joejulian/gio
1
0
Fork 1
mirror of https://git.sr.ht/~eliasnaur/gio synced 2026-07-02 16:06:19 +00:00
Code Issues Packages Projects Releases Wiki Activity

gpu: [compute] fix path gaps by eliminating redundant path points

Browse Source 
See https://github.com/linebender/piet-gpu/issues/62 for description
of the issue. The fix is the Gio copy of the piet-gpu fix.

Signed-off-by: Elias Naur <mail@eliasnaur.com>
This commit is contained in:
Elias Naur
2021-02-15 20:04:01 +01:00
parent b5d21b209c
commit 2feec23561
6 changed files with 105 additions and 158 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gpu/shaders/elements.comp
+32 -7
View File
@@ -62,6 +62,8 @@ uint state_flag_index(uint partition_ix) {
#define FLAG_SET_LINEWIDTH 1
#define FLAG_SET_BBOX 2
#define FLAG_RESET_BBOX 4
#define FLAG_START_PATH 8
#define FLAG_END_PATH 16
// This is almost like a monoid (the interaction between transformation and
// bounding boxes is approximate)
@@ -87,32 +89,45 @@ State combine_state(State a, State b) {
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)) | b.flags;
c.flags = (a.flags & (FLAG_SET_LINEWIDTH | FLAG_SET_BBOX | FLAG_START_PATH)) | b.flags;
c.flags |= (a.flags & FLAG_RESET_BBOX) >> 1;
c.flags |= (a.flags & FLAG_END_PATH) >> 1;
c.path_count = a.path_count + b.path_count;
c.pathseg_count = a.pathseg_count + b.pathseg_count;
c.tail = a.tail;
if ((a.flags & FLAG_END_PATH) != 0 && (b.flags & FLAG_START_PATH) == 0) {
c.tail = a.pathseg_count;
} else if ((b.flags & FLAG_START_PATH) != 0) {
c.tail = b.tail + a.pathseg_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);
uint tag_flags = Element_tag(ref);
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.tail = 0;
c.path_count = 0;
c.pathseg_count = 0;
switch (tag) {
// flags contain FLAG_END_PATH for segments last in their path.
uint flags = tag_flags >> 16;
switch (tag_flags & 0xffff) {
case Element_FillLine:
case Element_StrokeLine:
LineSeg line = Element_FillLine_read(ref);
c.bbox.xy = min(line.p0, line.p1);
c.bbox.zw = max(line.p0, line.p1);
c.pathseg_count = 1;
c.flags = flags;
break;
case Element_FillQuad:
case Element_StrokeQuad:
@@ -120,6 +135,7 @@ State map_element(ElementRef 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;
c.flags = flags;
break;
case Element_FillCubic:
case Element_StrokeCubic:
@@ -127,6 +143,7 @@ State map_element(ElementRef 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;
c.flags = flags;
break;
case Element_Fill:
case Element_FillImage:
@@ -166,6 +183,7 @@ shared vec2 sh_translate[WG_SIZE];
shared vec4 sh_bbox[WG_SIZE];
shared float sh_width[WG_SIZE];
shared uint sh_flags[WG_SIZE];
shared uint sh_tail[WG_SIZE];
shared uint sh_path_count[WG_SIZE];
shared uint sh_pathseg_count[WG_SIZE];
@@ -200,6 +218,7 @@ void main() {
sh_translate[gl_LocalInvocationID.x] = agg.translate;
sh_bbox[gl_LocalInvocationID.x] = agg.bbox;
sh_width[gl_LocalInvocationID.x] = agg.linewidth;
sh_tail[gl_LocalInvocationID.x] = agg.tail;
sh_flags[gl_LocalInvocationID.x] = agg.flags;
sh_path_count[gl_LocalInvocationID.x] = agg.path_count;
sh_pathseg_count[gl_LocalInvocationID.x] = agg.pathseg_count;
@@ -213,6 +232,7 @@ void main() {
other.bbox = sh_bbox[ix];
other.linewidth = sh_width[ix];
other.flags = sh_flags[ix];
other.tail = sh_tail[ix];
other.path_count = sh_path_count[ix];
other.pathseg_count = sh_pathseg_count[ix];
agg = combine_state(other, agg);
@@ -223,6 +243,7 @@ void main() {
sh_bbox[gl_LocalInvocationID.x] = agg.bbox;
sh_width[gl_LocalInvocationID.x] = agg.linewidth;
sh_flags[gl_LocalInvocationID.x] = agg.flags;
sh_tail[gl_LocalInvocationID.x] = agg.tail;
sh_path_count[gl_LocalInvocationID.x] = agg.path_count;
sh_pathseg_count[gl_LocalInvocationID.x] = agg.pathseg_count;
}
@@ -233,6 +254,7 @@ void main() {
exclusive.translate = vec2(0.0, 0.0);
exclusive.linewidth = 1.0; //TODO should be 0.0
exclusive.flags = 0;
exclusive.tail = 0;
exclusive.path_count = 0;
exclusive.pathseg_count = 0;
@@ -312,6 +334,7 @@ void main() {
other.bbox = sh_bbox[ix];
other.linewidth = sh_width[ix];
other.flags = sh_flags[ix];
other.tail = sh_tail[ix];
other.path_count = sh_path_count[ix];
other.pathseg_count = sh_pathseg_count[ix];
row = combine_state(row, other);
@@ -323,7 +346,9 @@ void main() {
// 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);
uint tag = Element_tag(this_ref);
uint tag_flags = Element_tag(this_ref);
uint tag = tag_flags & 0xffff;
uint flags = tag_flags >> 16;
switch (tag) {
case Element_FillLine:
case Element_StrokeLine:
@@ -334,7 +359,7 @@ void main() {
path_cubic.p0 = p0;
path_cubic.p1 = mix(p0, p1, 1.0 / 3.0);
path_cubic.p2 = mix(p1, p0, 1.0 / 3.0);
path_cubic.p3 = p1;
path_cubic.succ_ix = (flags & FLAG_END_PATH) == 0 ? st.pathseg_count : st.tail;
path_cubic.path_ix = st.path_count;
if (tag == Element_StrokeLine) {
path_cubic.stroke = get_linewidth(st);
@@ -358,7 +383,7 @@ void main() {
path_cubic.p0 = p0;
path_cubic.p1 = mix(p1, p0, 1.0 / 3.0);
path_cubic.p2 = mix(p1, p2, 1.0 / 3.0);
path_cubic.p3 = p2;
path_cubic.succ_ix = (flags & FLAG_END_PATH) == 0 ? st.pathseg_count : st.tail;
path_cubic.path_ix = st.path_count;
if (tag == Element_StrokeQuad) {
path_cubic.stroke = get_linewidth(st);
@@ -379,7 +404,7 @@ void main() {
path_cubic.p0 = st.mat.xy * cubic.p0.x + st.mat.zw * cubic.p0.y + st.translate;
path_cubic.p1 = st.mat.xy * cubic.p1.x + st.mat.zw * cubic.p1.y + st.translate;
path_cubic.p2 = st.mat.xy * cubic.p2.x + st.mat.zw * cubic.p2.y + st.translate;
path_cubic.p3 = st.mat.xy * cubic.p3.x + st.mat.zw * cubic.p3.y + st.translate;
path_cubic.succ_ix = (flags & FLAG_END_PATH) == 0 ? st.pathseg_count : st.tail;
path_cubic.path_ix = st.path_count;
if (tag == Element_StrokeCubic) {
path_cubic.stroke = get_linewidth(st);