// CNN v3 — Bottleneck
// AvgPool2x2(enc1) + Conv(16->16, 3x3, dilation=2) + ReLU  (no FiLM)
//
// Input:  enc1_tex_lo  (rgba32uint, 8xf16)  half-res  ch 0-7
//         enc1_tex_hi  (rgba32uint, 8xf16)  half-res  ch 8-15
// Output: bn_out_lo    (rgba32uint, 8xf16)  quarter-res
//         bn_out_hi    (rgba32uint, 8xf16)  quarter-res
//
// Weight layout (f16, OIHW + bias):
//   [0 .. 16*16*9)  conv: w[out][in][ky*3+kx]
//   [2304 .. +16)   bias: b[out]

#include "cnn_v3/common"

const BN_IN:      u32 = 16u;
const BN_OUT:     u32 = 16u;
const BN_DILATION: i32 = 2;

struct Params {
    weight_offset: u32,
    _pad0: u32, _pad1: u32, _pad2: u32,
}

@group(0) @binding(0) var enc1_tex_lo: texture_2d<u32>;
@group(0) @binding(1) var enc1_tex_hi: 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 bn_out_lo: texture_storage_2d<rgba32uint, write>;
@group(0) @binding(5) var bn_out_hi: texture_storage_2d<rgba32uint, write>;

fn load_enc1_avg(qcoord: vec2i, half_dims: vec2i) -> array<f32, 16> {
    let quart_dims = half_dims / 2;
    if (qcoord.x < 0 || qcoord.y < 0 || qcoord.x >= quart_dims.x || qcoord.y >= quart_dims.y) {
        return array<f32, 16>(0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.);
    }
    let base = qcoord * 2;
    var s: array<f32, BN_IN>;
    for (var dy: i32 = 0; dy < 2; dy++) {
        for (var dx: i32 = 0; dx < 2; dx++) {
            let hc = clamp(base + vec2i(dx, dy), vec2i(0), half_dims - vec2i(1));
            let lo = unpack_8ch(enc1_tex_lo, hc);
            let hi = unpack_8ch(enc1_tex_hi, hc);
            for (var i: u32 = 0u; i < 8u; i++) {
                s[i]      += lo[i];
                s[i + 8u] += hi[i];
            }
        }
    }
    for (var i: u32 = 0u; i < BN_IN; i++) { s[i] *= 0.25; }
    return s;
}

@compute @workgroup_size(8, 8)
fn bottleneck_main(@builtin(global_invocation_id) id: vec3u) {
    let half_dims  = vec2i(textureDimensions(enc1_tex_lo));
    let quart_dims = half_dims / 2;
    let coord      = vec2i(id.xy);
    if (coord.x >= quart_dims.x || coord.y >= quart_dims.y) { return; }

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

    for (var o: u32 = 0u; o < BN_OUT; o++) {
        var sum = get_w(wo, BN_OUT * BN_IN * 9u + o);  // bias
        for (var ky: i32 = -1; ky <= 1; ky++) {
            for (var kx: i32 = -1; kx <= 1; kx++) {
                let feat = load_enc1_avg(coord + vec2i(kx, ky) * BN_DILATION, half_dims);
                let ki   = u32(ky + 1) * 3u + u32(kx + 1);
                for (var i: u32 = 0u; i < BN_IN; i++) {
                    sum += get_w(wo, o * BN_IN * 9u + i * 9u + ki) * feat[i];
                }
            }
        }
        out[o] = max(0.0, sum);
    }

    textureStore(bn_out_lo, coord, vec4u(
        pack2x16float(vec2f(out[0], out[1])),
        pack2x16float(vec2f(out[2], out[3])),
        pack2x16float(vec2f(out[4], out[5])),
        pack2x16float(vec2f(out[6], out[7]))
    ));
    textureStore(bn_out_hi, coord, vec4u(
        pack2x16float(vec2f(out[8],  out[9])),
        pack2x16float(vec2f(out[10], out[11])),
        pack2x16float(vec2f(out[12], out[13])),
        pack2x16float(vec2f(out[14], out[15]))
    ));
}