faiss 0.1.7 → 0.2.3

data/vendor/faiss/faiss/gpu/utils/Timer.cpp

@@ -5,55 +5,51 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-
-#include <faiss/gpu/utils/Timer.h>
 #include <faiss/gpu/utils/DeviceUtils.h>
+#include <faiss/gpu/utils/Timer.h>
 #include <faiss/impl/FaissAssert.h>
 #include <chrono>
 
-namespace faiss { namespace gpu {
+namespace faiss {
+namespace gpu {
 
 KernelTimer::KernelTimer(cudaStream_t stream)
-    : startEvent_(0),
-      stopEvent_(0),
-      stream_(stream),
-      valid_(true) {
-  CUDA_VERIFY(cudaEventCreate(&startEvent_));
-  CUDA_VERIFY(cudaEventCreate(&stopEvent_));
-
-  CUDA_VERIFY(cudaEventRecord(startEvent_, stream_));
+        : startEvent_(0), stopEvent_(0), stream_(stream), valid_(true) {
+    CUDA_VERIFY(cudaEventCreate(&startEvent_));
+    CUDA_VERIFY(cudaEventCreate(&stopEvent_));
+
+    CUDA_VERIFY(cudaEventRecord(startEvent_, stream_));
 }
 
 KernelTimer::~KernelTimer() {
-  CUDA_VERIFY(cudaEventDestroy(startEvent_));
-  CUDA_VERIFY(cudaEventDestroy(stopEvent_));
+    CUDA_VERIFY(cudaEventDestroy(startEvent_));
+    CUDA_VERIFY(cudaEventDestroy(stopEvent_));
 }
 
-float
-KernelTimer::elapsedMilliseconds() {
-  FAISS_ASSERT(valid_);
+float KernelTimer::elapsedMilliseconds() {
+    FAISS_ASSERT(valid_);
 
-  CUDA_VERIFY(cudaEventRecord(stopEvent_, stream_));
-  CUDA_VERIFY(cudaEventSynchronize(stopEvent_));
+    CUDA_VERIFY(cudaEventRecord(stopEvent_, stream_));
+    CUDA_VERIFY(cudaEventSynchronize(stopEvent_));
 
-  auto time = 0.0f;
-  CUDA_VERIFY(cudaEventElapsedTime(&time, startEvent_, stopEvent_));
-  valid_ = false;
+    auto time = 0.0f;
+    CUDA_VERIFY(cudaEventElapsedTime(&time, startEvent_, stopEvent_));
+    valid_ = false;
 
-  return time;
+    return time;
 }
 
 CpuTimer::CpuTimer() {
-  start_ = std::chrono::steady_clock::now();
+    start_ = std::chrono::steady_clock::now();
 }
 
-float
-CpuTimer::elapsedMilliseconds() {
-  auto end = std::chrono::steady_clock::now();
+float CpuTimer::elapsedMilliseconds() {
+    auto end = std::chrono::steady_clock::now();
 
-  std::chrono::duration<float, std::milli> duration = end - start_;
+    std::chrono::duration<float, std::milli> duration = end - start_;
 
-  return duration.count();
+    return duration.count();
 }
 
-} } // namespace
+} // namespace gpu
+} // namespace faiss

data/vendor/faiss/faiss/gpu/utils/Timer.h

@@ -5,48 +5,49 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-
 #pragma once
 
 #include <cuda_runtime.h>
 #include <chrono>
 
-namespace faiss { namespace gpu {
+namespace faiss {
+namespace gpu {
 
 /// Utility class for timing execution of a kernel
 class KernelTimer {
- public:
-  /// Constructor starts the timer and adds an event into the current
-  /// device stream
-  KernelTimer(cudaStream_t stream = 0);
-
-  /// Destructor releases event resources
-  ~KernelTimer();
-
-  /// Adds a stop event then synchronizes on the stop event to get the
-  /// actual GPU-side kernel timings for any kernels launched in the
-  /// current stream. Returns the number of milliseconds elapsed.
-  /// Can only be called once.
-  float elapsedMilliseconds();
-
- private:
-  cudaEvent_t startEvent_;
-  cudaEvent_t stopEvent_;
-  cudaStream_t stream_;
-  bool valid_;
+   public:
+    /// Constructor starts the timer and adds an event into the current
+    /// device stream
+    KernelTimer(cudaStream_t stream = 0);
+
+    /// Destructor releases event resources
+    ~KernelTimer();
+
+    /// Adds a stop event then synchronizes on the stop event to get the
+    /// actual GPU-side kernel timings for any kernels launched in the
+    /// current stream. Returns the number of milliseconds elapsed.
+    /// Can only be called once.
+    float elapsedMilliseconds();
+
+   private:
+    cudaEvent_t startEvent_;
+    cudaEvent_t stopEvent_;
+    cudaStream_t stream_;
+    bool valid_;
 };
 
 /// CPU wallclock elapsed timer
 class CpuTimer {
- public:
-  /// Creates and starts a new timer
-  CpuTimer();
+   public:
+    /// Creates and starts a new timer
+    CpuTimer();
 
-  /// Returns elapsed time in milliseconds
-  float elapsedMilliseconds();
+    /// Returns elapsed time in milliseconds
+    float elapsedMilliseconds();
 
- private:
-  std::chrono::time_point<std::chrono::steady_clock> start_;
+   private:
+    std::chrono::time_point<std::chrono::steady_clock> start_;
 };
 
-} } // namespace
+} // namespace gpu
+} // namespace faiss

data/vendor/faiss/faiss/impl/AdditiveQuantizer.cpp

@@ -0,0 +1,270 @@
+/**
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+// -*- c++ -*-
+
+#include <faiss/impl/AdditiveQuantizer.h>
+#include <faiss/impl/FaissAssert.h>
+
+#include <cstddef>
+#include <cstdio>
+#include <cstring>
+#include <memory>
+#include <random>
+
+#include <algorithm>
+
+#include <faiss/utils/Heap.h>
+#include <faiss/utils/distances.h>
+#include <faiss/utils/hamming.h> // BitstringWriter
+#include <faiss/utils/utils.h>
+
+extern "C" {
+
+// general matrix multiplication
+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);
+}
+
+namespace {
+
+// c and a and b can overlap
+void fvec_add(size_t d, const float* a, const float* b, float* c) {
+    for (size_t i = 0; i < d; i++) {
+        c[i] = a[i] + b[i];
+    }
+}
+
+void fvec_add(size_t d, const float* a, float b, float* c) {
+    for (size_t i = 0; i < d; i++) {
+        c[i] = a[i] + b;
+    }
+}
+
+} // namespace
+
+namespace faiss {
+
+void AdditiveQuantizer::set_derived_values() {
+    tot_bits = 0;
+    is_byte_aligned = true;
+    codebook_offsets.resize(M + 1, 0);
+    for (int i = 0; i < M; i++) {
+        int nbit = nbits[i];
+        size_t k = 1 << nbit;
+        codebook_offsets[i + 1] = codebook_offsets[i] + k;
+        tot_bits += nbit;
+        if (nbit % 8 != 0) {
+            is_byte_aligned = false;
+        }
+    }
+    total_codebook_size = codebook_offsets[M];
+    // convert bits to bytes
+    code_size = (tot_bits + 7) / 8;
+}
+
+void AdditiveQuantizer::pack_codes(
+        size_t n,
+        const int32_t* codes,
+        uint8_t* packed_codes,
+        int64_t ld_codes) const {
+    if (ld_codes == -1) {
+        ld_codes = M;
+    }
+#pragma omp parallel for if (n > 1000)
+    for (int64_t i = 0; i < n; i++) {
+        const int32_t* codes1 = codes + i * ld_codes;
+        BitstringWriter bsw(packed_codes + i * code_size, code_size);
+        for (int m = 0; m < M; m++) {
+            bsw.write(codes1[m], nbits[m]);
+        }
+    }
+}
+
+void AdditiveQuantizer::decode(const uint8_t* code, float* x, size_t n) const {
+    FAISS_THROW_IF_NOT_MSG(
+            is_trained, "The additive quantizer is not trained yet.");
+
+    // standard additive quantizer decoding
+#pragma omp parallel for if (n > 1000)
+    for (int64_t i = 0; i < n; i++) {
+        BitstringReader bsr(code + i * code_size, code_size);
+        float* xi = x + i * d;
+        for (int m = 0; m < M; m++) {
+            int idx = bsr.read(nbits[m]);
+            const float* c = codebooks.data() + d * (codebook_offsets[m] + idx);
+            if (m == 0) {
+                memcpy(xi, c, sizeof(*x) * d);
+            } else {
+                fvec_add(d, xi, c, xi);
+            }
+        }
+    }
+}
+
+AdditiveQuantizer::~AdditiveQuantizer() {}
+
+/****************************************************************************
+ * Support for fast distance computations and search with additive quantizer
+ ****************************************************************************/
+
+void AdditiveQuantizer::compute_centroid_norms(float* norms) const {
+    size_t ntotal = (size_t)1 << tot_bits;
+    // TODO: make tree of partial sums
+#pragma omp parallel
+    {
+        std::vector<float> tmp(d);
+#pragma omp for
+        for (int64_t i = 0; i < ntotal; i++) {
+            decode_64bit(i, tmp.data());
+            norms[i] = fvec_norm_L2sqr(tmp.data(), d);
+        }
+    }
+}
+
+void AdditiveQuantizer::decode_64bit(idx_t bits, float* xi) const {
+    for (int m = 0; m < M; m++) {
+        idx_t idx = bits & (((size_t)1 << nbits[m]) - 1);
+        bits >>= nbits[m];
+        const float* c = codebooks.data() + d * (codebook_offsets[m] + idx);
+        if (m == 0) {
+            memcpy(xi, c, sizeof(*xi) * d);
+        } else {
+            fvec_add(d, xi, c, xi);
+        }
+    }
+}
+
+void AdditiveQuantizer::compute_LUT(size_t n, const float* xq, float* LUT)
+        const {
+    // in all cases, it is large matrix multiplication
+
+    FINTEGER ncenti = total_codebook_size;
+    FINTEGER di = d;
+    FINTEGER nqi = n;
+    float one = 1, zero = 0;
+
+    sgemm_("Transposed",
+           "Not transposed",
+           &ncenti,
+           &nqi,
+           &di,
+           &one,
+           codebooks.data(),
+           &di,
+           xq,
+           &di,
+           &zero,
+           LUT,
+           &ncenti);
+}
+
+namespace {
+
+void compute_inner_prod_with_LUT(
+        const AdditiveQuantizer& aq,
+        const float* LUT,
+        float* ips) {
+    size_t prev_size = 1;
+    for (int m = 0; m < aq.M; m++) {
+        const float* LUTm = LUT + aq.codebook_offsets[m];
+        int nb = aq.nbits[m];
+        size_t nc = (size_t)1 << nb;
+
+        if (m == 0) {
+            memcpy(ips, LUT, sizeof(*ips) * nc);
+        } else {
+            for (int64_t i = nc - 1; i >= 0; i--) {
+                float v = LUTm[i];
+                fvec_add(prev_size, ips, v, ips + i * prev_size);
+            }
+        }
+        prev_size *= nc;
+    }
+}
+
+} // anonymous namespace
+
+void AdditiveQuantizer::knn_exact_inner_product(
+        idx_t n,
+        const float* xq,
+        idx_t k,
+        float* distances,
+        idx_t* labels) const {
+    std::unique_ptr<float[]> LUT(new float[n * total_codebook_size]);
+    compute_LUT(n, xq, LUT.get());
+    size_t ntotal = (size_t)1 << tot_bits;
+
+#pragma omp parallel if (n > 100)
+    {
+        std::vector<float> dis(ntotal);
+#pragma omp for
+        for (idx_t i = 0; i < n; i++) {
+            const float* LUTi = LUT.get() + i * total_codebook_size;
+            compute_inner_prod_with_LUT(*this, LUTi, dis.data());
+            float* distances_i = distances + i * k;
+            idx_t* labels_i = labels + i * k;
+            minheap_heapify(k, distances_i, labels_i);
+            minheap_addn(k, distances_i, labels_i, dis.data(), nullptr, ntotal);
+            minheap_reorder(k, distances_i, labels_i);
+        }
+    }
+}
+
+void AdditiveQuantizer::knn_exact_L2(
+        idx_t n,
+        const float* xq,
+        idx_t k,
+        float* distances,
+        idx_t* labels,
+        const float* norms) const {
+    std::unique_ptr<float[]> LUT(new float[n * total_codebook_size]);
+    compute_LUT(n, xq, LUT.get());
+    std::unique_ptr<float[]> q_norms(new float[n]);
+    fvec_norms_L2sqr(q_norms.get(), xq, d, n);
+    size_t ntotal = (size_t)1 << tot_bits;
+
+#pragma omp parallel if (n > 100)
+    {
+        std::vector<float> dis(ntotal);
+#pragma omp for
+        for (idx_t i = 0; i < n; i++) {
+            const float* LUTi = LUT.get() + i * total_codebook_size;
+            float* distances_i = distances + i * k;
+            idx_t* labels_i = labels + i * k;
+
+            compute_inner_prod_with_LUT(*this, LUTi, dis.data());
+
+            // update distances using
+            // ||x - y||^2 = ||x||^2 + ||y||^2 - 2 * <x,y>
+
+            maxheap_heapify(k, distances_i, labels_i);
+            for (idx_t j = 0; j < ntotal; j++) {
+                float disj = q_norms[i] + norms[j] - 2 * dis[j];
+                if (disj < distances_i[0]) {
+                    heap_replace_top<CMax<float, int64_t>>(
+                            k, distances_i, labels_i, disj, j);
+                }
+            }
+            maxheap_reorder(k, distances_i, labels_i);
+        }
+    }
+}
+
+} // namespace faiss

data/vendor/faiss/faiss/impl/AdditiveQuantizer.h

@@ -0,0 +1,115 @@
+/**
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <cstdint>
+#include <vector>
+
+#include <faiss/Index.h>
+
+namespace faiss {
+
+/** Abstract structure for additive quantizers
+ *
+ * Different from the product quantizer in which the decoded vector is the
+ * concatenation of M sub-vectors, additive quantizers sum M sub-vectors
+ * to get the decoded vector.
+ */
+struct AdditiveQuantizer {
+    size_t d;                     ///< size of the input vectors
+    size_t M;                     ///< number of codebooks
+    std::vector<size_t> nbits;    ///< bits for each step
+    std::vector<float> codebooks; ///< codebooks
+
+    // derived values
+    std::vector<size_t> codebook_offsets;
+    size_t code_size;           ///< code size in bytes
+    size_t tot_bits;            ///< total number of bits
+    size_t total_codebook_size; ///< size of the codebook in vectors
+    bool is_byte_aligned;
+
+    bool verbose;    ///< verbose during training?
+    bool is_trained; ///< is trained or not
+
+    ///< compute derived values when d, M and nbits have been set
+    void set_derived_values();
+
+    ///< Train the additive quantizer
+    virtual void train(size_t n, const float* x) = 0;
+
+    /** Encode a set of vectors
+     *
+     * @param x      vectors to encode, size n * d
+     * @param codes  output codes, size n * code_size
+     */
+    virtual void compute_codes(const float* x, uint8_t* codes, size_t n)
+            const = 0;
+
+    /** pack a series of code to bit-compact format
+     *
+     * @param codes  codes to be packed, size n * code_size
+     * @param packed_codes output bit-compact codes
+     * @param ld_codes  leading dimension of codes
+     */
+    void pack_codes(
+            size_t n,
+            const int32_t* codes,
+            uint8_t* packed_codes,
+            int64_t ld_codes = -1) const;
+
+    /** Decode a set of vectors
+     *
+     * @param codes  codes to decode, size n * code_size
+     * @param x      output vectors, size n * d
+     */
+    void decode(const uint8_t* codes, float* x, size_t n) const;
+
+    /****************************************************************************
+     * Support for exhaustive distance computations with the centroids.
+     * Hence, the number of elements that can be enumerated is not too large.
+     ****************************************************************************/
+    using idx_t = Index::idx_t;
+
+    /// decoding function for a code in a 64-bit word
+    void decode_64bit(idx_t n, float* x) const;
+
+    /** Compute inner-product look-up tables. Used in the centroid search
+     * functions.
+     *
+     * @param xq     query vector, size (n, d)
+     * @param LUT    look-up table, size (n, total_codebook_size)
+     */
+    void compute_LUT(size_t n, const float* xq, float* LUT) const;
+
+    /// exact IP search
+    void knn_exact_inner_product(
+            idx_t n,
+            const float* xq,
+            idx_t k,
+            float* distances,
+            idx_t* labels) const;
+
+    /** For L2 search we need the L2 norms of the centroids
+     *
+     * @param norms    output norms table, size total_codebook_size
+     */
+    void compute_centroid_norms(float* norms) const;
+
+    /** Exact L2 search, with precomputed norms */
+    void knn_exact_L2(
+            idx_t n,
+            const float* xq,
+            idx_t k,
+            float* distances,
+            idx_t* labels,
+            const float* centroid_norms) const;
+
+    virtual ~AdditiveQuantizer();
+};
+
+}; // namespace faiss