS;
ElementRef ref = ElementRef(ix * Element_size);
th_state[0] = map_element(ref);
for (uint i = 1; i < N_ROWS; i++) {
// discussion question: would it be faster to load using more coherent patterns
// into thread memory? This is kinda strided.
th_state[i] = combine_state(th_state[i - 1], map_element(Element_index(ref, i)));
}
State agg = th_state[N_ROWS - 1];
sh_state[gl_LocalInvocationID.x] = agg;
for (uint i = 0; i < LG_WG_SIZE; i++) {
barrier();
if (gl_LocalInvocationID.x >= (1 << i)) {
State other = sh_state[gl_LocalInvocationID.x - (1 << i)];
agg = combine_state(other, agg);
}
barrier();
sh_state[gl_LocalInvocationID.x] = agg;
}
State exclusive;
exclusive.bbox = vec4(0.0, 0.0, 0.0, 0.0);
exclusive.mat = vec4(1.0, 0.0, 0.0, 1.0);
exclusive.translate = vec2(0.0, 0.0);
exclusive.linewidth = 1.0; //TODO should be 0.0
exclusive.flags = 0;
exclusive.path_count = 0;
exclusive.pathseg_count = 0;
exclusive.trans_count = 0;
// Publish aggregate for this partition
if (gl_LocalInvocationID.x == WG_SIZE - 1) {
// Note: with memory model, we'd want to generate the atomic store version of this.
State_write(state_aggregate_ref(part_ix), agg);
uint flag = FLAG_AGGREGATE_READY;
memoryBarrierBuffer();
if (part_ix == 0) {
State_write(state_prefix_ref(part_ix), agg);
flag = FLAG_PREFIX_READY;
}
state[state_flag_index(part_ix)] = flag;
if (part_ix != 0) {
// step 4 of paper: decoupled lookback
uint look_back_ix = part_ix - 1;
State their_agg;
uint their_ix = 0;
while (true) {
flag = state[state_flag_index(look_back_ix)];
if (flag == FLAG_PREFIX_READY) {
State their_prefix = State_read(state_prefix_ref(look_back_ix));
exclusive = combine_state(their_prefix, exclusive);
break;
} else if (flag == FLAG_AGGREGATE_READY) {
their_agg = State_read(state_aggregate_ref(look_back_ix));
exclusive = combine_state(their_agg, exclusive);
look_back_ix--;
their_ix = 0;
continue;
}
// else spin
// Unfortunately there's no guarantee of forward progress of other
// workgroups, so compute a bit of the aggregate before trying again.
// In the worst case, spinning stops when the aggregate is complete.
ElementRef ref = ElementRef((look_back_ix * PARTITION_SIZE + their_ix) * Element_size);
State s = map_element(ref);
if (their_ix == 0) {
their_agg = s;
} else {
their_agg = combine_state(their_agg, s);
}
their_ix++;
if (their_ix == PARTITION_SIZE) {
exclusive = combine_state(their_agg, exclusive);
if (look_back_ix == 0) {
break;
}
look_back_ix--;
their_ix = 0;
}
}
// step 5 of paper: compute inclusive prefix
State inclusive_prefix = combine_state(exclusive, agg);
sh_prefix = exclusive;
State_write(state_prefix_ref(part_ix), inclusive_prefix);
memoryBarrierBuffer();
flag = FLAG_PREFIX_READY;
state[state_flag_index(part_ix)] = flag;
}
}
barrier();
if (part_ix != 0) {
exclusive = sh_prefix;
}
State row = exclusive;
if (gl_LocalInvocationID.x > 0) {
State other = sh_state[gl_LocalInvocationID.x - 1];
row = combine_state(row, other);
}
for (uint i = 0; i < N_ROWS; i++) {
State st = combine_state(row, th_state[i]);
// Here we read again from the original scene. There may be
// gains to be had from stashing in shared memory or possibly
// registers (though register pressure is an issue).
ElementRef this_ref = Element_index(ref, i);
ElementTag tag = Element_tag(this_ref);
uint fill_mode = fill_mode_from_flags(st.flags >> LG_FILL_MODE);
bool is_stroke = fill_mode == MODE_STROKE;
switch (tag.tag) {
case Element_Line:
LineSeg line = Element_Line_read(this_ref);
PathCubic path_cubic;
path_cubic.p0 = line.p0;
path_cubic.p1 = mix(line.p0, line.p1, 1.0 / 3.0);
path_cubic.p2 = mix(line.p1, line.p0, 1.0 / 3.0);
path_cubic.p3 = line.p1;
path_cubic.path_ix = st.path_count;
path_cubic.trans_ix = st.trans_count;
if (is_stroke) {
path_cubic.stroke = get_linewidth(st);
} else {
path_cubic.stroke = vec2(0.0);
}
PathSegRef path_out_ref = PathSegRef(conf.pathseg_alloc.offset + (st.pathseg_count - 1) * PathSeg_size);
PathSeg_Cubic_write(conf.pathseg_alloc, path_out_ref, fill_mode, path_cubic);
break;
case Element_Quad:
QuadSeg quad = Element_Quad_read(this_ref);
path_cubic.p0 = quad.p0;
path_cubic.p1 = mix(quad.p1, quad.p0, 1.0 / 3.0);
path_cubic.p2 = mix(quad.p1, quad.p2, 1.0 / 3.0);
path_cubic.p3 = quad.p2;
path_cubic.path_ix = st.path_count;
path_cubic.trans_ix = st.trans_count;
if (is_stroke) {
path_cubic.stroke = get_linewidth(st);
} else {
path_cubic.stroke = vec2(0.0);
}
path_out_ref = PathSegRef(conf.pathseg_alloc.offset + (st.pathseg_count - 1) * PathSeg_size);
PathSeg_Cubic_write(conf.pathseg_alloc, path_out_ref, fill_mode, path_cubic);
break;
case Element_Cubic:
CubicSeg cubic = Element_Cubic_read(this_ref);
path_cubic.p0 = cubic.p0;
path_cubic.p1 = cubic.p1;
path_cubic.p2 = cubic.p2;
path_cubic.p3 = cubic.p3;
path_cubic.path_ix = st.path_count;
path_cubic.trans_ix = st.trans_count;
if (is_stroke) {
path_cubic.stroke = get_linewidth(st);
} else {
path_cubic.stroke = vec2(0.0);
}
path_out_ref = PathSegRef(conf.pathseg_alloc.offset + (st.pathseg_count - 1) * PathSeg_size);
PathSeg_Cubic_write(conf.pathseg_alloc, path_out_ref, fill_mode, path_cubic);
break;
case Element_FillColor:
FillColor fill = Element_FillColor_read(this_ref);
AnnoColor anno_fill;
anno_fill.rgba_color = fill.rgba_color;
if (is_stroke) {
vec2 lw = get_linewidth(st);
anno_fill.bbox = st.bbox + vec4(-lw, lw);
anno_fill.linewidth = st.linewidth * sqrt(abs(st.mat.x * st.mat.w - st.mat.y * st.mat.z));
} else {
anno_fill.bbox = st.bbox;
anno_fill.linewidth = 0.0;
}
AnnotatedRef out_ref = AnnotatedRef(conf.anno_alloc.offset + (st.path_count - 1) * Annotated_size);
Annotated_Color_write(conf.anno_alloc, out_ref, fill_mode, anno_fill);
break;
case Element_FillImage:
FillImage fill_img = Element_FillImage_read(this_ref);
AnnoImage anno_img;
anno_img.index = fill_img.index;
anno_img.offset = fill_img.offset;
if (is_stroke) {
vec2 lw = get_linewidth(st);
anno_img.bbox = st.bbox + vec4(-lw, lw);
anno_img.linewidth = st.linewidth * sqrt(abs(st.mat.x * st.mat.w - st.mat.y * st.mat.z));
} else {
anno_img.bbox = st.bbox;
anno_img.linewidth = 0.0;
}
out_ref = AnnotatedRef(conf.anno_alloc.offset + (st.path_count - 1) * Annotated_size);
Annotated_Image_write(conf.anno_alloc, out_ref, fill_mode, anno_img);
break;
case Element_BeginClip:
Clip begin_clip = Element_BeginClip_read(this_ref);
AnnoBeginClip anno_begin_clip;
// This is the absolute bbox, it's been transformed during encoding.
anno_begin_clip.bbox = begin_clip.bbox;
if (is_stroke) {
vec2 lw = get_linewidth(st);
anno_begin_clip.linewidth = st.linewidth * sqrt(abs(st.mat.x * st.mat.w - st.mat.y * st.mat.z));
} else {
anno_fill.linewidth = 0.0;
}
out_ref = AnnotatedRef(conf.anno_alloc.offset + (st.path_count - 1) * Annotated_size);
Annotated_BeginClip_write(conf.anno_alloc, out_ref, fill_mode, anno_begin_clip);
break;
case Element_EndClip:
Clip end_clip = Element_EndClip_read(this_ref);
// This bbox is expected to be the same as the begin one.
AnnoEndClip anno_end_clip = AnnoEndClip(end_clip.bbox);
out_ref = AnnotatedRef(conf.anno_alloc.offset + (st.path_count - 1) * Annotated_size);
Annotated_EndClip_write(conf.anno_alloc, out_ref, anno_end_clip);
break;
case Element_Transform:
TransformSeg transform = TransformSeg(st.mat, st.translate);
TransformSegRef trans_ref = TransformSegRef(conf.trans_alloc.offset + (st.trans_count - 1) * TransformSeg_size);
TransformSeg_write(conf.trans_alloc, trans_ref, transform);
break;
}
}
}
Reference in a new issue View git blame Copy permalink