faiss 0.2.0 → 0.2.4

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 +1,503 @@
+/**
+ * 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 <cstddef>
+#include <cstdio>
+#include <cstring>
+#include <memory>
+#include <random>
+
+#include <algorithm>
+
+#include <faiss/impl/FaissAssert.h>
+#include <faiss/utils/Heap.h>
+#include <faiss/utils/distances.h>
+#include <faiss/utils/hamming.h>
+#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 faiss {
+
+AdditiveQuantizer::AdditiveQuantizer(
+        size_t d,
+        const std::vector<size_t>& nbits,
+        Search_type_t search_type)
+        : d(d),
+          M(nbits.size()),
+          nbits(nbits),
+          verbose(false),
+          is_trained(false),
+          search_type(search_type) {
+    norm_max = norm_min = NAN;
+    code_size = 0;
+    tot_bits = 0;
+    total_codebook_size = 0;
+    only_8bit = false;
+    set_derived_values();
+}
+
+AdditiveQuantizer::AdditiveQuantizer()
+        : AdditiveQuantizer(0, std::vector<size_t>()) {}
+
+void AdditiveQuantizer::set_derived_values() {
+    tot_bits = 0;
+    only_8bit = 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 != 0) {
+            only_8bit = false;
+        }
+    }
+    total_codebook_size = codebook_offsets[M];
+    switch (search_type) {
+        case ST_decompress:
+        case ST_LUT_nonorm:
+        case ST_norm_from_LUT:
+            break; // nothing to add
+        case ST_norm_float:
+            tot_bits += 32;
+            break;
+        case ST_norm_qint8:
+        case ST_norm_cqint8:
+            tot_bits += 8;
+            break;
+        case ST_norm_qint4:
+        case ST_norm_cqint4:
+            tot_bits += 4;
+            break;
+    }
+
+    // convert bits to bytes
+    code_size = (tot_bits + 7) / 8;
+}
+
+namespace {
+
+// TODO
+// https://stackoverflow.com/questions/31631224/hacks-for-clamping-integer-to-0-255-and-doubles-to-0-0-1-0
+
+uint8_t encode_qint8(float x, float amin, float amax) {
+    float x1 = (x - amin) / (amax - amin) * 256;
+    int32_t xi = int32_t(floor(x1));
+
+    return xi < 0 ? 0 : xi > 255 ? 255 : xi;
+}
+
+uint8_t encode_qint4(float x, float amin, float amax) {
+    float x1 = (x - amin) / (amax - amin) * 16;
+    int32_t xi = int32_t(floor(x1));
+
+    return xi < 0 ? 0 : xi > 15 ? 15 : xi;
+}
+
+float decode_qint8(uint8_t i, float amin, float amax) {
+    return (i + 0.5) / 256 * (amax - amin) + amin;
+}
+
+float decode_qint4(uint8_t i, float amin, float amax) {
+    return (i + 0.5) / 16 * (amax - amin) + amin;
+}
+
+} // anonymous namespace
+
+uint32_t AdditiveQuantizer::encode_qcint(float x) const {
+    idx_t id;
+    qnorm.assign(idx_t(1), &x, &id, idx_t(1));
+    return uint32_t(id);
+}
+
+float AdditiveQuantizer::decode_qcint(uint32_t c) const {
+    return qnorm.get_xb()[c];
+}
+
+void AdditiveQuantizer::pack_codes(
+        size_t n,
+        const int32_t* codes,
+        uint8_t* packed_codes,
+        int64_t ld_codes,
+        const float* norms) const {
+    if (ld_codes == -1) {
+        ld_codes = M;
+    }
+    std::vector<float> norm_buf;
+    if (search_type == ST_norm_float || search_type == ST_norm_qint4 ||
+        search_type == ST_norm_qint8 || search_type == ST_norm_cqint8 ||
+        search_type == ST_norm_cqint4) {
+        if (!norms) {
+            norm_buf.resize(n);
+            std::vector<float> x_recons(n * d);
+            decode_unpacked(codes, x_recons.data(), n, ld_codes);
+            fvec_norms_L2sqr(norm_buf.data(), x_recons.data(), d, n);
+            norms = norm_buf.data();
+        }
+    }
+#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]);
+        }
+        switch (search_type) {
+            case ST_decompress:
+            case ST_LUT_nonorm:
+            case ST_norm_from_LUT:
+                break;
+            case ST_norm_float:
+                bsw.write(*(uint32_t*)&norms[i], 32);
+                break;
+            case ST_norm_qint8: {
+                uint8_t b = encode_qint8(norms[i], norm_min, norm_max);
+                bsw.write(b, 8);
+                break;
+            }
+            case ST_norm_qint4: {
+                uint8_t b = encode_qint4(norms[i], norm_min, norm_max);
+                bsw.write(b, 4);
+                break;
+            }
+            case ST_norm_cqint8: {
+                uint32_t b = encode_qcint(norms[i]);
+                bsw.write(b, 8);
+                break;
+            }
+            case ST_norm_cqint4: {
+                uint32_t b = encode_qcint(norms[i]);
+                bsw.write(b, 4);
+                break;
+            }
+        }
+    }
+}
+
+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);
+            }
+        }
+    }
+}
+
+void AdditiveQuantizer::decode_unpacked(
+        const int32_t* code,
+        float* x,
+        size_t n,
+        int64_t ld_codes) const {
+    FAISS_THROW_IF_NOT_MSG(
+            is_trained, "The additive quantizer is not trained yet.");
+
+    if (ld_codes == -1) {
+        ld_codes = M;
+    }
+
+    // standard additive quantizer decoding
+#pragma omp parallel for if (n > 1000)
+    for (int64_t i = 0; i < n; i++) {
+        const int32_t* codesi = code + i * ld_codes;
+        float* xi = x + i * d;
+        for (int m = 0; m < M; m++) {
+            int idx = codesi[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 in centroids
+ ****************************************************************************/
+
+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_centroids_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_centroids_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);
+        }
+    }
+}
+
+/****************************************************************************
+ * Support for fast distance computations in codes
+ ****************************************************************************/
+
+namespace {
+
+float accumulate_IPs(
+        const AdditiveQuantizer& aq,
+        BitstringReader& bs,
+        const uint8_t* codes,
+        const float* LUT) {
+    float accu = 0;
+    for (int m = 0; m < aq.M; m++) {
+        size_t nbit = aq.nbits[m];
+        int idx = bs.read(nbit);
+        accu += LUT[idx];
+        LUT += (uint64_t)1 << nbit;
+    }
+    return accu;
+}
+
+} // anonymous namespace
+
+template <>
+float AdditiveQuantizer::
+        compute_1_distance_LUT<true, AdditiveQuantizer::ST_LUT_nonorm>(
+                const uint8_t* codes,
+                const float* LUT) const {
+    BitstringReader bs(codes, code_size);
+    return accumulate_IPs(*this, bs, codes, LUT);
+}
+
+template <>
+float AdditiveQuantizer::
+        compute_1_distance_LUT<false, AdditiveQuantizer::ST_LUT_nonorm>(
+                const uint8_t* codes,
+                const float* LUT) const {
+    BitstringReader bs(codes, code_size);
+    return -accumulate_IPs(*this, bs, codes, LUT);
+}
+
+template <>
+float AdditiveQuantizer::
+        compute_1_distance_LUT<false, AdditiveQuantizer::ST_norm_float>(
+                const uint8_t* codes,
+                const float* LUT) const {
+    BitstringReader bs(codes, code_size);
+    float accu = accumulate_IPs(*this, bs, codes, LUT);
+    uint32_t norm_i = bs.read(32);
+    float norm2 = *(float*)&norm_i;
+    return norm2 - 2 * accu;
+}
+
+template <>
+float AdditiveQuantizer::
+        compute_1_distance_LUT<false, AdditiveQuantizer::ST_norm_cqint8>(
+                const uint8_t* codes,
+                const float* LUT) const {
+    BitstringReader bs(codes, code_size);
+    float accu = accumulate_IPs(*this, bs, codes, LUT);
+    uint32_t norm_i = bs.read(8);
+    float norm2 = decode_qcint(norm_i);
+    return norm2 - 2 * accu;
+}
+
+template <>
+float AdditiveQuantizer::
+        compute_1_distance_LUT<false, AdditiveQuantizer::ST_norm_cqint4>(
+                const uint8_t* codes,
+                const float* LUT) const {
+    BitstringReader bs(codes, code_size);
+    float accu = accumulate_IPs(*this, bs, codes, LUT);
+    uint32_t norm_i = bs.read(4);
+    float norm2 = decode_qcint(norm_i);
+    return norm2 - 2 * accu;
+}
+
+template <>
+float AdditiveQuantizer::
+        compute_1_distance_LUT<false, AdditiveQuantizer::ST_norm_qint8>(
+                const uint8_t* codes,
+                const float* LUT) const {
+    BitstringReader bs(codes, code_size);
+    float accu = accumulate_IPs(*this, bs, codes, LUT);
+    uint32_t norm_i = bs.read(8);
+    float norm2 = decode_qint8(norm_i, norm_min, norm_max);
+    return norm2 - 2 * accu;
+}
+
+template <>
+float AdditiveQuantizer::
+        compute_1_distance_LUT<false, AdditiveQuantizer::ST_norm_qint4>(
+                const uint8_t* codes,
+                const float* LUT) const {
+    BitstringReader bs(codes, code_size);
+    float accu = accumulate_IPs(*this, bs, codes, LUT);
+    uint32_t norm_i = bs.read(4);
+    float norm2 = decode_qint4(norm_i, norm_min, norm_max);
+    return norm2 - 2 * accu;
+}
+
+} // namespace faiss