Function edgetam_repvit_downsample_forward

sam3.cpp:4768–4791

Downsample: pre_block → spatial DW s=2 → channel expand → FFN + residual

// Downsample: pre_block → spatial DW s=2 → channel expand → FFN + residual
static struct ggml_tensor* edgetam_repvit_downsample_forward(struct ggml_context* ctx,
                                                              struct ggml_tensor* x,
                                                              const edgetam_repvit_downsample& ds) {
    // Pre-block at current channels
    x = edgetam_repvit_block_forward(ctx, x, ds.pre_block);

    // Spatial downsample: DW 3×3 conv, stride=2, pad=1
    auto* ds_w_f32 = (ds.spatial_w->type == GGML_TYPE_F32) ? ds.spatial_w
                     : ggml_cast(ctx, ds.spatial_w, GGML_TYPE_F32);
    x = ggml_conv_2d_dw_direct(ctx, ds_w_f32, x, 2, 2, 1, 1, 1, 1);
    x = edgetam_conv2d_bias(ctx, x, ds.spatial_b);

    // Channel expand: 1×1 conv (mul_mat path)
    x = edgetam_conv1x1_mulmat(ctx, x, ds.channel_w, ds.channel_b);

    // FFN + residual (mul_mat path for 1×1 convs)
    auto* ffn_in = x;
    auto* ffn = edgetam_conv1x1_mulmat(ctx, x, ds.ffn_conv1_w, ds.ffn_conv1_b);
    ffn = ggml_gelu(ctx, ffn);
    ffn = edgetam_conv1x1_mulmat(ctx, ffn, ds.ffn_conv2_w, ds.ffn_conv2_b);
    x = ggml_add(ctx, ffn, ffn_in);

    return x;
}

// Build the full RepViT backbone graph.
// Input: [W, H, 3, 1] image