power_play/src/kernel/kernel_flood.knl
2025-08-11 22:44:18 -05:00

61 lines
2.5 KiB
Plaintext

ConstantBuffer<K_FloodSig> sig : register(b0);
/* ========================== *
* Entry point
* ========================== */
Struct(CsInput)
{
Semantic(uint3, SV_DispatchThreadID);
};
[numthreads(8, 8, 1)]
K_ENTRY void cs(CsInput input)
{
uint2 id = input.SV_DispatchThreadID.xy;
uint2 tex_size = uint2(sig.tex_width, sig.tex_height);
if (id.x < tex_size.x && id.y < tex_size.y) {
Texture2D<Vec4> emittance_tex = GpuResourceFromUrid(sig.emittance_tex_urid);
RWTexture2D<uint2> read_flood_tex = GpuResourceFromUrid(sig.read_flood_tex_urid);
RWTexture2D<uint2> target_flood_tex = GpuResourceFromUrid(sig.target_flood_tex_urid);
int step_len = sig.step_len;
if (step_len == -1) {
/* Seed */
Vec4 emittance = emittance_tex[id];
uint2 seed = uint2(0xFFFF, 0xFFFF);
if (emittance.a > 0) {
seed = id;
}
target_flood_tex[id] = seed;
} else {
/* Flood */
Vec2I32 read_coords[9] = {
(Vec2I32)id + Vec2I32(-step_len, -step_len), /* top left */
(Vec2I32)id + Vec2I32(0 , -step_len), /* top center */
(Vec2I32)id + Vec2I32(+step_len, -step_len), /* top right */
(Vec2I32)id + Vec2I32(-step_len, 0 ), /* center left */
(Vec2I32)id + Vec2I32(0 , 0 ), /* center center */
(Vec2I32)id + Vec2I32(+step_len, 0 ), /* center right */
(Vec2I32)id + Vec2I32(-step_len, +step_len), /* bottom left */
(Vec2I32)id + Vec2I32(0 , +step_len), /* bottom center */
(Vec2I32)id + Vec2I32(+step_len, +step_len) /* bottom right */
};
uint2 closest_seed = uint2(0xFFFF, 0xFFFF);
u32 closest_seed_len_sq = 0xFFFFFFFF;
for (int i = 0; i < 9; ++i) {
Vec2I32 coord = read_coords[i];
if (coord.x >= 0 && coord.x < (int)tex_size.x && coord.y >= 0 && coord.y < (int)tex_size.y) {
uint2 seed = read_flood_tex[coord];
Vec2I32 dist_vec = (Vec2I32)id - (Vec2I32)seed;
u32 dist_len_sq = dot(dist_vec, dist_vec);
if (dist_len_sq < closest_seed_len_sq) {
closest_seed = seed;
closest_seed_len_sq = dist_len_sq;
}
}
}
target_flood_tex[id] = closest_seed;
}
}
}