#version 450

#extension GL_KHR_shader_subgroup_clustered: enable
#extension GL_EXT_shader_explicit_arithmetic_types_int8: enable
#extension GL_EXT_shader_explicit_arithmetic_types_int16: enable
#extension GL_EXT_shader_explicit_arithmetic_types_int64: enable
#extension GL_EXT_shader_explicit_arithmetic_types_float16: enable

layout (local_size_x = 8) in;

layout(binding = 0) buffer Buffers
{
    i8vec4 i8;
    u8vec4 u8;
    i16vec4 i16;
    u16vec4 u16;
    i64vec4 i64;
    u64vec4 u64;
    f16vec4 f16;
} data[4];

void main()
{
    uint invocation = (gl_SubgroupInvocationID + gl_SubgroupSize) % 4;

    data[invocation].i8.x   = subgroupClusteredAdd(data[0].i8.x, 1);
    data[invocation].i8.xy  = subgroupClusteredAdd(data[1].i8.xy, 1);
    data[invocation].i8.xyz = subgroupClusteredAdd(data[2].i8.xyz, 1);
    data[invocation].i8     = subgroupClusteredAdd(data[3].i8, 1);

    data[invocation].i8.x   = subgroupClusteredMul(data[0].i8.x, 1);
    data[invocation].i8.xy  = subgroupClusteredMul(data[1].i8.xy, 1);
    data[invocation].i8.xyz = subgroupClusteredMul(data[2].i8.xyz, 1);
    data[invocation].i8     = subgroupClusteredMul(data[3].i8, 1);

    data[invocation].i8.x   = subgroupClusteredMin(data[0].i8.x, 1);
    data[invocation].i8.xy  = subgroupClusteredMin(data[1].i8.xy, 1);
    data[invocation].i8.xyz = subgroupClusteredMin(data[2].i8.xyz, 1);
    data[invocation].i8     = subgroupClusteredMin(data[3].i8, 1);

    data[invocation].i8.x   = subgroupClusteredMax(data[0].i8.x, 1);
    data[invocation].i8.xy  = subgroupClusteredMax(data[1].i8.xy, 1);
    data[invocation].i8.xyz = subgroupClusteredMax(data[2].i8.xyz, 1);
    data[invocation].i8     = subgroupClusteredMax(data[3].i8, 1);

    data[invocation].i8.x   = subgroupClusteredAnd(data[0].i8.x, 1);
    data[invocation].i8.xy  = subgroupClusteredAnd(data[1].i8.xy, 1);
    data[invocation].i8.xyz = subgroupClusteredAnd(data[2].i8.xyz, 1);
    data[invocation].i8     = subgroupClusteredAnd(data[3].i8, 1);

    data[invocation].i8.x   = subgroupClusteredOr(data[0].i8.x, 1);
    data[invocation].i8.xy  = subgroupClusteredOr(data[1].i8.xy, 1);
    data[invocation].i8.xyz = subgroupClusteredOr(data[2].i8.xyz, 1);
    data[invocation].i8     = subgroupClusteredOr(data[3].i8, 1);

    data[invocation].i8.x   = subgroupClusteredXor(data[0].i8.x, 1);
    data[invocation].i8.xy  = subgroupClusteredXor(data[1].i8.xy, 1);
    data[invocation].i8.xyz = subgroupClusteredXor(data[2].i8.xyz, 1);
    data[invocation].i8     = subgroupClusteredXor(data[3].i8, 1);

    data[invocation].u8.x   = subgroupClusteredAdd(data[0].u8.x, 1);
    data[invocation].u8.xy  = subgroupClusteredAdd(data[1].u8.xy, 1);
    data[invocation].u8.xyz = subgroupClusteredAdd(data[2].u8.xyz, 1);
    data[invocation].u8     = subgroupClusteredAdd(data[3].u8, 1);

    data[invocation].u8.x   = subgroupClusteredMul(data[0].u8.x, 1);
    data[invocation].u8.xy  = subgroupClusteredMul(data[1].u8.xy, 1);
    data[invocation].u8.xyz = subgroupClusteredMul(data[2].u8.xyz, 1);
    data[invocation].u8     = subgroupClusteredMul(data[3].u8, 1);

    data[invocation].u8.x   = subgroupClusteredMin(data[0].u8.x, 1);
    data[invocation].u8.xy  = subgroupClusteredMin(data[1].u8.xy, 1);
    data[invocation].u8.xyz = subgroupClusteredMin(data[2].u8.xyz, 1);
    data[invocation].u8     = subgroupClusteredMin(data[3].u8, 1);

    data[invocation].u8.x   = subgroupClusteredMax(data[0].u8.x, 1);
    data[invocation].u8.xy  = subgroupClusteredMax(data[1].u8.xy, 1);
    data[invocation].u8.xyz = subgroupClusteredMax(data[2].u8.xyz, 1);
    data[invocation].u8     = subgroupClusteredMax(data[3].u8, 1);

    data[invocation].u8.x   = subgroupClusteredAnd(data[0].u8.x, 1);
    data[invocation].u8.xy  = subgroupClusteredAnd(data[1].u8.xy, 1);
    data[invocation].u8.xyz = subgroupClusteredAnd(data[2].u8.xyz, 1);
    data[invocation].u8     = subgroupClusteredAnd(data[3].u8, 1);

    data[invocation].u8.x   = subgroupClusteredOr(data[0].u8.x, 1);
    data[invocation].u8.xy  = subgroupClusteredOr(data[1].u8.xy, 1);
    data[invocation].u8.xyz = subgroupClusteredOr(data[2].u8.xyz, 1);
    data[invocation].u8     = subgroupClusteredOr(data[3].u8, 1);

    data[invocation].u8.x   = subgroupClusteredXor(data[0].u8.x, 1);
    data[invocation].u8.xy  = subgroupClusteredXor(data[1].u8.xy, 1);
    data[invocation].u8.xyz = subgroupClusteredXor(data[2].u8.xyz, 1);
    data[invocation].u8     = subgroupClusteredXor(data[3].u8, 1);

    data[invocation].i16.x   = subgroupClusteredAdd(data[0].i16.x, 1);
    data[invocation].i16.xy  = subgroupClusteredAdd(data[1].i16.xy, 1);
    data[invocation].i16.xyz = subgroupClusteredAdd(data[2].i16.xyz, 1);
    data[invocation].i16     = subgroupClusteredAdd(data[3].i16, 1);

    data[invocation].i16.x   = subgroupClusteredMul(data[0].i16.x, 1);
    data[invocation].i16.xy  = subgroupClusteredMul(data[1].i16.xy, 1);
    data[invocation].i16.xyz = subgroupClusteredMul(data[2].i16.xyz, 1);
    data[invocation].i16     = subgroupClusteredMul(data[3].i16, 1);

    data[invocation].i16.x   = subgroupClusteredMin(data[0].i16.x, 1);
    data[invocation].i16.xy  = subgroupClusteredMin(data[1].i16.xy, 1);
    data[invocation].i16.xyz = subgroupClusteredMin(data[2].i16.xyz, 1);
    data[invocation].i16     = subgroupClusteredMin(data[3].i16, 1);

    data[invocation].i16.x   = subgroupClusteredMax(data[0].i16.x, 1);
    data[invocation].i16.xy  = subgroupClusteredMax(data[1].i16.xy, 1);
    data[invocation].i16.xyz = subgroupClusteredMax(data[2].i16.xyz, 1);
    data[invocation].i16     = subgroupClusteredMax(data[3].i16, 1);

    data[invocation].i16.x   = subgroupClusteredAnd(data[0].i16.x, 1);
    data[invocation].i16.xy  = subgroupClusteredAnd(data[1].i16.xy, 1);
    data[invocation].i16.xyz = subgroupClusteredAnd(data[2].i16.xyz, 1);
    data[invocation].i16     = subgroupClusteredAnd(data[3].i16, 1);

    data[invocation].i16.x   = subgroupClusteredOr(data[0].i16.x, 1);
    data[invocation].i16.xy  = subgroupClusteredOr(data[1].i16.xy, 1);
    data[invocation].i16.xyz = subgroupClusteredOr(data[2].i16.xyz, 1);
    data[invocation].i16     = subgroupClusteredOr(data[3].i16, 1);

    data[invocation].i16.x   = subgroupClusteredXor(data[0].i16.x, 1);
    data[invocation].i16.xy  = subgroupClusteredXor(data[1].i16.xy, 1);
    data[invocation].i16.xyz = subgroupClusteredXor(data[2].i16.xyz, 1);
    data[invocation].i16     = subgroupClusteredXor(data[3].i16, 1);

    data[invocation].u16.x   = subgroupClusteredAdd(data[0].u16.x, 1);
    data[invocation].u16.xy  = subgroupClusteredAdd(data[1].u16.xy, 1);
    data[invocation].u16.xyz = subgroupClusteredAdd(data[2].u16.xyz, 1);
    data[invocation].u16     = subgroupClusteredAdd(data[3].u16, 1);

    data[invocation].u16.x   = subgroupClusteredMul(data[0].u16.x, 1);
    data[invocation].u16.xy  = subgroupClusteredMul(data[1].u16.xy, 1);
    data[invocation].u16.xyz = subgroupClusteredMul(data[2].u16.xyz, 1);
    data[invocation].u16     = subgroupClusteredMul(data[3].u16, 1);

    data[invocation].u16.x   = subgroupClusteredMin(data[0].u16.x, 1);
    data[invocation].u16.xy  = subgroupClusteredMin(data[1].u16.xy, 1);
    data[invocation].u16.xyz = subgroupClusteredMin(data[2].u16.xyz, 1);
    data[invocation].u16     = subgroupClusteredMin(data[3].u16, 1);

    data[invocation].u16.x   = subgroupClusteredMax(data[0].u16.x, 1);
    data[invocation].u16.xy  = subgroupClusteredMax(data[1].u16.xy, 1);
    data[invocation].u16.xyz = subgroupClusteredMax(data[2].u16.xyz, 1);
    data[invocation].u16     = subgroupClusteredMax(data[3].u16, 1);

    data[invocation].u16.x   = subgroupClusteredAnd(data[0].u16.x, 1);
    data[invocation].u16.xy  = subgroupClusteredAnd(data[1].u16.xy, 1);
    data[invocation].u16.xyz = subgroupClusteredAnd(data[2].u16.xyz, 1);
    data[invocation].u16     = subgroupClusteredAnd(data[3].u16, 1);

    data[invocation].u16.x   = subgroupClusteredOr(data[0].u16.x, 1);
    data[invocation].u16.xy  = subgroupClusteredOr(data[1].u16.xy, 1);
    data[invocation].u16.xyz = subgroupClusteredOr(data[2].u16.xyz, 1);
    data[invocation].u16     = subgroupClusteredOr(data[3].u16, 1);

    data[invocation].u16.x   = subgroupClusteredXor(data[0].u16.x, 1);
    data[invocation].u16.xy  = subgroupClusteredXor(data[1].u16.xy, 1);
    data[invocation].u16.xyz = subgroupClusteredXor(data[2].u16.xyz, 1);
    data[invocation].u16     = subgroupClusteredXor(data[3].u16, 1);

    data[invocation].i64.x   = subgroupClusteredAdd(data[0].i64.x, 1);
    data[invocation].i64.xy  = subgroupClusteredAdd(data[1].i64.xy, 1);
    data[invocation].i64.xyz = subgroupClusteredAdd(data[2].i64.xyz, 1);
    data[invocation].i64     = subgroupClusteredAdd(data[3].i64, 1);

    data[invocation].i64.x   = subgroupClusteredMul(data[0].i64.x, 1);
    data[invocation].i64.xy  = subgroupClusteredMul(data[1].i64.xy, 1);
    data[invocation].i64.xyz = subgroupClusteredMul(data[2].i64.xyz, 1);
    data[invocation].i64     = subgroupClusteredMul(data[3].i64, 1);

    data[invocation].i64.x   = subgroupClusteredMin(data[0].i64.x, 1);
    data[invocation].i64.xy  = subgroupClusteredMin(data[1].i64.xy, 1);
    data[invocation].i64.xyz = subgroupClusteredMin(data[2].i64.xyz, 1);
    data[invocation].i64     = subgroupClusteredMin(data[3].i64, 1);

    data[invocation].i64.x   = subgroupClusteredMax(data[0].i64.x, 1);
    data[invocation].i64.xy  = subgroupClusteredMax(data[1].i64.xy, 1);
    data[invocation].i64.xyz = subgroupClusteredMax(data[2].i64.xyz, 1);
    data[invocation].i64     = subgroupClusteredMax(data[3].i64, 1);

    data[invocation].i64.x   = subgroupClusteredAnd(data[0].i64.x, 1);
    data[invocation].i64.xy  = subgroupClusteredAnd(data[1].i64.xy, 1);
    data[invocation].i64.xyz = subgroupClusteredAnd(data[2].i64.xyz, 1);
    data[invocation].i64     = subgroupClusteredAnd(data[3].i64, 1);

    data[invocation].i64.x   = subgroupClusteredOr(data[0].i64.x, 1);
    data[invocation].i64.xy  = subgroupClusteredOr(data[1].i64.xy, 1);
    data[invocation].i64.xyz = subgroupClusteredOr(data[2].i64.xyz, 1);
    data[invocation].i64     = subgroupClusteredOr(data[3].i64, 1);

    data[invocation].i64.x   = subgroupClusteredXor(data[0].i64.x, 1);
    data[invocation].i64.xy  = subgroupClusteredXor(data[1].i64.xy, 1);
    data[invocation].i64.xyz = subgroupClusteredXor(data[2].i64.xyz, 1);
    data[invocation].i64     = subgroupClusteredXor(data[3].i64, 1);

    data[invocation].u64.x   = subgroupClusteredAdd(data[0].u64.x, 1);
    data[invocation].u64.xy  = subgroupClusteredAdd(data[1].u64.xy, 1);
    data[invocation].u64.xyz = subgroupClusteredAdd(data[2].u64.xyz, 1);
    data[invocation].u64     = subgroupClusteredAdd(data[3].u64, 1);

    data[invocation].u64.x   = subgroupClusteredMul(data[0].u64.x, 1);
    data[invocation].u64.xy  = subgroupClusteredMul(data[1].u64.xy, 1);
    data[invocation].u64.xyz = subgroupClusteredMul(data[2].u64.xyz, 1);
    data[invocation].u64     = subgroupClusteredMul(data[3].u64, 1);

    data[invocation].u64.x   = subgroupClusteredMin(data[0].u64.x, 1);
    data[invocation].u64.xy  = subgroupClusteredMin(data[1].u64.xy, 1);
    data[invocation].u64.xyz = subgroupClusteredMin(data[2].u64.xyz, 1);
    data[invocation].u64     = subgroupClusteredMin(data[3].u64, 1);

    data[invocation].u64.x   = subgroupClusteredMax(data[0].u64.x, 1);
    data[invocation].u64.xy  = subgroupClusteredMax(data[1].u64.xy, 1);
    data[invocation].u64.xyz = subgroupClusteredMax(data[2].u64.xyz, 1);
    data[invocation].u64     = subgroupClusteredMax(data[3].u64, 1);

    data[invocation].u64.x   = subgroupClusteredAnd(data[0].u64.x, 1);
    data[invocation].u64.xy  = subgroupClusteredAnd(data[1].u64.xy, 1);
    data[invocation].u64.xyz = subgroupClusteredAnd(data[2].u64.xyz, 1);
    data[invocation].u64     = subgroupClusteredAnd(data[3].u64, 1);

    data[invocation].u64.x   = subgroupClusteredOr(data[0].u64.x, 1);
    data[invocation].u64.xy  = subgroupClusteredOr(data[1].u64.xy, 1);
    data[invocation].u64.xyz = subgroupClusteredOr(data[2].u64.xyz, 1);
    data[invocation].u64     = subgroupClusteredOr(data[3].u64, 1);

    data[invocation].u64.x   = subgroupClusteredXor(data[0].u64.x, 1);
    data[invocation].u64.xy  = subgroupClusteredXor(data[1].u64.xy, 1);
    data[invocation].u64.xyz = subgroupClusteredXor(data[2].u64.xyz, 1);
    data[invocation].u64     = subgroupClusteredXor(data[3].u64, 1);

    data[invocation].f16.x   = subgroupClusteredAdd(data[0].f16.x, 1);
    data[invocation].f16.xy  = subgroupClusteredAdd(data[1].f16.xy, 1);
    data[invocation].f16.xyz = subgroupClusteredAdd(data[2].f16.xyz, 1);
    data[invocation].f16     = subgroupClusteredAdd(data[3].f16, 1);

    data[invocation].f16.x   = subgroupClusteredMul(data[0].f16.x, 1);
    data[invocation].f16.xy  = subgroupClusteredMul(data[1].f16.xy, 1);
    data[invocation].f16.xyz = subgroupClusteredMul(data[2].f16.xyz, 1);
    data[invocation].f16     = subgroupClusteredMul(data[3].f16, 1);

    data[invocation].f16.x   = subgroupClusteredMin(data[0].f16.x, 1);
    data[invocation].f16.xy  = subgroupClusteredMin(data[1].f16.xy, 1);
    data[invocation].f16.xyz = subgroupClusteredMin(data[2].f16.xyz, 1);
    data[invocation].f16     = subgroupClusteredMin(data[3].f16, 1);

    data[invocation].f16.x   = subgroupClusteredMax(data[0].f16.x, 1);
    data[invocation].f16.xy  = subgroupClusteredMax(data[1].f16.xy, 1);
    data[invocation].f16.xyz = subgroupClusteredMax(data[2].f16.xyz, 1);
    data[invocation].f16     = subgroupClusteredMax(data[3].f16, 1);
}