#version 450

#extension GL_KHR_shader_subgroup_shuffle_relative: 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   = subgroupShuffleUp(data[0].i8.x,    invocation);
    data[invocation].i8.xy  = subgroupShuffleUp(data[1].i8.xy,   invocation);
    data[invocation].i8.xyz = subgroupShuffleUp(data[2].i8.xyz,  invocation);
    data[invocation].i8     = subgroupShuffleUp(data[3].i8,      invocation);

    data[invocation].i8.x   = subgroupShuffleDown(data[0].i8.x,    invocation);
    data[invocation].i8.xy  = subgroupShuffleDown(data[1].i8.xy,   invocation);
    data[invocation].i8.xyz = subgroupShuffleDown(data[2].i8.xyz,  invocation);
    data[invocation].i8     = subgroupShuffleDown(data[3].i8,      invocation);

    data[invocation].u8.x   = subgroupShuffleUp(data[0].u8.x,    invocation);
    data[invocation].u8.xy  = subgroupShuffleUp(data[1].u8.xy,   invocation);
    data[invocation].u8.xyz = subgroupShuffleUp(data[2].u8.xyz,  invocation);
    data[invocation].u8     = subgroupShuffleUp(data[3].u8,      invocation);

    data[invocation].u8.x   = subgroupShuffleDown(data[0].u8.x,    invocation);
    data[invocation].u8.xy  = subgroupShuffleDown(data[1].u8.xy,   invocation);
    data[invocation].u8.xyz = subgroupShuffleDown(data[2].u8.xyz,  invocation);
    data[invocation].u8     = subgroupShuffleDown(data[3].u8,      invocation);

    data[invocation].i16.x   = subgroupShuffleUp(data[0].i16.x,    invocation);
    data[invocation].i16.xy  = subgroupShuffleUp(data[1].i16.xy,   invocation);
    data[invocation].i16.xyz = subgroupShuffleUp(data[2].i16.xyz,  invocation);
    data[invocation].i16     = subgroupShuffleUp(data[3].i16,      invocation);

    data[invocation].i16.x   = subgroupShuffleDown(data[0].i16.x,    invocation);
    data[invocation].i16.xy  = subgroupShuffleDown(data[1].i16.xy,   invocation);
    data[invocation].i16.xyz = subgroupShuffleDown(data[2].i16.xyz,  invocation);
    data[invocation].i16     = subgroupShuffleDown(data[3].i16,      invocation);

    data[invocation].u16.x   = subgroupShuffleUp(data[0].u16.x,    invocation);
    data[invocation].u16.xy  = subgroupShuffleUp(data[1].u16.xy,   invocation);
    data[invocation].u16.xyz = subgroupShuffleUp(data[2].u16.xyz,  invocation);
    data[invocation].u16     = subgroupShuffleUp(data[3].u16,      invocation);

    data[invocation].u16.x   = subgroupShuffleDown(data[0].u16.x,    invocation);
    data[invocation].u16.xy  = subgroupShuffleDown(data[1].u16.xy,   invocation);
    data[invocation].u16.xyz = subgroupShuffleDown(data[2].u16.xyz,  invocation);
    data[invocation].u16     = subgroupShuffleDown(data[3].u16,      invocation);

    data[invocation].i64.x   = subgroupShuffleUp(data[0].i64.x,    invocation);
    data[invocation].i64.xy  = subgroupShuffleUp(data[1].i64.xy,   invocation);
    data[invocation].i64.xyz = subgroupShuffleUp(data[2].i64.xyz,  invocation);
    data[invocation].i64     = subgroupShuffleUp(data[3].i64,      invocation);

    data[invocation].i64.x   = subgroupShuffleDown(data[0].i64.x,    invocation);
    data[invocation].i64.xy  = subgroupShuffleDown(data[1].i64.xy,   invocation);
    data[invocation].i64.xyz = subgroupShuffleDown(data[2].i64.xyz,  invocation);
    data[invocation].i64     = subgroupShuffleDown(data[3].i64,      invocation);

    data[invocation].u64.x   = subgroupShuffleUp(data[0].u64.x,    invocation);
    data[invocation].u64.xy  = subgroupShuffleUp(data[1].u64.xy,   invocation);
    data[invocation].u64.xyz = subgroupShuffleUp(data[2].u64.xyz,  invocation);
    data[invocation].u64     = subgroupShuffleUp(data[3].u64,      invocation);

    data[invocation].u64.x   = subgroupShuffleDown(data[0].u64.x,    invocation);
    data[invocation].u64.xy  = subgroupShuffleDown(data[1].u64.xy,   invocation);
    data[invocation].u64.xyz = subgroupShuffleDown(data[2].u64.xyz,  invocation);
    data[invocation].u64     = subgroupShuffleDown(data[3].u64,      invocation);

    data[invocation].f16.x   = subgroupShuffleUp(data[0].f16.x,    invocation);
    data[invocation].f16.xy  = subgroupShuffleUp(data[1].f16.xy,   invocation);
    data[invocation].f16.xyz = subgroupShuffleUp(data[2].f16.xyz,  invocation);
    data[invocation].f16     = subgroupShuffleUp(data[3].f16,      invocation);

    data[invocation].f16.x   = subgroupShuffleDown(data[0].f16.x,    invocation);
    data[invocation].f16.xy  = subgroupShuffleDown(data[1].f16.xy,   invocation);
    data[invocation].f16.xyz = subgroupShuffleDown(data[2].f16.xyz,  invocation);
    data[invocation].f16     = subgroupShuffleDown(data[3].f16,      invocation);
}