Function edgetam_encode_image

sam3.cpp:4833–4985

Full EdgeTAM image encoding: preprocess → RepViT → FPN → state

// Full EdgeTAM image encoding: preprocess → RepViT → FPN → state
static bool edgetam_encode_image(sam3_state& state,
                                  const sam3_model& model,
                                  const sam3_image& image) {
    auto t_start = std::chrono::high_resolution_clock::now();
    const auto& hp = model.hparams;
    const int img_size = sam3_eff_img_size(state, hp);

    fprintf(stderr, "%s: encoding %dx%d image (EdgeTAM RepViT)\n", __func__,
            image.width, image.height);

    state.orig_width = image.width;
    state.orig_height = image.height;

    // ── Preprocess (same ImageNet normalization as SAM2) ─────────────────
    auto img_data = sam2_preprocess_image(image, img_size);

    // ── Build graph ──────────────────────────────────────────────────────
    const size_t buf_size = ggml_tensor_overhead() * 16384 + ggml_graph_overhead() * 2;
    struct ggml_init_params gparams = {
        /*.mem_size   =*/buf_size,
        /*.mem_buffer =*/nullptr,
        /*.no_alloc   =*/true,
    };
    auto* ctx0 = ggml_init(gparams);
    if (!ctx0) {
        fprintf(stderr, "%s: failed to init compute context\n", __func__);
        return false;
    }

    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);

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

    // Build EdgeTAM FPN neck (optimized: mul_mat for 1×1, unified [W,H,C] layout)
    struct ggml_tensor* fpn_outs[4] = {};
    edgetam_build_fpn_neck_graph(ctx0, stage_outs, model, fpn_outs);

    // Mark FPN outputs
    int n_fpn = 4 - hp.scalp;
    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]);
    }

    // Build computation graph
    auto* graph = ggml_new_graph_custom(ctx0, 32768, false);
    for (int i = 0; i < n_fpn; ++i) {
        ggml_build_forward_expand(graph, fpn_outs[i]);
    }

    // ── Allocate + compute ───────────────────────────────────────────────
    auto* galloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));