Function sam3_profile_edgetam_encode

sam3.cpp:5096–5422

}

bool sam3_profile_edgetam_encode(const sam3_model& model,
                                  const sam3_image& image,
                                  int n_threads,
                                  int n_warmup,
                                  int n_iter) {
    const auto& hp = model.hparams;
    if (!hp.is_edgetam()) {
        fprintf(stderr, "%s: model is not EdgeTAM\n", __func__);
        return false;
    }

    const int img_size = hp.img_size;
    fprintf(stderr, "%s: profiling EdgeTAM RepViT+FPN on %dx%d image\n",
            __func__, img_size, img_size);
    fprintf(stderr, "  warmup=%d, iterations=%d, threads=%d\n",
            n_warmup, n_iter, n_threads);
    fprintf(stderr, "  backend: %s\n", ggml_backend_name(model.backend));

    // Stage config
    fprintf(stderr, "  stages: [%d, %d, %d, %d] blocks, channels: [%d, %d, %d, %d]\n",
            hp.repvit_stages[0], hp.repvit_stages[1],
            hp.repvit_stages[2], hp.repvit_stages[3],
            hp.repvit_channels[0], hp.repvit_channels[1],
            hp.repvit_channels[2], hp.repvit_channels[3]);

    // ── Preprocess image ────────────────────────────────────────────────
    auto img_data = sam2_preprocess_image(image, img_size);

    // ════════════════════════════════════════════════════════════════════
    // PART 1: Full graph — op summary + total timing
    // ════════════════════════════════════════════════════════════════════
    fprintf(stderr, "\n=== FULL GRAPH (RepViT + FPN) ===\n");
    {
        const size_t buf_size = ggml_tensor_overhead() * 16384 + ggml_graph_overhead() * 2;
        struct ggml_init_params gp = {buf_size, nullptr, true};
        auto* ctx0 = ggml_init(gp);

        auto* inp = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, img_size, img_size, 3, 1);
        ggml_set_name(inp, "input_image");
        ggml_set_input(inp);

        struct ggml_tensor* stage_outs[4] = {};
        edgetam_build_repvit_graph(ctx0, inp, model, stage_outs);

        struct ggml_tensor* fpn_outs[4] = {};
        edgetam_build_fpn_neck_graph(ctx0, stage_outs, model, fpn_outs);

        int n_fpn = 4 - hp.scalp;
        auto* graph = ggml_new_graph_custom(ctx0, 32768, false);
        for (int i = 0; i < n_fpn; ++i) {
            char name[64];
            snprintf(name, sizeof(name), "fpn_out_%d", i);
            ggml_set_name(fpn_outs[i], name);
            ggml_set_output(fpn_outs[i]);
            ggml_build_forward_expand(graph, fpn_outs[i]);
        }

        sam3_profile_print_op_summary(graph, "Full RepViT+FPN");