vello/piet-gpu/shader/gen/binning.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 BinInstanceRef
{
    uint offset;
};

struct BinInstance
{
    uint element_ix;
};

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;
    uint n_trans;
    uint n_path;
    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_84, constant uint& v_84BufferSize)
{
    Alloc param = alloc;
    uint param_1 = offset;
    if (!touch_mem(param, param_1))
    {
        return 0u;
    }
    uint v = v_84.memory[offset];
    return v;
}

static inline __attribute__((always_inline))
AnnotatedTag Annotated_tag(thread const Alloc& a, thread const AnnotatedRef& ref, device Memory& v_84, constant uint& v_84BufferSize)
{
    Alloc param = a;
    uint param_1 = ref.offset >> uint(2);
    uint tag_and_flags = read_mem(param, param_1, v_84, v_84BufferSize);
    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_84, constant uint& v_84BufferSize)
{
    uint ix = ref.offset >> uint(2);
    Alloc param = a;
    uint param_1 = ix + 0u;
    uint raw0 = read_mem(param, param_1, v_84, v_84BufferSize);
    Alloc param_2 = a;
    uint param_3 = ix + 1u;
    uint raw1 = read_mem(param_2, param_3, v_84, v_84BufferSize);
    Alloc param_4 = a;
    uint param_5 = ix + 2u;
    uint raw2 = read_mem(param_4, param_5, v_84, v_84BufferSize);
    Alloc param_6 = a;
    uint param_7 = ix + 3u;
    uint raw3 = read_mem(param_6, param_7, v_84, v_84BufferSize);
    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_84, constant uint& v_84BufferSize)
{
    Alloc param = a;
    AnnoEndClipRef param_1 = AnnoEndClipRef{ ref.offset + 4u };
    return AnnoEndClip_read(param, param_1, v_84, v_84BufferSize);
}

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_84, constant uint& v_84BufferSize)
{
    uint _90 = atomic_fetch_add_explicit((device atomic_uint*)&v_84.mem_offset, size, memory_order_relaxed);
    uint offset = _90;
    MallocResult r;
    r.failed = (offset + size) > uint(int((v_84BufferSize - 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 _119 = atomic_fetch_max_explicit((device atomic_uint*)&v_84.mem_error, 1u, memory_order_relaxed);
        return r;
    }
    return r;
}

static inline __attribute__((always_inline))
void write_mem(thread const Alloc& alloc, thread const uint& offset, thread const uint& val, device Memory& v_84, constant uint& v_84BufferSize)
{
    Alloc param = alloc;
    uint param_1 = offset;
    if (!touch_mem(param, param_1))
    {
        return;
    }
    v_84.memory[offset] = val;
}

static inline __attribute__((always_inline))
void BinInstance_write(thread const Alloc& a, thread const BinInstanceRef& ref, thread const BinInstance& s, device Memory& v_84, constant uint& v_84BufferSize)
{
    uint ix = ref.offset >> uint(2);
    Alloc param = a;
    uint param_1 = ix + 0u;
    uint param_2 = s.element_ix;
    write_mem(param, param_1, param_2, v_84, v_84BufferSize);
}

kernel void main0(constant uint* spvBufferSizeConstants [[buffer(25)]], device Memory& v_84 [[buffer(0)]], const device ConfigBuf& _253 [[buffer(1)]], uint3 gl_WorkGroupID [[threadgroup_position_in_grid]], uint3 gl_LocalInvocationID [[thread_position_in_threadgroup]])
{
    threadgroup uint bitmaps[8][256];
    threadgroup short sh_alloc_failed;
    threadgroup uint count[8][256];
    threadgroup Alloc sh_chunk_alloc[256];
    constant uint& v_84BufferSize = spvBufferSizeConstants[0];
    uint my_n_elements = _253.conf.n_elements;
    uint my_partition = gl_WorkGroupID.x;
    for (uint i = 0u; i < 8u; i++)
    {
        bitmaps[i][gl_LocalInvocationID.x] = 0u;
    }
    if (gl_LocalInvocationID.x == 0u)
    {
        sh_alloc_failed = short(false);
    }
    threadgroup_barrier(mem_flags::mem_threadgroup);
    uint element_ix = (my_partition * 256u) + gl_LocalInvocationID.x;
    AnnotatedRef ref = AnnotatedRef{ _253.conf.anno_alloc.offset + (element_ix * 40u) };
    uint tag = 0u;
    if (element_ix < my_n_elements)
    {
        Alloc param;
        param.offset = _253.conf.anno_alloc.offset;
        AnnotatedRef param_1 = ref;
        tag = Annotated_tag(param, param_1, v_84, v_84BufferSize).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 = _253.conf.anno_alloc.offset;
            AnnotatedRef param_3 = ref;
            AnnoEndClip clip = Annotated_EndClip_read(param_2, param_3, v_84, v_84BufferSize);
            x0 = int(floor(clip.bbox.x * 0.00390625));
            y0 = int(floor(clip.bbox.y * 0.00390625));
            x1 = int(ceil(clip.bbox.z * 0.00390625));
            y1 = int(ceil(clip.bbox.w * 0.00390625));
            break;
        }
    }
    uint width_in_bins = ((_253.conf.width_in_tiles + 16u) - 1u) / 16u;
    uint height_in_bins = ((_253.conf.height_in_tiles + 16u) - 1u) / 16u;
    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;
    int y = y0;
    uint my_slice = gl_LocalInvocationID.x / 32u;
    uint my_mask = 1u << (gl_LocalInvocationID.x & 31u);
    while (y < y1)
    {
        uint _437 = atomic_fetch_or_explicit((threadgroup atomic_uint*)&bitmaps[my_slice][(uint(y) * width_in_bins) + uint(x)], my_mask, memory_order_relaxed);
        x++;
        if (x == x1)
        {
            x = x0;
            y++;
        }
    }
    threadgroup_barrier(mem_flags::mem_threadgroup);
    uint element_count = 0u;
    for (uint i_1 = 0u; i_1 < 8u; i_1++)
    {
        element_count += uint(int(popcount(bitmaps[i_1][gl_LocalInvocationID.x])));
        count[i_1][gl_LocalInvocationID.x] = element_count;
    }
    uint param_4 = 0u;
    uint param_5 = 0u;
    bool param_6 = true;
    Alloc chunk_alloc = new_alloc(param_4, param_5, param_6);
    if (element_count != 0u)
    {
        uint param_7 = element_count * 4u;
        MallocResult _487 = malloc(param_7, v_84, v_84BufferSize);
        MallocResult chunk = _487;
        chunk_alloc = chunk.alloc;
        sh_chunk_alloc[gl_LocalInvocationID.x] = chunk_alloc;
        if (chunk.failed)
        {
            sh_alloc_failed = short(true);
        }
    }
    uint out_ix = (_253.conf.bin_alloc.offset >> uint(2)) + (((my_partition * 256u) + gl_LocalInvocationID.x) * 2u);
    Alloc param_8;
    param_8.offset = _253.conf.bin_alloc.offset;
    uint param_9 = out_ix;
    uint param_10 = element_count;
    write_mem(param_8, param_9, param_10, v_84, v_84BufferSize);
    Alloc param_11;
    param_11.offset = _253.conf.bin_alloc.offset;
    uint param_12 = out_ix + 1u;
    uint param_13 = chunk_alloc.offset;
    write_mem(param_11, param_12, param_13, v_84, v_84BufferSize);
    threadgroup_barrier(mem_flags::mem_threadgroup);
    bool _543;
    if (!bool(sh_alloc_failed))
    {
        _543 = v_84.mem_error != 0u;
    }
    else
    {
        _543 = bool(sh_alloc_failed);
    }
    if (_543)
    {
        return;
    }
    x = x0;
    y = y0;
    while (y < y1)
    {
        uint bin_ix = (uint(y) * width_in_bins) + uint(x);
        uint out_mask = bitmaps[my_slice][bin_ix];
        if ((out_mask & my_mask) != 0u)
        {
            uint idx = uint(int(popcount(out_mask & (my_mask - 1u))));
            if (my_slice > 0u)
            {
                idx += count[my_slice - 1u][bin_ix];
            }
            Alloc out_alloc = sh_chunk_alloc[bin_ix];
            uint out_offset = out_alloc.offset + (idx * 4u);
            Alloc param_14 = out_alloc;
            BinInstanceRef param_15 = BinInstanceRef{ out_offset };
            BinInstance param_16 = BinInstance{ element_ix };
            BinInstance_write(param_14, param_15, param_16, v_84, v_84BufferSize);
        }
        x++;
        if (x == x1)
        {
            x = x0;
            y++;
        }
    }
}