Narek/external index storage

CMakeLists.txt

@@ -63,18 +63,11 @@ else ()
 endif ()
 
 # Core compilation settings affecting "index.hpp"
-target_compile_definitions(
-    ${USEARCH_TARGET_NAME} INTERFACE $<$<NOT:$<BOOL:${USEARCH_USE_OPENMP}>>:USEARCH_USE_OPENMP=0>
-)
+target_compile_definitions(${USEARCH_TARGET_NAME} INTERFACE "USEARCH_USE_OPENMP=$<BOOL:${USEARCH_USE_OPENMP}>")
 
 # Supplementary compilation settings affecting "index_plugins.hpp"
-target_compile_definitions(
-    ${USEARCH_TARGET_NAME} INTERFACE $<$<NOT:$<BOOL:${USEARCH_USE_FP16LIB}>>:USEARCH_USE_FP16LIB=1>
-)
-
-target_compile_definitions(
-    ${USEARCH_TARGET_NAME} INTERFACE $<$<NOT:$<BOOL:${USEARCH_USE_SIMSIMD}>>:USEARCH_USE_SIMSIMD=0>
-)
+target_compile_definitions(${USEARCH_TARGET_NAME} INTERFACE "USEARCH_USE_FP16LIB=$<BOOL:${USEARCH_USE_FP16LIB}>")
+target_compile_definitions(${USEARCH_TARGET_NAME} INTERFACE "USEARCH_USE_SIMSIMD=$<BOOL:${USEARCH_USE_SIMSIMD}>")
 
 target_include_directories(
     ${USEARCH_TARGET_NAME} ${USEARCH_SYSTEM_INCLUDE} INTERFACE $<BUILD_INTERFACE:${USEARCH_INCLUDE_BUILD_DIR}>
@@ -296,14 +289,12 @@ function (setup_target TARGET_NAME)
     endif ()
 
     # Core compilation settings affecting "index.hpp"
-    target_compile_definitions(${TARGET_NAME} PRIVATE $<$<NOT:$<BOOL:${USEARCH_USE_OPENMP}>>:USEARCH_USE_OPENMP=0>)
+    target_compile_definitions(${TARGET_NAME} PRIVATE "USEARCH_USE_OPENMP=$<BOOL:${USEARCH_USE_OPENMP}>")
 
     # Supplementary compilation settings affecting "index_plugins.hpp"
-    target_compile_definitions(
-        ${TARGET_NAME} PRIVATE $<$<NOT:$<BOOL:${USEARCH_USE_FP16LIB}>>:USEARCH_USE_FP16LIB=1>
-    )
+    target_compile_definitions(${TARGET_NAME} PRIVATE "USEARCH_USE_FP16LIB=$<BOOL:${USEARCH_USE_FP16LIB}>")
+    target_compile_definitions(${TARGET_NAME} PRIVATE "USEARCH_USE_SIMSIMD=$<BOOL:${USEARCH_USE_SIMSIMD}>")
 
-    target_compile_definitions(${TARGET_NAME} PRIVATE $<$<NOT:$<BOOL:${USEARCH_USE_SIMSIMD}>>:USEARCH_USE_SIMSIMD=0>)
 
 endfunction ()
 

c/test.c

@@ -32,6 +32,7 @@ usearch_init_options_t create_options(size_t const dimensions) {
     opts.dimensions = dimensions;
     opts.expansion_add = 40;    // 40 in faiss
     opts.expansion_search = 16; // 10 in faiss
+    opts.multi = false;
     opts.metric_kind = usearch_metric_ip_k;
     opts.metric = NULL;
     opts.quantization = usearch_scalar_f32_k;

cpp/bench.cpp

@@ -42,8 +42,6 @@
 #include <omp.h> // `omp_set_num_threads()`
 #endif
 
-#include <simsimd/simsimd.h>
-
 #include <usearch/index_dense.hpp>
 
 using namespace unum::usearch;
@@ -615,4 +613,4 @@ int main(int argc, char** argv) {
         run_punned<index_dense_gt<default_key_t, std::uint32_t>>(dataset, args, config, limits);
 
     return 0;
-}
+}

cpp/test.cpp

@@ -5,12 +5,12 @@
 #include <cassert>   // `assert`
 #include <random>    // `std::default_random_engine`
 #include <stdexcept>
-#include <unordered_map>
 #include <vector> // for std::vector
 
 #include <usearch/index.hpp>
 #include <usearch/index_dense.hpp>
 #include <usearch/index_plugins.hpp>
+#include <usearch/std_storage.hpp>
 
 using namespace unum::usearch;
 using namespace unum;
@@ -77,7 +77,7 @@ void test_cosine(index_at& index, std::vector<std::vector<scalar_at>> const& vec
     expect((index.stats(0).nodes == 3));
 
     // Check if clustering endpoint compiles
-    index.cluster(vector_first, 0, args...);
+    // index.cluster(vector_first, 0, args...);
 
     // Try removals and replacements
     if constexpr (punned_ak) {
@@ -141,8 +141,8 @@ void test_cosine(index_at& index, std::vector<std::vector<scalar_at>> const& vec
         index.get(key_second, vec_recovered_from_view.data());
         expect(std::equal(vector_second, vector_second + dimensions, vec_recovered_from_view.data()));
 
-        auto compaction_result = index.compact();
-        expect(bool(compaction_result));
+        // auto compaction_result = index.compact();
+        // expect(bool(compaction_result));
     }
 
     expect(index.memory_usage() > 0);
@@ -155,14 +155,15 @@ void test_cosine(index_at& index, std::vector<std::vector<scalar_at>> const& vec
     }
 }
 
-template <typename scalar_at, typename key_at, typename slot_at> //
+template <typename storage_at, typename scalar_at, typename key_at, typename slot_at> //
 void test_cosine(std::size_t collection_size, std::size_t dimensions) {
 
+    using storage_t = storage_at;
     using scalar_t = scalar_at;
     using vector_key_t = key_at;
     using slot_t = slot_at;
 
-    using index_typed_t = index_gt<float, vector_key_t, slot_t>;
+    using index_typed_t = index_gt<storage_t, float, vector_key_t, slot_t>;
     using member_cref_t = typename index_typed_t::member_cref_t;
     using member_citerator_t = typename index_typed_t::member_citerator_t;
 
@@ -197,15 +198,17 @@ void test_cosine(std::size_t collection_size, std::size_t dimensions) {
         std::printf("- templates with connectivity %zu \n", connectivity);
         metric_t metric{&matrix, dimensions};
         index_config_t config(connectivity);
-        index_typed_t index_typed(config);
+        storage_t storage{config};
+        index_typed_t index_typed_tmp(&storage, config);
+        index_typed_t index_typed = std::move(index_typed_tmp);
         test_cosine<false>(index_typed, matrix, metric);
     }
 
     // Type-punned:
     for (bool multi : {false, true}) {
         for (std::size_t connectivity : {3, 13, 50}) {
             std::printf("- punned with connectivity %zu \n", connectivity);
-            using index_t = index_dense_gt<vector_key_t, slot_t>;
+            using index_t = index_dense_gt<vector_key_t, slot_t, storage_t>;
             metric_punned_t metric(dimensions, metric_kind_t::cos_k, scalar_kind<scalar_at>());
             index_dense_config_t config(connectivity);
             config.multi = multi;
@@ -310,9 +313,24 @@ int main(int, char**) {
     for (std::size_t collection_size : {10, 500})
         for (std::size_t dimensions : {97, 256}) {
             std::printf("Indexing %zu vectors with cos: <float, std::int64_t, std::uint32_t> \n", collection_size);
-            test_cosine<float, std::int64_t, std::uint32_t>(collection_size, dimensions);
-            std::printf("Indexing %zu vectors with cos: <float, std::int64_t, uint40_t> \n", collection_size);
-            test_cosine<float, std::int64_t, uint40_t>(collection_size, dimensions);
+            using key_t = std::int64_t;
+            {
+                using slot_t = std::uint32_t;
+                using storage_v2_t = storage_v2_at<key_t, slot_t>;
+                using std_storage_t = std_storage_at<key_t, slot_t>;
+
+                test_cosine<storage_v2_t, float, std::int64_t, std::uint32_t>(collection_size, dimensions);
+                test_cosine<std_storage_t, float, std::int64_t, std::uint32_t>(collection_size, dimensions);
+            }
+            {
+                using slot_t = uint40_t;
+                using storage_v2_t = storage_v2_at<key_t, slot_t>;
+                using std_storage_t = std_storage_at<key_t, slot_t>;
+
+                std::printf("Indexing %zu vectors with cos: <float, std::int64_t, uint40_t> \n", collection_size);
+                test_cosine<storage_v2_t, float, key_t, slot_t>(collection_size, dimensions);
+                test_cosine<std_storage_t, float, key_t, slot_t>(collection_size, dimensions);
+            }
         }
 
     for (std::size_t connectivity : {3, 13, 50})

docs/benchmarks.md

@@ -55,7 +55,7 @@ Also worth noting, 8-bit quantization results in almost no quantization loss and
 
 Within this repository you will find two commonly used utilities:
 
-- `cpp/bench.cpp` the produces the `bench` binary for broad USearch benchmarks.
+- `cpp/bench.cpp` the produces the `bench_cpp` binary for broad USearch benchmarks.
 - `python/bench.py` and `python/bench.ipynb` for interactive charts against FAISS.
 
 To achieve best highest results we suggest compiling locally for the target architecture.
@@ -64,17 +64,19 @@ To achieve best highest results we suggest compiling locally for the target arch
 cmake -B ./build_release \
     -DCMAKE_BUILD_TYPE=Release \
     -DUSEARCH_USE_OPENMP=1 \
+    -DUSEARCH_USE_SIMSIMD=1 \
+    -DUSEARCH_USE_FP16LIB=0 \
     -DUSEARCH_USE_JEMALLOC=1 && \
     make -C ./build_release -j
 
-./build_release/bench --help
+./build_release/bench_cpp --help
 ```
 
 Which would print the following instructions.
 
 ```txt
 SYNOPSIS
-        ./build_release/bench [--vectors <path>] [--queries <path>] [--neighbors <path>] [-b] [-j
+        ./build_release/bench_cpp [--vectors <path>] [--queries <path>] [--neighbors <path>] [-b] [-j
                               <integer>] [-c <integer>] [--expansion-add <integer>]
                               [--expansion-search <integer>] [--native|--f16quant|--i8quant]
                               [--ip|--l2sq|--cos|--haversine] [-h]
@@ -115,12 +117,12 @@ OPTIONS
 Here is an example of running the C++ benchmark:
 
 ```sh
-./build_release/bench \
+./build_release/bench_cpp \
     --vectors datasets/wiki_1M/base.1M.fbin \
     --queries datasets/wiki_1M/query.public.100K.fbin \
     --neighbors datasets/wiki_1M/groundtruth.public.100K.ibin
 
-./build_release/bench \
+./build_release/bench_cpp \
     --vectors datasets/t2i_1B/base.1B.fbin \
     --queries datasets/t2i_1B/query.public.100K.fbin \
     --neighbors datasets/t2i_1B/groundtruth.public.100K.ibin \
@@ -205,17 +207,17 @@ With `perf`:
 
 ```sh
 # Pass environment variables with `-E`, and `-d` for details
-sudo -E perf stat -d ./build_release/bench ...
-sudo -E perf mem -d ./build_release/bench ...
+sudo -E perf stat -d ./build_release/bench_cpp ...
+sudo -E perf mem -d ./build_release/bench_cpp ...
 # Sample on-CPU functions for the specified command, at 1 Kilo Hertz:
-sudo -E perf record -F 1000 ./build_release/bench ...
-perf record -d -e arm_spe// -- ./build_release/bench ..
+sudo -E perf record -F 1000 ./build_release/bench_cpp ...
+perf record -d -e arm_spe// -- ./build_release/bench_cpp ..
 ```
 
 ### Caches
 
 ```sh
-sudo perf stat -e 'faults,dTLB-loads,dTLB-load-misses,cache-misses,cache-references' ./build_release/bench ...
+sudo perf stat -e 'faults,dTLB-loads,dTLB-load-misses,cache-misses,cache-references' ./build_release/bench_cpp ...
 ```
 
 Typical output on a 1M vectors dataset is:
@@ -242,4 +244,3 @@ sudo sysctl -w vm.nr_hugepages=2048
 sudo reboot
 sudo cat /proc/sys/vm/nr_hugepages
 ```
-