Faiss
 All Classes Namespaces Functions Variables Typedefs Enumerations Enumerator Friends
PerfFlat.cu
1 /**
2  * Copyright (c) Facebook, Inc. and its affiliates.
3  *
4  * This source code is licensed under the MIT license found in the
5  * LICENSE file in the root directory of this source tree.
6  */
7 
8 
9 #include "../../IndexFlat.h"
10 #include "../../utils.h"
11 #include "../GpuIndexFlat.h"
12 #include "IndexWrapper.h"
13 #include "../test/TestUtils.h"
14 #include "../utils/DeviceTensor.cuh"
15 #include "../utils/DeviceUtils.h"
16 #include "../utils/HostTensor.cuh"
17 #include "../utils/Timer.h"
18 #include <gflags/gflags.h>
19 #include <map>
20 #include <memory>
21 #include <vector>
22 
23 #include <cuda_profiler_api.h>
24 
25 DEFINE_bool(l2, true, "L2 or inner product");
26 DEFINE_int32(k, 3, "final number of closest results returned");
27 DEFINE_int32(num, 128, "# of vecs");
28 DEFINE_int32(dim, 128, "# of dimensions");
29 DEFINE_int32(num_queries, 3, "number of query vectors");
30 DEFINE_bool(diff, true, "show exact distance + index output discrepancies");
31 DEFINE_bool(use_float16, false, "use encodings in float16");
32 DEFINE_bool(use_float16_math, false, "perform math in float16");
33 DEFINE_bool(transposed, false, "store vectors transposed");
34 DEFINE_int64(seed, -1, "specify random seed");
35 DEFINE_int32(num_gpus, 1, "number of gpus to use");
36 DEFINE_int64(pinned_mem, 0, "pinned memory allocation to use");
37 DEFINE_bool(cpu, true, "run the CPU code for timing and comparison");
38 DEFINE_bool(use_unified_mem, false, "use Pascal unified memory for the index");
39 
40 using namespace faiss::gpu;
41 
42 int main(int argc, char** argv) {
43  gflags::ParseCommandLineFlags(&argc, &argv, true);
44 
45  cudaProfilerStop();
46 
47  auto seed = FLAGS_seed != -1L ? FLAGS_seed : time(nullptr);
48  printf("using seed %ld\n", seed);
49 
50  auto numQueries = FLAGS_num_queries;
51 
52  auto index = std::unique_ptr<faiss::IndexFlat>(
53  new faiss::IndexFlat(FLAGS_dim, FLAGS_l2 ?
54  faiss::METRIC_L2 : faiss::METRIC_INNER_PRODUCT));
55 
56  HostTensor<float, 2, true> vecs({FLAGS_num, FLAGS_dim});
57  faiss::float_rand(vecs.data(), vecs.numElements(), seed);
58 
59  index->add(FLAGS_num, vecs.data());
60 
61  printf("Database: dim %d num vecs %d\n", FLAGS_dim, FLAGS_num);
62  printf("%s lookup: %d queries, total k %d\n",
63  FLAGS_l2 ? "L2" : "IP",
64  numQueries, FLAGS_k);
65  printf("float16 encoding %s\n", FLAGS_use_float16 ? "enabled" : "disabled");
66  printf("transposed storage %s\n", FLAGS_transposed ? "enabled" : "disabled");
67 
68  // Convert to GPU index
69  printf("Copying index to %d GPU(s)...\n", FLAGS_num_gpus);
70 
71  auto initFn = [&index](faiss::gpu::GpuResources* res, int dev) ->
72  std::unique_ptr<faiss::gpu::GpuIndexFlat> {
73  ((faiss::gpu::StandardGpuResources*) res)->setPinnedMemory(
74  FLAGS_pinned_mem);
75 
76  GpuIndexFlatConfig config;
77  config.device = dev;
78  config.useFloat16 = FLAGS_use_float16;
79  config.useFloat16Accumulator = FLAGS_use_float16_math;
80  config.storeTransposed = FLAGS_transposed;
81  config.memorySpace = FLAGS_use_unified_mem ?
82  MemorySpace::Unified : MemorySpace::Device;
83 
84  auto p = std::unique_ptr<faiss::gpu::GpuIndexFlat>(
85  new faiss::gpu::GpuIndexFlat(res, index.get(), config));
86  return p;
87  };
88 
89  IndexWrapper<faiss::gpu::GpuIndexFlat> gpuIndex(FLAGS_num_gpus, initFn);
90  printf("copy done\n");
91 
92  // Build query vectors
93  HostTensor<float, 2, true> cpuQuery({numQueries, FLAGS_dim});
94  faiss::float_rand(cpuQuery.data(), cpuQuery.numElements(), seed);
95 
96  // Time faiss CPU
97  HostTensor<float, 2, true> cpuDistances({numQueries, FLAGS_k});
98  HostTensor<faiss::Index::idx_t, 2, true> cpuIndices({numQueries, FLAGS_k});
99 
100  if (FLAGS_cpu) {
101  float cpuTime = 0.0f;
102 
103  CpuTimer timer;
104  index->search(numQueries,
105  cpuQuery.data(),
106  FLAGS_k,
107  cpuDistances.data(),
108  cpuIndices.data());
109 
110  cpuTime = timer.elapsedMilliseconds();
111  printf("CPU time %.3f ms\n", cpuTime);
112  }
113 
114  HostTensor<float, 2, true> gpuDistances({numQueries, FLAGS_k});
115  HostTensor<faiss::Index::idx_t, 2, true> gpuIndices({numQueries, FLAGS_k});
116 
117  CUDA_VERIFY(cudaProfilerStart());
118  faiss::gpu::synchronizeAllDevices();
119 
120  float gpuTime = 0.0f;
121 
122  // Time GPU
123  {
124  CpuTimer timer;
125 
126  gpuIndex.getIndex()->search(cpuQuery.getSize(0),
127  cpuQuery.data(),
128  FLAGS_k,
129  gpuDistances.data(),
130  gpuIndices.data());
131 
132  // There is a device -> host copy above, so no need to time
133  // additional synchronization with the GPU
134  gpuTime = timer.elapsedMilliseconds();
135  }
136 
137  CUDA_VERIFY(cudaProfilerStop());
138  printf("GPU time %.3f ms\n", gpuTime);
139 
140  if (FLAGS_cpu) {
141  compareLists(cpuDistances.data(), cpuIndices.data(),
142  gpuDistances.data(), gpuIndices.data(),
143  numQueries, FLAGS_k,
144  "", true, FLAGS_diff, false);
145  }
146 
147  CUDA_VERIFY(cudaDeviceSynchronize());
148 
149  return 0;
150 }
float elapsedMilliseconds()
Returns elapsed time in milliseconds.
Definition: Timer.cpp:50
CPU wallclock elapsed timer.
Definition: Timer.h:40
bool useFloat16
Whether or not data is stored as float16.
Definition: GpuIndexFlat.h:33
int device
GPU device on which the index is resident.
Definition: GpuIndex.h:25
MemorySpace memorySpace
Definition: GpuIndex.h:30