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