faiss 0.6.1 → 0.6.3

data/vendor/faiss/faiss/utils/quantize_lut.cpp

@@ -75,7 +75,7 @@ void round_uint8_per_column(
             max_span = span;
         }
     }
-    float a = 255 / max_span;
+    float a = max_span > 0 ? 255.0f / max_span : 0.0f;
     float b = 0;
     for (size_t i = 0; i < n; i++) {
         b += mins[i];
@@ -111,7 +111,7 @@ void round_uint8_per_column_multi(
             max_span = span;
         }
     }
-    float a = 255 / max_span;
+    float a = max_span > 0 ? 255.0f / max_span : 0.0f;
     float b = 0;
     for (size_t i = 0; i < n; i++) {
         b += mins[i];
@@ -143,7 +143,8 @@ void quantize_LUT_and_bias(
         float* b_out) {
     float a, b;
     if (!bias) {
-        FAISS_THROW_IF_NOT(!lut_is_3d);
+        FAISS_THROW_IF_MSG(
+                lut_is_3d, "3d LUT is not supported when bias is null");
         std::vector<float> mins(M);
         float max_span_LUT = -HUGE_VAL, max_span_dis = 0;
         b = 0;
@@ -154,7 +155,12 @@ void quantize_LUT_and_bias(
             max_span_dis += span;
             b += mins[i];
         }
-        a = std::min(255 / max_span_LUT, 65535 / max_span_dis);
+        a = std::min(
+                max_span_LUT > 0 ? 255.0f / max_span_LUT : HUGE_VALF,
+                max_span_dis > 0 ? 65535.0f / max_span_dis : HUGE_VALF);
+        if (!std::isfinite(a)) {
+            a = 0.0f;
+        }
 
         for (size_t i = 0; i < M; i++) {
             round_tab(LUT + i * ksub, ksub, a, mins[i], LUTq + i * ksub);
@@ -174,7 +180,12 @@ void quantize_LUT_and_bias(
             max_span_dis += span;
             b += mins[i];
         }
-        a = std::min(255 / max_span_LUT, 65535 / max_span_dis);
+        a = std::min(
+                max_span_LUT > 0 ? 255.0f / max_span_LUT : HUGE_VALF,
+                max_span_dis > 0 ? 65535.0f / max_span_dis : HUGE_VALF);
+        if (!std::isfinite(a)) {
+            a = 0.0f;
+        }
         b += bias_min;
 
         for (size_t i = 0; i < M; i++) {
@@ -208,7 +219,12 @@ void quantize_LUT_and_bias(
             b = std::min(b, b2j);
         }
 
-        a = std::min(255 / max_span_LUT, 65535 / max_span_dis);
+        a = std::min(
+                max_span_LUT > 0 ? 255.0f / max_span_LUT : HUGE_VALF,
+                max_span_dis > 0 ? 65535.0f / max_span_dis : HUGE_VALF);
+        if (!std::isfinite(a)) {
+            a = 0.0f;
+        }
 
         ij = 0;
         size_t ij_2 = 0;
@@ -256,7 +272,7 @@ void quantize_LUT_and_bias(
             max_span = std::max(max_span, span);
             b += mins[i];
         }
-        a = 255 / max_span;
+        a = max_span > 0 ? 255.0f / max_span : 0.0f;
         ij = 0;
         size_t ij_2 = 0;
         for (size_t j = 0; j < nprobe; j++) {
@@ -305,7 +321,12 @@ void aq_quantize_LUT_and_bias(
         max_span_dis += (i >= M - M_norm ? span * norm_scale : span);
         b += mins[i];
     }
-    a = std::min(255 / max_span_LUT, 65535 / max_span_dis);
+    a = std::min(
+            max_span_LUT > 0 ? 255.0f / max_span_LUT : HUGE_VALF,
+            max_span_dis > 0 ? 65535.0f / max_span_dis : HUGE_VALF);
+    if (!std::isfinite(a)) {
+        a = 0.0f;
+    }
     b += bias_min;
 
     for (size_t i = 0; i < M; i++) {

data/vendor/faiss/faiss/utils/rabitq_simd.h

@@ -7,6 +7,7 @@
 
 #pragma once
 
+#include <algorithm>
 #include <cstddef>
 #include <cstdint>
 #include <cstring>
@@ -16,6 +17,15 @@
 
 namespace faiss::rabitq {
 
+/// SIMD levels with RaBitQ query/LUT quantization implementations.
+constexpr int RABITQ_QUANTIZATION_SIMD_LEVELS = (1 << int(SIMDLevel::NONE)) |
+        (1 << int(SIMDLevel::AVX2)) | (1 << int(SIMDLevel::AVX512));
+
+struct BitwiseAndDotProductResult {
+    uint64_t dot_product;
+    uint64_t popcount;
+};
+
 /**
  * Compute dot product between query and binary data using popcount on AND.
  *
@@ -32,6 +42,17 @@ uint64_t bitwise_and_dot_product(
         size_t size,
         size_t qb);
 
+/**
+ * Compute bitwise_and_dot_product(query, data, size, qb) and popcount(data,
+ * size) in one pass over data.
+ */
+template <SIMDLevel SL = SINGLE_SIMD_LEVEL>
+BitwiseAndDotProductResult bitwise_and_dot_product_with_popcount(
+        const uint8_t* query,
+        const uint8_t* data,
+        size_t size,
+        size_t qb);
+
 /**
  * Compute dot product between query and binary data using popcount on XOR.
  *
@@ -58,6 +79,56 @@ uint64_t bitwise_xor_dot_product(
 template <SIMDLevel SL = SINGLE_SIMD_LEVEL>
 uint64_t popcount(const uint8_t* data, size_t size);
 
+/**
+ * Rearrange per-dimension quantized query codes into bit-plane layout.
+ *
+ * @p rotated_qq holds one qb-bit code per dimension (one byte each, value in
+ * [0, 2^qb)). @p out receives qb bit-planes of ((d + 7) / 8) bytes each:
+ * bit-plane j packs bit j of code i into bit (i % 8) of byte (i / 8). @p out
+ * must have room for qb * ((d + 7) / 8) bytes and is fully overwritten.
+ *
+ * @param rotated_qq  per-dimension codes (d bytes)
+ * @param d           dimensionality
+ * @param qb          quantization bits per dimension (1..8)
+ * @param out         bit-plane output buffer (qb * ((d + 7) / 8) bytes)
+ */
+template <SIMDLevel SL = SINGLE_SIMD_LEVEL>
+void rearrange_bit_planes(
+        const uint8_t* rotated_qq,
+        size_t d,
+        size_t qb,
+        uint8_t* out);
+
+/// Find min/max of one 16-entry FastScan LUT row.
+template <SIMDLevel SL = SINGLE_SIMD_LEVEL>
+void lut_minmax_16(const float* tab, float& mn, float& mx);
+
+/// Find min/max of an arbitrary-length float vector. For n == 0, mn and mx
+/// are left unchanged.
+template <SIMDLevel SL = SINGLE_SIMD_LEVEL>
+void minmax_values(const float* values, size_t n, float& mn, float& mx);
+
+/// Quantize one 16-entry FastScan LUT row with non-negative half-up rounding.
+template <SIMDLevel SL = SINGLE_SIMD_LEVEL>
+void lut_quantize_16_to_uint8(
+        const float* tab,
+        float mn,
+        float a,
+        uint8_t* out);
+
+/// Quantize rotated query values and accumulate query correction terms.
+template <SIMDLevel SL = SINGLE_SIMD_LEVEL>
+void quantize_query_values(
+        const float* rq,
+        size_t d,
+        float v_min,
+        float inv_delta,
+        uint8_t max_code,
+        bool centered,
+        uint8_t* rqq,
+        size_t& sum_qq,
+        int64_t& sum2_signed_odd_int);
+
 // NONE specializations — scalar fallbacks
 
 template <>
@@ -85,6 +156,35 @@ inline uint64_t bitwise_and_dot_product<SIMDLevel::NONE>(
     return sum;
 }
 
+template <>
+inline BitwiseAndDotProductResult bitwise_and_dot_product_with_popcount<
+        SIMDLevel::NONE>(
+        const uint8_t* query,
+        const uint8_t* data,
+        size_t size,
+        size_t qb) {
+    uint64_t dot_product = 0;
+    uint64_t popcount_sum = 0;
+    size_t offset = 0;
+    for (size_t step = 64 / 8; offset + step <= size; offset += step) {
+        const auto yv = *(const uint64_t*)(data + offset);
+        popcount_sum += popcount64(yv);
+        for (int j = 0; j < qb; j++) {
+            const auto qv = *(const uint64_t*)(query + j * size + offset);
+            dot_product += popcount64(qv & yv) << j;
+        }
+    }
+    for (; offset < size; ++offset) {
+        const auto yv = *(data + offset);
+        popcount_sum += popcount32(yv);
+        for (int j = 0; j < qb; j++) {
+            const auto qv = *(query + j * size + offset);
+            dot_product += popcount32(qv & yv) << j;
+        }
+    }
+    return {dot_product, popcount_sum};
+}
+
 template <>
 inline uint64_t bitwise_xor_dot_product<SIMDLevel::NONE>(
         const uint8_t* query,
@@ -125,6 +225,108 @@ inline uint64_t popcount<SIMDLevel::NONE>(const uint8_t* data, size_t size) {
     return sum;
 }
 
+template <>
+inline void rearrange_bit_planes<SIMDLevel::NONE>(
+        const uint8_t* rotated_qq,
+        size_t d,
+        size_t qb,
+        uint8_t* out) {
+    const size_t offset = (d + 7) / 8;
+    memset(out, 0, offset * qb);
+    for (size_t idim = 0; idim < d; idim++) {
+        for (size_t iv = 0; iv < qb; iv++) {
+            const bool bit = ((rotated_qq[idim] & (1 << iv)) != 0);
+            out[iv * offset + idim / 8] |= bit ? (1 << (idim % 8)) : 0;
+        }
+    }
+}
+
+inline uint8_t round_clamped_byte_scalar(float x, uint8_t max_code) {
+    if (x <= 0.0f) {
+        return 0;
+    }
+    if (x >= max_code) {
+        return max_code;
+    }
+    return static_cast<uint8_t>(static_cast<int>(x + 0.5f));
+}
+
+inline uint8_t round_nonnegative_byte_scalar(float x) {
+    return round_clamped_byte_scalar(x, 255);
+}
+
+template <>
+inline void lut_minmax_16<SIMDLevel::NONE>(
+        const float* tab,
+        float& mn,
+        float& mx) {
+    mn = tab[0];
+    mx = tab[0];
+    for (size_t s = 1; s < 16; s++) {
+        mn = std::min(mn, tab[s]);
+        mx = std::max(mx, tab[s]);
+    }
+}
+
+template <>
+inline void minmax_values<SIMDLevel::NONE>(
+        const float* values,
+        size_t n,
+        float& mn,
+        float& mx) {
+    if (n == 0) {
+        return;
+    }
+    mn = values[0];
+    mx = values[0];
+    for (size_t i = 1; i < n; i++) {
+        mn = std::min(mn, values[i]);
+        mx = std::max(mx, values[i]);
+    }
+}
+
+template <>
+inline void lut_quantize_16_to_uint8<SIMDLevel::NONE>(
+        const float* tab,
+        float mn,
+        float a,
+        uint8_t* out) {
+    for (size_t s = 0; s < 16; s++) {
+        out[s] = round_nonnegative_byte_scalar(a * (tab[s] - mn));
+    }
+}
+
+template <>
+inline void quantize_query_values<SIMDLevel::NONE>(
+        const float* rq,
+        size_t d,
+        float v_min,
+        float inv_delta,
+        uint8_t max_code,
+        bool centered,
+        uint8_t* rqq,
+        size_t& sum_qq,
+        int64_t& sum2_signed_odd_int) {
+    if (centered) {
+        for (size_t i = 0; i < d; i++) {
+            const uint8_t v_qq = round_clamped_byte_scalar(
+                    (rq[i] - v_min) * inv_delta, max_code);
+            rqq[i] = v_qq;
+            sum_qq += v_qq;
+
+            const int64_t signed_odd_int = int64_t(v_qq) * 2 - max_code;
+            sum2_signed_odd_int += signed_odd_int * signed_odd_int;
+        }
+    } else {
+        for (size_t i = 0; i < d; i++) {
+            const uint8_t v_qq = round_clamped_byte_scalar(
+                    (rq[i] - v_min) * inv_delta, max_code);
+            rqq[i] = v_qq;
+            sum_qq += v_qq;
+        }
+    }
+}
+
 } // namespace faiss::rabitq
 
 /*********************************************************

data/vendor/faiss/faiss/utils/simd_impl/distances_avx2.cpp

@@ -11,7 +11,6 @@
 
 #include <faiss/impl/AuxIndexStructures.h>
 #include <faiss/impl/ResultHandler.h>
-#include <faiss/utils/distances_fused/distances_fused.h>
 #include <faiss/utils/simd_impl/exhaustive_L2sqr_blas_cmax.h>
 
 #ifndef FINTEGER

data/vendor/faiss/faiss/utils/simd_impl/distances_avx512.cpp

@@ -9,6 +9,35 @@
 
 #include <immintrin.h>
 
+#include <memory>
+
+#include <faiss/impl/AuxIndexStructures.h>
+#include <faiss/impl/ResultHandler.h>
+#include <faiss/utils/distances_fused/distances_fused.h>
+#include <faiss/utils/simd_impl/exhaustive_L2sqr_blas_cmax.h>
+
+#ifndef FINTEGER
+#define FINTEGER long
+#endif
+
+extern "C" {
+
+int sgemm_(
+        const char* transa,
+        const char* transb,
+        FINTEGER* m,
+        FINTEGER* n,
+        FINTEGER* k,
+        const float* alpha,
+        const float* a,
+        FINTEGER* lda,
+        const float* b,
+        FINTEGER* ldb,
+        float* beta,
+        float* c,
+        FINTEGER* ldc);
+}
+
 #define THE_SIMD_LEVEL SIMDLevel::AVX512
 #include <faiss/utils/simd_impl/distances_autovec-inl.h>
 // NOLINTNEXTLINE(facebook-hte-InlineHeader)
@@ -954,21 +983,6 @@ size_t fvec_L2sqr_ny_nearest<SIMDLevel::AVX512>(
             &fvec_L2sqr_ny_nearest_D8<SIMDLevel::AVX512>);
 }
 
-template <>
-size_t fvec_L2sqr_ny_nearest_y_transposed<SIMDLevel::AVX512>(
-        float* distances_tmp_buffer,
-        const float* x,
-        const float* y,
-        const float* y_sqlen,
-        size_t d,
-        size_t d_offset,
-        size_t ny) {
-    return fvec_L2sqr_ny_nearest_y_transposed<SIMDLevel::NONE>(
-            distances_tmp_buffer, x, y, y_sqlen, d, d_offset, ny);
-}
-
-// TODO: Following functions are not used in the current codebase. Check AVX2 ,
-// respective implementation has been used
 template <size_t DIM>
 size_t fvec_L2sqr_ny_nearest_y_transposed_D(
         float* /* distances_tmp_buffer */,
@@ -1082,6 +1096,33 @@ size_t fvec_L2sqr_ny_nearest_y_transposed_D(
     return current_min_index;
 }
 
+template <>
+size_t fvec_L2sqr_ny_nearest_y_transposed<SIMDLevel::AVX512>(
+        float* distances_tmp_buffer,
+        const float* x,
+        const float* y,
+        const float* y_sqlen,
+        size_t d,
+        size_t d_offset,
+        size_t ny) {
+    // optimized for a few special cases
+#define DISPATCH(dval)                                     \
+    case dval:                                             \
+        return fvec_L2sqr_ny_nearest_y_transposed_D<dval>( \
+                distances_tmp_buffer, x, y, y_sqlen, d_offset, ny);
+
+    switch (d) {
+        DISPATCH(1)
+        DISPATCH(2)
+        DISPATCH(4)
+        DISPATCH(8)
+        default:
+            return fvec_L2sqr_ny_nearest_y_transposed<SIMDLevel::NONE>(
+                    distances_tmp_buffer, x, y, y_sqlen, d, d_offset, ny);
+    }
+#undef DISPATCH
+}
+
 template <>
 int fvec_madd_and_argmin<SIMDLevel::AVX512>(
         size_t n,
@@ -1092,4 +1133,211 @@ int fvec_madd_and_argmin<SIMDLevel::AVX512>(
     return fvec_madd_and_argmin_sse(n, a, bf, b, c);
 }
 
+template <>
+void exhaustive_L2sqr_blas_cmax<SIMDLevel::AVX512>(
+        const float* x,
+        const float* y,
+        size_t d,
+        size_t nx,
+        size_t ny,
+        Top1BlockResultHandler<CMax<float, int64_t>>& res,
+        const float* y_norms) {
+    // BLAS does not like empty matrices
+    if (nx == 0 || ny == 0) {
+        return;
+    }
+
+    /* block sizes */
+    const size_t bs_x = distance_compute_blas_query_bs;
+    const size_t bs_y = distance_compute_blas_database_bs;
+    std::unique_ptr<float[]> ip_block(new float[bs_x * bs_y]);
+    std::unique_ptr<float[]> x_norms(new float[nx]);
+    std::unique_ptr<float[]> del2;
+
+    fvec_norms_L2sqr(x_norms.get(), x, d, nx);
+
+    if (!y_norms) {
+        float* y_norms2 = new float[ny];
+        del2.reset(y_norms2);
+        fvec_norms_L2sqr(y_norms2, y, d, ny);
+        y_norms = y_norms2;
+    }
+
+    for (size_t i0 = 0; i0 < nx; i0 += bs_x) {
+        size_t i1 = i0 + bs_x;
+        if (i1 > nx) {
+            i1 = nx;
+        }
+
+        res.begin_multiple(i0, i1);
+
+        for (size_t j0 = 0; j0 < ny; j0 += bs_y) {
+            size_t j1 = j0 + bs_y;
+            if (j1 > ny) {
+                j1 = ny;
+            }
+            /* compute the actual dot products */
+            {
+                float one = 1, zero = 0;
+                FINTEGER nyi = j1 - j0, nxi = i1 - i0, di = d;
+                sgemm_("Transpose",
+                       "Not transpose",
+                       &nyi,
+                       &nxi,
+                       &di,
+                       &one,
+                       y + j0 * d,
+                       &di,
+                       x + i0 * d,
+                       &di,
+                       &zero,
+                       ip_block.get(),
+                       &nyi);
+            }
+#pragma omp parallel for schedule(static) if ((i1 - i0) >= 16)
+            for (int64_t i = static_cast<int64_t>(i0);
+                 i < static_cast<int64_t>(i1);
+                 i++) {
+                float* ip_line = ip_block.get() + (i - i0) * (j1 - j0);
+
+                _mm_prefetch((const char*)ip_line, _MM_HINT_NTA);
+                _mm_prefetch((const char*)(ip_line + 16), _MM_HINT_NTA);
+
+                // constant
+                const __m512 mul_minus2 = _mm512_set1_ps(-2);
+
+                // Track 16 min distances + 16 min indices.
+                // All the distances tracked do not take x_norms[i]
+                // into account in order to get rid of extra
+                // _mm512_add_ps(x_norms[i], ...) instructions
+                // in distance computations.
+                __m512 min_distances =
+                        _mm512_set1_ps(res.dis_tab[i] - x_norms[i]);
+
+                // these indices are local and are relative to j0.
+                // so, value 0 means j0.
+                __m512i min_indices = _mm512_set1_epi32(0);
+
+                __m512i current_indices = _mm512_setr_epi32(
+                        0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+                const __m512i indices_delta = _mm512_set1_epi32(16);
+
+                // current j index
+                size_t idx_j = 0;
+                size_t count = j1 - j0;
+
+                // process 32 elements per loop
+                for (; idx_j < (count / 32) * 32; idx_j += 32, ip_line += 32) {
+                    _mm_prefetch((const char*)(ip_line + 32), _MM_HINT_NTA);
+                    _mm_prefetch((const char*)(ip_line + 48), _MM_HINT_NTA);
+
+                    // load values for norms
+                    const __m512 y_norm_0 =
+                            _mm512_loadu_ps(y_norms + idx_j + j0 + 0);
+                    const __m512 y_norm_1 =
+                            _mm512_loadu_ps(y_norms + idx_j + j0 + 16);
+
+                    // load values for dot products
+                    const __m512 ip_0 = _mm512_loadu_ps(ip_line + 0);
+                    const __m512 ip_1 = _mm512_loadu_ps(ip_line + 16);
+
+                    // compute dis = y_norm[j] - 2 * dot(x_norm[i], y_norm[j]).
+                    // x_norm[i] was dropped off because it is a constant for a
+                    // given i.
+                    __m512 distances_0 =
+                            _mm512_fmadd_ps(ip_0, mul_minus2, y_norm_0);
+                    __m512 distances_1 =
+                            _mm512_fmadd_ps(ip_1, mul_minus2, y_norm_1);
+
+                    // compare the new distances to the min distances
+                    // for each of the first group of 16 AVX512 components.
+                    const __mmask16 comparison_0 = _mm512_cmp_ps_mask(
+                            min_distances, distances_0, _CMP_LE_OS);
+
+                    // update min distances and indices with closest vectors if
+                    // needed.
+                    min_distances = _mm512_mask_blend_ps(
+                            comparison_0, distances_0, min_distances);
+                    min_indices = _mm512_mask_blend_epi32(
+                            comparison_0, current_indices, min_indices);
+                    current_indices =
+                            _mm512_add_epi32(current_indices, indices_delta);
+
+                    // compare the new distances to the min distances
+                    // for each of the second group of 16 AVX512 components.
+                    const __mmask16 comparison_1 = _mm512_cmp_ps_mask(
+                            min_distances, distances_1, _CMP_LE_OS);
+
+                    // update min distances and indices with closest vectors if
+                    // needed.
+                    min_distances = _mm512_mask_blend_ps(
+                            comparison_1, distances_1, min_distances);
+                    min_indices = _mm512_mask_blend_epi32(
+                            comparison_1, current_indices, min_indices);
+                    current_indices =
+                            _mm512_add_epi32(current_indices, indices_delta);
+                }
+
+                // dump values and find the minimum distance / minimum index
+                float min_distances_scalar[16];
+                uint32_t min_indices_scalar[16];
+                _mm512_storeu_ps(min_distances_scalar, min_distances);
+                _mm512_storeu_si512(
+                        (__m512i*)(min_indices_scalar), min_indices);
+
+                float current_min_distance = res.dis_tab[i];
+                uint32_t current_min_index = res.ids_tab[i];
+
+                // if two indices are represented with equal distance values,
+                // then the index with the min value is returned.
+                for (size_t jv = 0; jv < 16; jv++) {
+                    // add missing x_norms[i]
+                    float distance_candidate =
+                            min_distances_scalar[jv] + x_norms[i];
+
+                    // negative values can occur for identical vectors
+                    //    due to roundoff errors.
+                    if (distance_candidate < 0) {
+                        distance_candidate = 0;
+                    }
+
+                    int64_t index_candidate = min_indices_scalar[jv] + j0;
+
+                    if (current_min_distance > distance_candidate) {
+                        current_min_distance = distance_candidate;
+                        current_min_index = index_candidate;
+                    } else if (
+                            current_min_distance == distance_candidate &&
+                            current_min_index > index_candidate) {
+                        current_min_index = index_candidate;
+                    }
+                }
+
+                // process leftovers
+                for (; idx_j < count; idx_j++, ip_line++) {
+                    float ip = *ip_line;
+                    float dis = x_norms[i] + y_norms[idx_j + j0] - 2 * ip;
+                    // negative values can occur for identical vectors
+                    //    due to roundoff errors.
+                    if (dis < 0) {
+                        dis = 0;
+                    }
+
+                    if (current_min_distance > dis) {
+                        current_min_distance = dis;
+                        current_min_index = idx_j + j0;
+                    }
+                }
+
+                //
+                res.add_result(i, current_min_distance, current_min_index);
+            }
+        }
+        // Does nothing for SingleBestResultHandler, but
+        // keeping the call for the consistency.
+        res.end_multiple();
+        InterruptCallback::check();
+    }
+}
+
 } // namespace faiss