usearch 0.1.10 → 0.2.2

package/src/advanced.hpp

@@ -1,7 +1,18 @@
 #pragma once
-#include <functional>
-#include <thread>
-#include <vector>
+
+#if defined(WIN32) || defined(_WIN32) || defined(__WIN32__) || defined(__NT__)
+#define _USE_MATH_DEFINES
+
+#define WINDOWS
+#endif
+
+#include <stdlib.h> // `aligned_alloc`
+
+#include <functional> // `std::function`
+#include <numeric>    // `std::iota`
+#include <stdexcept>  // `std::invalid_argument`
+#include <thread>     // `std::thread`
+#include <vector>     // `std::vector`
 
 #if __linux__
 #include <sys/auxv.h>
@@ -42,11 +53,15 @@ template <typename at> inline at clamp(at v, at lo, at hi) noexcept {
     return usearch::clamp(v, lo, hi, std::less<at>{});
 }
 
-class i8q100_converted_t;
-class f16_converted_t;
+class f8_bits_t;
+class f16_bits_t;
 
 inline float f16_to_f32(std::uint16_t u16) noexcept {
-#if defined(__aarch64__)
+#if defined(__AVX512F__)
+    _Float16 f16;
+    std::memcpy(&f16, &u16, sizeof(std::uint16_t));
+    return float(f16);
+#elif defined(__aarch64__)
     __fp16 f16;
     std::memcpy(&f16, &u16, sizeof(std::uint16_t));
     return float(f16);
@@ -56,7 +71,12 @@ inline float f16_to_f32(std::uint16_t u16) noexcept {
 }
 
 inline std::uint16_t f32_to_f16(float f32) noexcept {
-#if defined(__aarch64__)
+#if defined(__AVX512F__)
+    _Float16 f16 = _Float16(f32);
+    std::uint16_t u16;
+    std::memcpy(&u16, &f16, sizeof(std::uint16_t));
+    return u16;
+#elif defined(__aarch64__)
     __fp16 f16 = __fp16(f32);
     std::uint16_t u16;
     std::memcpy(&u16, &f16, sizeof(std::uint16_t));
@@ -66,81 +86,89 @@ inline std::uint16_t f32_to_f16(float f32) noexcept {
 #endif
 }
 
-class f16_converted_t {
+class f16_bits_t {
     std::uint16_t uint16_{};
 
   public:
-    inline f16_converted_t() noexcept : uint16_(0) {}
-    inline f16_converted_t(f16_converted_t&&) = default;
-    inline f16_converted_t& operator=(f16_converted_t&&) = default;
-    inline f16_converted_t(f16_converted_t const&) = default;
-    inline f16_converted_t& operator=(f16_converted_t const&) = default;
+    inline f16_bits_t() noexcept : uint16_(0) {}
+    inline f16_bits_t(f16_bits_t&&) = default;
+    inline f16_bits_t& operator=(f16_bits_t&&) = default;
+    inline f16_bits_t(f16_bits_t const&) = default;
+    inline f16_bits_t& operator=(f16_bits_t const&) = default;
 
     inline operator float() const noexcept { return f16_to_f32(uint16_); }
 
-    inline f16_converted_t(i8q100_converted_t) noexcept;
-    inline f16_converted_t(float v) noexcept : uint16_(f32_to_f16(v)) {}
-    inline f16_converted_t(double v) noexcept : uint16_(f32_to_f16(v)) {}
-
-    inline f16_converted_t operator+(f16_converted_t other) const noexcept { return {float(*this) + float(other)}; }
-    inline f16_converted_t operator-(f16_converted_t other) const noexcept { return {float(*this) - float(other)}; }
-    inline f16_converted_t operator*(f16_converted_t other) const noexcept { return {float(*this) * float(other)}; }
-    inline f16_converted_t operator/(f16_converted_t other) const noexcept { return {float(*this) / float(other)}; }
-    inline f16_converted_t operator+(float other) const noexcept { return {float(*this) + other}; }
-    inline f16_converted_t operator-(float other) const noexcept { return {float(*this) - other}; }
-    inline f16_converted_t operator*(float other) const noexcept { return {float(*this) * other}; }
-    inline f16_converted_t operator/(float other) const noexcept { return {float(*this) / other}; }
-    inline f16_converted_t operator+(double other) const noexcept { return {float(*this) + other}; }
-    inline f16_converted_t operator-(double other) const noexcept { return {float(*this) - other}; }
-    inline f16_converted_t operator*(double other) const noexcept { return {float(*this) * other}; }
-    inline f16_converted_t operator/(double other) const noexcept { return {float(*this) / other}; }
-
-    inline f16_converted_t& operator+=(float v) noexcept {
+    inline f16_bits_t(f8_bits_t) noexcept;
+    inline f16_bits_t(float v) noexcept : uint16_(f32_to_f16(v)) {}
+    inline f16_bits_t(double v) noexcept : uint16_(f32_to_f16(v)) {}
+
+    inline f16_bits_t operator+(f16_bits_t other) const noexcept { return {float(*this) + float(other)}; }
+    inline f16_bits_t operator-(f16_bits_t other) const noexcept { return {float(*this) - float(other)}; }
+    inline f16_bits_t operator*(f16_bits_t other) const noexcept { return {float(*this) * float(other)}; }
+    inline f16_bits_t operator/(f16_bits_t other) const noexcept { return {float(*this) / float(other)}; }
+    inline f16_bits_t operator+(float other) const noexcept { return {float(*this) + other}; }
+    inline f16_bits_t operator-(float other) const noexcept { return {float(*this) - other}; }
+    inline f16_bits_t operator*(float other) const noexcept { return {float(*this) * other}; }
+    inline f16_bits_t operator/(float other) const noexcept { return {float(*this) / other}; }
+    inline f16_bits_t operator+(double other) const noexcept { return {float(*this) + other}; }
+    inline f16_bits_t operator-(double other) const noexcept { return {float(*this) - other}; }
+    inline f16_bits_t operator*(double other) const noexcept { return {float(*this) * other}; }
+    inline f16_bits_t operator/(double other) const noexcept { return {float(*this) / other}; }
+
+    inline f16_bits_t& operator+=(float v) noexcept {
         uint16_ = f32_to_f16(v + f16_to_f32(uint16_));
         return *this;
     }
 
-    inline f16_converted_t& operator-=(float v) noexcept {
+    inline f16_bits_t& operator-=(float v) noexcept {
         uint16_ = f32_to_f16(v - f16_to_f32(uint16_));
         return *this;
     }
 
-    inline f16_converted_t& operator*=(float v) noexcept {
+    inline f16_bits_t& operator*=(float v) noexcept {
         uint16_ = f32_to_f16(v * f16_to_f32(uint16_));
         return *this;
     }
 
-    inline f16_converted_t& operator/=(float v) noexcept {
+    inline f16_bits_t& operator/=(float v) noexcept {
         uint16_ = f32_to_f16(v / f16_to_f32(uint16_));
         return *this;
     }
 };
 
-class i8q100_converted_t {
+/**
+ *  @brief  Numeric type for uniformly-distributed floating point
+ *          values within [-1,1] range, quantized to integers [-100,100].
+ */
+class f8_bits_t {
     std::int8_t int8_{};
 
   public:
-    inline i8q100_converted_t() noexcept : int8_(0) {}
-    inline i8q100_converted_t(i8q100_converted_t&&) = default;
-    inline i8q100_converted_t& operator=(i8q100_converted_t&&) = default;
-    inline i8q100_converted_t(i8q100_converted_t const&) = default;
-    inline i8q100_converted_t& operator=(i8q100_converted_t const&) = default;
-
-    inline operator float() const noexcept { return float(int8_) / 100.f; }
-    inline operator f16_converted_t() const noexcept { return float(int8_) / 100.f; }
-    inline operator double() const noexcept { return double(int8_) / 100.0; }
-    inline operator std::int8_t() const noexcept { return int8_; }
-    inline operator std::int16_t() const noexcept { return int8_; }
-    inline operator std::int32_t() const noexcept { return int8_; }
-    inline operator std::int64_t() const noexcept { return int8_; }
-
-    inline i8q100_converted_t(float v) noexcept
-        : int8_(usearch::clamp<std::int8_t>(static_cast<std::int8_t>(v * 100.f), -100, 100)) {}
-    inline i8q100_converted_t(double v) noexcept
-        : int8_(usearch::clamp<std::int8_t>(static_cast<std::int8_t>(v * 100.0), -100, 100)) {}
+    constexpr static float divisor_k = 100.f;
+    constexpr static std::int8_t min_k = -100;
+    constexpr static std::int8_t max_k = 100;
+
+    inline f8_bits_t() noexcept : int8_(0) {}
+    inline f8_bits_t(f8_bits_t&&) = default;
+    inline f8_bits_t& operator=(f8_bits_t&&) = default;
+    inline f8_bits_t(f8_bits_t const&) = default;
+    inline f8_bits_t& operator=(f8_bits_t const&) = default;
+
+    inline operator float() const noexcept { return float(int8_) / divisor_k; }
+    inline operator f16_bits_t() const noexcept { return float(int8_) / divisor_k; }
+    inline operator double() const noexcept { return double(int8_) / divisor_k; }
+    inline explicit operator std::int8_t() const noexcept { return int8_; }
+    inline explicit operator std::int16_t() const noexcept { return int8_; }
+    inline explicit operator std::int32_t() const noexcept { return int8_; }
+    inline explicit operator std::int64_t() const noexcept { return int8_; }
+
+    inline f8_bits_t(float v) noexcept
+        : int8_(usearch::clamp<std::int8_t>(static_cast<std::int8_t>(v * divisor_k), min_k, max_k)) {}
+    inline f8_bits_t(double v) noexcept
+        : int8_(usearch::clamp<std::int8_t>(static_cast<std::int8_t>(v * divisor_k), min_k, max_k)) {}
 };
 
-inline f16_converted_t::f16_converted_t(i8q100_converted_t v) noexcept : f16_converted_t(float(v)) {}
+inline f16_bits_t::f16_bits_t(f8_bits_t v) noexcept : f16_bits_t(float(v)) {}
 
 struct uuid_t {
     std::uint8_t octets[16];
@@ -162,7 +190,7 @@ void multithreaded(std::size_t threads, std::size_t tasks, callback_at&& callbac
     for (std::size_t thread_idx = 0; thread_idx != threads; ++thread_idx) {
         threads_pool.emplace_back([=]() {
             for (std::size_t task_idx = thread_idx * tasks_per_thread;
-                 task_idx < std::min(tasks, thread_idx * tasks_per_thread + tasks_per_thread); ++task_idx)
+                 task_idx < (std::min)(tasks, thread_idx * tasks_per_thread + tasks_per_thread); ++task_idx)
                 callback(thread_idx, task_idx);
         });
     }
@@ -200,7 +228,7 @@ enum class accuracy_t {
     f32_k,
     f16_k,
     f64_k,
-    i8q100_k,
+    f8_k,
 };
 
 inline char const* accuracy_name(accuracy_t accuracy) noexcept {
@@ -208,7 +236,7 @@ inline char const* accuracy_name(accuracy_t accuracy) noexcept {
     case accuracy_t::f32_k: return "f32";
     case accuracy_t::f16_k: return "f16";
     case accuracy_t::f64_k: return "f64";
-    case accuracy_t::i8q100_k: return "i8q100";
+    case accuracy_t::f8_k: return "f8";
     default: return "";
     }
 }
@@ -226,10 +254,10 @@ inline accuracy_t accuracy_from_name(char const* name, std::size_t len) {
         accuracy = accuracy_t::f64_k;
     else if (str_equals(name, len, "f16"))
         accuracy = accuracy_t::f16_k;
-    else if (str_equals(name, len, "i8q100"))
-        accuracy = accuracy_t::i8q100_k;
+    else if (str_equals(name, len, "f8"))
+        accuracy = accuracy_t::f8_k;
     else
-        throw std::runtime_error("Unknown type, choose: f32, f16, f64, i8q100");
+        throw std::invalid_argument("Unknown type, choose: f32, f16, f64, f8");
     return accuracy;
 }
 
@@ -237,16 +265,24 @@ template <typename index_at>
 inline index_at index_from_name( //
     char const* name, std::size_t len, std::size_t dimensions, accuracy_t accuracy, config_t const& config) {
 
-    if (str_equals(name, len, "l2_sq") || str_equals(name, len, "euclidean_sq"))
+    if (str_equals(name, len, "l2_sq") || str_equals(name, len, "euclidean_sq")) {
+        if (dimensions == 0)
+            throw std::invalid_argument("The number of dimensions must be positive");
         return index_at::l2(dimensions, accuracy, config);
-    else if (str_equals(name, len, "ip") || str_equals(name, len, "inner") || str_equals(name, len, "dot"))
+    } else if (str_equals(name, len, "ip") || str_equals(name, len, "inner") || str_equals(name, len, "dot")) {
+        if (dimensions == 0)
+            throw std::invalid_argument("The number of dimensions must be positive");
         return index_at::ip(dimensions, accuracy, config);
-    else if (str_equals(name, len, "cos") || str_equals(name, len, "angular"))
+    } else if (str_equals(name, len, "cos") || str_equals(name, len, "angular")) {
+        if (dimensions == 0)
+            throw std::invalid_argument("The number of dimensions must be positive");
         return index_at::cos(dimensions, accuracy, config);
-    else if (str_equals(name, len, "haversine"))
+    } else if (str_equals(name, len, "haversine")) {
+        if (dimensions != 2 && dimensions != 0)
+            throw std::invalid_argument("The number of dimensions must be equal to two");
         return index_at::haversine(accuracy, config);
-    else
-        throw std::runtime_error("Unknown distance, choose: l2_sq, ip, cos, hamming, jaccard");
+    } else
+        throw std::invalid_argument("Unknown distance, choose: l2_sq, ip, cos, hamming, jaccard");
     return {};
 }
 
@@ -274,7 +310,7 @@ inline std::size_t bytes_per_scalar(accuracy_t accuracy) noexcept {
     case accuracy_t::f32_k: return 4;
     case accuracy_t::f16_k: return 2;
     case accuracy_t::f64_k: return 8;
-    case accuracy_t::i8q100_k: return 1;
+    case accuracy_t::f8_k: return 1;
     default: return 0;
     }
 }
@@ -309,26 +345,10 @@ template <> struct cast_gt<f64_t, f64_t> {
     bool operator()(byte_t const*, std::size_t, byte_t*) noexcept { return false; }
 };
 
-template <> struct cast_gt<f16_converted_t, f16_converted_t> {
+template <> struct cast_gt<f16_bits_t, f16_bits_t> {
     bool operator()(byte_t const*, std::size_t, byte_t*) noexcept { return false; }
 };
 
-struct ip_i8q100_converted_t {
-    float operator()(i8q100_converted_t const* a, i8q100_converted_t const* b, std::size_t d) const noexcept {
-        std::int64_t ab = 0;
-#if defined(__GNUC__)
-#pragma GCC ivdep
-#elif defined(__clang__)
-#pragma clang loop vectorize(enable)
-#elif defined(_OPENMP)
-#pragma omp simd reduction(+ : ab)
-#endif
-        for (std::size_t i = 0; i < d; ++i)
-            ab += std::int16_t(a[i]) * std::int16_t(b[i]);
-        return 1.f - ab / 10000.f;
-    }
-};
-
 /**
  *  @brief  Oversimplified type-punned index for equidimensional floating-point
  *          vectors with automatic down-casting and hardware acceleration.
@@ -359,10 +379,10 @@ class auto_index_gt {
     accuracy_t accuracy_ = accuracy_t::f32_k;
     isa_t acceleration_ = isa_t::auto_k;
 
-    std::unique_ptr<index_t> index_;
-    mutable std::vector<byte_t> cast_buffer_;
+    index_t* index_{};
+    mutable std::vector<byte_t> cast_buffer_{};
     struct casts_t {
-        cast_t from_i8q100{};
+        cast_t from_f8{};
         cast_t from_f16{};
         cast_t from_f32{};
         cast_t from_f64{};
@@ -371,22 +391,42 @@ class auto_index_gt {
     punned_stateful_metric_t root_metric_;
     config_t root_config_;
 
+    mutable std::vector<std::size_t> available_threads_;
+    mutable std::mutex available_threads_mutex_;
+
   public:
     auto_index_gt() = default;
-    auto_index_gt(auto_index_gt&& other)
-        : dimensions_(other.dimensions_), casted_vector_bytes_(other.casted_vector_bytes_), accuracy_(other.accuracy_),
-          acceleration_(other.acceleration_), index_(std::move(other.index_)),
-          cast_buffer_(std::move(other.cast_buffer_)), casts_(std::move(other.casts_)),
-          root_metric_(std::move(other.root_metric_)), root_config_(std::move(other.root_config_)) {}
+    auto_index_gt(auto_index_gt&& other) { swap(other); }
+    auto_index_gt& operator=(auto_index_gt&& other) noexcept {
+        swap(other);
+        return *this;
+    }
+
+    ~auto_index_gt() { aligned_index_free(index_); }
+
+    void swap(auto_index_gt& other) noexcept {
+        std::swap(dimensions_, other.dimensions_);
+        std::swap(casted_vector_bytes_, other.casted_vector_bytes_);
+        std::swap(accuracy_, other.accuracy_);
+        std::swap(acceleration_, other.acceleration_);
+        std::swap(index_, other.index_);
+        std::swap(cast_buffer_, other.cast_buffer_);
+        std::swap(casts_, other.casts_);
+        std::swap(root_metric_, other.root_metric_);
+        std::swap(root_config_, other.root_config_);
+        std::swap(available_threads_, other.available_threads_);
+    }
 
     std::size_t dimensions() const noexcept { return dimensions_; }
     std::size_t connectivity() const noexcept { return index_->connectivity(); }
     std::size_t size() const noexcept { return index_->size(); }
     std::size_t capacity() const noexcept { return index_->capacity(); }
+    config_t const& config() const noexcept { return root_config_; }
+    void clear() noexcept { return index_->clear(); }
 
     isa_t acceleration() const noexcept { return acceleration_; }
     std::size_t concurrency() const noexcept {
-        return std::min(root_config_.max_threads_add, root_config_.max_threads_search);
+        return (std::min)(root_config_.max_threads_add, root_config_.max_threads_search);
     }
 
     void save(char const* path) const { index_->save(path); }
@@ -395,13 +435,21 @@ class auto_index_gt {
     void reserve(std::size_t capacity) { index_->reserve(capacity); }
 
     // clang-format off
-    void add(label_t label, f16_converted_t const* vector, std::size_t thread = 0, bool copy = true) { return add(label, vector, thread, copy, casts_.from_f16); }
-    void add(label_t label, f32_t const* vector, std::size_t thread = 0, bool copy = true) { return add(label, vector, thread, copy, casts_.from_f32); }
-    void add(label_t label, f64_t const* vector, std::size_t thread = 0, bool copy = true) { return add(label, vector, thread, copy, casts_.from_f64); }
-    
-    std::size_t search(f16_converted_t const* vector, std::size_t wanted, label_t* matches, distance_t* distances, std::size_t thread = 0) const { return search(vector, wanted, matches, distances, thread, casts_.from_f16); }
-    std::size_t search(f32_t const* vector, std::size_t wanted, label_t* matches, distance_t* distances, std::size_t thread = 0) const { return search(vector, wanted, matches, distances, thread, casts_.from_f32); }
-    std::size_t search(f64_t const* vector, std::size_t wanted, label_t* matches, distance_t* distances, std::size_t thread = 0) const { return search(vector, wanted, matches, distances, thread, casts_.from_f64); }
+    void add(label_t label, f16_bits_t const* vector) { return add(label, vector, true, casts_.from_f16); }
+    void add(label_t label, f32_t const* vector) { return add(label, vector, true, casts_.from_f32); }
+    void add(label_t label, f64_t const* vector) { return add(label, vector, true, casts_.from_f64); }
+
+    void add(label_t label, f16_bits_t const* vector, std::size_t thread, bool copy) { return add(label, vector, thread, copy, casts_.from_f16); }
+    void add(label_t label, f32_t const* vector, std::size_t thread, bool copy) { return add(label, vector, thread, copy, casts_.from_f32); }
+    void add(label_t label, f64_t const* vector, std::size_t thread, bool copy) { return add(label, vector, thread, copy, casts_.from_f64); }
+
+    std::size_t search(f16_bits_t const* vector, std::size_t wanted, label_t* matches, distance_t* distances) const { return search(vector, wanted, matches, distances, casts_.from_f16); }
+    std::size_t search(f32_t const* vector, std::size_t wanted, label_t* matches, distance_t* distances) const { return search(vector, wanted, matches, distances, casts_.from_f32); }
+    std::size_t search(f64_t const* vector, std::size_t wanted, label_t* matches, distance_t* distances) const { return search(vector, wanted, matches, distances, casts_.from_f64); }
+
+    std::size_t search(f16_bits_t const* vector, std::size_t wanted, label_t* matches, distance_t* distances, std::size_t thread) const { return search(vector, wanted, matches, distances, thread, casts_.from_f16); }
+    std::size_t search(f32_t const* vector, std::size_t wanted, label_t* matches, distance_t* distances, std::size_t thread) const { return search(vector, wanted, matches, distances, thread, casts_.from_f32); }
+    std::size_t search(f64_t const* vector, std::size_t wanted, label_t* matches, distance_t* distances, std::size_t thread) const { return search(vector, wanted, matches, distances, thread, casts_.from_f64); }
 
     static auto_index_gt ip(std::size_t dimensions, accuracy_t accuracy = accuracy_t::f16_k, config_t config = {}) { return make(dimensions, accuracy, ip_metric(dimensions, accuracy), make_casts(accuracy), config); }
     static auto_index_gt l2(std::size_t dimensions, accuracy_t accuracy = accuracy_t::f16_k, config_t config = {}) { return make(dimensions, accuracy, l2_metric(dimensions, accuracy), make_casts(accuracy), config); }
@@ -427,12 +475,41 @@ class auto_index_gt {
 
         result.root_metric_ = root_metric_;
         result.root_config_ = root_config_;
-        result.index_.reset(new index_t(result.root_config_, result.root_metric_));
+        index_t* raw = aligned_index_alloc();
+        new (raw) index_t(root_config_, root_metric_);
+        result.index_ = raw;
 
         return result;
     }
 
   private:
+    static index_t* aligned_index_alloc() noexcept {
+        return (index_t*)aligned_alloc(64, 64 * divide_round_up<64>(sizeof(index_t)));
+    }
+
+    static void aligned_index_free(index_t* raw) noexcept { free(raw); }
+
+    struct thread_lock_t {
+        auto_index_gt const& parent;
+        std::size_t thread_id;
+
+        ~thread_lock_t() { parent.thread_unlock(thread_id); }
+    };
+
+    thread_lock_t thread_lock() const {
+        available_threads_mutex_.lock();
+        std::size_t thread_id = available_threads_.back();
+        available_threads_.pop_back();
+        available_threads_mutex_.unlock();
+        return {*this, thread_id};
+    }
+
+    void thread_unlock(std::size_t thread_id) const {
+        available_threads_mutex_.lock();
+        available_threads_.push_back(thread_id);
+        available_threads_mutex_.unlock();
+    }
+
     template <typename scalar_at>
     void add(label_t label, scalar_at const* vector, std::size_t thread, bool copy, cast_t const& cast) {
         byte_t const* vector_data = reinterpret_cast<byte_t const*>(vector);
@@ -463,29 +540,54 @@ class auto_index_gt {
         return index_->search({vector_data, vector_bytes}, wanted, matches, distances, thread);
     }
 
+    template <typename scalar_at> void add(label_t label, scalar_at const* vector, bool copy, cast_t const& cast) {
+        thread_lock_t lock = thread_lock();
+        return add(label, vector, lock.thread_id, copy, cast);
+    }
+
+    template <typename scalar_at>
+    std::size_t search(                              //
+        scalar_at const* vector, std::size_t wanted, //
+        label_t* matches, distance_t* distances,     //
+        cast_t const& cast) const {
+        thread_lock_t lock = thread_lock();
+        return search(vector, wanted, matches, distances, lock.thread_id, cast);
+    }
+
     static auto_index_gt make(                            //
         std::size_t dimensions, accuracy_t accuracy,      //
         metric_and_meta_t metric_and_meta, casts_t casts, //
         config_t config) {
 
-        std::size_t max_threads = std::max(config.max_threads_add, config.max_threads_search);
+        std::size_t hardware_threads = std::thread::hardware_concurrency();
+        config.max_threads_add = config.max_threads_add ? config.max_threads_add : hardware_threads;
+        config.max_threads_search = config.max_threads_search ? config.max_threads_search : hardware_threads;
+        std::size_t max_threads = (std::max)(config.max_threads_add, config.max_threads_search);
         auto_index_gt result;
         result.dimensions_ = dimensions;
         result.accuracy_ = accuracy;
         result.casted_vector_bytes_ = bytes_per_scalar(accuracy) * dimensions;
         result.cast_buffer_.resize(max_threads * result.casted_vector_bytes_);
         result.casts_ = casts;
-        result.index_.reset(new index_t(config, metric_and_meta.metric));
         result.acceleration_ = metric_and_meta.acceleration;
         result.root_metric_ = metric_and_meta.metric;
         result.root_config_ = config;
+
+        // Fill the thread IDs.
+        result.available_threads_.resize(max_threads);
+        std::iota(result.available_threads_.begin(), result.available_threads_.end(), 0ul);
+
+        // Availible since C11, but only C++17, so we use the C version.
+        index_t* raw = aligned_index_alloc();
+        new (raw) index_t(config, metric_and_meta.metric);
+        result.index_ = raw;
         return result;
     }
 
     template <typename to_scalar_at> static casts_t make_casts() {
         casts_t result;
-        result.from_i8q100 = cast_gt<i8q100_converted_t, to_scalar_at>{};
-        result.from_f16 = cast_gt<f16_converted_t, to_scalar_at>{};
+        result.from_f8 = cast_gt<f8_bits_t, to_scalar_at>{};
+        result.from_f16 = cast_gt<f16_bits_t, to_scalar_at>{};
         result.from_f32 = cast_gt<f32_t, to_scalar_at>{};
         result.from_f64 = cast_gt<f64_t, to_scalar_at>{};
         return result;
@@ -495,8 +597,8 @@ class auto_index_gt {
         switch (accuracy) {
         case accuracy_t::f64_k: return make_casts<f64_t>();
         case accuracy_t::f32_k: return make_casts<f32_t>();
-        case accuracy_t::f16_k: return make_casts<f16_converted_t>();
-        case accuracy_t::i8q100_k: return make_casts<i8q100_converted_t>();
+        case accuracy_t::f16_k: return make_casts<f16_bits_t>();
+        case accuracy_t::f8_k: return make_casts<f8_bits_t>();
         default: return {};
         }
     }
@@ -515,7 +617,7 @@ class auto_index_gt {
         if (dimensions % 4 == 0)
             return {
                 pun_metric<simsimd_f32_t>([](simsimd_f32_t const* a, simsimd_f32_t const* b, size_t d) noexcept {
-                    return 1 - simsimd_dot_f32x4avx2(a, b, d);
+                    return 1.f - simsimd_dot_f32x4avx2(a, b, d);
                 }),
                 isa_t::avx2_k,
             };
@@ -523,14 +625,14 @@ class auto_index_gt {
         if (supports_arm_sve())
             return {
                 pun_metric<simsimd_f32_t>([](simsimd_f32_t const* a, simsimd_f32_t const* b, size_t d) noexcept {
-                    return 1 - simsimd_dot_f32sve(a, b, d);
+                    return 1.f - simsimd_dot_f32sve(a, b, d);
                 }),
                 isa_t::sve_k,
             };
         if (dimensions % 4 == 0)
             return {
                 pun_metric<simsimd_f32_t>([](simsimd_f32_t const* a, simsimd_f32_t const* b, size_t d) noexcept {
-                    return 1 - simsimd_dot_f32x4neon(a, b, d);
+                    return 1.f - simsimd_dot_f32x4neon(a, b, d);
                 }),
                 isa_t::neon_k,
             };
@@ -543,31 +645,39 @@ class auto_index_gt {
         (void)dimensions;
 #if defined(USEARCH_USE_SIMD)
 #if defined(__x86_64__)
+        if (dimensions % 32 == 0)
+            return {
+                pun_metric<simsimd_f16_t>([](simsimd_f16_t const* a, simsimd_f16_t const* b, size_t d) noexcept {
+                    return 1.f - simsimd_dot_f16x16avx512(a, b, d);
+                }),
+                isa_t::avx512_k,
+            };
 #elif defined(__aarch64__)
         if (supports_arm_sve())
             return {
                 pun_metric<simsimd_f16_t>([](simsimd_f16_t const* a, simsimd_f16_t const* b, size_t d) noexcept {
-                    return 1 - simsimd_dot_f16sve(a, b, d);
+                    return 1.f - simsimd_dot_f16sve(a, b, d);
                 }),
                 isa_t::sve_k,
             };
         if (dimensions % 8 == 0)
             return {
                 pun_metric<simsimd_f16_t>([](simsimd_f16_t const* a, simsimd_f16_t const* b, size_t d) noexcept {
-                    return 1 - simsimd_dot_f16x8neon(a, b, d);
+                    return 1.f - simsimd_dot_f16x8neon(a, b, d);
                 }),
                 isa_t::neon_k,
             };
 #endif
 #endif
-        return {pun_metric<f16_converted_t>(ip_gt<f16_converted_t>{}), isa_t::auto_k};
+        return {pun_metric<f16_bits_t>(ip_gt<f16_bits_t, f32_t>{}), isa_t::auto_k};
     }
 
     static metric_and_meta_t ip_metric(std::size_t dimensions, accuracy_t accuracy) {
         switch (accuracy) {
-        case accuracy_t::i8q100_k: return {pun_metric<i8q100_converted_t>(ip_i8q100_converted_t{}), isa_t::auto_k};
+            // The tow most common numeric types for the most common metric have  afew optimized versions:
         case accuracy_t::f16_k: return ip_metric_f16(dimensions);
         case accuracy_t::f32_k: return ip_metric_f32(dimensions);
+        case accuracy_t::f8_k: return {pun_metric<f8_bits_t>(ip_gt<f8_bits_t, f32_t>{}), isa_t::auto_k};
         case accuracy_t::f64_k: return {pun_metric<f64_t>(ip_gt<f64_t>{}), isa_t::auto_k};
         default: return {};
         }
@@ -575,18 +685,18 @@ class auto_index_gt {
 
     static metric_and_meta_t l2_metric(std::size_t, accuracy_t accuracy) {
         switch (accuracy) {
-        case accuracy_t::i8q100_k: return {};
-        case accuracy_t::f16_k: return {pun_metric<f16_converted_t>(l2_squared_gt<f16_converted_t>{}), isa_t::auto_k};
-        case accuracy_t::f32_k: return {pun_metric<f32_t>(l2_squared_gt<f32_t>{}), isa_t::auto_k};
-        case accuracy_t::f64_k: return {pun_metric<f64_t>(l2_squared_gt<f64_t>{}), isa_t::auto_k};
+        case accuracy_t::f8_k: return {pun_metric<f8_bits_t>(l2sq_gt<f8_bits_t, f32_t>{}), isa_t::auto_k};
+        case accuracy_t::f16_k: return {pun_metric<f16_bits_t>(l2sq_gt<f16_bits_t, f32_t>{}), isa_t::auto_k};
+        case accuracy_t::f32_k: return {pun_metric<f32_t>(l2sq_gt<f32_t>{}), isa_t::auto_k};
+        case accuracy_t::f64_k: return {pun_metric<f64_t>(l2sq_gt<f64_t>{}), isa_t::auto_k};
         default: return {};
         }
     }
 
     static metric_and_meta_t cos_metric(std::size_t, accuracy_t accuracy) {
         switch (accuracy) {
-        case accuracy_t::i8q100_k: return {};
-        case accuracy_t::f16_k: return {pun_metric<f16_converted_t>(cos_gt<f16_converted_t>{}), isa_t::auto_k};
+        case accuracy_t::f8_k: return {pun_metric<f8_bits_t>(cos_gt<f8_bits_t, f32_t>{}), isa_t::auto_k};
+        case accuracy_t::f16_k: return {pun_metric<f16_bits_t>(cos_gt<f16_bits_t, f32_t>{}), isa_t::auto_k};
         case accuracy_t::f32_k: return {pun_metric<f32_t>(cos_gt<f32_t>{}), isa_t::auto_k};
         case accuracy_t::f64_k: return {pun_metric<f64_t>(cos_gt<f64_t>{}), isa_t::auto_k};
         default: return {};
@@ -595,8 +705,8 @@ class auto_index_gt {
 
     static metric_and_meta_t haversine_metric(accuracy_t accuracy) {
         switch (accuracy) {
-        case accuracy_t::i8q100_k: return {};
-        case accuracy_t::f16_k: return {pun_metric<f16_converted_t>(haversine_gt<f16_converted_t>{}), isa_t::auto_k};
+        case accuracy_t::f8_k: return {pun_metric<f8_bits_t>(haversine_gt<f8_bits_t, f32_t>{}), isa_t::auto_k};
+        case accuracy_t::f16_k: return {pun_metric<f16_bits_t>(haversine_gt<f16_bits_t, f32_t>{}), isa_t::auto_k};
         case accuracy_t::f32_k: return {pun_metric<f32_t>(haversine_gt<f32_t>{}), isa_t::auto_k};
         case accuracy_t::f64_k: return {pun_metric<f64_t>(haversine_gt<f64_t>{}), isa_t::auto_k};
         default: return {};