faiss 0.4.3 → 0.5.0

data/vendor/faiss/faiss/index_factory.cpp

@@ -27,10 +27,12 @@
 #include <faiss/IndexIVFAdditiveQuantizer.h>
 #include <faiss/IndexIVFAdditiveQuantizerFastScan.h>
 #include <faiss/IndexIVFFlat.h>
+#include <faiss/IndexIVFFlatPanorama.h>
 #include <faiss/IndexIVFPQ.h>
 #include <faiss/IndexIVFPQFastScan.h>
 #include <faiss/IndexIVFPQR.h>
 #include <faiss/IndexIVFRaBitQ.h>
+#include <faiss/IndexIVFRaBitQFastScan.h>
 #include <faiss/IndexIVFSpectralHash.h>
 #include <faiss/IndexLSH.h>
 #include <faiss/IndexLattice.h>
@@ -39,6 +41,7 @@
 #include <faiss/IndexPQFastScan.h>
 #include <faiss/IndexPreTransform.h>
 #include <faiss/IndexRaBitQ.h>
+#include <faiss/IndexRaBitQFastScan.h>
 #include <faiss/IndexRefine.h>
 #include <faiss/IndexRowwiseMinMax.h>
 #include <faiss/IndexScalarQuantizer.h>
@@ -49,6 +52,9 @@
 #include <faiss/IndexBinaryHNSW.h>
 #include <faiss/IndexBinaryHash.h>
 #include <faiss/IndexBinaryIVF.h>
+#include <faiss/IndexIDMap.h>
+#include <algorithm>
+#include <cctype>
 #include <string>
 
 namespace faiss {
@@ -326,6 +332,10 @@ IndexIVF* parse_IndexIVF(
     if (match("FlatDedup")) {
         return new IndexIVFFlatDedup(get_q(), d, nlist, mt, own_il);
     }
+    if (match("FlatPanorama([0-9]+)?")) {
+        int nlevels = mres_to_int(sm[1], 8); // default to 8 levels
+        return new IndexIVFFlatPanorama(get_q(), d, nlist, nlevels, mt, own_il);
+    }
     if (match(sq_pattern)) {
         return new IndexIVFScalarQuantizer(
                 get_q(),
@@ -450,6 +460,10 @@ IndexIVF* parse_IndexIVF(
     if (match(rabitq_pattern)) {
         return new IndexIVFRaBitQ(get_q(), d, nlist, mt, own_il);
     }
+    if (match("RaBitQfs(_[0-9]+)?")) {
+        int bbs = mres_to_int(sm[1], 32, 1);
+        return new IndexIVFRaBitQFastScan(get_q(), d, nlist, mt, bbs, own_il);
+    }
     return nullptr;
 }
 
@@ -676,6 +690,12 @@ Index* parse_other_indexes(
         return new IndexRaBitQ(d, metric);
     }
 
+    // IndexRaBitQFastScan
+    if (match("RaBitQfs(_[0-9]+)?")) {
+        int bbs = mres_to_int(sm[1], 32, 1);
+        return new IndexRaBitQFastScan(d, metric, bbs);
+    }
+
     return nullptr;
 }
 
@@ -934,6 +954,28 @@ IndexBinary* index_binary_factory(
         bool own_invlists) {
     IndexBinary* index = nullptr;
 
+    std::smatch sm;
+    std::string desc_str(description);
+
+    // Handle IDMap2 and IDMap wrappers (prefix or suffix)
+    if (re_match(desc_str, "(.+),IDMap2", sm) ||
+        re_match(desc_str, "IDMap2,(.+)", sm)) {
+        IndexBinary* sub_index =
+                index_binary_factory(d, sm[1].str().c_str(), own_invlists);
+        IndexBinaryIDMap2* idmap2 = new IndexBinaryIDMap2(sub_index);
+        idmap2->own_fields = true;
+        return idmap2;
+    }
+
+    if (re_match(desc_str, "(.+),IDMap", sm) ||
+        re_match(desc_str, "IDMap,(.+)", sm)) {
+        IndexBinary* sub_index =
+                index_binary_factory(d, sm[1].str().c_str(), own_invlists);
+        IndexBinaryIDMap* idmap = new IndexBinaryIDMap(sub_index);
+        idmap->own_fields = true;
+        return idmap;
+    }
+
     int ncentroids = -1;
     int M, nhash, b;
 
@@ -959,7 +1001,7 @@ IndexBinary* index_binary_factory(
     } else if (sscanf(description, "BHash%d", &b) == 1) {
         index = new IndexBinaryHash(d, b);
 
-    } else if (std::string(description) == "BFlat") {
+    } else if (desc_str == "BFlat") {
         index = new IndexBinaryFlat(d);
 
     } else {

data/vendor/faiss/faiss/index_factory.h

@@ -12,7 +12,7 @@
 
 namespace faiss {
 
-/** Build and index with the sequence of processing steps described in
+/** Build an index with the sequence of processing steps described in
  *  the string. */
 Index* index_factory(
         int d,

data/vendor/faiss/faiss/index_io.h

@@ -16,7 +16,7 @@
  * object that abstracts the medium.
  *
  * The read functions return objects that should be deallocated with
- * delete. All references within these objectes are owned by the
+ * delete. All references within these objects are owned by the
  * object.
  */
 

data/vendor/faiss/faiss/invlists/InvertedLists.cpp

@@ -346,6 +346,211 @@ void ArrayInvertedLists::permute_invlists(const idx_t* map) {
 
 ArrayInvertedLists::~ArrayInvertedLists() {}
 
+/***********************************************
+ * ArrayInvertedListsPanorama implementation
+ **********************************************/
+
+ArrayInvertedListsPanorama::ArrayInvertedListsPanorama(
+        size_t nlist,
+        size_t code_size,
+        size_t n_levels)
+        : ArrayInvertedLists(nlist, code_size),
+          n_levels(n_levels),
+          level_width(
+                  (((code_size / sizeof(float)) + n_levels - 1) / n_levels) *
+                  sizeof(float)) {
+    FAISS_THROW_IF_NOT(n_levels > 0);
+    FAISS_THROW_IF_NOT(code_size % sizeof(float) == 0);
+    FAISS_THROW_IF_NOT_MSG(
+            !use_iterator,
+            "IndexIVFFlatPanorama does not support iterators, use vanilla IndexIVFFlat instead");
+    FAISS_ASSERT(level_width % sizeof(float) == 0);
+
+    cum_sums.resize(nlist);
+}
+
+const float* ArrayInvertedListsPanorama::get_cum_sums(size_t list_no) const {
+    assert(list_no < nlist);
+    return cum_sums[list_no].data();
+}
+
+size_t ArrayInvertedListsPanorama::add_entries(
+        size_t list_no,
+        size_t n_entry,
+        const idx_t* ids_in,
+        const uint8_t* code) {
+    assert(list_no < nlist);
+    size_t o = ids[list_no].size();
+
+    ids[list_no].resize(o + n_entry);
+    memcpy(&ids[list_no][o], ids_in, sizeof(ids_in[0]) * n_entry);
+
+    size_t new_size = o + n_entry;
+    size_t num_batches = (new_size + kBatchSize - 1) / kBatchSize;
+    codes[list_no].resize(num_batches * kBatchSize * code_size);
+    cum_sums[list_no].resize(num_batches * kBatchSize * (n_levels + 1));
+
+    copy_codes_to_level_layout(list_no, o, n_entry, code);
+    compute_cumulative_sums(list_no, o, n_entry, code);
+
+    return o;
+}
+
+void ArrayInvertedListsPanorama::update_entries(
+        size_t list_no,
+        size_t offset,
+        size_t n_entry,
+        const idx_t* ids_in,
+        const uint8_t* code) {
+    assert(list_no < nlist);
+    assert(n_entry + offset <= ids[list_no].size());
+
+    memcpy(&ids[list_no][offset], ids_in, sizeof(ids_in[0]) * n_entry);
+    copy_codes_to_level_layout(list_no, offset, n_entry, code);
+    compute_cumulative_sums(list_no, offset, n_entry, code);
+}
+
+void ArrayInvertedListsPanorama::resize(size_t list_no, size_t new_size) {
+    ids[list_no].resize(new_size);
+
+    size_t num_batches = (new_size + kBatchSize - 1) / kBatchSize;
+    codes[list_no].resize(num_batches * kBatchSize * code_size);
+    cum_sums[list_no].resize(num_batches * kBatchSize * (n_levels + 1));
+}
+
+const uint8_t* ArrayInvertedListsPanorama::get_single_code(
+        size_t list_no,
+        size_t offset) const {
+    assert(list_no < nlist);
+    assert(offset < ids[list_no].size());
+
+    uint8_t* recons_buffer = new uint8_t[code_size];
+
+    const uint8_t* codes_base = codes[list_no].data();
+
+    size_t batch_no = offset / kBatchSize;
+    size_t pos_in_batch = offset % kBatchSize;
+    size_t batch_offset = batch_no * kBatchSize * code_size;
+
+    for (size_t level = 0; level < n_levels; level++) {
+        size_t level_offset = level * level_width * kBatchSize;
+        const uint8_t* src = codes_base + batch_offset + level_offset +
+                pos_in_batch * level_width;
+        uint8_t* dest = recons_buffer + level * level_width;
+        size_t copy_size =
+                std::min(level_width, code_size - level * level_width);
+        memcpy(dest, src, copy_size);
+    }
+
+    return recons_buffer;
+}
+
+void ArrayInvertedListsPanorama::release_codes(
+        size_t list_no,
+        const uint8_t* codes) const {
+    // Only delete if it's heap-allocated (from get_single_code).
+    // If it's from get_codes (raw storage), it will be codes[list_no].data()
+    if (codes != this->codes[list_no].data()) {
+        delete[] codes;
+    }
+}
+
+InvertedListsIterator* ArrayInvertedListsPanorama::get_iterator(
+        size_t /* list_no */,
+        void* /* inverted_list_context */) const {
+    FAISS_THROW_MSG(
+            "IndexIVFFlatPanorama does not support iterators, use vanilla IndexIVFFlat instead");
+    return nullptr;
+}
+
+void ArrayInvertedListsPanorama::compute_cumulative_sums(
+        size_t list_no,
+        size_t offset,
+        size_t n_entry,
+        const uint8_t* code) {
+    // Cast to float* is safe here as we guarantee codes are always float
+    // vectors for `IndexIVFFlatPanorama` (verified by the constructor).
+    const float* vectors = reinterpret_cast<const float*>(code);
+    const size_t d = code_size / sizeof(float);
+
+    std::vector<float> suffix_sums(d + 1);
+
+    for (size_t entry_idx = 0; entry_idx < n_entry; entry_idx++) {
+        size_t current_pos = offset + entry_idx;
+        size_t batch_no = current_pos / kBatchSize;
+        size_t pos_in_batch = current_pos % kBatchSize;
+
+        const float* vector = vectors + entry_idx * d;
+
+        // Compute suffix sums of squared values.
+        suffix_sums[d] = 0.0f;
+        for (int j = d - 1; j >= 0; j--) {
+            float squared_val = vector[j] * vector[j];
+            suffix_sums[j] = suffix_sums[j + 1] + squared_val;
+        }
+
+        // Store cumulative sums in batch-oriented layout.
+        size_t cumsum_batch_offset = batch_no * kBatchSize * (n_levels + 1);
+        float* cumsum_base = cum_sums[list_no].data();
+
+        const size_t level_width_floats = level_width / sizeof(float);
+        for (size_t level = 0; level < n_levels; level++) {
+            size_t start_idx = level * level_width_floats;
+            size_t cumsum_offset =
+                    cumsum_batch_offset + level * kBatchSize + pos_in_batch;
+            if (start_idx < d) {
+                cumsum_base[cumsum_offset] = sqrt(suffix_sums[start_idx]);
+            } else {
+                cumsum_base[cumsum_offset] = 0.0f;
+            }
+        }
+
+        // Last level sum is always 0.
+        size_t cumsum_offset =
+                cumsum_batch_offset + n_levels * kBatchSize + pos_in_batch;
+        cumsum_base[cumsum_offset] = 0.0f;
+    }
+}
+
+// Helper method to copy codes into level-oriented batch layout at a given
+// offset in the list.
+void ArrayInvertedListsPanorama::copy_codes_to_level_layout(
+        size_t list_no,
+        size_t offset,
+        size_t n_entry,
+        const uint8_t* code) {
+    uint8_t* codes_base = codes[list_no].data();
+    size_t current_pos = offset;
+    for (size_t entry_idx = 0; entry_idx < n_entry;) {
+        // Determine which batch we're in and position within that batch.
+        size_t batch_no = current_pos / kBatchSize;
+        size_t pos_in_batch = current_pos % kBatchSize;
+        size_t entries_in_this_batch =
+                std::min(n_entry - entry_idx, kBatchSize - pos_in_batch);
+
+        // Copy entries into level-oriented layout for this batch.
+        size_t batch_offset = batch_no * kBatchSize * code_size;
+        for (size_t level = 0; level < n_levels; level++) {
+            size_t level_offset = level * level_width * kBatchSize;
+            size_t start_byte = level * level_width;
+            size_t copy_size =
+                    std::min(level_width, code_size - level * level_width);
+
+            for (size_t i = 0; i < entries_in_this_batch; i++) {
+                const uint8_t* src =
+                        code + (entry_idx + i) * code_size + start_byte;
+                uint8_t* dest = codes_base + batch_offset + level_offset +
+                        (pos_in_batch + i) * level_width;
+
+                memcpy(dest, src, copy_size);
+            }
+        }
+
+        entry_idx += entries_in_this_batch;
+        current_pos += entries_in_this_batch;
+    }
+}
+
 /*****************************************************************
  * Meta-inverted list implementations
  *****************************************************************/

data/vendor/faiss/faiss/invlists/InvertedLists.h

@@ -276,6 +276,68 @@ struct ArrayInvertedLists : InvertedLists {
     ~ArrayInvertedLists() override;
 };
 
+/// Level-oriented storage as defined in the IVFFlat section of Panorama
+/// (https://www.arxiv.org/pdf/2510.00566).
+struct ArrayInvertedListsPanorama : ArrayInvertedLists {
+    static constexpr size_t kBatchSize = 128;
+    std::vector<MaybeOwnedVector<float>> cum_sums;
+    const size_t n_levels;
+    const size_t level_width; // in code units
+
+    ArrayInvertedListsPanorama(size_t nlist, size_t code_size, size_t n_levels);
+
+    const float* get_cum_sums(size_t list_no) const;
+
+    size_t add_entries(
+            size_t list_no,
+            size_t n_entry,
+            const idx_t* ids,
+            const uint8_t* code) override;
+
+    void update_entries(
+            size_t list_no,
+            size_t offset,
+            size_t n_entry,
+            const idx_t* ids,
+            const uint8_t* code) override;
+
+    void resize(size_t list_no, size_t new_size) override;
+
+    /// Panorama's layout make it impractical to support iterators as defined
+    /// by Faiss (i.e. `InvertedListsIterator` API). The iterator would require
+    /// to allocate and reassemble the vector at each call.
+    /// Hence, we override this method to throw an error, this effectively
+    /// disables the `iterate_codes` and `iterate_codes_range` methods.
+    InvertedListsIterator* get_iterator(
+            size_t list_no,
+            void* inverted_list_context = nullptr) const override;
+
+    /// Reconstructs a single code from level-oriented storage to flat format.
+    const uint8_t* get_single_code(size_t list_no, size_t offset)
+            const override;
+
+    /// Frees codes returned by `get_single_code`.
+    void release_codes(size_t list_no, const uint8_t* codes) const override;
+
+   private:
+    /// Helper method to copy codes into level-oriented batch layout at a given
+    /// offset in the list.
+    void copy_codes_to_level_layout(
+            size_t list_no,
+            size_t offset,
+            size_t n_entry,
+            const uint8_t* code);
+
+    /// Helper method to compute the cumulative sums of the codes.
+    /// The cumsums also follow the level-oriented batch layout to minimize the
+    /// number of random memory accesses.
+    void compute_cumulative_sums(
+            size_t list_no,
+            size_t offset,
+            size_t n_entry,
+            const uint8_t* code);
+};
+
 /*****************************************************************
  * Meta-inverted lists
  *

data/vendor/faiss/faiss/utils/AlignedTable.h

@@ -25,7 +25,7 @@ inline bool is_aligned_pointer(const void* x) {
 }
 
 // class that manages suitably aligned arrays for SIMD
-// T should be a POV type. The default alignment is 32 for AVX
+// T should be a POD type. The default alignment is 32 for AVX
 template <class T, int A = 32>
 struct AlignedTableTightAlloc {
     T* ptr;

data/vendor/faiss/faiss/utils/Heap.cpp

@@ -139,7 +139,7 @@ void HeapArray<C>::per_line_extrema(T* out_val, TI* out_ids) const {
     }
 }
 
-// explicit instanciations
+// explicit instantiations
 
 template struct HeapArray<CMin<float, int64_t>>;
 template struct HeapArray<CMax<float, int64_t>>;
@@ -238,7 +238,7 @@ void merge_knn_results(
     }
 }
 
-// explicit instanciations
+// explicit instantiations
 #define INSTANTIATE(C, distance_t)                                \
     template void merge_knn_results<int64_t, C<distance_t, int>>( \
             size_t,                                               \

data/vendor/faiss/faiss/utils/Heap.h

@@ -150,7 +150,7 @@ inline void heap_replace_top(
     bh_ids[i] = id;
 }
 
-/* Partial instanciation for heaps with TI = int64_t */
+/* Partial instantiation for heaps with TI = int64_t */
 
 template <typename T>
 inline void minheap_pop(size_t k, T* bh_val, int64_t* bh_ids) {
@@ -393,7 +393,7 @@ inline void heap_addn(
     }
 }
 
-/* Partial instanciation for heaps with TI = int64_t */
+/* Partial instantiation for heaps with TI = int64_t */
 
 template <typename T>
 inline void minheap_addn(
@@ -489,7 +489,7 @@ struct HeapArray {
         return val + key * k;
     }
 
-    /// Correspponding identifiers
+    /// Corresponding identifiers
     TI* get_ids(size_t key) {
         return ids + key * k;
     }

data/vendor/faiss/faiss/utils/NeuralNet.cpp

@@ -71,7 +71,7 @@ Tensor2DTemplate<T> Tensor2DTemplate<T>::column(size_t j) const {
     return out;
 }
 
-// explicit template instanciation
+// explicit template instantiation
 template struct Tensor2DTemplate<float>;
 template struct Tensor2DTemplate<int32_t>;
 

data/vendor/faiss/faiss/utils/NeuralNet.h

@@ -75,7 +75,7 @@ struct Embedding {
 };
 
 /// Feed forward layer that expands to a hidden dimension, applies a ReLU non
-/// linearity and maps back to the orignal dimension
+/// linearity and maps back to the original dimension
 struct FFN {
     Linear linear1, linear2;
 
@@ -103,7 +103,7 @@ struct QINCoStep {
         return residual_blocks[i];
     }
 
-    /** encode a set of vectors x with intial estimate xhat. Optionally return
+    /** encode a set of vectors x with initial estimate xhat. Optionally return
      * the delta to be added to xhat to form the new xhat */
     nn::Int32Tensor2D encode(
             const nn::Tensor2D& xhat,
@@ -141,7 +141,7 @@ struct QINCo : NeuralNetCodec {
 
     nn::Int32Tensor2D encode(const nn::Tensor2D& x) const override;
 
-    virtual ~QINCo() {}
+    virtual ~QINCo() override {}
 };
 
 } // namespace faiss

data/vendor/faiss/faiss/utils/approx_topk/approx_topk.h

@@ -50,8 +50,8 @@
 //        for j in range(0, NBUCKETS):
 //          idx = beam * n + i * NBUCKETS + j
 //          if distances[idx] < local_min_distances[j]:
-//            local_min_distances[i] = distances[idx]
-//            local_min_indices[i] = indices[idx]
+//            local_min_distances[j] = distances[idx]
+//            local_min_indices[j] = indices[idx]
 //
 //    for j in range(0, NBUCKETS):
 //      heap.push(local_min_distances[j], local_min_indices[j])

data/vendor/faiss/faiss/utils/approx_topk/avx2-inl.h

@@ -106,7 +106,7 @@ struct HeapWithBuckets<CMax<float, int>, NBUCKETS, N> {
                                 distance_candidate,
                                 _CMP_LE_OS);
 
-                        // // blend seems to be slower that min
+                        // // blend seems to be slower than min
                         // const __m256 min_distances_new = _mm256_blendv_ps(
                         //         distance_candidate,
                         //         min_distances_i[j][p],
@@ -120,7 +120,7 @@ struct HeapWithBuckets<CMax<float, int>, NBUCKETS, N> {
                                                 min_indices_i[j][p]),
                                         comparison));
 
-                        // // blend seems to be slower that min
+                        // // blend seems to be slower than min
                         // const __m256 max_distances_new = _mm256_blendv_ps(
                         //         min_distances_i[j][p],
                         //         distance_candidate,

data/vendor/faiss/faiss/utils/approx_topk/mode.h

@@ -21,7 +21,7 @@
 /// It seems that only the limited number of combinations are
 /// meaningful, because of the limited supply of SIMD registers.
 /// Also, certain combinations, such as B32_D1 and B16_D1, were concluded
-/// to be not very precise in benchmarks, so ones were not introduced.
+/// to be not very precise in benchmarks, so they were not introduced.
 ///
 /// TODO: Consider d-ary SIMD heap.
 

data/vendor/faiss/faiss/utils/distances.h

@@ -324,7 +324,7 @@ void knn_inner_product(
  *  vector y, for the L2 distance
  * @param x    query vectors, size nx * d
  * @param y    database vectors, size ny * d
- * @param res  result heap strcture, which also provides k. Sorted on output
+ * @param res  result heap structure, which also provides k. Sorted on output
  * @param y_norm2    (optional) norms for the y vectors (nullptr or size ny)
  * @param sel  search in this subset of vectors
  */
@@ -389,7 +389,7 @@ void knn_inner_products_by_idx(
  * @param x    query vectors, size nx * d
  * @param y    database vectors, size (max(ids) + 1) * d
  * @param subset subset of database vectors to consider, size (nx, nsubset)
- * @param res  rIDesult structure
+ * @param res  result structure
  * @param ld_subset stride for the subset array. -1: use nsubset, 0: all queries
  * process the same subset
  */

data/vendor/faiss/faiss/utils/extra_distances-inl.h

@@ -5,6 +5,8 @@
  * LICENSE file in the root directory of this source tree.
  */
 
+#pragma once
+
 /** In this file are the implementations of extra metrics beyond L2
  *  and inner product */
 
@@ -188,7 +190,7 @@ inline float VectorDistance<METRIC_GOWER>::operator()(
 
 /***************************************************************************
  * Dispatching function that takes a metric type and a consumer object
- * the consumer object should contain a retun type T and a operation template
+ * the consumer object should contain a return type T and a operation template
  * function f() that is called to perform the operation. The first argument
  * of the function is the VectorDistance object. The rest are passed in as is.
  **************************************************************************/

data/vendor/faiss/faiss/utils/hamming-inl.h

@@ -5,6 +5,8 @@
  * LICENSE file in the root directory of this source tree.
  */
 
+#pragma once
+
 namespace faiss {
 
 // BitstringWriter and BitstringReader functions

data/vendor/faiss/faiss/utils/hamming.cpp

@@ -257,12 +257,13 @@ void hammings_knn_mc(
 
     std::vector<HCounterState<HammingComputer>> cs;
     for (size_t i = 0; i < na; ++i) {
-        cs.push_back(HCounterState<HammingComputer>(
-                all_counters.data() + i * nBuckets,
-                all_ids_per_dis.get() + i * nBuckets * k,
-                a + i * bytes_per_code,
-                8 * bytes_per_code,
-                k));
+        cs.push_back(
+                HCounterState<HammingComputer>(
+                        all_counters.data() + i * nBuckets,
+                        all_ids_per_dis.get() + i * nBuckets * k,
+                        a + i * bytes_per_code,
+                        8 * bytes_per_code,
+                        k));
     }
 
     const size_t block_size = hamming_batch_size;

data/vendor/faiss/faiss/utils/hamming.h

@@ -14,7 +14,7 @@
  * fvecs2bitvecs).
  *
  * User-defined type hamdis_t is used for distances because at this time
- * it is still uncler clear how we will need to balance
+ * it is still unclear clear how we will need to balance
  * - flexibility in vector size (may need 16- or even 8-bit vectors)
  * - memory usage
  * - cache-misses when dealing with large volumes of data (fewer bits is better)

data/vendor/faiss/faiss/utils/hamming_distance/common.h

@@ -30,8 +30,7 @@ inline int popcount64(uint64_t x) {
 // This table was moved from .cpp to .h file, because
 // otherwise it was causing compilation errors while trying to
 // compile swig modules on Windows.
-// todo for C++17: switch to 'inline constexpr'
-static constexpr uint8_t hamdis_tab_ham_bytes[256] = {
+inline constexpr uint8_t hamdis_tab_ham_bytes[256] = {
         0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4, 1, 2, 2, 3, 2, 3, 3, 4,
         2, 3, 3, 4, 3, 4, 4, 5, 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
         2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, 1, 2, 2, 3, 2, 3, 3, 4,

data/vendor/faiss/faiss/utils/partitioning.cpp

@@ -140,7 +140,7 @@ typename C::T partition_fuzzy_median3(
     using T = typename C::T;
 
     // here we use bissection with a median of 3 to find the threshold and
-    // compress the arrays afterwards. So it's a n*log(n) algoirithm rather than
+    // compress the arrays afterwards. So it's a n*log(n) algorithm rather than
     // qselect's O(n) but it avoids shuffling around the array.
 
     FAISS_THROW_IF_NOT(n >= 3);
@@ -350,7 +350,7 @@ int simd_compress_array(
         }
     }
 
-    // handle remaining, only striclty lt ones.
+    // handle remaining, only strictly lt ones.
     for (; i0 + 15 < n; i0 += 16) {
         simd16uint16 v(vals + i0);
         simd16uint16 max2 = max_func<C>(v, thr16);
@@ -506,7 +506,7 @@ uint16_t simd_partition_fuzzy_with_bounds(
 
     uint64_t t2 = get_cy();
 
-    partition_stats.bissect_cycles += t1 - t0;
+    partition_stats.bisect_cycles += t1 - t0;
     partition_stats.compress_cycles += t2 - t1;
 
     return thresh;
@@ -662,7 +662,7 @@ uint16_t simd_partition_fuzzy_with_bounds_histogram(
         }
     }
 
-    IFV printf("end bissection: thresh=%d q=%ld n_eq=%ld\n", thresh, q, n_eq);
+    IFV printf("end bisection: thresh=%d q=%ld n_eq=%ld\n", thresh, q, n_eq);
 
     if (!C::is_max) {
         if (n_eq == 0) {
@@ -762,7 +762,7 @@ typename C::T partition_fuzzy(
             vals, ids, n, q_min, q_max, q_out);
 }
 
-// explicit template instanciations
+// explicit template instantiations
 
 template float partition_fuzzy<CMin<float, int64_t>>(
         float* vals,