faiss 0.6.1 → 0.6.3

data/vendor/faiss/faiss/VectorTransform.cpp

@@ -47,6 +47,19 @@ int sgemm_(
         float* c,
         FINTEGER* ldc);
 
+int sgemv_(
+        const char* trans,
+        FINTEGER* m,
+        FINTEGER* n,
+        float* alpha,
+        const float* a,
+        FINTEGER* lda,
+        const float* x,
+        FINTEGER* incx,
+        float* beta,
+        float* y,
+        FINTEGER* incy);
+
 int dgemm_(
         const char* transa,
         const char* transb,
@@ -192,6 +205,25 @@ void LinearTransform::apply_noalloc(idx_t n, const float* x, float* xt) const {
 
     float one = 1;
     FINTEGER nbiti = d_out, ni = static_cast<FINTEGER>(n), di = d_in;
+    if (n == 1) {
+        FINTEGER onei = 1;
+        // Avoid GEMM packing overhead for single-vector transforms. GEMV is
+        // mathematically equivalent but may not be bit-exact with GEMM because
+        // BLAS implementations can use different accumulation orders.
+        sgemv_("Transposed",
+               &di,
+               &nbiti,
+               &one,
+               A.data(),
+               &di,
+               x,
+               &onei,
+               &c_factor,
+               xt,
+               &onei);
+        return;
+    }
+
     sgemm_("Transposed",
            "Not transposed",
            &nbiti,
@@ -1064,8 +1096,7 @@ ITQTransform::ITQTransform(int din, int dout, bool do_pca_in)
 }
 
 void ITQTransform::train(idx_t n, const float* x_in) {
-    FAISS_THROW_IF_NOT_MSG(
-            !is_trained, "ITQTransform has already been trained");
+    FAISS_THROW_IF_MSG(is_trained, "ITQTransform has already been trained");
 
     size_t max_train_points = std::max(d_in * max_train_per_dim, 32768);
     const float* x =

data/vendor/faiss/faiss/clone_index.cpp

@@ -18,10 +18,12 @@
 #include <faiss/IndexBinaryFlat.h>
 #include <faiss/IndexBinaryHNSW.h>
 #include <faiss/IndexBinaryIVF.h>
+#include <faiss/IndexEDEN.h>
 #include <faiss/IndexFlat.h>
 #include <faiss/IndexHNSW.h>
 #include <faiss/IndexIVF.h>
 #include <faiss/IndexIVFAdditiveQuantizerFastScan.h>
+#include <faiss/IndexIVFEDEN.h>
 #include <faiss/IndexIVFFlat.h>
 #include <faiss/IndexIVFFlatPanorama.h>
 #include <faiss/IndexIVFPQ.h>
@@ -77,6 +79,7 @@ VectorTransform* Cloner::clone_VectorTransform(const VectorTransform* vt) {
     TRYCLONE(PCAMatrix, vt)
     TRYCLONE(ITQMatrix, vt)
     TRYCLONE(RandomRotationMatrix, vt)
+    TRYCLONE(HadamardRotation, vt)
     TRYCLONE(LinearTransform, vt) {
         FAISS_THROW_MSG("clone not supported for this type of VectorTransform");
     }
@@ -98,6 +101,7 @@ IndexIVF* Cloner::clone_IndexIVF(const IndexIVF* ivf) {
     TRYCLONE(IndexIVFResidualQuantizerFastScan, ivf)
     TRYCLONE(IndexIVFPQFastScan, ivf)
 
+    TRYCLONE(IndexIVFEDEN, ivf)
     TRYCLONE(IndexIVFRaBitQFastScan, ivf)
     TRYCLONE(IndexIVFRaBitQ, ivf)
 
@@ -287,6 +291,7 @@ Index* Cloner::clone_Index(const Index* index) {
     TRYCLONE(IndexLattice, index)
     TRYCLONE(IndexRandom, index)
     TRYCLONE(IndexPQFastScan, index)
+    TRYCLONE(IndexEDEN, index)
 
     TRYCLONE(IndexScalarQuantizer, index)
     TRYCLONE(MultiIndexQuantizer, index)

data/vendor/faiss/faiss/cppcontrib/SaDecodeKernels.h

@@ -313,7 +313,7 @@
 #ifdef __AVX2__
 #include <faiss/cppcontrib/sa_decode/Level2-avx2-inl.h>
 #include <faiss/cppcontrib/sa_decode/PQ-avx2-inl.h>
-#elif defined(__ARM_NEON)
+#elif defined(__ARM_NEON) || defined(__aarch64__)
 #include <faiss/cppcontrib/sa_decode/Level2-neon-inl.h>
 #include <faiss/cppcontrib/sa_decode/PQ-neon-inl.h>
 #else