Function sam3_segment_pvs

sam3.cpp:10765–11115

*****************************************************************************/

sam3_result sam3_segment_pvs(sam3_state& state,
                             const sam3_model& model,
                             const sam3_pvs_params& params) {
#if SAM3_LOG_LEVEL >= 1
    auto t_start = std::chrono::high_resolution_clock::now();
#endif
    const auto& hp = model.hparams;
    const int D = hp.sam_embed_dim;                      // 256
    const int H = sam3_eff_feat_size(state, hp);
    const int num_mask_tokens = hp.sam_n_multimask + 1;  // 4
    const int eff_img_size = sam3_eff_img_size(state, hp);
    sam3_result result;

    // ── Validate ─────────────────────────────────────────────────────────
    if (!state.neck_trk[0]) {
        fprintf(stderr, "%s: image not encoded — call sam3_encode_image first\n", __func__);
        return result;
    }
    if (params.pos_points.empty() && !params.use_box) {
        fprintf(stderr, "%s: no prompts provided (need at least one point or box)\n", __func__);
        return result;
    }

    SAM3_LOG(2, "%s: %zu pos points, %zu neg points, box=%s, multimask=%s\n",
             __func__, params.pos_points.size(), params.neg_points.size(),
             params.use_box ? "yes" : "no", params.multimask ? "yes" : "no");

    // ── Build computation graph ──────────────────────────────────────────
    const size_t buf_size = ggml_tensor_overhead() * 8192 + ggml_graph_overhead() * 2;
    struct ggml_init_params gparams = {
        /*.mem_size   =*/buf_size,
        /*.mem_buffer =*/nullptr,
        /*.no_alloc   =*/true,
    };
    struct ggml_context* ctx0 = ggml_init(gparams);
    if (!ctx0) {
        fprintf(stderr, "%s: failed to init compute context\n", __func__);
        return result;
    }

    // ── SAM prompt encoder (CPU pre-compute + input tensors) ─────────────
    auto pe_out = sam3_build_sam_pe(ctx0, params, D, H);

    // ── Create fresh input tensors for tracker features ──────────────────
    // CRITICAL: Do NOT use state.neck_trk[*] directly in graph operations
    // (like ggml_add). They are tensors from a previous graph, and
    // ggml_build_forward_expand traces all ancestors — which would pull in
    // the ENTIRE ViT + neck recomputation (2500+ nodes, ~37 seconds).
    // Instead: create fresh input tensors and copy data from state on CPU.
    const int H0 = H * 4;  // 288
    const int H1 = H * 2;  // 144

    auto* image_feats = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, D, H, H, 1);
    ggml_set_name(image_feats, "sam_dec_image_feats");
    ggml_set_input(image_feats);

    auto* feat_s0 = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, D, H0, H0, 1);
    ggml_set_name(feat_s0, "pvs_feat_s0");