vello/piet-gpu/shader/binning.comp

// SPDX-License-Identifier: Apache-2.0 OR MIT OR Unlicense

// The binning stage of the pipeline.
//
// Each workgroup processes N_TILE paths.
// Each thread processes one path and calculates a N_TILE_X x N_TILE_Y coverage mask
// based on the path bounding box to bin the paths.

#version 450
#extension GL_GOOGLE_include_directive : enable

#include "mem.h"
#include "setup.h"

layout(local_size_x = N_TILE, local_size_y = 1) in;

layout(set = 0, binding = 1) readonly buffer ConfigBuf {
    Config conf;
};

#include "annotated.h"
#include "bins.h"

// scale factors useful for converting coordinates to bins
#define SX (1.0 / float(N_TILE_X * TILE_WIDTH_PX))
#define SY (1.0 / float(N_TILE_Y * TILE_HEIGHT_PX))

// Constant not available in GLSL. Also consider uintBitsToFloat(0x7f800000)
#define INFINITY (1.0 / 0.0)

// Note: cudaraster has N_TILE + 1 to cut down on bank conflicts.
// Bitmaps are sliced (256bit into 8 (N_SLICE) 32bit submaps)
shared uint bitmaps[N_SLICE][N_TILE];
shared uint count[N_SLICE][N_TILE];
shared Alloc sh_chunk_alloc[N_TILE];
shared bool sh_alloc_failed;

void main() {
    uint my_n_elements = conf.n_elements;
    uint my_partition = gl_WorkGroupID.x;

    for (uint i = 0; i < N_SLICE; i++) {
        bitmaps[i][gl_LocalInvocationID.x] = 0;
    }
    if (gl_LocalInvocationID.x == 0) {
        sh_alloc_failed = false;
    }
    barrier();

    // Read inputs and determine coverage of bins
    uint element_ix = my_partition * N_TILE + gl_LocalInvocationID.x;
    AnnotatedRef ref = AnnotatedRef(conf.anno_alloc.offset + element_ix * Annotated_size);
    uint tag = Annotated_Nop;
    if (element_ix < my_n_elements) {
        tag = Annotated_tag(conf.anno_alloc, ref).tag;
    }
    int x0 = 0, y0 = 0, x1 = 0, y1 = 0;
    switch (tag) {
    case Annotated_Color:
    case Annotated_Image:
    case Annotated_BeginClip:
    case Annotated_EndClip:
        // Note: we take advantage of the fact that these drawing elements
        // have the bbox at the same place in their layout.
        AnnoEndClip clip = Annotated_EndClip_read(conf.anno_alloc, ref);
        x0 = int(floor(clip.bbox.x * SX));
        y0 = int(floor(clip.bbox.y * SY));
        x1 = int(ceil(clip.bbox.z * SX));
        y1 = int(ceil(clip.bbox.w * SY));
        break;
    }

    // At this point, we run an iterator over the coverage area,
    // trying to keep divergence low.
    // Right now, it's just a bbox, but we'll get finer with
    // segments.
    uint width_in_bins = (conf.width_in_tiles + N_TILE_X - 1)/N_TILE_X;
    uint height_in_bins = (conf.height_in_tiles + N_TILE_Y - 1)/N_TILE_Y;
    x0 = clamp(x0, 0, int(width_in_bins));
    x1 = clamp(x1, x0, int(width_in_bins));
    y0 = clamp(y0, 0, int(height_in_bins));
    y1 = clamp(y1, y0, int(height_in_bins));
    if (x0 == x1) y1 = y0;
    int x = x0, y = y0;
    uint my_slice = gl_LocalInvocationID.x / 32;
    uint my_mask = 1 << (gl_LocalInvocationID.x & 31);
    while (y < y1) {
        atomicOr(bitmaps[my_slice][y * width_in_bins + x], my_mask);
        x++;
        if (x == x1) {
            x = x0;
            y++;
        }
    }

    barrier();
    // Allocate output segments.
    uint element_count = 0;
    for (uint i = 0; i < N_SLICE; i++) {
        element_count += bitCount(bitmaps[i][gl_LocalInvocationID.x]);
        count[i][gl_LocalInvocationID.x] = element_count;
    }
    // element_count is number of elements covering bin for this invocation.
    Alloc chunk_alloc = new_alloc(0, 0, true);
    if (element_count != 0) {
        // TODO: aggregate atomic adds (subgroup is probably fastest)
        MallocResult chunk = malloc(element_count * BinInstance_size);
        chunk_alloc = chunk.alloc;
        sh_chunk_alloc[gl_LocalInvocationID.x] = chunk_alloc;
        if (chunk.failed) {
            sh_alloc_failed = true;
        }
    }
    // Note: it might be more efficient for reading to do this in the
    // other order (each bin is a contiguous sequence of partitions)
    uint out_ix = (conf.bin_alloc.offset >> 2) + (my_partition * N_TILE + gl_LocalInvocationID.x) * 2;
    write_mem(conf.bin_alloc, out_ix, element_count);
    write_mem(conf.bin_alloc, out_ix + 1, chunk_alloc.offset);

    barrier();
    if (sh_alloc_failed || mem_error != NO_ERROR) {
        return;
    }

    // Use similar strategy as Laine & Karras paper; loop over bbox of bins
    // touched by this element
    x = x0;
    y = y0;
    while (y < y1) {
        uint bin_ix = y * width_in_bins + x;
        uint out_mask = bitmaps[my_slice][bin_ix];
        if ((out_mask & my_mask) != 0) {
            uint idx = bitCount(out_mask & (my_mask - 1));
            if (my_slice > 0) {
                idx += count[my_slice - 1][bin_ix];
            }
            Alloc out_alloc = sh_chunk_alloc[bin_ix];
            uint out_offset = out_alloc.offset + idx * BinInstance_size;
            BinInstance_write(out_alloc, BinInstanceRef(out_offset), BinInstance(element_ix));
        }
        x++;
        if (x == x1) {
            x = x0;
            y++;
        }
    }
}