#version 450 #extension GL_KHR_shader_subgroup_quad: 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 = subgroupQuadBroadcast(data[0].i8.x, 1); data[invocation].i8.xy = subgroupQuadBroadcast(data[1].i8.xy, 1); data[invocation].i8.xyz = subgroupQuadBroadcast(data[2].i8.xyz, 1); data[invocation].i8 = subgroupQuadBroadcast(data[3].i8, 1); data[invocation].i8.x = subgroupQuadSwapHorizontal(data[0].i8.x); data[invocation].i8.xy = subgroupQuadSwapHorizontal(data[1].i8.xy); data[invocation].i8.xyz = subgroupQuadSwapHorizontal(data[2].i8.xyz); data[invocation].i8 = subgroupQuadSwapHorizontal(data[3].i8); data[invocation].i8.x = subgroupQuadSwapVertical(data[0].i8.x); data[invocation].i8.xy = subgroupQuadSwapVertical(data[1].i8.xy); data[invocation].i8.xyz = subgroupQuadSwapVertical(data[2].i8.xyz); data[invocation].i8 = subgroupQuadSwapVertical(data[3].i8); data[invocation].i8.x = subgroupQuadSwapDiagonal(data[0].i8.x); data[invocation].i8.xy = subgroupQuadSwapDiagonal(data[1].i8.xy); data[invocation].i8.xyz = subgroupQuadSwapDiagonal(data[2].i8.xyz); data[invocation].i8 = subgroupQuadSwapDiagonal(data[3].i8); data[invocation].u8.x = subgroupQuadBroadcast(data[0].u8.x, 1); data[invocation].u8.xy = subgroupQuadBroadcast(data[1].u8.xy, 1); data[invocation].u8.xyz = subgroupQuadBroadcast(data[2].u8.xyz, 1); data[invocation].u8 = subgroupQuadBroadcast(data[3].u8, 1); data[invocation].u8.x = subgroupQuadSwapHorizontal(data[0].u8.x); data[invocation].u8.xy = subgroupQuadSwapHorizontal(data[1].u8.xy); data[invocation].u8.xyz = subgroupQuadSwapHorizontal(data[2].u8.xyz); data[invocation].u8 = subgroupQuadSwapHorizontal(data[3].u8); data[invocation].u8.x = subgroupQuadSwapVertical(data[0].u8.x); data[invocation].u8.xy = subgroupQuadSwapVertical(data[1].u8.xy); data[invocation].u8.xyz = subgroupQuadSwapVertical(data[2].u8.xyz); data[invocation].u8 = subgroupQuadSwapVertical(data[3].u8); data[invocation].u8.x = subgroupQuadSwapDiagonal(data[0].u8.x); data[invocation].u8.xy = subgroupQuadSwapDiagonal(data[1].u8.xy); data[invocation].u8.xyz = subgroupQuadSwapDiagonal(data[2].u8.xyz); data[invocation].u8 = subgroupQuadSwapDiagonal(data[3].u8); data[invocation].i16.x = subgroupQuadBroadcast(data[0].i16.x, 1); data[invocation].i16.xy = subgroupQuadBroadcast(data[1].i16.xy, 1); data[invocation].i16.xyz = subgroupQuadBroadcast(data[2].i16.xyz, 1); data[invocation].i16 = subgroupQuadBroadcast(data[3].i16, 1); data[invocation].i16.x = subgroupQuadSwapHorizontal(data[0].i16.x); data[invocation].i16.xy = subgroupQuadSwapHorizontal(data[1].i16.xy); data[invocation].i16.xyz = subgroupQuadSwapHorizontal(data[2].i16.xyz); data[invocation].i16 = subgroupQuadSwapHorizontal(data[3].i16); data[invocation].i16.x = subgroupQuadSwapVertical(data[0].i16.x); data[invocation].i16.xy = subgroupQuadSwapVertical(data[1].i16.xy); data[invocation].i16.xyz = subgroupQuadSwapVertical(data[2].i16.xyz); data[invocation].i16 = subgroupQuadSwapVertical(data[3].i16); data[invocation].i16.x = subgroupQuadSwapDiagonal(data[0].i16.x); data[invocation].i16.xy = subgroupQuadSwapDiagonal(data[1].i16.xy); data[invocation].i16.xyz = subgroupQuadSwapDiagonal(data[2].i16.xyz); data[invocation].i16 = subgroupQuadSwapDiagonal(data[3].i16); data[invocation].u16.x = subgroupQuadBroadcast(data[0].u16.x, 1); data[invocation].u16.xy = subgroupQuadBroadcast(data[1].u16.xy, 1); data[invocation].u16.xyz = subgroupQuadBroadcast(data[2].u16.xyz, 1); data[invocation].u16 = subgroupQuadBroadcast(data[3].u16, 1); data[invocation].u16.x = subgroupQuadSwapHorizontal(data[0].u16.x); data[invocation].u16.xy = subgroupQuadSwapHorizontal(data[1].u16.xy); data[invocation].u16.xyz = subgroupQuadSwapHorizontal(data[2].u16.xyz); data[invocation].u16 = subgroupQuadSwapHorizontal(data[3].u16); data[invocation].u16.x = subgroupQuadSwapVertical(data[0].u16.x); data[invocation].u16.xy = subgroupQuadSwapVertical(data[1].u16.xy); data[invocation].u16.xyz = subgroupQuadSwapVertical(data[2].u16.xyz); data[invocation].u16 = subgroupQuadSwapVertical(data[3].u16); data[invocation].u16.x = subgroupQuadSwapDiagonal(data[0].u16.x); data[invocation].u16.xy = subgroupQuadSwapDiagonal(data[1].u16.xy); data[invocation].u16.xyz = subgroupQuadSwapDiagonal(data[2].u16.xyz); data[invocation].u16 = subgroupQuadSwapDiagonal(data[3].u16); data[invocation].i64.x = subgroupQuadBroadcast(data[0].i64.x, 1); data[invocation].i64.xy = subgroupQuadBroadcast(data[1].i64.xy, 1); data[invocation].i64.xyz = subgroupQuadBroadcast(data[2].i64.xyz, 1); data[invocation].i64 = subgroupQuadBroadcast(data[3].i64, 1); data[invocation].i64.x = subgroupQuadSwapHorizontal(data[0].i64.x); data[invocation].i64.xy = subgroupQuadSwapHorizontal(data[1].i64.xy); data[invocation].i64.xyz = subgroupQuadSwapHorizontal(data[2].i64.xyz); data[invocation].i64 = subgroupQuadSwapHorizontal(data[3].i64); data[invocation].i64.x = subgroupQuadSwapVertical(data[0].i64.x); data[invocation].i64.xy = subgroupQuadSwapVertical(data[1].i64.xy); data[invocation].i64.xyz = subgroupQuadSwapVertical(data[2].i64.xyz); data[invocation].i64 = subgroupQuadSwapVertical(data[3].i64); data[invocation].i64.x = subgroupQuadSwapDiagonal(data[0].i64.x); data[invocation].i64.xy = subgroupQuadSwapDiagonal(data[1].i64.xy); data[invocation].i64.xyz = subgroupQuadSwapDiagonal(data[2].i64.xyz); data[invocation].i64 = subgroupQuadSwapDiagonal(data[3].i64); data[invocation].u64.x = subgroupQuadBroadcast(data[0].u64.x, 1); data[invocation].u64.xy = subgroupQuadBroadcast(data[1].u64.xy, 1); data[invocation].u64.xyz = subgroupQuadBroadcast(data[2].u64.xyz, 1); data[invocation].u64 = subgroupQuadBroadcast(data[3].u64, 1); data[invocation].u64.x = subgroupQuadSwapHorizontal(data[0].u64.x); data[invocation].u64.xy = subgroupQuadSwapHorizontal(data[1].u64.xy); data[invocation].u64.xyz = subgroupQuadSwapHorizontal(data[2].u64.xyz); data[invocation].u64 = subgroupQuadSwapHorizontal(data[3].u64); data[invocation].u64.x = subgroupQuadSwapVertical(data[0].u64.x); data[invocation].u64.xy = subgroupQuadSwapVertical(data[1].u64.xy); data[invocation].u64.xyz = subgroupQuadSwapVertical(data[2].u64.xyz); data[invocation].u64 = subgroupQuadSwapVertical(data[3].u64); data[invocation].u64.x = subgroupQuadSwapDiagonal(data[0].u64.x); data[invocation].u64.xy = subgroupQuadSwapDiagonal(data[1].u64.xy); data[invocation].u64.xyz = subgroupQuadSwapDiagonal(data[2].u64.xyz); data[invocation].u64 = subgroupQuadSwapDiagonal(data[3].u64); data[invocation].f16.x = subgroupQuadBroadcast(data[0].f16.x, 1); data[invocation].f16.xy = subgroupQuadBroadcast(data[1].f16.xy, 1); data[invocation].f16.xyz = subgroupQuadBroadcast(data[2].f16.xyz, 1); data[invocation].f16 = subgroupQuadBroadcast(data[3].f16, 1); data[invocation].f16.x = subgroupQuadSwapHorizontal(data[0].f16.x); data[invocation].f16.xy = subgroupQuadSwapHorizontal(data[1].f16.xy); data[invocation].f16.xyz = subgroupQuadSwapHorizontal(data[2].f16.xyz); data[invocation].f16 = subgroupQuadSwapHorizontal(data[3].f16); data[invocation].f16.x = subgroupQuadSwapVertical(data[0].f16.x); data[invocation].f16.xy = subgroupQuadSwapVertical(data[1].f16.xy); data[invocation].f16.xyz = subgroupQuadSwapVertical(data[2].f16.xyz); data[invocation].f16 = subgroupQuadSwapVertical(data[3].f16); data[invocation].f16.x = subgroupQuadSwapDiagonal(data[0].f16.x); data[invocation].f16.xy = subgroupQuadSwapDiagonal(data[1].f16.xy); data[invocation].f16.xyz = subgroupQuadSwapDiagonal(data[2].f16.xyz); data[invocation].f16 = subgroupQuadSwapDiagonal(data[3].f16); }