#version 450

#extension GL_KHR_shader_subgroup_rotate: enable
#extension GL_EXT_shader_subgroup_extended_types_int8: enable
#extension GL_EXT_shader_subgroup_extended_types_int16: enable
#extension GL_EXT_shader_subgroup_extended_types_int64: enable
#extension GL_EXT_shader_subgroup_extended_types_float16: enable

layout (local_size_x = 8) in;

readonly buffer roblock
{
    uint delta;
} ro;

layout(binding = 0) buffer Buffers
{
    i8vec4 i8;
    u8vec4 u8;
    i16vec4 i16;
    u16vec4 u16;
    i64vec4 i64;
    u64vec4 u64;
    f16vec4 f16;
} data[4];

void main()
{
    uint delta = ro.delta;

    data[delta].i8.x   = subgroupRotate(data[0].i8.x,    delta);
    data[delta].i8.xy  = subgroupRotate(data[1].i8.xy,   delta);
    data[delta].i8.xyz = subgroupRotate(data[2].i8.xyz,  delta);
    data[delta].i8     = subgroupRotate(data[3].i8,      delta);

    data[delta].i8.x   = subgroupClusteredRotate(data[0].i8.x,    delta, 1);
    data[delta].i8.xy  = subgroupClusteredRotate(data[1].i8.xy,   delta, 1);
    data[delta].i8.xyz = subgroupClusteredRotate(data[2].i8.xyz,  delta, 1);
    data[delta].i8     = subgroupClusteredRotate(data[3].i8,      delta, 1);

    data[delta].u8.x   = subgroupRotate(data[0].u8.x,    delta);
    data[delta].u8.xy  = subgroupRotate(data[1].u8.xy,   delta);
    data[delta].u8.xyz = subgroupRotate(data[2].u8.xyz,  delta);
    data[delta].u8     = subgroupRotate(data[3].u8,      delta);

    data[delta].u8.x   = subgroupClusteredRotate(data[0].u8.x,    delta, 1);
    data[delta].u8.xy  = subgroupClusteredRotate(data[1].u8.xy,   delta, 1);
    data[delta].u8.xyz = subgroupClusteredRotate(data[2].u8.xyz,  delta, 1);
    data[delta].u8     = subgroupClusteredRotate(data[3].u8,      delta, 1);

    data[delta].i16.x   = subgroupRotate(data[0].i16.x,    delta);
    data[delta].i16.xy  = subgroupRotate(data[1].i16.xy,   delta);
    data[delta].i16.xyz = subgroupRotate(data[2].i16.xyz,  delta);
    data[delta].i16     = subgroupRotate(data[3].i16,      delta);

    data[delta].i16.x   = subgroupClusteredRotate(data[0].i16.x,    delta, 1);
    data[delta].i16.xy  = subgroupClusteredRotate(data[1].i16.xy,   delta, 1);
    data[delta].i16.xyz = subgroupClusteredRotate(data[2].i16.xyz,  delta, 1);
    data[delta].i16     = subgroupClusteredRotate(data[3].i16,      delta, 1);

    data[delta].u16.x   = subgroupRotate(data[0].u16.x,    delta);
    data[delta].u16.xy  = subgroupRotate(data[1].u16.xy,   delta);
    data[delta].u16.xyz = subgroupRotate(data[2].u16.xyz,  delta);
    data[delta].u16     = subgroupRotate(data[3].u16,      delta);

    data[delta].u16.x   = subgroupClusteredRotate(data[0].u16.x,    delta, 1);
    data[delta].u16.xy  = subgroupClusteredRotate(data[1].u16.xy,   delta, 1);
    data[delta].u16.xyz = subgroupClusteredRotate(data[2].u16.xyz,  delta, 1);
    data[delta].u16     = subgroupClusteredRotate(data[3].u16,      delta, 1);

    data[delta].i64.x   = subgroupRotate(data[0].i64.x,    delta);
    data[delta].i64.xy  = subgroupRotate(data[1].i64.xy,   delta);
    data[delta].i64.xyz = subgroupRotate(data[2].i64.xyz,  delta);
    data[delta].i64     = subgroupRotate(data[3].i64,      delta);

    data[delta].i64.x   = subgroupClusteredRotate(data[0].i64.x,    delta, 1);
    data[delta].i64.xy  = subgroupClusteredRotate(data[1].i64.xy,   delta, 1);
    data[delta].i64.xyz = subgroupClusteredRotate(data[2].i64.xyz,  delta, 1);
    data[delta].i64     = subgroupClusteredRotate(data[3].i64,      delta, 1);

    data[delta].u64.x   = subgroupRotate(data[0].u64.x,    delta);
    data[delta].u64.xy  = subgroupRotate(data[1].u64.xy,   delta);
    data[delta].u64.xyz = subgroupRotate(data[2].u64.xyz,  delta);
    data[delta].u64     = subgroupRotate(data[3].u64,      delta);

    data[delta].u64.x   = subgroupClusteredRotate(data[0].u64.x,    delta, 1);
    data[delta].u64.xy  = subgroupClusteredRotate(data[1].u64.xy,   delta, 1);
    data[delta].u64.xyz = subgroupClusteredRotate(data[2].u64.xyz,  delta, 1);
    data[delta].u64     = subgroupClusteredRotate(data[3].u64,      delta, 1);

    data[delta].f16.x   = subgroupRotate(data[0].f16.x,    delta);
    data[delta].f16.xy  = subgroupRotate(data[1].f16.xy,   delta);
    data[delta].f16.xyz = subgroupRotate(data[2].f16.xyz,  delta);
    data[delta].f16     = subgroupRotate(data[3].f16,      delta);

    data[delta].f16.x   = subgroupClusteredRotate(data[0].f16.x,    delta, 1);
    data[delta].f16.xy  = subgroupClusteredRotate(data[1].f16.xy,   delta, 1);
    data[delta].f16.xyz = subgroupClusteredRotate(data[2].f16.xyz,  delta, 1);
    data[delta].f16     = subgroupClusteredRotate(data[3].f16,      delta, 1);
}