usearch 0.4.0 → 0.8.1

package/src/test.cpp

@@ -0,0 +1,62 @@
+/**
+ * @brief A trivial test.
+ */
+#include <cassert>
+
+#include "punned.hpp"
+
+using namespace unum::usearch;
+using namespace unum;
+
+using point_id_t = std::int64_t;
+
+template <typename scalar_at, typename index_at> void test3d(index_at&& index) {
+    using span_t = span_gt<scalar_at const>;
+    using distance_t = typename index_at::distance_t;
+
+    scalar_at vec[3] = {10, 20, 15};
+
+    index.reserve(10);
+    index.add(42, span_t{&vec[0], 3ul});
+
+    point_id_t matched_labels[10] = {0};
+    distance_t matched_distances[10] = {0};
+    std::size_t matched_count = index.search(span_t{&vec[0], 3ul}, 5).dump_to(matched_labels, matched_distances);
+
+    assert(matched_count == 1);
+    assert(matched_labels[0] == 42);
+    assert(std::abs(matched_distances[0]) < 0.01);
+
+    index.save("tmp.usearch");
+
+    // Search again over reconstructed index
+    index.load("tmp.usearch");
+    matched_count = index.search(span_t{&vec[0], 3ul}, 5).dump_to(matched_labels, matched_distances);
+    assert(matched_count == 1);
+    assert(matched_labels[0] == 42);
+    assert(std::abs(matched_distances[0]) < 0.01);
+
+    // Search again over mapped index
+    index.view("tmp.usearch");
+    matched_count = index.search(span_t{&vec[0], 3ul}, 5).dump_to(matched_labels, matched_distances);
+    assert(matched_count == 1);
+    assert(matched_labels[0] == 42);
+    assert(std::abs(matched_distances[0]) < 0.01);
+}
+
+int main(int, char**) {
+
+    test3d<float>(index_gt<cos_gt<float>, point_id_t, std::uint32_t, float>{});
+    test3d<float>(index_gt<l2sq_gt<float>, point_id_t, std::uint32_t, float>{});
+
+    test3d<double>(index_gt<cos_gt<double>, point_id_t, std::uint32_t, double>{});
+    test3d<double>(index_gt<l2sq_gt<double>, point_id_t, std::uint32_t, double>{});
+
+    test3d<float>(punned_small_t::cos(3));
+    test3d<float>(punned_small_t::l2sq(3));
+
+    test3d<double>(punned_small_t::cos(3));
+    test3d<double>(punned_small_t::l2sq(3));
+
+    return 0;
+}

package/swift/Test.swift

@@ -10,7 +10,7 @@ import USearch
 
 @available(iOS 13, macOS 10.15, tvOS 13.0, watchOS 6.0, *)
 func testUnit() throws {
-    let index = Index.l2(dimensions: 4, connectivity: 8)
+    let index = Index.l2sq(dimensions: 4, connectivity: 8)
     let vectorA: [Float32] = [0.3, 0.5, 1.2, 1.4]
     let vectorB: [Float32] = [0.4, 0.2, 1.2, 1.1]
     index.add(label: 42, vector: vectorA[...])

package/wolfram/lib.cpp

@@ -1,9 +1,9 @@
 #include "WolframLibrary.h"
-#include "advanced.hpp"
+#include "punned.hpp"
 
 using namespace unum::usearch;
 using distance_t = punned_distance_t;
-using native_index_t = auto_index_gt<int>;
+using punned_t = punned_gt<int>;
 using span_t = span_gt<float>;
 
 EXTERN_C DLLEXPORT int WolframLibrary_initialize(WolframLibraryData libData) { return LIBRARY_NO_ERROR; }
@@ -25,10 +25,10 @@ EXTERN_C DLLEXPORT int IndexCreate(WolframLibraryData libData, mint Argc, MArgum
         config.max_threads_search = 4;
 
         accuracy_t accuracy = accuracy_from_name(accuracy_cstr, strlen(accuracy_cstr));
-        native_index_t index = index_from_name<native_index_t>( //
+        punned_t index = index_from_name<punned_t>( //
             metric_cstr, strlen(metric_cstr), dimensions, accuracy, config);
 
-        native_index_t* result_ptr = new native_index_t(std::move(index));
+        punned_t* result_ptr = new punned_t(std::move(index));
         MArgument_setInteger(Res, (long)result_ptr);
 
     } catch (...) {
@@ -42,7 +42,7 @@ EXTERN_C DLLEXPORT int IndexCreate(WolframLibraryData libData, mint Argc, MArgum
 
 EXTERN_C DLLEXPORT int IndexSave(WolframLibraryData libData, mint Argc, MArgument* Args, MArgument Res) {
     char* path_cstr = nullptr;
-    native_index_t* c_ptr = (native_index_t*)MArgument_getUTF8String(Args[0]);
+    punned_t* c_ptr = (punned_t*)MArgument_getUTF8String(Args[0]);
     try {
         path_cstr = MArgument_getUTF8String(Args[1]);
         c_ptr->save(path_cstr);
@@ -55,7 +55,7 @@ EXTERN_C DLLEXPORT int IndexSave(WolframLibraryData libData, mint Argc, MArgumen
 
 EXTERN_C DLLEXPORT int IndexLoad(WolframLibraryData libData, mint Argc, MArgument* Args, MArgument Res) {
     char* path_cstr = nullptr;
-    native_index_t* c_ptr = (native_index_t*)MArgument_getUTF8String(Args[0]);
+    punned_t* c_ptr = (punned_t*)MArgument_getUTF8String(Args[0]);
     try {
         path_cstr = MArgument_getUTF8String(Args[1]);
         c_ptr->load(path_cstr);
@@ -68,7 +68,7 @@ EXTERN_C DLLEXPORT int IndexLoad(WolframLibraryData libData, mint Argc, MArgumen
 
 EXTERN_C DLLEXPORT int IndexView(WolframLibraryData libData, mint Argc, MArgument* Args, MArgument Res) {
     char* path_cstr = nullptr;
-    native_index_t* c_ptr = (native_index_t*)MArgument_getUTF8String(Args[0]);
+    punned_t* c_ptr = (punned_t*)MArgument_getUTF8String(Args[0]);
     try {
         path_cstr = MArgument_getUTF8String(Args[1]);
         c_ptr->view(path_cstr);
@@ -80,37 +80,37 @@ EXTERN_C DLLEXPORT int IndexView(WolframLibraryData libData, mint Argc, MArgumen
 }
 
 EXTERN_C DLLEXPORT int IndexDestroy(WolframLibraryData libData, mint Argc, MArgument* Args, MArgument Res) {
-    delete (native_index_t*)MArgument_getUTF8String(Args[0]);
+    delete (punned_t*)MArgument_getUTF8String(Args[0]);
     return LIBRARY_NO_ERROR;
 }
 
 EXTERN_C DLLEXPORT int IndexSize(WolframLibraryData libData, mint Argc, MArgument* Args, MArgument Res) {
-    size_t res = ((native_index_t*)MArgument_getUTF8String(Args[0]))->size();
+    size_t res = ((punned_t*)MArgument_getUTF8String(Args[0]))->size();
     MArgument_setInteger(Res, res);
     return LIBRARY_NO_ERROR;
 }
 
 EXTERN_C DLLEXPORT int IndexConnectivity(WolframLibraryData libData, mint Argc, MArgument* Args, MArgument Res) {
-    size_t res = ((native_index_t*)MArgument_getUTF8String(Args[0]))->connectivity();
+    size_t res = ((punned_t*)MArgument_getUTF8String(Args[0]))->connectivity();
     MArgument_setInteger(Res, res);
     return LIBRARY_NO_ERROR;
 }
 
 EXTERN_C DLLEXPORT int IndexDimensions(WolframLibraryData libData, mint Argc, MArgument* Args, MArgument Res) {
-    size_t res = ((native_index_t*)MArgument_getUTF8String(Args[0]))->dimensions();
+    size_t res = ((punned_t*)MArgument_getUTF8String(Args[0]))->dimensions();
     MArgument_setInteger(Res, res);
     return LIBRARY_NO_ERROR;
 }
 
 EXTERN_C DLLEXPORT int IndexCapacity(WolframLibraryData libData, mint Argc, MArgument* Args, MArgument Res) {
-    size_t res = ((native_index_t*)MArgument_getUTF8String(Args[0]))->capacity();
+    size_t res = ((punned_t*)MArgument_getUTF8String(Args[0]))->capacity();
     MArgument_setInteger(Res, res);
     return LIBRARY_NO_ERROR;
 }
 
 EXTERN_C DLLEXPORT int IndexAdd(WolframLibraryData libData, mint Argc, MArgument* Args, MArgument Res) {
     char* path_cstr = nullptr;
-    native_index_t* c_ptr = (native_index_t*)MArgument_getUTF8String(Args[0]);
+    punned_t* c_ptr = (punned_t*)MArgument_getUTF8String(Args[0]);
     float* vector_data = nullptr;
     try {
         int label = MArgument_getInteger(Args[1]);
@@ -126,7 +126,7 @@ EXTERN_C DLLEXPORT int IndexAdd(WolframLibraryData libData, mint Argc, MArgument
 }
 
 EXTERN_C DLLEXPORT int IndexSearch(WolframLibraryData libData, mint Argc, MArgument* Args, MArgument Res) {
-    native_index_t* c_ptr = (native_index_t*)MArgument_getUTF8String(Args[0]);
+    punned_t* c_ptr = (punned_t*)MArgument_getUTF8String(Args[0]);
     MTensor matches;
     mint dims[] = {1};
     int wanted = MArgument_getInteger(Args[2]);

package/python/bench.py

@@ -1,146 +0,0 @@
-import time
-import numpy as np
-import fire
-
-import numpy as np
-from numba import cfunc, types, carray
-
-from faiss import IndexHNSWFlat, IndexIVFPQ
-from usearch.index import Index
-from usearch.io import load_matrix
-
-
-def measure(f) -> float:
-    a = time.time_ns()
-    result = f()
-    b = time.time_ns()
-    c = b - a
-    secs = c / (10 ** 9)
-    print(f'- Took: {secs:.2f} seconds')
-    return secs, result
-
-
-def bench_faiss(index, vectors: np.array, queries: np.array, neighbors: np.array):
-    dt, _ = measure(lambda: index.add(vectors))
-    print(f'- Performance: {vectors.shape[0]/dt:.2f} insertions/s')
-
-    dt, results = measure(lambda: index.search(queries, neighbors.shape[1]))
-    print(f'- Performance: {queries.shape[0]/dt:.2f} queries/s')
-
-    recall_at_one = 0.0
-    matches = results[1]
-    for i in range(queries.shape[0]):
-        recall_at_one += neighbors[i, 0] in matches[i, :]
-    recall_at_one /= queries.shape[0]
-    print(f'- Recall@1: {recall_at_one * 100:.2f} %')
-
-
-def bench_usearch(index, vectors: np.array, queries: np.array, neighbors: np.array):
-    labels = np.arange(vectors.shape[0], dtype=np.longlong)
-    assert len(index) == 0
-    dt, _ = measure(lambda: index.add(labels, vectors))
-    assert len(index) == vectors.shape[0]
-    print(f'- Performance: {vectors.shape[0]/dt:.2f} insertions/s')
-
-    dt, results = measure(lambda: index.search(queries, neighbors.shape[1]))
-    print(f'- Performance: {queries.shape[0]/dt:.2f} queries/s')
-
-    recall_at_one = 0.0
-    matches = results[0]
-    for i in range(queries.shape[0]):
-        recall_at_one += neighbors[i, 0] in matches[i, :]
-    recall_at_one /= queries.shape[0]
-    print(f'- Recall@1: {recall_at_one * 100:.2f} %')
-
-# Showcases how to use Numba to JIT-compile similarity measures for USearch.
-# https://numba.readthedocs.io/en/stable/reference/jit-compilation.html#c-callbacks
-
-
-signature_f32 = types.float32(
-    types.CPointer(types.float32),
-    types.CPointer(types.float32),
-    types.size_t, types.size_t)
-
-
-@cfunc(signature_f32)
-def inner_product_f32(a, b, n, m):
-    a_array = carray(a, n)
-    b_array = carray(b, n)
-    c = types.float32(0)
-    for i in range(n):
-        c += a_array[i] * b_array[i]
-    return types.float32(1 - c)
-
-
-def main(
-    vectors: str,
-    queries: str,
-    neighbors: str,
-    connectivity: int = 16,
-    expansion_add: int = 128,
-    expansion_search: int = 64,
-):
-
-    vectors_mat = load_matrix(vectors)
-    queries_mat = load_matrix(queries)
-    neighbors_mat = load_matrix(neighbors)
-    dim = vectors_mat.shape[1]
-
-    print('USearch: HNSW')
-    index = Index(
-        ndim=dim,
-        expansion_add=expansion_add,
-        expansion_search=expansion_search,
-        connectivity=connectivity,
-    )
-    bench_usearch(index, vectors_mat, queries_mat, neighbors_mat)
-
-    print('USearch: HNSW + Numba JIT')
-    index = Index(
-        ndim=dim,
-        expansion_add=expansion_add,
-        expansion_search=expansion_search,
-        connectivity=connectivity,
-        metric_pointer=inner_product_f32.address,
-    )
-    bench_usearch(index, vectors_mat, queries_mat, neighbors_mat)
-
-    print('USearch: HNSW + Half Precision')
-    index = Index(
-        ndim=dim,
-        dtype='f16',
-        expansion_add=expansion_add,
-        expansion_search=expansion_search,
-        connectivity=connectivity,
-    )
-    bench_usearch(index, vectors_mat, queries_mat, neighbors_mat)
-
-    print('USearch: HNSW + Quarter Precision')
-    index = Index(
-        ndim=dim,
-        dtype='f8',
-        expansion_add=expansion_add,
-        expansion_search=expansion_search,
-        connectivity=connectivity,
-    )
-    bench_usearch(index, vectors_mat, queries_mat, neighbors_mat)
-
-    print('FAISS: HNSW')
-    index = IndexHNSWFlat(dim, connectivity)
-    index.hnsw.efSearch = expansion_search
-    index.hnsw.efConstruction = expansion_add
-    bench_faiss(index, vectors_mat, queries_mat, neighbors_mat)
-
-    print('FAISS: HNSW + IVFPQ')
-    nlist = 10000  # Number of inverted lists (number of partitions or cells).
-    nsegment = 16  # Number of segments for PQ (number of subquantizers).
-    nbit = 8       # Number of bits to encode each segment.
-    coarse_quantizer = IndexHNSWFlat(dim, connectivity)
-    index = IndexIVFPQ(coarse_quantizer, dim, nlist, nsegment, nbit)
-    index.train(vectors_mat)
-    index.nprobe = 10
-    bench_faiss(index, vectors_mat, queries_mat, neighbors_mat)
-
-
-if __name__ == '__main__':
-    fire.Fire(main)