// CNN v3 — Decoder level 1
// NearestUp2x(bottleneck) + cat(enc1_skip) -> Conv(16->4, 3x3, zero-pad) + FiLM + ReLU
//
// Inputs: bottleneck_tex (rgba32uint, 8xf16)  quarter-res
//         enc1_tex       (rgba32uint, 8xf16)  half-res    (skip connection)
// Output: dec1_out       (rgba16float, 4ch)   half-res    (dispatch at half-res dims)
//
// Weight layout (f16, OIHW + bias):
//   [0 .. 16*4*9)  conv: w[out][in][ky][kx]   (in=16: 8 bottleneck + 8 enc1 skip)
//   [576 .. +4)    bias: b[out]

#include "cnn_v3/common"

const DEC1_IN:  u32 = 16u;
const DEC1_OUT: u32 = 4u;

struct Params {
    weight_offset: u32,
    _pad: vec3u,
    gamma: vec4f,
    beta:  vec4f,
}

@group(0) @binding(0) var bottleneck_tex: texture_2d<u32>;
@group(0) @binding(1) var enc1_tex: texture_2d<u32>;
@group(0) @binding(2) var<storage, read> weights: array<u32>;
@group(0) @binding(3) var<uniform> params: Params;
@group(0) @binding(4) var dec1_out: texture_storage_2d<rgba16float, write>;

// Load 16 concatenated channels at half-res coord hcoord:
//   ch 0-7:  bottleneck nearest-up (bottleneck_tex[hcoord/2])
//   ch 8-15: enc1 skip             (enc1_tex[hcoord])
// Returns zeros for OOB hcoord (zero-padding for the conv).
fn load_dec1_concat(hcoord: vec2i, half_dims: vec2i) -> array<f32, 16> {
    if (hcoord.x < 0 || hcoord.y < 0 || hcoord.x >= half_dims.x || hcoord.y >= half_dims.y) {
        return array<f32, 16>(0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.);
    }
    let quart_dims = half_dims / 2;
    let qc = clamp(hcoord / 2, vec2i(0), quart_dims - vec2i(1));
    let b  = unpack_8ch(bottleneck_tex, qc);
    let s  = unpack_8ch(enc1_tex, hcoord);
    return array<f32, 16>(
        b[0], b[1], b[2], b[3], b[4], b[5], b[6], b[7],
        s[0], s[1], s[2], s[3], s[4], s[5], s[6], s[7]
    );
}

@compute @workgroup_size(8, 8)
fn dec1_main(@builtin(global_invocation_id) id: vec3u) {
    let half_dims = vec2i(textureDimensions(enc1_tex));
    let coord     = vec2i(id.xy);
    if (coord.x >= half_dims.x || coord.y >= half_dims.y) { return; }

    let wo = params.weight_offset;
    var out: array<f32, DEC1_OUT>;

    for (var o: u32 = 0u; o < DEC1_OUT; o++) {
        var sum = get_w(wo, DEC1_OUT * DEC1_IN * 9u + o);  // bias
        for (var ky: i32 = -1; ky <= 1; ky++) {
            for (var kx: i32 = -1; kx <= 1; kx++) {
                let feat = load_dec1_concat(coord + vec2i(kx, ky), half_dims);
                let ki   = u32(ky + 1) * 3u + u32(kx + 1);
                for (var i: u32 = 0u; i < DEC1_IN; i++) {
                    sum += get_w(wo, o * DEC1_IN * 9u + i * 9u + ki) * feat[i];
                }
            }
        }
        out[o] = max(0.0, params.gamma[o] * sum + params.beta[o]);
    }

    textureStore(dec1_out, coord, vec4f(out[0], out[1], out[2], out[3]));
}