• matthijs's avatar
    sync with FB version 2017-11-22 · 250a3d3f
    matthijs authored
    various bugfixes from github issues
    kmean with some frozen centroids
    GPU better tiling for large flat datasets
    default AVX for vector ops
    250a3d3f
Distance.cu 16.3 KB
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492
/**
 * Copyright (c) 2015-present, Facebook, Inc.
 * All rights reserved.
 *
 * This source code is licensed under the BSD+Patents license found in the
 * LICENSE file in the root directory of this source tree.
 */

// Copyright 2004-present Facebook. All Rights Reserved.

#include "Distance.cuh"
#include "BroadcastSum.cuh"
#include "L2Norm.cuh"
#include "L2Select.cuh"
#include "../../FaissAssert.h"
#include "../GpuResources.h"
#include "../utils/DeviceUtils.h"
#include "../utils/Limits.cuh"
#include "../utils/MatrixMult.cuh"
#include "../utils/BlockSelectKernel.cuh"

#include <memory>
#include <thrust/fill.h>
#include <thrust/for_each.h>
#include <thrust/device_ptr.h>
#include <thrust/execution_policy.h>

namespace faiss { namespace gpu {

namespace {

template <typename T>
Tensor<T, 2, true> sliceCentroids(Tensor<T, 2, true>& centroids,
                                  Tensor<T, 2, true>* centroidsTransposed,
                                  int startCentroid,
                                  int num) {
  if (startCentroid == 0 && num == centroids.getSize(0)) {
    if (centroidsTransposed) {
      return *centroidsTransposed;
    } else {
      return centroids;
    }
  }

  if (centroidsTransposed) {
    // (dim, num)
    return centroidsTransposed->narrow(1, startCentroid, num);
  } else {
    return centroids.narrow(0, startCentroid, num);
  }
}

// For each chunk of k indices, increment the index by chunk * increment
template <typename T>
__global__ void incrementIndex(Tensor<T, 2, true> indices,
                               int k,
                               int increment) {
  for (int i = threadIdx.x; i < k; i += blockDim.x) {
    indices[blockIdx.y][blockIdx.x * k + i] += blockIdx.x * increment;
  }
}

// Used to update result indices in distance computation where the number of
// centroids is high, and is tiled
template <typename T>
void runIncrementIndex(Tensor<T, 2, true>& indices,
                       int k,
                       int increment,
                       cudaStream_t stream) {
  dim3 grid(indices.getSize(1) / k, indices.getSize(0));
  int block = std::min(k, 512);

  // should be exact
  FAISS_ASSERT(grid.x * k == indices.getSize(1));

  incrementIndex<<<grid, block, 0, stream>>>(indices, k, increment);

  cudaDeviceSynchronize();
}

// If the inner size (dim) of the vectors is small, we want a larger query tile
// size, like 1024

void chooseTileSize(int numQueries,
                    int numCentroids,
                    int dim,
                    int elementSize,
                    size_t tempMemAvailable,
                    int& tileRows,
                    int& tileCols) {
  // The matrix multiplication should be large enough to be efficient, but if it
  // is too large, we seem to lose efficiency as opposed to double-streaming.
  // Each tile size here defines 1/2 of the memory use due to double streaming.
  // We ignore available temporary memory, as that is adjusted independently by
  // the user and can thus meet these requirements (or not).
  // For <= 4 GB GPUs, prefer 512 MB of usage.
  // For <= 8 GB GPUs, prefer 768 MB of usage.
  // Otherwise, prefer 1 GB of usage.
  auto totalMem = getCurrentDeviceProperties().totalGlobalMem;

  int targetUsage = 0;

  if (totalMem <= ((size_t) 4) * 1024 * 1024 * 1024) {
    targetUsage = 512 * 1024 * 1024;
  } else if (totalMem <= ((size_t) 8) * 1024 * 1024 * 1024) {
    targetUsage = 768 * 1024 * 1024;
  } else {
    targetUsage = 1024 * 1024 * 1024;
  }

  targetUsage /= 2 * elementSize;

  // 512 seems to be a batch size sweetspot for float32.
  // If we are on float16, increase to 512.
  // If the k size (vec dim) of the matrix multiplication is small (<= 32),
  // increase to 1024.
  int preferredTileRows = 512;
  if (dim <= 32) {
    preferredTileRows = 1024;
  }

  tileRows = std::min(preferredTileRows, numQueries);

  // tileCols is the remainder size
  tileCols = std::min(targetUsage / preferredTileRows, numCentroids);
}

}

template <typename T>
void runDistance(bool computeL2,
                 GpuResources* resources,
                 Tensor<T, 2, true>& centroids,
                 Tensor<T, 2, true>* centroidsTransposed,
                 Tensor<T, 1, true>* centroidNorms,
                 Tensor<T, 2, true>& queries,
                 int k,
                 Tensor<T, 2, true>& outDistances,
                 Tensor<int, 2, true>& outIndices,
                 bool useHgemm,
                 bool ignoreOutDistances) {
  FAISS_ASSERT(outDistances.getSize(0) == queries.getSize(0));
  FAISS_ASSERT(outIndices.getSize(0) == queries.getSize(0));
  FAISS_ASSERT(outDistances.getSize(1) == k);
  FAISS_ASSERT(outIndices.getSize(1) == k);

  auto& mem = resources->getMemoryManagerCurrentDevice();
  auto defaultStream = resources->getDefaultStreamCurrentDevice();

  // If we're quering against a 0 sized set, just return empty results
  if (centroids.numElements() == 0) {
    thrust::fill(thrust::cuda::par.on(defaultStream),
                 outDistances.data(), outDistances.end(),
                 Limits<T>::getMax());

    thrust::fill(thrust::cuda::par.on(defaultStream),
                 outIndices.data(), outIndices.end(),
                 -1);

    return;
  }

  // L2: If ||c||^2 is not pre-computed, calculate it
  DeviceTensor<T, 1, true> cNorms;
  if (computeL2 && !centroidNorms) {
    cNorms = std::move(DeviceTensor<T, 1, true>(
                       mem,
                       {centroids.getSize(0)}, defaultStream));
    runL2Norm(centroids, cNorms, true, defaultStream);
    centroidNorms = &cNorms;
  }

  //
  // Prepare norm vector ||q||^2; ||c||^2 is already pre-computed
  //
  int qNormSize[1] = {queries.getSize(0)};
  DeviceTensor<T, 1, true> queryNorms(mem, qNormSize, defaultStream);

  // ||q||^2
  if (computeL2) {
    runL2Norm(queries, queryNorms, true, defaultStream);
  }

  // By default, aim to use up to 512 MB of memory for the processing, with both
  // number of queries and number of centroids being at least 512.
  int tileRows = 0;
  int tileCols = 0;
  chooseTileSize(queries.getSize(0),
                 centroids.getSize(0),
                 queries.getSize(1),
                 sizeof(T),
                 mem.getSizeAvailable(),
                 tileRows,
                 tileCols);

  int numColTiles = utils::divUp(centroids.getSize(0), tileCols);

  FAISS_ASSERT(k <= centroids.getSize(0));
  FAISS_ASSERT(k <= 1024); // select limitation

  // Temporary output memory space we'll use
  DeviceTensor<T, 2, true> distanceBuf1(
    mem, {tileRows, tileCols}, defaultStream);
  DeviceTensor<T, 2, true> distanceBuf2(
    mem, {tileRows, tileCols}, defaultStream);
  DeviceTensor<T, 2, true>* distanceBufs[2] =
    {&distanceBuf1, &distanceBuf2};

  DeviceTensor<T, 2, true> outDistanceBuf1(
    mem, {tileRows, numColTiles * k}, defaultStream);
  DeviceTensor<T, 2, true> outDistanceBuf2(
    mem, {tileRows, numColTiles * k}, defaultStream);
  DeviceTensor<T, 2, true>* outDistanceBufs[2] =
    {&outDistanceBuf1, &outDistanceBuf2};

  DeviceTensor<int, 2, true> outIndexBuf1(
    mem, {tileRows, numColTiles * k}, defaultStream);
  DeviceTensor<int, 2, true> outIndexBuf2(
    mem, {tileRows, numColTiles * k}, defaultStream);
  DeviceTensor<int, 2, true>* outIndexBufs[2] =
    {&outIndexBuf1, &outIndexBuf2};

  auto streams = resources->getAlternateStreamsCurrentDevice();
  streamWait(streams, {defaultStream});

  int curStream = 0;

  // Tile over the input queries
  for (int i = 0; i < queries.getSize(0); i += tileRows) {
    int curQuerySize = std::min(tileRows, queries.getSize(0) - i);

    auto outDistanceView =
      outDistances.narrow(0, i, curQuerySize);
    auto outIndexView =
      outIndices.narrow(0, i, curQuerySize);

    auto queryView =
      queries.narrow(0, i, curQuerySize);
    auto queryNormNiew =
      queryNorms.narrow(0, i, curQuerySize);

    auto outDistanceBufRowView =
      outDistanceBufs[curStream]->narrow(0, 0, curQuerySize);
    auto outIndexBufRowView =
      outIndexBufs[curStream]->narrow(0, 0, curQuerySize);

    // Tile over the centroids
    for (int j = 0; j < centroids.getSize(0); j += tileCols) {
      int curCentroidSize = std::min(tileCols, centroids.getSize(0) - j);

      int curColTile = j / tileCols;

      auto centroidsView =
        sliceCentroids(centroids, centroidsTransposed, j, curCentroidSize);

      auto distanceBufView = distanceBufs[curStream]->
        narrow(0, 0, curQuerySize).narrow(1, 0, curCentroidSize);

      auto outDistanceBufColView =
        outDistanceBufRowView.narrow(1, k * curColTile, k);
      auto outIndexBufColView =
        outIndexBufRowView.narrow(1, k * curColTile, k);

      // L2: distance is ||c||^2 - 2qc + ||q||^2, we compute -2qc
      // IP: just compute qc
      // (query id x dim) x (centroid id, dim)' = (query id, centroid id)
      runMatrixMult(distanceBufView, false,
                    queryView, false,
                    centroidsView,
                    centroidsTransposed ? false : true,
                    computeL2 ? -2.0f : 1.0f, 0.0f, useHgemm,
                    resources->getBlasHandleCurrentDevice(),
                    streams[curStream]);

      if (computeL2) {
        // For L2 distance, we use this fused kernel that performs both
        // adding ||c||^2 to -2qc and k-selection, so we only need two
        // passes (one write by the gemm, one read here) over the huge
        // region of output memory
        //
        // If we aren't tiling along the number of centroids, we can perform the
        // output work directly
        if (tileCols == centroids.getSize(0)) {
          // Write into the final output
          runL2SelectMin(distanceBufView,
                         *centroidNorms,
                         outDistanceView,
                         outIndexView,
                         k,
                         streams[curStream]);

          if (!ignoreOutDistances) {
            // expand (query id) to (query id, k) by duplicating along rows
            // top-k ||c||^2 - 2qc + ||q||^2 in the form (query id, k)
            runSumAlongRows(queryNormNiew, outDistanceView, streams[curStream]);
          }
        } else {
          auto centroidNormsView =
            centroidNorms->narrow(0, j, curCentroidSize);

          // Write into our intermediate output
          runL2SelectMin(distanceBufView,
                         centroidNormsView,
                         outDistanceBufColView,
                         outIndexBufColView,
                         k,
                         streams[curStream]);

          if (!ignoreOutDistances) {
            // expand (query id) to (query id, k) by duplicating along rows
            // top-k ||c||^2 - 2qc + ||q||^2 in the form (query id, k)
            runSumAlongRows(queryNormNiew,
                            outDistanceBufColView,
                            streams[curStream]);
          }
        }
      } else {
        // For IP, just k-select the output for this tile
        if (tileCols == centroids.getSize(0)) {
          // Write into the final output
          runBlockSelect(distanceBufView,
                         outDistanceView,
                         outIndexView,
                         true, k, streams[curStream]);
        } else {
          // Write into the intermediate output
          runBlockSelect(distanceBufView,
                         outDistanceBufColView,
                         outIndexBufColView,
                         true, k, streams[curStream]);
        }
      }
    }

    // As we're finished with processing a full set of centroids, perform the
    // final k-selection
    if (tileCols != centroids.getSize(0)) {
      // The indices are tile-relative; for each tile of k, we need to add
      // tileCols to the index
      runIncrementIndex(outIndexBufRowView, k, tileCols, streams[curStream]);

      runBlockSelectPair(outDistanceBufRowView,
                         outIndexBufRowView,
                         outDistanceView,
                         outIndexView,
                         computeL2 ? false : true, k, streams[curStream]);
    }

    curStream = (curStream + 1) % 2;
  }

  // Have the desired ordering stream wait on the multi-stream
  streamWait({defaultStream}, streams);
}

template <typename T>
void runL2Distance(GpuResources* resources,
                   Tensor<T, 2, true>& centroids,
                   Tensor<T, 2, true>* centroidsTransposed,
                   Tensor<T, 1, true>* centroidNorms,
                   Tensor<T, 2, true>& queries,
                   int k,
                   Tensor<T, 2, true>& outDistances,
                   Tensor<int, 2, true>& outIndices,
                   bool useHgemm,
                   bool ignoreOutDistances = false) {
  runDistance<T>(true, // L2
                 resources,
                 centroids,
                 centroidsTransposed,
                 centroidNorms,
                 queries,
                 k,
                 outDistances,
                 outIndices,
                 useHgemm,
                 ignoreOutDistances);
}

template <typename T>
void runIPDistance(GpuResources* resources,
                   Tensor<T, 2, true>& centroids,
                   Tensor<T, 2, true>* centroidsTransposed,
                   Tensor<T, 2, true>& queries,
                   int k,
                   Tensor<T, 2, true>& outDistances,
                   Tensor<int, 2, true>& outIndices,
                   bool useHgemm) {
  runDistance<T>(false, // IP
                 resources,
                 centroids,
                 centroidsTransposed,
                 nullptr,
                 queries,
                 k,
                 outDistances,
                 outIndices,
                 useHgemm,
                 false);
}

//
// Instantiations of the distance templates
//

void
runIPDistance(GpuResources* resources,
              Tensor<float, 2, true>& vectors,
              Tensor<float, 2, true>* vectorsTransposed,
              Tensor<float, 2, true>& queries,
              int k,
              Tensor<float, 2, true>& outDistances,
              Tensor<int, 2, true>& outIndices) {
  runIPDistance<float>(resources,
                       vectors,
                       vectorsTransposed,
                       queries,
                       k,
                       outDistances,
                       outIndices,
                       false);
}

#ifdef FAISS_USE_FLOAT16
void
runIPDistance(GpuResources* resources,
              Tensor<half, 2, true>& vectors,
              Tensor<half, 2, true>* vectorsTransposed,
              Tensor<half, 2, true>& queries,
              int k,
              Tensor<half, 2, true>& outDistances,
              Tensor<int, 2, true>& outIndices,
              bool useHgemm) {
  runIPDistance<half>(resources,
                      vectors,
                      vectorsTransposed,
                      queries,
                      k,
                      outDistances,
                      outIndices,
                      useHgemm);
}
#endif

void
runL2Distance(GpuResources* resources,
              Tensor<float, 2, true>& vectors,
              Tensor<float, 2, true>* vectorsTransposed,
              Tensor<float, 1, true>* vectorNorms,
              Tensor<float, 2, true>& queries,
              int k,
              Tensor<float, 2, true>& outDistances,
              Tensor<int, 2, true>& outIndices,
              bool ignoreOutDistances) {
  runL2Distance<float>(resources,
                       vectors,
                       vectorsTransposed,
                       vectorNorms,
                       queries,
                       k,
                       outDistances,
                       outIndices,
                       false,
                       ignoreOutDistances);
}

#ifdef FAISS_USE_FLOAT16
void
runL2Distance(GpuResources* resources,
              Tensor<half, 2, true>& vectors,
              Tensor<half, 2, true>* vectorsTransposed,
              Tensor<half, 1, true>* vectorNorms,
              Tensor<half, 2, true>& queries,
              int k,
              Tensor<half, 2, true>& outDistances,
              Tensor<int, 2, true>& outIndices,
              bool useHgemm,
              bool ignoreOutDistances) {
  runL2Distance<half>(resources,
                      vectors,
                      vectorsTransposed,
                      vectorNorms,
                      queries,
                      k,
                      outDistances,
                      outIndices,
                      useHgemm,
                      ignoreOutDistances);
}
#endif

} } // namespace