Merge branch 'main' into feature/capped-invlists

Author: Tickdack

faiss/CMakeLists.txt

@@ -296,12 +296,13 @@ set(FAISS_HEADERS
   utils/rabitq_simd.h
   utils/random.h
   utils/sorting.h
-  utils/simdlib.h
-  utils/simdlib_avx2.h
-  utils/simdlib_avx512.h
-  utils/simdlib_emulated.h
-  utils/simdlib_neon.h
-  utils/simdlib_ppc64.h
+  impl/simdlib/simdlib.h
+  impl/simdlib/simdlib_dispatch.h
+  impl/simdlib/simdlib_avx2.h
+  impl/simdlib/simdlib_avx512.h
+  impl/simdlib/simdlib_emulated.h
+  impl/simdlib/simdlib_neon.h
+  impl/simdlib/simdlib_ppc64.h
   utils/utils.h
   utils/simd_levels.h
   utils/distances_fused/avx512.h
@@ -321,6 +322,7 @@ set(FAISS_HEADERS
   utils/hamming_distance/avx2-inl.h
   utils/hamming_distance/avx512-inl.h
   utils/simd_impl/distances_autovec-inl.h
+  utils/simd_impl/distances_simdlib256.h
   utils/simd_impl/distances_sse-inl.h
 )
 
@@ -459,6 +461,14 @@ if(FAISS_OPT_LEVEL STREQUAL "dd")
   endif()
 endif()
 
+# NEON is mandatory on ARM64 — ensure COMPILE_SIMD_ARM_NEON is always defined
+# and NEON source files are always compiled into the main faiss target.
+# (On x86, AVX2/AVX512 are optional and only compiled per opt_level above.)
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "(aarch64|arm64|ARM64)")
+  target_compile_definitions(faiss PRIVATE COMPILE_SIMD_ARM_NEON)
+  target_sources(faiss PRIVATE ${FAISS_SIMD_NEON_SRC})
+endif()
+
 if(FAISS_ENABLE_SVS)
   find_package(svs_runtime 0.2.0 REQUIRED)
 

faiss/IndexIVFPQFastScan.cpp

@@ -15,10 +15,10 @@
 
 #include <faiss/impl/AuxIndexStructures.h>
 #include <faiss/impl/FaissAssert.h>
+#include <faiss/impl/simdlib/simdlib_dispatch.h>
 #include <faiss/utils/Heap.h>
 #include <faiss/utils/distances.h>
 #include <faiss/utils/extra_distances.h>
-#include <faiss/utils/simdlib.h>
 
 #include <faiss/invlists/BlockInvertedLists.h>
 
@@ -349,7 +349,7 @@ struct IVFPQFastScanScanner : InvertedListScanner {
         const float* x = index.by_residual ? residual.data() : this->xi;
         float accu = 0;
         // implemented for all vector distances, although only L2 and IP are
-        // suppored by FastScan
+        // supported by FastScan
         with_VectorDistance(pq.dsub, index.metric_type, 0.0, [&](auto vd) {
             int m;
             for (m = 0; m + 1 < pq.M; m += 2) {

faiss/IndexRaBitQFastScan.h

@@ -16,8 +16,8 @@
 #include <faiss/impl/RaBitQUtils.h>
 #include <faiss/impl/RaBitQuantizer.h>
 #include <faiss/impl/simd_result_handlers.h>
+#include <faiss/impl/simdlib/simdlib_dispatch.h>
 #include <faiss/utils/Heap.h>
-#include <faiss/utils/simdlib.h>
 
 namespace faiss {
 

faiss/clone_index.cpp

@@ -102,8 +102,8 @@ IndexIVF* Cloner::clone_IndexIVF(const IndexIVF* ivf) {
     TRYCLONE(IndexIVFRaBitQ, ivf)
 
     TRYCLONE(IndexIVFFlatDedup, ivf)
-    TRYCLONE(IndexIVFFlat, ivf)
     TRYCLONE(IndexIVFFlatPanorama, ivf)
+    TRYCLONE(IndexIVFFlat, ivf)
 
     TRYCLONE(IndexIVFSpectralHash, ivf)
 

faiss/impl/LookupTableScaler.h

@@ -10,7 +10,7 @@
 #include <cstdint>
 #include <cstdlib>
 
-#include <faiss/utils/simdlib.h>
+#include <faiss/impl/simdlib/simdlib_dispatch.h>
 
 /*******************************************
  * The Scaler objects are used to specialize the handling of the

faiss/impl/residual_quantizer_encode_steps.cpp

@@ -11,9 +11,9 @@
 #include <faiss/impl/FaissAssert.h>
 #include <faiss/impl/ResidualQuantizer.h>
 #include <faiss/impl/simd_dispatch.h>
+#include <faiss/impl/simdlib/simdlib_dispatch.h>
 #include <faiss/utils/Heap.h>
 #include <faiss/utils/distances.h>
-#include <faiss/utils/simdlib.h>
 #include <faiss/utils/utils.h>
 
 #include <faiss/utils/approx_topk/approx_topk.h>

faiss/impl/scalar_quantizer/codecs.h

@@ -8,8 +8,8 @@
 #pragma once
 
 #include <faiss/impl/ScalarQuantizer.h>
+#include <faiss/impl/simdlib/simdlib_dispatch.h>
 #include <faiss/utils/simd_levels.h>
-#include <faiss/utils/simdlib.h>
 
 namespace faiss {
 

faiss/impl/scalar_quantizer/distance_computers.h

@@ -8,8 +8,8 @@
 #pragma once
 
 #include <faiss/impl/ScalarQuantizer.h>
+#include <faiss/impl/simdlib/simdlib_dispatch.h>
 #include <faiss/utils/simd_levels.h>
-#include <faiss/utils/simdlib.h>
 
 namespace faiss {
 

faiss/impl/scalar_quantizer/quantizers.h

@@ -8,10 +8,10 @@
 #pragma once
 
 #include <faiss/impl/ScalarQuantizer.h>
+#include <faiss/impl/simdlib/simdlib_dispatch.h>
 #include <faiss/utils/bf16.h>
 #include <faiss/utils/fp16.h>
 #include <faiss/utils/simd_levels.h>
-#include <faiss/utils/simdlib.h>
 
 namespace faiss {
 

faiss/impl/scalar_quantizer/scanners.h

@@ -11,8 +11,8 @@
 #pragma once
 
 #include <faiss/impl/ScalarQuantizer.h>
+#include <faiss/impl/simdlib/simdlib_dispatch.h>
 #include <faiss/utils/simd_levels.h>
-#include <faiss/utils/simdlib.h>
 
 #include <faiss/impl/simd_dispatch.h>