blob: 49775ebdcfb234a32a517a904fff0d35b1311c2c [file] [log] [blame]
// SPDX-License-Identifier: Apache-2.0 OR MIT OR Unlicense
// The main phase for the bbox union computation
#version 450
#define LG_N_SEQ 0
#define N_SEQ (1 << LG_N_SEQ)
#define LG_WG_SIZE 9
#define WG_SIZE (1 << LG_WG_SIZE)
#define PART_SIZE (WG_SIZE * N_SEQ)
layout(local_size_x = WG_SIZE, local_size_y = 1) in;
// The bicyclic monoid
struct Bic {
uint a;
uint b;
};
struct BicBbox {
Bic bic;
uint pad2;
uint pad3;
vec4 bbox;
};
Bic bic_combine(Bic x, Bic y) {
uint m = min(x.b, y.a);
return Bic(x.a + y.a - m, x.b + y.b - m);
}
// Node type values
#define OPEN_PAREN 0
#define CLOSE_PAREN 1
struct Node {
uint node_type;
uint pad1;
uint pad2;
uint pad3;
vec4 bbox;
};
#define BBOX_IDENTITY vec4(1e9, 1e9, -1e9, -1e9)
vec4 bbox_union(vec4 a, vec4 b) {
return vec4(min(a.xy, b.xy), max(a.zw, b.zw));
}
layout(binding = 0) readonly buffer InBuf {
Node[] inbuf;
};
layout(binding = 1) readonly buffer BicBuf {
BicBbox[] bicbuf;
};
layout(binding = 2) readonly buffer StackBuf {
vec4[] stack;
};
layout(binding = 3) buffer OutBuf {
vec4[] outbuf;
};
shared Bic sh_bic[WG_SIZE * 2 - 2];
shared vec4 sh_stack[WG_SIZE];
shared vec4 sh_bbox[WG_SIZE * 2 - 2];
void main() {
uint th = gl_LocalInvocationID.x;
// materialize stack snapshot as of the start of this partition
Bic bic = Bic(0, 0);
vec4 bbox = BBOX_IDENTITY;
if (th < gl_WorkGroupID.x) {
bic = bicbuf[th].bic;
bbox = bicbuf[th].bbox;
}
sh_bic[th] = bic;
sh_bbox[th] = bbox;
for (uint i = 0; i < LG_WG_SIZE; i++) {
barrier();
uint other_ix = th + (1u << i);
if (other_ix < WG_SIZE) {
bic = bic_combine(bic, sh_bic[other_ix]);
bbox = bbox_union(bbox, sh_bbox[other_ix]);
}
barrier();
sh_bic[th] = bic;
sh_bbox[th] = bbox;
}
barrier();
uint size = sh_bic[0].b;
// binary search in stack to do stream compaction
uint sp = PART_SIZE - 1 - th;
uint ix = 0;
for (uint i = 0; i < LG_WG_SIZE; i++) {
uint probe = ix + (uint(PART_SIZE / 2) >> i);
if (sp < sh_bic[probe].b) {
ix = probe;
}
}
// ix is the largest value such that sp < sh_bic[ix].b (if any)
uint b = sh_bic[ix].b;
if (sp < b) {
vec4 bbox = stack[ix * PART_SIZE + b - sp - 1];
if (ix + 1 < WG_SIZE) {
bbox = bbox_union(bbox, sh_bbox[ix + 1]);
}
sh_stack[th] = bbox;
}
barrier();
// Do tree reduction of bicyclic semigroups and bounding boxes
Node inp = inbuf[gl_GlobalInvocationID.x];
uint node_type = inp.node_type;
bic = Bic(uint(node_type == CLOSE_PAREN), uint(node_type == OPEN_PAREN));
sh_bic[th] = bic;
sh_bbox[th] = inp.bbox;
uint inbase = 0;
for (uint i = 0; i < LG_WG_SIZE - 1; i++) {
uint outbase = 2 * WG_SIZE - (1u << (LG_WG_SIZE - i));
barrier();
if (th < (1u << (LG_WG_SIZE - 1 - i))) {
sh_bic[outbase + th] = bic_combine(sh_bic[inbase + th * 2], sh_bic[inbase + th * 2 + 1]);
sh_bbox[outbase + th] = bbox_union(sh_bbox[inbase + th * 2], sh_bbox[inbase + th * 2 + 1]);
}
inbase = outbase;
}
barrier();
// Search for predecessor node and aggregate bounding boxes
ix = th;
bbox = inp.bbox;
bic = Bic(0, 0);
if (node_type == CLOSE_PAREN) {
// only aggregate bbox for nodes with children; ascribe to close paren
// Note: if we also maintained the link to the open paren, we could
// store the bbox there. This will depend on the use case.
uint j = 0;
while (j < LG_WG_SIZE) {
uint base = 2 * WG_SIZE - (2u << (LG_WG_SIZE - j));
if (((ix >> j) & 1) != 0) {
Bic test = bic_combine(sh_bic[base + (ix >> j) - 1], bic);
if (test.b > 0) {
break;
}
bic = test;
bbox = bbox_union(sh_bbox[base + (ix >> j) - 1], bbox);
ix -= 1u << j;
}
j++;
}
if (ix > 0) {
while (j > 0) {
j--;
uint base = 2 * WG_SIZE - (2u << (LG_WG_SIZE - j));
Bic test = bic_combine(sh_bic[base + (ix >> j) - 1], bic);
if (test.b == 0) {
bic = test;
bbox = bbox_union(sh_bbox[base + (ix >> j) - 1], bbox);
ix -= 1u << j;
}
}
}
// ix is the smallest value such that reduce(ix..th).b == 0
if (ix == 0 && bic.a < size) {
bbox = bbox_union(sh_stack[PART_SIZE - 1 - bic.a], bbox);
}
}
outbuf[gl_GlobalInvocationID.x] = bbox;
}