Function LevenshteinDistances_call

python/stringzillas.c:433–574

}

static PyObject *LevenshteinDistances_call(LevenshteinDistances *self, PyObject *args, PyObject *kwargs) {
    PyObject *a_obj = NULL, *b_obj = NULL, *device_obj = NULL, *out_obj = NULL;

    static char *kwlist[] = {"a", "b", "device", "out", NULL};
    if (!PyArg_ParseTupleAndKeywords(args, kwargs, "OO|OO", kwlist, &a_obj, &b_obj, &device_obj, &out_obj)) return NULL;

    DeviceScope *device_scope = NULL;
    if (device_obj != NULL && device_obj != Py_None) {
        if (!PyObject_TypeCheck(device_obj, &DeviceScopeType)) {
            PyErr_SetString(PyExc_TypeError, "device must be a DeviceScope instance");
            return NULL;
        }
        device_scope = (DeviceScope *)device_obj;
    }

    szs_device_scope_t device_handle = device_scope ? device_scope->handle : default_device_scope;
    sz_size_t kernel_input_size = 0;
    void *kernel_a_texts_punned = NULL;
    void *kernel_b_texts_punned = NULL;
    sz_size_t *kernel_results = NULL;
    sz_size_t kernel_results_stride = sizeof(sz_size_t);
    sz_status_t (*kernel_punned)(szs_levenshtein_distances_t, szs_device_scope_t, void *, void *, sz_size_t *,
                                 sz_size_t, char const **) = NULL;

    // Swap allocators only when using CUDA with a GPU device (inputs must be unified)
    if (requires_unified_memory(self->capabilities))
        if (!try_swap_to_unified_allocator(a_obj) || !try_swap_to_unified_allocator(b_obj)) return NULL;

    // Handle 32-bit tape inputs
    sz_sequence_u32tape_t a_u32tape, b_u32tape;
    sz_bool_t a_is_u32tape = sz_py_export_strings_as_u32tape( //
        a_obj, &a_u32tape.data, &a_u32tape.offsets, &a_u32tape.count);
    sz_bool_t b_is_u32tape = sz_py_export_strings_as_u32tape( //
        b_obj, &b_u32tape.data, &b_u32tape.offsets, &b_u32tape.count);
    if (a_is_u32tape && b_is_u32tape) {
        if (a_u32tape.count != b_u32tape.count) {
            PyErr_SetString(PyExc_ValueError, "Input sequences must have the same length");
            return NULL;
        }

        kernel_input_size = a_u32tape.count;
        kernel_punned = szs_levenshtein_distances_u32tape;
        kernel_a_texts_punned = &a_u32tape;
        kernel_b_texts_punned = &b_u32tape;
    }

    // Handle 64-bit tape inputs
    sz_sequence_u64tape_t a_u64tape, b_u64tape;
    sz_bool_t a_is_u64tape = !a_is_u32tape && sz_py_export_strings_as_u64tape( //
                                                  a_obj, &a_u64tape.data, &a_u64tape.offsets, &a_u64tape.count);
    sz_bool_t b_is_u64tape = !b_is_u32tape && sz_py_export_strings_as_u64tape( //
                                                  b_obj, &b_u64tape.data, &b_u64tape.offsets, &b_u64tape.count);
    if (a_is_u64tape && b_is_u64tape) {
        if (a_u64tape.count != b_u64tape.count) {
            PyErr_SetString(PyExc_ValueError, "Input sequences must have the same length");
            return NULL;
        }
        kernel_input_size = a_u64tape.count;