| struct f_inputs { |
| uint tint_local_index : SV_GroupIndex; |
| }; |
| |
| |
| cbuffer cbuffer_u : register(b0) { |
| uint4 u[4]; |
| }; |
| groupshared float4x3 w; |
| float4x3 v(uint start_byte_offset) { |
| return float4x3(asfloat(u[(start_byte_offset / 16u)].xyz), asfloat(u[((16u + start_byte_offset) / 16u)].xyz), asfloat(u[((32u + start_byte_offset) / 16u)].xyz), asfloat(u[((48u + start_byte_offset) / 16u)].xyz)); |
| } |
| |
| void f_inner(uint tint_local_index) { |
| if ((tint_local_index < 1u)) { |
| w = float4x3((0.0f).xxx, (0.0f).xxx, (0.0f).xxx, (0.0f).xxx); |
| } |
| GroupMemoryBarrierWithGroupSync(); |
| w = v(0u); |
| w[1u] = asfloat(u[0u].xyz); |
| w[1u] = asfloat(u[0u].xyz).zxy; |
| w[0u].y = asfloat(u[1u].x); |
| } |
| |
| [numthreads(1, 1, 1)] |
| void f(f_inputs inputs) { |
| f_inner(inputs.tint_local_index); |
| } |
| |