-
Notifications
You must be signed in to change notification settings - Fork 3.7k
/
PQCodeDistances-inl.cuh
561 lines (478 loc) · 18.3 KB
/
PQCodeDistances-inl.cuh
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
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
/**
* 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.
*/
#include <faiss/gpu/impl/BroadcastSum.cuh>
#include <faiss/gpu/impl/Distance.cuh>
#include <faiss/gpu/impl/L2Norm.cuh>
#include <faiss/gpu/utils/ConversionOperators.cuh>
#include <faiss/gpu/utils/DeviceDefs.cuh>
#include <faiss/gpu/utils/DeviceUtils.h>
#include <faiss/gpu/utils/Float16.cuh>
#include <faiss/gpu/utils/MatrixMult.cuh>
#include <faiss/gpu/utils/PtxUtils.cuh>
#include <faiss/gpu/utils/StaticUtils.h>
#include <faiss/gpu/utils/Transpose.cuh>
namespace faiss { namespace gpu {
// Kernel responsible for calculating distance from residual vector to
// each product quantizer code centroid
template <typename OutCodeT,
typename CentroidT,
int DimsPerSubQuantizer,
bool L2Distance>
__global__ void
__launch_bounds__(288, 4)
pqCodeDistances(Tensor<float, 2, true> queries,
int queriesPerBlock,
Tensor<CentroidT, 2, true> coarseCentroids,
Tensor<float, 3, true> pqCentroids,
Tensor<int, 2, true> topQueryToCentroid,
// (query id)(coarse)(subquantizer)(code) -> dist
Tensor<OutCodeT, 4, true> outCodeDistances) {
const auto numSubQuantizers = pqCentroids.getSize(0);
const auto dimsPerSubQuantizer = pqCentroids.getSize(1);
assert(DimsPerSubQuantizer == dimsPerSubQuantizer);
const auto codesPerSubQuantizer = pqCentroids.getSize(2);
bool isLoadingThread = threadIdx.x >= codesPerSubQuantizer;
int loadingThreadId = threadIdx.x - codesPerSubQuantizer;
extern __shared__ float smem[];
// Each thread calculates a single code
float subQuantizerData[DimsPerSubQuantizer];
auto code = threadIdx.x;
auto subQuantizer = blockIdx.y;
// Each thread will load the pq centroid data for the code that it
// is processing
#pragma unroll
for (int i = 0; i < DimsPerSubQuantizer; ++i) {
subQuantizerData[i] = pqCentroids[subQuantizer][i][code].ldg();
}
// Where we store our query vector
float* smemQuery = smem;
// Where we store our residual vector; this is double buffered so we
// can be loading the next one while processing the current one
float* smemResidual1 = &smemQuery[DimsPerSubQuantizer];
float* smemResidual2 = &smemResidual1[DimsPerSubQuantizer];
// Where we pre-load the coarse centroid IDs
int* coarseIds = (int*) &smemResidual2[DimsPerSubQuantizer];
// Each thread is calculating the distance for a single code,
// performing the reductions locally
// Handle multiple queries per block
auto startQueryId = blockIdx.x * queriesPerBlock;
auto numQueries = queries.getSize(0) - startQueryId;
if (numQueries > queriesPerBlock) {
numQueries = queriesPerBlock;
}
for (int query = 0; query < numQueries; ++query) {
auto queryId = startQueryId + query;
auto querySubQuantizer =
queries[queryId][subQuantizer * DimsPerSubQuantizer].data();
// Load current query vector
for (int i = threadIdx.x; i < DimsPerSubQuantizer; i += blockDim.x) {
smemQuery[i] = querySubQuantizer[i];
}
// Load list of coarse centroids found
for (int i = threadIdx.x;
i < topQueryToCentroid.getSize(1); i += blockDim.x) {
coarseIds[i] = topQueryToCentroid[queryId][i];
}
// We need coarseIds below
// FIXME: investigate loading separately, so we don't need this
__syncthreads();
// Preload first buffer of residual data
if (isLoadingThread) {
for (int i = loadingThreadId;
i < DimsPerSubQuantizer;
i += blockDim.x - codesPerSubQuantizer) {
auto coarseId = coarseIds[0];
// In case NaNs were in the original query data
coarseId = coarseId == -1 ? 0 : coarseId;
auto coarseCentroidSubQuantizer =
coarseCentroids[coarseId][subQuantizer * dimsPerSubQuantizer].data();
if (L2Distance) {
smemResidual1[i] = smemQuery[i] -
ConvertTo<float>::to(coarseCentroidSubQuantizer[i]);
} else {
smemResidual1[i] =
ConvertTo<float>::to(coarseCentroidSubQuantizer[i]);
}
}
}
// The block walks the list for a single query
for (int coarse = 0; coarse < topQueryToCentroid.getSize(1); ++coarse) {
// Wait for smemResidual1 to be loaded
__syncthreads();
if (isLoadingThread) {
// Preload second buffer of residual data
for (int i = loadingThreadId;
i < DimsPerSubQuantizer;
i += blockDim.x - codesPerSubQuantizer) {
// FIXME: try always making this centroid id 0 so we can
// terminate
if (coarse != (topQueryToCentroid.getSize(1) - 1)) {
auto coarseId = coarseIds[coarse + 1];
// In case NaNs were in the original query data
coarseId = coarseId == -1 ? 0 : coarseId;
auto coarseCentroidSubQuantizer =
coarseCentroids[coarseId]
[subQuantizer * dimsPerSubQuantizer].data();
if (L2Distance) {
smemResidual2[i] = smemQuery[i] -
ConvertTo<float>::to(coarseCentroidSubQuantizer[i]);
} else {
smemResidual2[i] =
ConvertTo<float>::to(coarseCentroidSubQuantizer[i]);
}
}
}
} else {
// These are the processing threads
float dist = 0.0f;
constexpr int kUnroll = 4;
constexpr int kRemainder = DimsPerSubQuantizer % kUnroll;
constexpr int kRemainderBase = DimsPerSubQuantizer - kRemainder;
float vals[kUnroll];
// Calculate residual - pqCentroid for each dim that we're
// processing
// Unrolled loop
if (L2Distance) {
#pragma unroll
for (int i = 0; i < DimsPerSubQuantizer / kUnroll; ++i) {
#pragma unroll
for (int j = 0; j < kUnroll; ++j) {
vals[j] = smemResidual1[i * kUnroll + j];
}
#pragma unroll
for (int j = 0; j < kUnroll; ++j) {
vals[j] -= subQuantizerData[i * kUnroll + j];
}
#pragma unroll
for (int j = 0; j < kUnroll; ++j) {
vals[j] *= vals[j];
}
#pragma unroll
for (int j = 0; j < kUnroll; ++j) {
dist += vals[j];
}
}
} else {
// Inner product: query slice against the reconstructed sub-quantizer
// for this coarse cell (query o (centroid + subQCentroid))
#pragma unroll
for (int i = 0; i < DimsPerSubQuantizer / kUnroll; ++i) {
#pragma unroll
for (int j = 0; j < kUnroll; ++j) {
vals[j] = smemResidual1[i * kUnroll + j];
}
#pragma unroll
for (int j = 0; j < kUnroll; ++j) {
vals[j] += subQuantizerData[i * kUnroll + j];
}
#pragma unroll
for (int j = 0; j < kUnroll; ++j) {
vals[j] *= smemQuery[i * kUnroll + j];
}
#pragma unroll
for (int j = 0; j < kUnroll; ++j) {
dist += vals[j];
}
}
}
// Remainder loop
if (L2Distance) {
#pragma unroll
for (int j = 0; j < kRemainder; ++j) {
vals[j] = smemResidual1[kRemainderBase + j];
}
#pragma unroll
for (int j = 0; j < kRemainder; ++j) {
vals[j] -= subQuantizerData[kRemainderBase + j];
}
#pragma unroll
for (int j = 0; j < kRemainder; ++j) {
vals[j] *= vals[j];
}
} else {
// Inner product
// Inner product: query slice against the reconstructed sub-quantizer
// for this coarse cell (query o (centroid + subQCentroid))
#pragma unroll
for (int j = 0; j < kRemainder; ++j) {
vals[j] = smemResidual1[kRemainderBase + j];
}
#pragma unroll
for (int j = 0; j < kRemainder; ++j) {
vals[j] += subQuantizerData[kRemainderBase + j];
}
#pragma unroll
for (int j = 0; j < kRemainder; ++j) {
vals[j] *= smemQuery[kRemainderBase + j];
}
}
#pragma unroll
for (int j = 0; j < kRemainder; ++j) {
dist += vals[j];
}
// We have the distance for our code; write it out
outCodeDistances[queryId][coarse][subQuantizer][code] =
ConvertTo<OutCodeT>::to(dist);
} // !isLoadingThread
// Swap residual buffers
float* tmp = smemResidual1;
smemResidual1 = smemResidual2;
smemResidual2 = tmp;
}
}
}
template <typename CentroidT>
__global__ void
residualVector(Tensor<float, 2, true> queries,
Tensor<CentroidT, 2, true> coarseCentroids,
Tensor<int, 2, true> topQueryToCentroid,
int numSubDim,
// output is transposed:
// (sub q)(query id)(centroid id)(sub dim)
Tensor<float, 4, true> residual) {
// block x is query id
// block y is centroid id
// thread x is dim
auto queryId = blockIdx.x;
auto centroidId = blockIdx.y;
int realCentroidId = topQueryToCentroid[queryId][centroidId];
for (int dim = threadIdx.x; dim < queries.getSize(1); dim += blockDim.x) {
float q = queries[queryId][dim];
float c = ConvertTo<float>::to(coarseCentroids[realCentroidId][dim]);
residual[dim / numSubDim][queryId][centroidId][dim % numSubDim] = q - c;
}
}
template <typename CentroidT>
void
runResidualVector(Tensor<float, 3, true>& pqCentroids,
Tensor<float, 2, true>& queries,
Tensor<CentroidT, 2, true>& coarseCentroids,
Tensor<int, 2, true>& topQueryToCentroid,
Tensor<float, 4, true>& residual,
cudaStream_t stream) {
auto grid =
dim3(topQueryToCentroid.getSize(0), topQueryToCentroid.getSize(1));
auto block = dim3(std::min(queries.getSize(1), getMaxThreadsCurrentDevice()));
residualVector<<<grid, block, 0, stream>>>(
queries, coarseCentroids, topQueryToCentroid, pqCentroids.getSize(1),
residual);
CUDA_TEST_ERROR();
}
template <typename CentroidT>
void
runPQCodeDistancesMM(Tensor<float, 3, true>& pqCentroids,
Tensor<float, 2, true>& queries,
Tensor<CentroidT, 2, true>& coarseCentroids,
Tensor<int, 2, true>& topQueryToCentroid,
NoTypeTensor<4, true>& outCodeDistances,
bool useFloat16Lookup,
DeviceMemory& mem,
cublasHandle_t handle,
cudaStream_t stream) {
// Calculate (q - c) residual vector
// (sub q)(query id)(centroid id)(sub dim)
DeviceTensor<float, 4, true> residual(
mem,
{pqCentroids.getSize(0),
topQueryToCentroid.getSize(0),
topQueryToCentroid.getSize(1),
pqCentroids.getSize(1)},
stream);
runResidualVector(pqCentroids, queries,
coarseCentroids, topQueryToCentroid,
residual, stream);
// Calculate ||q - c||^2
DeviceTensor<float, 1, true> residualNorms(
mem,
{pqCentroids.getSize(0) *
topQueryToCentroid.getSize(0) *
topQueryToCentroid.getSize(1)},
stream);
auto residualView2 = residual.view<2>(
{pqCentroids.getSize(0) *
topQueryToCentroid.getSize(0) *
topQueryToCentroid.getSize(1),
pqCentroids.getSize(1)});
runL2Norm(residualView2, true, residualNorms, true, stream);
// Perform a batch MM:
// (sub q) x {(q * c)(sub dim) x (sub dim)(code)} =>
// (sub q) x {(q * c)(code)}
auto residualView3 = residual.view<3>(
{pqCentroids.getSize(0),
topQueryToCentroid.getSize(0) * topQueryToCentroid.getSize(1),
pqCentroids.getSize(1)});
DeviceTensor<float, 3, true> residualDistance(
mem,
{pqCentroids.getSize(0),
topQueryToCentroid.getSize(0) * topQueryToCentroid.getSize(1),
pqCentroids.getSize(2)},
stream);
runIteratedMatrixMult(residualDistance, false,
residualView3, false,
pqCentroids, false,
-2.0f, 0.0f,
handle,
stream);
// Sum ||q - c||^2 along rows
auto residualDistanceView2 = residualDistance.view<2>(
{pqCentroids.getSize(0) *
topQueryToCentroid.getSize(0) *
topQueryToCentroid.getSize(1),
pqCentroids.getSize(2)});
runSumAlongRows(residualNorms, residualDistanceView2, false, stream);
Tensor<float, 4, true> outCodeDistancesF;
DeviceTensor<float, 4, true> outCodeDistancesFloatMem;
if (useFloat16Lookup) {
outCodeDistancesFloatMem = DeviceTensor<float, 4, true>(
mem, {outCodeDistances.getSize(0),
outCodeDistances.getSize(1),
outCodeDistances.getSize(2),
outCodeDistances.getSize(3)},
stream);
outCodeDistancesF = outCodeDistancesFloatMem;
} else {
outCodeDistancesF = outCodeDistances.toTensor<float>();
}
// Transpose -2(sub q)(q * c)(code) to -2(q * c)(sub q)(code) (which
// is where we build our output distances)
auto outCodeDistancesView = outCodeDistancesF.view<3>(
{topQueryToCentroid.getSize(0) * topQueryToCentroid.getSize(1),
outCodeDistances.getSize(2),
outCodeDistances.getSize(3)});
runTransposeAny(residualDistance, 0, 1, outCodeDistancesView, stream);
// Calculate code norms per each sub-dim
// (sub q)(sub dim)(code) is pqCentroids
// transpose to (sub q)(code)(sub dim)
DeviceTensor<float, 3, true> pqCentroidsTranspose(
mem,
{pqCentroids.getSize(0), pqCentroids.getSize(2), pqCentroids.getSize(1)},
stream);
runTransposeAny(pqCentroids, 1, 2, pqCentroidsTranspose, stream);
auto pqCentroidsTransposeView = pqCentroidsTranspose.view<2>(
{pqCentroids.getSize(0) * pqCentroids.getSize(2),
pqCentroids.getSize(1)});
DeviceTensor<float, 1, true> pqCentroidsNorm(
mem,
{pqCentroids.getSize(0) * pqCentroids.getSize(2)},
stream);
runL2Norm(pqCentroidsTransposeView, true, pqCentroidsNorm, true, stream);
// View output as (q * c)(sub q * code), and add centroid norm to
// each row
auto outDistancesCodeViewCols = outCodeDistancesView.view<2>(
{topQueryToCentroid.getSize(0) * topQueryToCentroid.getSize(1),
outCodeDistances.getSize(2) * outCodeDistances.getSize(3)});
runSumAlongColumns(pqCentroidsNorm, outDistancesCodeViewCols, stream);
if (useFloat16Lookup) {
// Need to convert back
auto outCodeDistancesH = outCodeDistances.toTensor<half>();
convertTensor<float, half, 4>(stream,
outCodeDistancesF,
outCodeDistancesH);
}
}
template <typename CentroidT>
void
runPQCodeDistances(Tensor<float, 3, true>& pqCentroids,
Tensor<float, 2, true>& queries,
Tensor<CentroidT, 2, true>& coarseCentroids,
Tensor<int, 2, true>& topQueryToCentroid,
NoTypeTensor<4, true>& outCodeDistances,
bool l2Distance,
bool useFloat16Lookup,
cudaStream_t stream) {
const auto numSubQuantizers = pqCentroids.getSize(0);
const auto dimsPerSubQuantizer = pqCentroids.getSize(1);
const auto codesPerSubQuantizer = pqCentroids.getSize(2);
// FIXME: tune
// Reuse of pq centroid data is based on both # of queries * nprobe,
// and we should really be tiling in both dimensions
constexpr int kQueriesPerBlock = 8;
auto grid = dim3(utils::divUp(queries.getSize(0), kQueriesPerBlock),
numSubQuantizers);
// Reserve one block of threads for double buffering
// FIXME: probably impractical for large # of dims?
auto loadingThreads = utils::roundUp(dimsPerSubQuantizer, kWarpSize);
auto block = dim3(codesPerSubQuantizer + loadingThreads);
auto smem = (3 * dimsPerSubQuantizer) * sizeof(float)
+ topQueryToCentroid.getSize(1) * sizeof(int);
#define RUN_CODE(DIMS, L2) \
do { \
if (useFloat16Lookup) { \
auto outCodeDistancesT = outCodeDistances.toTensor<half>(); \
\
pqCodeDistances<half, CentroidT, DIMS, L2><<<grid, block, smem, stream>>>( \
queries, kQueriesPerBlock, \
coarseCentroids, pqCentroids, \
topQueryToCentroid, outCodeDistancesT); \
} else { \
auto outCodeDistancesT = outCodeDistances.toTensor<float>(); \
\
pqCodeDistances<float, CentroidT, DIMS, L2><<<grid, block, smem, stream>>>( \
queries, kQueriesPerBlock, \
coarseCentroids, pqCentroids, \
topQueryToCentroid, outCodeDistancesT); \
} \
} while (0)
#define CODE_L2(DIMS) \
do { \
if (l2Distance) { \
RUN_CODE(DIMS, true); \
} else { \
RUN_CODE(DIMS, false); \
} \
} while (0)
switch (dimsPerSubQuantizer) {
case 1:
CODE_L2(1);
break;
case 2:
CODE_L2(2);
break;
case 3:
CODE_L2(3);
break;
case 4:
CODE_L2(4);
break;
case 6:
CODE_L2(6);
break;
case 8:
CODE_L2(8);
break;
case 10:
CODE_L2(10);
break;
case 12:
CODE_L2(12);
break;
case 16:
CODE_L2(16);
break;
case 20:
CODE_L2(20);
break;
case 24:
CODE_L2(24);
break;
case 28:
CODE_L2(28);
break;
case 32:
CODE_L2(32);
break;
// FIXME: larger sizes require too many registers - we need the
// MM implementation working
default:
FAISS_THROW_MSG("Too many dimensions (>32) per subquantizer "
"not currently supported");
}
#undef RUN_CODE
#undef CODE_L2
CUDA_TEST_ERROR();
}
} } // namespace