vello/piet-gpu/shader/gen/tile_alloc.msl

#pragma clang diagnostic ignored "-Wmissing-prototypes"
#pragma clang diagnostic ignored "-Wunused-variable"

#include <metal_stdlib>
#include <simd/simd.h>
#include <metal_atomic>

using namespace metal;

struct Alloc
{
    uint offset;
};

struct MallocResult
{
    Alloc alloc;
    bool failed;
};

struct AnnoEndClipRef
{
    uint offset;
};

struct AnnoEndClip
{
    float4 bbox;
};

struct AnnotatedRef
{
    uint offset;
};

struct AnnotatedTag
{
    uint tag;
    uint flags;
};

struct PathRef
{
    uint offset;
};

struct TileRef
{
    uint offset;
};

struct Path
{
    uint4 bbox;
    TileRef tiles;
};

struct Memory
{
    uint mem_offset;
    uint mem_error;
    uint memory[1];
};

struct Alloc_1
{
    uint offset;
};

struct Config
{
    uint n_elements;
    uint n_pathseg;
    uint width_in_tiles;
    uint height_in_tiles;
    Alloc_1 tile_alloc;
    Alloc_1 bin_alloc;
    Alloc_1 ptcl_alloc;
    Alloc_1 pathseg_alloc;
    Alloc_1 anno_alloc;
    Alloc_1 trans_alloc;
    Alloc_1 bbox_alloc;
    Alloc_1 drawmonoid_alloc;
    Alloc_1 clip_alloc;
    Alloc_1 clip_bic_alloc;
    Alloc_1 clip_stack_alloc;
    Alloc_1 clip_bbox_alloc;
    uint n_trans;
    uint n_path;
    uint n_clip;
    uint trans_offset;
    uint linewidth_offset;
    uint pathtag_offset;
    uint pathseg_offset;
};

struct ConfigBuf
{
    Config conf;
};

constant uint3 gl_WorkGroupSize [[maybe_unused]] = uint3(256u, 1u, 1u);

static inline __attribute__((always_inline))
bool touch_mem(thread const Alloc& alloc, thread const uint& offset)
{
    return true;
}

static inline __attribute__((always_inline))
uint read_mem(thread const Alloc& alloc, thread const uint& offset, device Memory& v_92, constant uint& v_92BufferSize)
{
    Alloc param = alloc;
    uint param_1 = offset;
    if (!touch_mem(param, param_1))
    {
        return 0u;
    }
    uint v = v_92.memory[offset];
    return v;
}

static inline __attribute__((always_inline))
AnnotatedTag Annotated_tag(thread const Alloc& a, thread const AnnotatedRef& ref, device Memory& v_92, constant uint& v_92BufferSize)
{
    Alloc param = a;
    uint param_1 = ref.offset >> uint(2);
    uint tag_and_flags = read_mem(param, param_1, v_92, v_92BufferSize);
    return AnnotatedTag{ tag_and_flags & 65535u, tag_and_flags >> uint(16) };
}

static inline __attribute__((always_inline))
AnnoEndClip AnnoEndClip_read(thread const Alloc& a, thread const AnnoEndClipRef& ref, device Memory& v_92, constant uint& v_92BufferSize)
{
    uint ix = ref.offset >> uint(2);
    Alloc param = a;
    uint param_1 = ix + 0u;
    uint raw0 = read_mem(param, param_1, v_92, v_92BufferSize);
    Alloc param_2 = a;
    uint param_3 = ix + 1u;
    uint raw1 = read_mem(param_2, param_3, v_92, v_92BufferSize);
    Alloc param_4 = a;
    uint param_5 = ix + 2u;
    uint raw2 = read_mem(param_4, param_5, v_92, v_92BufferSize);
    Alloc param_6 = a;
    uint param_7 = ix + 3u;
    uint raw3 = read_mem(param_6, param_7, v_92, v_92BufferSize);
    AnnoEndClip s;
    s.bbox = float4(as_type<float>(raw0), as_type<float>(raw1), as_type<float>(raw2), as_type<float>(raw3));
    return s;
}

static inline __attribute__((always_inline))
AnnoEndClip Annotated_EndClip_read(thread const Alloc& a, thread const AnnotatedRef& ref, device Memory& v_92, constant uint& v_92BufferSize)
{
    Alloc param = a;
    AnnoEndClipRef param_1 = AnnoEndClipRef{ ref.offset + 4u };
    return AnnoEndClip_read(param, param_1, v_92, v_92BufferSize);
}

static inline __attribute__((always_inline))
Alloc new_alloc(thread const uint& offset, thread const uint& size, thread const bool& mem_ok)
{
    Alloc a;
    a.offset = offset;
    return a;
}

static inline __attribute__((always_inline))
MallocResult malloc(thread const uint& size, device Memory& v_92, constant uint& v_92BufferSize)
{
    uint _98 = atomic_fetch_add_explicit((device atomic_uint*)&v_92.mem_offset, size, memory_order_relaxed);
    uint offset = _98;
    MallocResult r;
    r.failed = (offset + size) > uint(int((v_92BufferSize - 8) / 4) * 4);
    uint param = offset;
    uint param_1 = size;
    bool param_2 = !r.failed;
    r.alloc = new_alloc(param, param_1, param_2);
    if (r.failed)
    {
        uint _127 = atomic_fetch_max_explicit((device atomic_uint*)&v_92.mem_error, 1u, memory_order_relaxed);
        return r;
    }
    return r;
}

static inline __attribute__((always_inline))
Alloc slice_mem(thread const Alloc& a, thread const uint& offset, thread const uint& size)
{
    return Alloc{ a.offset + offset };
}

static inline __attribute__((always_inline))
void write_mem(thread const Alloc& alloc, thread const uint& offset, thread const uint& val, device Memory& v_92, constant uint& v_92BufferSize)
{
    Alloc param = alloc;
    uint param_1 = offset;
    if (!touch_mem(param, param_1))
    {
        return;
    }
    v_92.memory[offset] = val;
}

static inline __attribute__((always_inline))
void Path_write(thread const Alloc& a, thread const PathRef& ref, thread const Path& s, device Memory& v_92, constant uint& v_92BufferSize)
{
    uint ix = ref.offset >> uint(2);
    Alloc param = a;
    uint param_1 = ix + 0u;
    uint param_2 = s.bbox.x | (s.bbox.y << uint(16));
    write_mem(param, param_1, param_2, v_92, v_92BufferSize);
    Alloc param_3 = a;
    uint param_4 = ix + 1u;
    uint param_5 = s.bbox.z | (s.bbox.w << uint(16));
    write_mem(param_3, param_4, param_5, v_92, v_92BufferSize);
    Alloc param_6 = a;
    uint param_7 = ix + 2u;
    uint param_8 = s.tiles.offset;
    write_mem(param_6, param_7, param_8, v_92, v_92BufferSize);
}

kernel void main0(constant uint* spvBufferSizeConstants [[buffer(25)]], device Memory& v_92 [[buffer(0)]], const device ConfigBuf& _305 [[buffer(1)]], uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]], uint3 gl_GlobalInvocationID [[thread_position_in_grid]])
{
    threadgroup uint sh_tile_count[256];
    threadgroup MallocResult sh_tile_alloc;
    constant uint& v_92BufferSize = spvBufferSizeConstants[0];
    uint th_ix = gl_LocalInvocationID.x;
    uint element_ix = gl_GlobalInvocationID.x;
    PathRef path_ref = PathRef{ _305.conf.tile_alloc.offset + (element_ix * 12u) };
    AnnotatedRef ref = AnnotatedRef{ _305.conf.anno_alloc.offset + (element_ix * 40u) };
    uint tag = 0u;
    if (element_ix < _305.conf.n_elements)
    {
        Alloc param;
        param.offset = _305.conf.anno_alloc.offset;
        AnnotatedRef param_1 = ref;
        tag = Annotated_tag(param, param_1, v_92, v_92BufferSize).tag;
    }
    int x0 = 0;
    int y0 = 0;
    int x1 = 0;
    int y1 = 0;
    switch (tag)
    {
        case 1u:
        case 2u:
        case 3u:
        case 4u:
        case 5u:
        {
            Alloc param_2;
            param_2.offset = _305.conf.anno_alloc.offset;
            AnnotatedRef param_3 = ref;
            AnnoEndClip clip = Annotated_EndClip_read(param_2, param_3, v_92, v_92BufferSize);
            x0 = int(floor(clip.bbox.x * 0.0625));
            y0 = int(floor(clip.bbox.y * 0.0625));
            x1 = int(ceil(clip.bbox.z * 0.0625));
            y1 = int(ceil(clip.bbox.w * 0.0625));
            break;
        }
    }
    x0 = clamp(x0, 0, int(_305.conf.width_in_tiles));
    y0 = clamp(y0, 0, int(_305.conf.height_in_tiles));
    x1 = clamp(x1, 0, int(_305.conf.width_in_tiles));
    y1 = clamp(y1, 0, int(_305.conf.height_in_tiles));
    Path path;
    path.bbox = uint4(uint(x0), uint(y0), uint(x1), uint(y1));
    uint tile_count = uint((x1 - x0) * (y1 - y0));
    if (tag == 5u)
    {
        tile_count = 0u;
    }
    sh_tile_count[th_ix] = tile_count;
    uint total_tile_count = tile_count;
    for (uint i = 0u; i < 8u; i++)
    {
        threadgroup_barrier(mem_flags::mem_threadgroup);
        if (th_ix >= (1u << i))
        {
            total_tile_count += sh_tile_count[th_ix - (1u << i)];
        }
        threadgroup_barrier(mem_flags::mem_threadgroup);
        sh_tile_count[th_ix] = total_tile_count;
    }
    if (th_ix == 255u)
    {
        uint param_4 = total_tile_count * 8u;
        MallocResult _476 = malloc(param_4, v_92, v_92BufferSize);
        sh_tile_alloc = _476;
    }
    threadgroup_barrier(mem_flags::mem_threadgroup);
    MallocResult alloc_start = sh_tile_alloc;
    bool _487;
    if (!alloc_start.failed)
    {
        _487 = v_92.mem_error != 0u;
    }
    else
    {
        _487 = alloc_start.failed;
    }
    if (_487)
    {
        return;
    }
    if (element_ix < _305.conf.n_elements)
    {
        uint _500;
        if (th_ix > 0u)
        {
            _500 = sh_tile_count[th_ix - 1u];
        }
        else
        {
            _500 = 0u;
        }
        uint tile_subix = _500;
        Alloc param_5 = alloc_start.alloc;
        uint param_6 = 8u * tile_subix;
        uint param_7 = 8u * tile_count;
        Alloc tiles_alloc = slice_mem(param_5, param_6, param_7);
        path.tiles = TileRef{ tiles_alloc.offset };
        Alloc param_8;
        param_8.offset = _305.conf.tile_alloc.offset;
        PathRef param_9 = path_ref;
        Path param_10 = path;
        Path_write(param_8, param_9, param_10, v_92, v_92BufferSize);
    }
    uint total_count = sh_tile_count[255] * 2u;
    uint start_ix = alloc_start.alloc.offset >> uint(2);
    for (uint i_1 = th_ix; i_1 < total_count; i_1 += 256u)
    {
        Alloc param_11 = alloc_start.alloc;
        uint param_12 = start_ix + i_1;
        uint param_13 = 0u;
        write_mem(param_11, param_12, param_13, v_92, v_92BufferSize);
    }
}