bblean 0.6.0b1__cp313-cp313-macosx_10_13_universal2.whl

bblean/csrc/README.md

@@ -0,0 +1 @@
+# C++ extensions for accellerated similarity calculations

bblean/csrc/similarity.cpp

@@ -0,0 +1,521 @@
+#include <pybind11/numpy.h>
+#include <pybind11/pybind11.h>
+#include <pybind11/stl.h>
+
+#include <algorithm>
+#include <cmath>
+#include <cstddef>
+#include <cstdint>
+#include <iostream>
+#include <numeric>
+#include <optional>
+#include <stdexcept>
+#include <vector>
+
+// Scalar popcount intrinsics:
+#if defined(__SSE_4_2__) || defined(_M_SSE4_2)
+// Compiler-portable, but *not available in systems that do not have SSE*
+// (which should be almost no CPUs nowadays)
+// Not actually vector instructions, they just live in the SSE header
+// Should be *exactly as fast* as __(builtin_)popcnt(ll) (compile to the same
+// code)
+//
+// nmmintrin.h is the SSE4.2 intrinsics (only) header for all compilers
+// NOTE: This ifdef is probably overkill, almost all cases should be covered by
+// the GCC|Clang|MSVC ifdefs, but it doesn't hurt to add it
+#include <nmmintrin.h>
+#define POPCOUNT_32 _mm_popcnt_u32
+#define POPCOUNT_64 _mm_popcnt_u64
+#elif defined(_MSC_VER)
+// Windows (MSVC compiler)
+#include <intrin.h>
+#define POPCOUNT_32 __popcnt
+#define POPCOUNT_64 __popcnt64
+#elif defined(__GNUC__) || defined(__clang__)
+// GCC | Clang
+#define POPCOUNT_32 __builtin_popcount
+#define POPCOUNT_64 __builtin_popcountll
+#else
+// If popcnt is not hardware supported numpy rolls out its own hand-coded
+// version, fail for simplicity since it is not worth it to support those archs
+#error "Popcount not supported in target architecture"
+#endif
+
+// TODO: See if worth it to use vector popcount intrinsics (AVX-512, only some
+// CPU) like jt_sim_packed
+namespace py = pybind11;
+
+template <typename T>
+using CArrayForcecast =
+    py::array_t<T, py::array::c_style | py::array::forcecast>;
+
+auto is_8byte_aligned(const py::array_t<uint8_t>& a) -> bool {
+    // Convert between ptr and integer requires reinterpret
+    return reinterpret_cast<std::uintptr_t>(a.data()) % alignof(uint64_t) == 0;
+}
+
+auto print_8byte_alignment_check(const py::array_t<uint8_t>& arr) -> void {
+    py::print("arr buf addr: ", reinterpret_cast<std::uintptr_t>(arr.data()));
+    py::print("uint64_t alignment requirement: ", alignof(uint64_t));
+    py::print("Is 8-byte aligned: ", is_8byte_aligned(arr));
+}
+
+uint32_t _popcount_1d(const py::array_t<uint8_t>& arr) {
+    if (arr.ndim() != 1) {
+        throw std::runtime_error("Input array must be 1-dimensional");
+    }
+#ifdef DEBUG_LOGS
+    print_8byte_alignment_check(arr);
+#endif
+    uint32_t count{0};  // Output scalar
+    py::ssize_t steps = arr.shape(0);
+    if (is_8byte_aligned(arr) and (steps % 64 == 0)) {
+#ifdef DEBUG_LOGS
+        py::print("DEBUG: _popcount_1d fn triggered uint64 + popcount 64");
+#endif
+        // Aligned to 64-bit boundary, interpret as uint64_t
+        steps /= sizeof(uint64_t);
+        auto in_cptr = static_cast<const uint64_t*>(arr.request().ptr);
+        for (py::ssize_t i{0}; i != steps; ++i) {  // not auto-vec by GCC
+            count += POPCOUNT_64(in_cptr[i]);
+        }
+        return count;
+    }
+
+#ifdef DEBUG_LOGS
+    py::print("DEBUG: _popcount_1d fn triggered uint8 + popcount 32");
+#endif
+    // Misaligned, loop over bytes
+    auto in_cptr = arr.data();
+    for (py::ssize_t i{0}; i != steps; ++i) {  // not auto-vec by GCC
+        count += POPCOUNT_32(in_cptr[i]);      // uint8 promoted to uint32
+    }
+    return count;
+}
+
+// TODO: Currently this is pretty slow unless hitting the "uint64_t" branch,
+// maybe two pass approach? first compute all popcounts, then sum (Numpy does
+// this). Maybe the additions could be auto-vec?
+py::array_t<uint32_t> _popcount_2d(const CArrayForcecast<uint8_t>& arr) {
+    if (arr.ndim() != 2) {
+        throw std::runtime_error("Input array must be 2-dimensional");
+    }
+    const py::ssize_t n_samples = arr.shape(0);
+
+    auto out = py::array_t<uint32_t>(n_samples);
+    auto out_ptr = out.mutable_data();
+    std::memset(out_ptr, 0, out.nbytes());
+
+#ifdef DEBUG_LOGS
+    print_8byte_alignment_check(arr);
+#endif
+    py::ssize_t steps = arr.shape(1);
+    if (is_8byte_aligned(arr) and (steps % 64 == 0)) {
+#ifdef DEBUG_LOGS
+        py::print("DEBUG: _popcount_2d fn triggered uint64 + popcount 64");
+#endif
+        // Aligned to 64-bit boundary, interpret as uint64_t
+        steps /= sizeof(uint64_t);
+        auto in_cptr = static_cast<const uint64_t*>(arr.request().ptr);
+        for (py::ssize_t i{0}; i != n_samples; ++i) {  // not auto-vec by GCC
+            const uint64_t* row_cptr = in_cptr + i * steps;
+            for (py::ssize_t j{0}; j != steps; ++j) {  // not auto-vec by GCC
+                out_ptr[i] += POPCOUNT_64(row_cptr[j]);
+            }
+        }
+        return out;
+    }
+
+#ifdef DEBUG_LOGS
+    py::print("DEBUG: _popcount_2d fn triggered uint8 + popcount 32");
+#endif
+    // Misaligned, loop over bytes
+    auto in_cptr = arr.data();
+    for (py::ssize_t i{0}; i != n_samples; ++i) {  // not auto-vec by GCC
+        const uint8_t* row_cptr = in_cptr + i * steps;
+        for (py::ssize_t j{0}; j != steps; ++j) {  // not auto-vec by GCC
+            out_ptr[i] += POPCOUNT_32(row_cptr[j]);
+        }
+    }
+    return out;
+}
+
+// The BitToByte table has shape (256, 8), and holds, for each
+// value in the range 0-255, a row with the 8 associated bits as uint8_t values
+constexpr std::array<std::array<uint8_t, 8>, 256> makeByteToBitsLookupTable() {
+    std::array<std::array<uint8_t, 8>, 256> byteToBits{};
+    for (int i{0}; i != 256; ++i) {
+        for (int b{0}; b != 8; ++b) {
+            // Shift right by b and, and fetch the least-significant-bit by
+            // and'ng with 1 = 000...1
+            byteToBits[i][7 - b] = (i >> b) & 1;
+        }
+    }
+    return byteToBits;
+}
+
+constexpr auto BYTE_TO_BITS = makeByteToBitsLookupTable();
+
+py::array_t<uint8_t> _nochecks_unpack_fingerprints_1d(
+    const CArrayForcecast<uint8_t>& packed_fps,
+    std::optional<py::ssize_t> n_features_opt = std::nullopt) {
+    py::ssize_t n_bytes = packed_fps.shape(0);
+    py::ssize_t n_features = n_features_opt.value_or(n_bytes * 8);
+    if (n_features % 8 != 0) {
+        throw std::runtime_error("Only n_features divisible by 8 is supported");
+    }
+    auto out = py::array_t<uint8_t>(n_features);
+    auto out_ptr = out.mutable_data();
+    auto in_cptr = packed_fps.data();
+    for (py::ssize_t j{0}; j != n_features; j += 8) {  // not auto-vec by GCC
+        // Copy the next 8 uint8 values in one go
+        std::memcpy(out_ptr + j, BYTE_TO_BITS[in_cptr[j / 8]].data(), 8);
+    }
+    return out;
+}
+
+py::array_t<uint8_t> _nochecks_unpack_fingerprints_2d(
+    const CArrayForcecast<uint8_t>& packed_fps,
+    std::optional<py::ssize_t> n_features_opt = std::nullopt) {
+    py::ssize_t n_samples = packed_fps.shape(0);
+    py::ssize_t n_bytes = packed_fps.shape(1);
+    py::ssize_t n_features = n_features_opt.value_or(n_bytes * 8);
+    if (n_features % 8 != 0) {
+        throw std::runtime_error("Only features divisible by 8 is supported");
+    }
+    auto out = py::array_t<uint8_t>({n_samples, n_features});
+    // Unchecked accessors (benchmarked and there is no real advantage to using
+    // ptrs)
+    auto acc_in = packed_fps.unchecked<2>();
+    auto acc_out = out.mutable_unchecked<2>();
+
+    for (py::ssize_t i{0}; i != n_samples; ++i) {  // not auto-vec by GCC
+        for (py::ssize_t j{0}; j != n_features;
+             j += 8) {  // not auto-vec by GCC
+            // Copy the next 8 uint8 values in one go
+            std::memcpy(&acc_out(i, j), BYTE_TO_BITS[acc_in(i, j / 8)].data(),
+                        8);
+        }
+    }
+    return out;
+}
+
+// Wrapper over _nochecks_unpack_fingerprints that performs ndim checks
+py::array_t<uint8_t> unpack_fingerprints(
+    const CArrayForcecast<uint8_t>& packed_fps,
+    std::optional<py::ssize_t> n_features_opt = std::nullopt) {
+    if (packed_fps.ndim() == 1) {
+        return _nochecks_unpack_fingerprints_1d(packed_fps, n_features_opt);
+    }
+    if (packed_fps.ndim() == 2) {
+        return _nochecks_unpack_fingerprints_2d(packed_fps, n_features_opt);
+    }
+    throw std::runtime_error("Input array must be 1- or 2-dimensional");
+}
+
+template <typename T>
+py::array_t<uint8_t> centroid_from_sum(const CArrayForcecast<T>& linear_sum,
+                                       int64_t n_samples, bool pack = true) {
+    if (linear_sum.ndim() != 1) {
+        throw std::runtime_error("linear_sum must be 1-dimensional");
+    }
+
+    py::ssize_t n_features = linear_sum.shape(0);
+    auto linear_sum_cptr = linear_sum.data();
+
+    py::array_t<uint8_t> centroid_unpacked(n_features);
+    auto centroid_unpacked_ptr = centroid_unpacked.mutable_data();
+    if (n_samples <= 1) {
+        for (int i{0}; i != n_features;
+             ++i) {  // yes auto-vec by GCC (versioned due to possible alias)
+            // Cast not required, but added for clarity since this is a
+            // narrowing conversion. if n_samples <= 1 then linear_sum is
+            // guaranteed to have a value that a uint8_t can hold (it should be
+            // 0 or 1)
+            // memcpy not possible due to the required cast
+            centroid_unpacked_ptr[i] = static_cast<uint8_t>(linear_sum_cptr[i]);
+        }
+    } else {
+        auto threshold = n_samples * 0.5;
+        for (int i{0}; i != n_features; ++i) {  // not auto-vec by GCC
+            centroid_unpacked_ptr[i] =
+                (linear_sum_cptr[i] >= threshold) ? 1 : 0;
+        }
+    }
+
+    if (not pack) {
+        return centroid_unpacked;
+    }
+
+    auto centroid_unpacked_cptr = centroid_unpacked.data();
+    int n_bytes = (n_features + 7) / 8;
+    auto centroid_packed = py::array_t<uint8_t>(n_bytes);
+    auto centroid_packed_ptr = centroid_packed.mutable_data();
+    std::memset(centroid_packed_ptr, 0, centroid_packed.nbytes());
+
+    // Slower than numpy, due to lack of SIMD
+    // The following loop is *marginally slower* (benchmkd') than the
+    // implemented one: for (int i{0}; i != n_features; ++i) {
+    //    if (centroid_unpacked_cptr[i]) {
+    //        centroid_packed_ptr[i / 8] |= (1 << (7 - (i % 8)));
+    //    }
+    //  }
+    //  TODO: Check if GCC is auto-vectorizing
+    for (int i{0}, stride{0}; i != n_bytes; i++, stride += 8) {
+        for (int b{0}; b != 8; ++b) {
+            centroid_packed_ptr[i] <<= 1;
+            centroid_packed_ptr[i] |= centroid_unpacked_cptr[stride + b];
+        }
+    }
+    return centroid_packed;
+}
+
+double jt_isim_from_sum(const CArrayForcecast<uint64_t>& linear_sum,
+                        int64_t n_objects) {
+    if (n_objects < 2) {
+        PyErr_WarnEx(PyExc_RuntimeWarning,
+                     "Invalid n_objects in isim. Expected n_objects >= 2", 1);
+        return std::numeric_limits<double>::quiet_NaN();
+    }
+    if (linear_sum.ndim() != 1) {
+        throw std::runtime_error("linear_sum must be a 1D array");
+    }
+    py::ssize_t n_features = linear_sum.shape(0);
+
+    auto in_cptr = linear_sum.data();
+    uint64_t sum_kq{0};
+    for (py::ssize_t i{0}; i != n_features; ++i) {  // yes auto-vec by GCC
+        sum_kq += in_cptr[i];
+    }
+
+    if (sum_kq == 0) {
+        return 1.0;
+    }
+
+    uint64_t sum_kqsq{0};
+    for (py::ssize_t i{0}; i != n_features; ++i) {  // yes auto-vec by GCC
+        sum_kqsq += in_cptr[i] * in_cptr[i];
+    }
+    auto a = (sum_kqsq - sum_kq) / 2.0;
+    return a / ((a + (n_objects * sum_kq)) - sum_kqsq);
+}
+
+// Contraint: T must be uint64_t or uint8_t
+template <typename T>
+void _calc_arr_vec_jt(const py::array_t<uint8_t>& arr,
+                      const py::array_t<uint8_t>& vec,
+                      const py::ssize_t n_samples, const py::ssize_t n_features,
+                      const uint32_t vec_popcount,
+                      const py::array_t<uint32_t>& cardinalities,
+                      py::array_t<double>& out) {
+    const py::ssize_t steps = n_features / sizeof(T);
+    auto arr_cptr = static_cast<const T*>(arr.request().ptr);
+    auto vec_cptr = static_cast<const T*>(vec.request().ptr);
+    auto card_cptr = cardinalities.data();
+    auto out_ptr = out.mutable_data();
+
+    for (py::ssize_t i{0}; i != n_samples; ++i) {  // not auto-vec by GCC
+        const T* arr_row_cptr = arr_cptr + i * steps;
+        uint32_t intersection{0};
+        for (py::ssize_t j{0}; j != steps; ++j) {  // not auto-vec by GCC
+            if constexpr (std::is_same_v<T, uint64_t>) {
+                intersection += POPCOUNT_64(arr_row_cptr[j] & vec_cptr[j]);
+            } else {
+                intersection += POPCOUNT_32(arr_row_cptr[j] & vec_cptr[j]);
+            }
+        }
+        auto denominator = card_cptr[i] + vec_popcount - intersection;
+        // Cast is technically unnecessary since std::max promotes to double,
+        // but added here for clarity (should compile to nop)
+        out_ptr[i] =
+            intersection / std::max(static_cast<double>(denominator), 1.0);
+    }
+}
+
+// # NOTE: This function is the bottleneck for bb compute calculations
+// In this function, _popcount_2d takes around ~25% of the time, _popcount_1d
+// around 5%. The internal loop with the popcounts is also quite heavy.
+// TODO: Investigate simple SIMD vectorization of these loops
+// TODO: Does this function return a copy?
+py::array_t<double> jt_sim_packed_precalc_cardinalities(
+    const py::array_t<uint8_t>& arr, const py::array_t<uint8_t>& vec,
+    const py::array_t<uint32_t>& cardinalities) {
+    py::ssize_t n_samples = arr.shape(0);
+    py::ssize_t n_features = arr.shape(1);
+    if (arr.ndim() != 2 || vec.ndim() != 1) {
+        throw std::runtime_error("arr must be 2D, vec must be 1D");
+    }
+    if (n_features != vec.shape(0)) {
+        throw std::runtime_error(
+            "Shapes should be (N, F) for arr and (F,) for vec");
+    }
+    auto out = py::array_t<double>(n_samples);
+
+    if (is_8byte_aligned(arr) and is_8byte_aligned(vec) and
+        (n_features % 64 == 0)) {
+#ifdef DEBUG_LOGS
+        py::print("DEBUG: jt_sim_packed fn triggered uint64 + popcount 64");
+#endif
+        // Aligned to 64-bit boundary, interpret as uint64_t
+        _calc_arr_vec_jt<uint64_t>(arr, vec, n_samples, n_features,
+                                   _popcount_1d(vec), cardinalities, out);
+        return out;
+    }
+
+#ifdef DEBUG_LOGS
+    py::print("DEBUG: jt_sim_packed fn triggered uint8 + popcount 32");
+#endif
+    // Misaligned, loop over bytes
+    _calc_arr_vec_jt<uint8_t>(arr, vec, n_samples, n_features,
+                              _popcount_1d(vec), cardinalities, out);
+    return out;
+}
+
+py::array_t<double> _jt_sim_arr_vec_packed(const py::array_t<uint8_t>& arr,
+                                  const py::array_t<uint8_t>& vec) {
+    return jt_sim_packed_precalc_cardinalities(arr, vec, _popcount_2d(arr));
+}
+
+// NOTE: This is only *slightly* faster for C++ than numpy, **only if the
+// array is uint8_t** if the array is uint64 already, it is slower
+template <typename T>
+py::array_t<uint64_t> add_rows(const CArrayForcecast<T>& arr) {
+    if (arr.ndim() != 2) {
+        throw std::runtime_error("Input array must be 2-dimensional");
+    }
+    auto arr_ptr = arr.data();
+    auto out = py::array_t<uint64_t>(arr.shape(1));
+    auto out_ptr = out.mutable_data();
+    std::memset(out_ptr, 0, out.nbytes());
+    py::ssize_t n_samples = arr.shape(0);
+    py::ssize_t n_features = arr.shape(1);
+    // Check GCC / CLang vectorize this
+    for (py::ssize_t i = 0; i < n_samples; ++i) {
+        const uint8_t* arr_row_ptr = arr_ptr + i * n_features;
+        for (py::ssize_t j = 0; j < n_features; ++j) {
+            out_ptr[j] += static_cast<uint64_t>(arr_row_ptr[j]);
+        }
+    }
+    return out;
+}
+
+double jt_isim_unpacked_u8(const CArrayForcecast<uint8_t>& arr) {
+    return jt_isim_from_sum(add_rows<uint8_t>(arr), arr.shape(0));
+}
+
+double jt_isim_packed_u8(
+    const CArrayForcecast<uint8_t>& arr,
+    std::optional<py::ssize_t> n_features_opt = std::nullopt) {
+    return jt_isim_from_sum(add_rows<uint8_t>(unpack_fingerprints(arr, n_features_opt)),
+                            arr.shape(0));
+}
+
+py::tuple jt_most_dissimilar_packed(
+    CArrayForcecast<uint8_t> fps_packed,
+    std::optional<py::ssize_t> n_features_opt = std::nullopt) {
+    if (fps_packed.ndim() != 2) {
+        throw std::runtime_error("Input array must be 2-dimensional");
+    }
+    py::ssize_t n_samples = fps_packed.shape(0);
+    py::ssize_t n_features_packed = fps_packed.shape(1);
+
+    auto fps_unpacked =
+        _nochecks_unpack_fingerprints_2d(fps_packed, n_features_opt);
+    py::ssize_t n_features_unpacked = fps_unpacked.shape(1);
+
+    auto linear_sum = py::array_t<uint64_t>(n_features_unpacked);
+    auto linear_sum_ptr = linear_sum.mutable_data();
+    std::memset(linear_sum_ptr, 0, linear_sum.nbytes());
+
+    // TODO: This sum could be vectorized manually or automatically
+    auto fps_unpacked_ptr = fps_unpacked.data();
+    for (py::ssize_t i{0}; i != n_samples; ++i) {
+        const uint8_t* row_cptr = fps_unpacked_ptr + i * n_features_unpacked;
+        for (py::ssize_t j{0}; j != n_features_unpacked;
+             ++j) {  // yes auto-vec by GCC (versioned due to possible alias)
+            linear_sum_ptr[j] += row_cptr[j];
+        }
+    }
+
+    auto centroid_packed =
+        centroid_from_sum<uint64_t>(linear_sum, n_samples, true);
+    auto cardinalities = _popcount_2d(fps_packed);
+
+    auto sims_cent = jt_sim_packed_precalc_cardinalities(
+        fps_packed, centroid_packed, cardinalities);
+    auto sims_cent_ptr = sims_cent.data();
+
+    auto fps_packed_cptr = fps_packed.data();
+
+    // argmin
+    py::ssize_t fp1_idx = std::distance(
+        sims_cent_ptr,
+        std::min_element(sims_cent_ptr, sims_cent_ptr + n_samples));
+    auto fp1_packed = py::array_t<uint8_t>(
+        n_features_packed, fps_packed_cptr + fp1_idx * n_features_packed);
+
+    auto sims_fp1 = jt_sim_packed_precalc_cardinalities(fps_packed, fp1_packed,
+                                                        cardinalities);
+    auto sims_fp1_ptr = sims_fp1.data();
+
+    // argmin
+    py::ssize_t fp2_idx = std::distance(
+        sims_fp1_ptr, std::min_element(sims_fp1_ptr, sims_fp1_ptr + n_samples));
+    auto fp2_packed = py::array_t<uint8_t>(
+        n_features_packed, fps_packed_cptr + fp2_idx * n_features_packed);
+
+    auto sims_fp2 = jt_sim_packed_precalc_cardinalities(fps_packed, fp2_packed,
+                                                        cardinalities);
+
+    return py::make_tuple(fp1_idx, fp2_idx, sims_fp1, sims_fp2);
+}
+
+PYBIND11_MODULE(_cpp_similarity, m) {
+    m.doc() = "Optimized molecular similarity calculators (C++ extensions)";
+
+    // Only bound for debugging purposes
+    m.def("_nochecks_unpack_fingerprints_2d", &_nochecks_unpack_fingerprints_2d,
+          "Unpack packed fingerprints", py::arg("a"),
+          py::arg("n_features") = std::nullopt);
+    m.def("_nochecks_unpack_fingerprints_1d", &_nochecks_unpack_fingerprints_1d,
+          "Unpack packed fingerprints", py::arg("a"),
+          py::arg("n_features") = std::nullopt);
+
+    // NOTE: There are some gains from using this fn but only ~3%, so don't warn
+    // for now if this fails, and don't expose it
+    m.def("unpack_fingerprints", &unpack_fingerprints,
+          "Unpack packed fingerprints", py::arg("a"),
+          py::arg("n_features") = std::nullopt);
+
+    // NOTE: pybind11's dynamic dispatch is *significantly* more
+    // expensive than casting to uint64_t always
+    // still this function is *barely* faster than python if no casts are
+    // needed, and slightly slower if casts are needed so it is not useful
+    // outside the C++ code, and it should not be exposed by default in any
+    // module (only for internal use and debugging)
+    m.def("centroid_from_sum", &centroid_from_sum<uint64_t>,
+          "centroid calculation", py::arg("linear_sum"), py::arg("n_samples"),
+          py::arg("pack") = true);
+
+    m.def("_popcount_2d", &_popcount_2d, "2D popcount", py::arg("a"));
+    m.def("_popcount_1d", &_popcount_1d, "1D popcount", py::arg("a"));
+    m.def("add_rows", &add_rows<uint8_t>, "add_rows", py::arg("arr"));
+
+    // API
+    m.def("jt_isim_from_sum", &jt_isim_from_sum,
+          "iSIM Tanimoto calculation from sum", py::arg("c_total"),
+          py::arg("n_objects"));
+    m.def("jt_isim_packed_u8", &jt_isim_packed_u8, "iSIM Tanimoto calculation",
+          py::arg("arr"), py::arg("n_features") = std::nullopt);
+    m.def("jt_isim_unpacked_u8", &jt_isim_unpacked_u8,
+          "iSIM Tanimoto calculation", py::arg("arr"));
+
+    m.def("_jt_sim_arr_vec_packed", &_jt_sim_arr_vec_packed,
+          "Tanimoto similarity between a matrix of packed fps and a single "
+          "packed fp",
+          py::arg("arr"), py::arg("vec"));
+    m.def("jt_most_dissimilar_packed", &jt_most_dissimilar_packed,
+          "Finds two fps in a packed fp array that are the most "
+          "Tanimoto-dissimilar",
+          py::arg("Y"), py::arg("n_features") = std::nullopt);
+}