#version 460

#define MAX_VER  81
#define MAX_PRIM 32

#define BARRIER() \
    memoryBarrierShared(); \
    barrier();

#extension GL_NV_mesh_shader : enable

layout(local_size_x = 32) in;

layout(max_vertices=MAX_VER) out;
layout(max_primitives=MAX_PRIM) out;
layout(triangles) out;

// test use of builtins in mesh shaders:

void main()
{
    uint iid = gl_LocalInvocationID.x;
    uint gid = gl_WorkGroupID.x;

    gl_MeshVerticesNV[iid].gl_Position = vec4(1.0);
    gl_MeshVerticesNV[iid].gl_PointSize = 2.0;
    gl_MeshVerticesNV[iid].gl_ClipDistance[3] = 3.0;
    gl_MeshVerticesNV[iid].gl_CullDistance[2] = 4.0;

    BARRIER();

    gl_MeshVerticesNV[iid+1].gl_Position = gl_MeshVerticesNV[iid].gl_Position;
    gl_MeshVerticesNV[iid+1].gl_PointSize = gl_MeshVerticesNV[iid].gl_PointSize;
    gl_MeshVerticesNV[iid+1].gl_ClipDistance[3] = gl_MeshVerticesNV[iid].gl_ClipDistance[3];
    gl_MeshVerticesNV[iid+1].gl_CullDistance[2] = gl_MeshVerticesNV[iid].gl_CullDistance[2];

    BARRIER();

    gl_MeshPrimitivesNV[iid].gl_PrimitiveID = 6;
    gl_MeshPrimitivesNV[iid].gl_Layer = 7;
    gl_MeshPrimitivesNV[iid].gl_ViewportIndex = 8;
    gl_MeshPrimitivesNV[iid].gl_ViewportMask[0] = 9;

    BARRIER();

    gl_MeshPrimitivesNV[iid+1].gl_PrimitiveID = gl_MeshPrimitivesNV[iid].gl_PrimitiveID;
    gl_MeshPrimitivesNV[iid+1].gl_Layer = gl_MeshPrimitivesNV[iid].gl_Layer;
    gl_MeshPrimitivesNV[iid+1].gl_ViewportIndex = gl_MeshPrimitivesNV[iid].gl_ViewportIndex;
    gl_MeshPrimitivesNV[iid+1].gl_ViewportMask[0] = gl_MeshPrimitivesNV[iid].gl_ViewportMask[0];

    BARRIER();

    // check bound limits
    gl_PrimitiveIndicesNV[0] = 257; // should truncate 257 -> 1
    gl_PrimitiveIndicesNV[(MAX_PRIM * 3) - 1] = 2;
    gl_PrimitiveIndicesNV[gid] = gl_PrimitiveIndicesNV[gid-1];

    // writes 4 indices at offset gl_DrawID
    writePackedPrimitiveIndices4x8NV(gl_DrawID, 0x01020304);

    gl_PrimitiveCountNV = MAX_PRIM * 3;

    BARRIER();
}