vello/piet-gpu/shader/kernel1.comp

// This is "kernel 1" in a 4-kernel pipeline. It traverses the scene graph
// and outputs "instances" (references to item + translation) for each item
// that intersects the tilegroup.
//
// This implementation is simplistic and leaves a lot of performance on the
// table. A fancier implementation would use threadgroup shared memory or
// subgroups (or possibly both) to parallelize the reading of the input and
// the computation of tilegroup intersection.
//
// In addition, there are some features currently missing. One is the use of
// a bump allocator to extend the current fixed allocation. Another is support
// for clipping.

#version 450
#extension GL_GOOGLE_include_directive : enable

// It's possible we should lay this out with x and do our own math.
layout(local_size_x = 1, local_size_y = 32) in;

layout(set = 0, binding = 0) readonly buffer SceneBuf {
    uint[] scene;
};

layout(set = 0, binding = 1) buffer TilegroupBuf {
    uint[] tilegroup;
};

layout(set = 0, binding = 2) buffer AllocBuf {
    uint alloc;
};

#include "scene.h"
#include "tilegroup.h"

#include "setup.h"

#define MAX_STACK 8

struct StackElement {
    PietItemRef group;
    uint index;
    vec2 offset;
};

void main() {
    StackElement stack[MAX_STACK];
    uint stack_ix = 0;
    uint tilegroup_ix = gl_GlobalInvocationID.y * WIDTH_IN_TILEGROUPS + gl_GlobalInvocationID.x;
    TileGroupRef tg_ref = TileGroupRef(tilegroup_ix * TILEGROUP_INITIAL_ALLOC);
    uint tg_limit = tg_ref.offset + TILEGROUP_INITIAL_ALLOC - 2 * TileGroup_size;
    vec2 xy0 = vec2(gl_GlobalInvocationID.xy) * vec2(TILEGROUP_WIDTH_PX, TILEGROUP_HEIGHT_PX);
    PietItemRef root = PietItemRef(0);
    SimpleGroup group = PietItem_Group_read(root);
    StackElement tos = StackElement(root, 0, group.offset.xy);

    while (true) {
        if (tos.index < group.n_items) {
            Bbox bbox = Bbox_read(Bbox_index(group.bboxes, tos.index));
            vec4 bb = vec4(bbox.bbox) + tos.offset.xyxy;
            bool hit = max(bb.x, xy0.x) < min(bb.z, xy0.x + float(TILEGROUP_WIDTH_PX))
                && max(bb.y, xy0.y) < min(bb.w, xy0.y + float(TILEGROUP_HEIGHT_PX));
            bool is_group = false;
            if (hit) {
                PietItemRef item_ref = PietItem_index(group.items, tos.index);
                is_group = PietItem_tag(item_ref) == PietItem_Group;
            }
            if (hit && !is_group) {
                PietItemRef item_ref = PietItem_index(group.items, tos.index);
                Instance ins = Instance(item_ref.offset, tos.offset);
                if (tg_ref.offset > tg_limit) {
                    // Allocation exceeded; do atomic bump alloc.
                    uint new_tg = atomicAdd(alloc, TILEGROUP_INITIAL_ALLOC);
                    Jump jump = Jump(new_tg);
                    TileGroup_Jump_write(tg_ref, jump);
                    tg_ref = TileGroupRef(new_tg);
                    tg_limit = tg_ref.offset + TILEGROUP_INITIAL_ALLOC - 2 * TileGroup_size;
                }
                TileGroup_Instance_write(tg_ref, ins);
                tg_ref.offset += TileGroup_size;
            }
            if (is_group) {
                PietItemRef item_ref = PietItem_index(group.items, tos.index);
                tos.index++;
                if (tos.index < group.n_items) {
                    stack[stack_ix++] = tos;
                }
                group = PietItem_Group_read(item_ref);
                tos = StackElement(item_ref, 0, tos.offset + group.offset.xy);
            } else {
                tos.index++;
            }
        } else {
            // processed all items in this group; pop the stack
            if (stack_ix == 0) {
                break;
            }
            tos = stack[--stack_ix];
            group = PietItem_Group_read(tos.group);
        }
    }
    TileGroup_End_write(tg_ref);
}