DD: give PQCodeDistanceScalar a SIMD level so the AVX2 path cannot run AVX512 code (#5719)

Summary:
Pull Request resolved: https://github.com/facebookresearch/faiss/pull/5719

`PQCodeDistanceScalar<PQDecoderT>` is defined in a header and instantiated in every per-SIMD translation unit (`avx2.cpp` with `-mavx2`, `avx512.cpp` with `-mavx512*`, the generic TU). All copies have the same symbol, so the linker keeps one. In a dynamic-dispatch x86 build it kept the `-mavx512*` copy: `PQCodeDistanceScalar<PQDecoderGeneric>::distance_four_codes` contained 93 `zmm` instructions and was called from `PQCodeDistance<PQDecoderGeneric, SIMDLevel::AVX2>`. On a CPU without AVX512 (AMD Milan, T1_MLN), IVFPQ search with a non-8-bit PQ (`PQDecoderGeneric`, also `PQDecoder16`) on the AVX2 path raises SIGILL.

Found by running the faiss C++ tests under the vendored `qemu-x86_64`, which has no AVX512: `IVFPQFastScan.*` (`test_fast_scan_distance_to_code`) died with SIGILL at `vmovdqa64 ...,%zmm0` in that function. Same class as S672695 (AVX512 code reached on a CPU without it); found before any non-AVX512 host was added to the Cogwheel suite.

Fix: template `PQCodeDistanceScalar` on `SIMDLevel` as well, as the DD migration guide prescribes for code compiled per level, so each TU's instantiation is a distinct symbol. `PQCodeDistance<D, SL>` passes its level; the NONE and ARM_NEON specializations in `pq_code_distance-generic.h` pass theirs. Applied to both definitions (`impl/pq_code_distance/pq_code_distance-inl.h` and `utils/pq_code_distance.h`). `SL` defaults to `SIMDLevel::NONE` so code outside faiss that names only the decoder keeps compiling: `unicorn/features/encoders/PQFSTable.h` uses `faiss::PQCodeDistanceScalar<PQDecoderT>`, and without the default 15 Unicorn targets failed to build on this diff.

Reviewed By: mnorris11

Differential Revision: D124399620

fbshipit-source-id: ece1baf0e1eba39712f881976dea33fe22b265c1

faiss/impl/pq_code_distance/pq_code_distance-generic.h

@@ -20,8 +20,8 @@ inline float pq_code_distance_8bit_single_impl<SIMDLevel::NONE>(
         size_t M,
         const float* sim_table,
         const uint8_t* code) {
-    return PQCodeDistanceScalar<PQDecoder8>::distance_single_code(
+    return PQCodeDistanceScalar<PQDecoder8, SIMDLevel::NONE>::
-            M, 8, sim_table, code);
+            distance_single_code(M, 8, sim_table, code);
 }
 
 // NOLINTNEXTLINE(facebook-hte-MisplacedTemplateSpecialization)
@@ -37,7 +37,7 @@ inline void pq_code_distance_8bit_four_impl<SIMDLevel::NONE>(
         float& result1,
         float& result2,
         float& result3) {
-    PQCodeDistanceScalar<PQDecoder8>::distance_four_codes(
+    PQCodeDistanceScalar<PQDecoder8, SIMDLevel::NONE>::distance_four_codes(
             M,
             8,
             sim_table,
@@ -60,8 +60,8 @@ inline float pq_code_distance_8bit_single_impl<SIMDLevel::ARM_NEON>(
         size_t M,
         const float* sim_table,
         const uint8_t* code) {
-    return PQCodeDistanceScalar<PQDecoder8>::distance_single_code(
+    return PQCodeDistanceScalar<PQDecoder8, SIMDLevel::ARM_NEON>::
-            M, 8, sim_table, code);
+            distance_single_code(M, 8, sim_table, code);
 }
 
 // NOLINTNEXTLINE(facebook-hte-MisplacedTemplateSpecialization)
@@ -77,7 +77,7 @@ inline void pq_code_distance_8bit_four_impl<SIMDLevel::ARM_NEON>(
         float& result1,
         float& result2,
         float& result3) {
-    PQCodeDistanceScalar<PQDecoder8>::distance_four_codes(
+    PQCodeDistanceScalar<PQDecoder8, SIMDLevel::ARM_NEON>::distance_four_codes(
             M,
             8,
             sim_table,

faiss/impl/pq_code_distance/pq_code_distance-inl.h

@@ -62,10 +62,14 @@ namespace pq_code_distance {
  *********************************************************************/
 
 /// Scalar PQ code distance implementation.
-/// Templated only on decoder type, independent of SIMD level.
 /// Used directly by non-PQDecoder8 decoders (PQDecoder16,
 /// PQDecoderGeneric) and as fallback for PQDecoder8 at NONE/NEON.
-template <typename PQDecoderT>
+// SL keeps each SIMD-level TU's instantiation a distinct symbol. Without it
+// the linker merges them and may keep one compiled with -mavx512*, which
+// then runs on the AVX2 path and raises SIGILL on CPUs without AVX512.
+// The default keeps callers outside faiss that name only the decoder (e.g.
+// unicorn/features/encoders/PQFSTable.h) compiling; faiss passes SL itself.
+template <typename PQDecoderT, SIMDLevel SL = SIMDLevel::NONE>
 struct PQCodeDistanceScalar {
     using PQDecoder = PQDecoderT;
 
@@ -168,7 +172,7 @@ struct PQCodeDistance {
         if constexpr (std::is_same_v<PQDecoderT, PQDecoder8>) {
             return pq_code_distance_8bit_single_impl<SL>(M, sim_table, code);
         } else {
-            return PQCodeDistanceScalar<PQDecoderT>::distance_single_code(
+            return PQCodeDistanceScalar<PQDecoderT, SL>::distance_single_code(
                     M, nbits, sim_table, code);
         }
     }
@@ -198,7 +202,7 @@ struct PQCodeDistance {
                     result2,
                     result3);
         } else {
-            PQCodeDistanceScalar<PQDecoderT>::distance_four_codes(
+            PQCodeDistanceScalar<PQDecoderT, SL>::distance_four_codes(
                     M,
                     nbits,
                     sim_table,

faiss/utils/pq_code_distance.h

@@ -53,10 +53,14 @@ namespace pq_code_distance {
  *********************************************************************/
 
 /// Scalar PQ code distance implementation.
-/// Templated only on decoder type, independent of SIMD level.
 /// Used directly by non-PQDecoder8 decoders (PQDecoder16,
 /// PQDecoderGeneric) and as fallback for PQDecoder8 at NONE/NEON.
-template <typename PQDecoderT>
+// SL keeps each SIMD-level TU's instantiation a distinct symbol. Without it
+// the linker merges them and may keep one compiled with -mavx512*, which
+// then runs on the AVX2 path and raises SIGILL on CPUs without AVX512.
+// The default keeps callers outside faiss that name only the decoder (e.g.
+// unicorn/features/encoders/PQFSTable.h) compiling; faiss passes SL itself.
+template <typename PQDecoderT, SIMDLevel SL = SIMDLevel::NONE>
 struct PQCodeDistanceScalar {
     using PQDecoder = PQDecoderT;
 
@@ -160,7 +164,7 @@ struct PQCodeDistance {
         if constexpr (std::is_same_v<PQDecoderT, PQDecoder8>) {
             return pq_code_distance_single_impl<SL>(M, nbits, sim_table, code);
         } else {
-            return PQCodeDistanceScalar<PQDecoderT>::distance_single_code(
+            return PQCodeDistanceScalar<PQDecoderT, SL>::distance_single_code(
                     M, nbits, sim_table, code);
         }
     }
@@ -191,7 +195,7 @@ struct PQCodeDistance {
                     result2,
                     result3);
         } else {
-            PQCodeDistanceScalar<PQDecoderT>::distance_four_codes(
+            PQCodeDistanceScalar<PQDecoderT, SL>::distance_four_codes(
                     M,
                     nbits,
                     sim_table,