forked from pytorch/FBGEMM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
EmbeddingSpMDMNBitRowWiseSparseBenchmark.cc
405 lines (369 loc) · 12 KB
/
EmbeddingSpMDMNBitRowWiseSparseBenchmark.cc
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
/*
* Copyright (c) Meta Platforms, Inc. and affiliates.
* All rights reserved.
*
* This source code is licensed under the BSD-style license found in the
* LICENSE file in the root directory of this source tree.
*/
#if defined(__x86_64__) || defined(__i386__) || \
(defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86)))
#include <immintrin.h>
#endif
#include <algorithm>
#include <cassert>
#include <chrono>
#include <cmath>
#include <cstdint>
#include <iomanip>
#include <iostream>
#include <map>
#include <numeric>
#include <random>
#include <set>
#include <vector>
#include "./BenchUtils.h"
#include "fbgemm/Fbgemm.h"
#include "fbgemm/FbgemmConvert.h"
#include "src/RefImplementations.h"
using namespace std;
using namespace fbgemm;
void print_fused_table(int rows, int embedding_dim, const uint8_t* table) {
for (int i = 0; i < rows; i++) {
std::cout << "row: " << i << " : " << std::endl;
for (int ii = 0; ii < embedding_dim; ii++) {
std::cout << (int)table[i * (embedding_dim + 2 * sizeof(float)) + ii]
<< ",";
}
std::cout << std::endl;
}
}
static vector<vector<int>> GetInputs_() {
vector<vector<int>> input_dims = {
// batch size, number of rows of table, emb dim , avg lengthl
// TODO: Add more inputs
// Use these -- but they are slow.
{10, 4000000, 32, 100},
{10, 4000000, 64, 100},
{10, 4000000, 128, 100},
{10, 4000000, 256, 100},
// Use these for debugging
// {2, 16, 128, 10},
// {10, 4000, 128, 100},
// {10, 4000, 128, 100},
// {10, 4000, 128, 100},
};
return input_dims;
}
int run_benchmark(
int bit_rate,
int batch_size,
int num_rows,
int embedding_dim,
int average_len,
bool normalize_by_lengths,
bool use_32_bit_indices = false,
bool prefetch = false) {
// Generate mapping table
default_random_engine generator;
constexpr float sparsity = 0.7;
vector<int32_t> mapping_table(num_rows);
bernoulli_distribution row_prune_dist(sparsity);
int num_compressed_rows = 0;
for (int i = 0; i < num_rows; ++i) {
if (row_prune_dist(generator)) {
// pruned
mapping_table[i] = -1;
} else {
mapping_table[i] = num_compressed_rows;
++num_compressed_rows;
}
}
// Create embedding table
int num_elem_per_byte = 8 / bit_rate;
int fused_embedding_dim =
(embedding_dim + num_elem_per_byte - 1) / num_elem_per_byte +
2 * sizeof(float16);
normal_distribution<float> embedding_distribution;
vector<uint8_t> fused_embedding_table(
num_compressed_rows * fused_embedding_dim);
for (int i = 0; i < num_compressed_rows; i++) {
for (int ii = 0;
ii < (embedding_dim + num_elem_per_byte - 1) / num_elem_per_byte;
ii++) {
fused_embedding_table[i * fused_embedding_dim + ii] = 2;
}
float16* scale_bias = reinterpret_cast<float16*>(
&fused_embedding_table[i * fused_embedding_dim] +
(embedding_dim + num_elem_per_byte - 1) / num_elem_per_byte);
float scale = 2.0f;
float bias = 1.0f;
FloatToFloat16_ref(&scale, scale_bias, 1, true /* clip */);
FloatToFloat16_ref(&bias, scale_bias + 1, 1, true /* clip */);
}
// print_fused_table(num_rows, embedding_dim, fused_embedding_table);
// Generate lengths
uniform_int_distribution<int> length_distribution(
1, std::min(2 * average_len + 1, num_rows));
vector<int> offsets(batch_size + 1);
offsets[0] = 0;
for (int i = 0; i < batch_size; ++i) {
offsets[i + 1] = offsets[i] + length_distribution(generator);
}
// Compute the number of indices
int lengths_sum = offsets[batch_size];
// Generate indices
vector<int64_t> indices;
vector<int32_t> indices_32;
vector<int> container(num_rows);
map<int64_t, set<int>> dedup_map; // index -> set(output index)
// please note we generate unique indices
for (int i = 0; i < batch_size; ++i) {
iota(container.begin(), container.end(), 0);
shuffle(container.begin(), container.end(), generator);
copy(
container.begin(),
container.begin() + (offsets[i + 1] - offsets[i]),
back_inserter(indices));
}
copy(begin(indices), end(indices), back_inserter(indices_32));
// Compute the number of valid indices
int num_valid_indices = 0;
for (int index : indices) {
if (mapping_table[index] != -1) {
++num_valid_indices;
}
}
cout << "lengths_sum " << lengths_sum << " num_valid_indices "
<< num_valid_indices << endl;
// Generate weights
vector<float> weights(lengths_sum);
for (int i = 0; i < lengths_sum; ++i) {
weights[i] = embedding_distribution(generator);
}
vector<float> output_sls_ref(batch_size * embedding_dim);
vector<float> output_slws_ref(output_sls_ref.size()),
output_sls(output_sls_ref.size()), output_slws(output_sls_ref.size());
constexpr int NUM_WARMUP = 4;
constexpr int NUM_ITER = 10;
// Only counts the number of bytes for reading embedding table and ignore
// others. Should be good enough as long as embdding_dim is big enough.
constexpr int CACHE_LINE_SIZE = 64;
double bytes = lengths_sum * 2 *
(use_32_bit_indices ? sizeof(int32_t) : sizeof(int64_t)) +
num_valid_indices * fused_embedding_dim;
double bytes_padded = lengths_sum *
((use_32_bit_indices ? sizeof(int32_t) : sizeof(int64_t)) +
CACHE_LINE_SIZE) +
num_valid_indices * CACHE_LINE_SIZE *
static_cast<int>(
(fused_embedding_dim + CACHE_LINE_SIZE - 1) / CACHE_LINE_SIZE);
for (bool has_weight : {false, true}) {
vector<float>& output_ref = has_weight ? output_slws_ref : output_sls_ref;
bool success = false, success_ref = false;
for (int i = 0; i < NUM_WARMUP + NUM_ITER; ++i) {
if (use_32_bit_indices) {
success_ref = EmbeddingSpMDMNBitRowWiseSparse_ref(
bit_rate,
embedding_dim,
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
indices_32.data(),
mapping_table.data(),
offsets.data(),
has_weight ? weights.data() : nullptr,
normalize_by_lengths,
output_ref.data());
} else {
success_ref = EmbeddingSpMDMNBitRowWiseSparse_ref(
bit_rate,
embedding_dim,
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
indices.data(),
mapping_table.data(),
offsets.data(),
has_weight ? weights.data() : nullptr,
normalize_by_lengths,
output_ref.data());
}
}
vector<float>& output = has_weight ? output_slws : output_sls;
auto kernel_32 = GenerateEmbeddingSpMDMNBitRowWiseSparse<int32_t>(
bit_rate,
embedding_dim,
has_weight,
normalize_by_lengths,
prefetch ? 16 : 0);
auto kernel_64 = GenerateEmbeddingSpMDMNBitRowWiseSparse<int64_t>(
bit_rate,
embedding_dim,
has_weight,
normalize_by_lengths,
prefetch ? 16 : 0);
for (bool flush_cache : {false, true}) {
double t = measureWithWarmup(
[&]() {
if (use_32_bit_indices) {
success = kernel_32(
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
indices_32.data(),
offsets.data(),
has_weight ? weights.data() : nullptr,
output.data(),
mapping_table.data());
} else {
success = kernel_64(
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
indices.data(),
offsets.data(),
has_weight ? weights.data() : nullptr,
output.data(),
mapping_table.data());
}
},
NUM_WARMUP,
NUM_ITER,
[&]() {
if (flush_cache) {
cache_evict(fused_embedding_table);
cache_evict(indices);
cache_evict(indices_32);
cache_evict(offsets);
cache_evict(weights);
cache_evict(output);
}
});
// printMatrix(
// matrix_op_t::NoTranspose,
// output.data(),
// batch_size,
// embedding_dim,
// embedding_dim,
// "");
// printMatrix(
// matrix_op_t::NoTranspose,
// output_ref.data(),
// batch_size,
// embedding_dim,
// embedding_dim,
// "");
// Check correctness
if (!flush_cache) {
if (success != success_ref) {
assert(
false && "ERROR: refernce impl and JIT imp did not both succeed");
} else if (success) {
for (size_t i = 0; i < output.size(); ++i) {
assert(fabs(output[i] - output_ref[i]) < 1e-3);
if (fabs(output[i] - output_ref[i]) >= 1e-3) {
cout << i << " " << output[i] << " " << output_ref[i] << endl;
}
}
}
}
if (has_weight) {
cout << setw(16) << "SLW(WEIGHTED) ";
} else {
cout << setw(16) << "SLS ";
}
if (flush_cache) {
cout << setw(20) << "cache flushed";
} else {
cout << setw(20) << "cache not flushed";
}
if (prefetch) {
cout << setw(16) << "prefetch on";
} else {
cout << setw(16) << "prefetch off";
}
cout << setw(8) << "b/w" << setw(10) << bytes / 1e9 / t << " GB/s"
<< setw(20) << "effective b/w: " << setw(16)
<< bytes_padded / 1e9 / t << "GB/s" << setw(8) << " time "
<< setw(16) << t << endl;
} // flush_cache
} // has_weight
return 0;
}
int main() {
int batch_size;
int num_rows;
int embedding_dim;
int average_len;
vector<vector<int>> inputs(GetInputs_());
for (int bit_rate : {2, 4}) {
for (auto& input : inputs) {
assert(input.size() > 3);
batch_size = input[0];
num_rows = input[1];
embedding_dim = input[2];
average_len = input[3];
cout << "bit_rate" << setw(6) << bit_rate << "batch size" << setw(6)
<< batch_size << setw(10) << "num rows" << setw(16) << num_rows
<< setw(10) << "emb dim" << setw(6) << embedding_dim << setw(16)
<< "avg length" << setw(6) << average_len << endl;
// args: batch sz, num rows, emb dim, avg len, normalize, use 32b,
// prefetch
cout << "64 bit indices, ";
run_benchmark(
bit_rate,
batch_size,
num_rows,
embedding_dim,
average_len,
false); // normalize_by_lengths
cout << "64 bit indices with prefetching, ";
run_benchmark(
bit_rate,
batch_size,
num_rows,
embedding_dim,
average_len,
false, // normalize_by_lengths
false, // use_32_bit_indices
true); // prefetch
cout << "32 bit indices, ";
run_benchmark(
bit_rate,
batch_size,
num_rows,
embedding_dim,
average_len,
false, // normalize_by_lengths
true); // use_32_bit_indices
cout << "32 bit indices with prefetching, ";
run_benchmark(
bit_rate,
batch_size,
num_rows,
embedding_dim,
average_len,
false, // normalize_by_lengths
true, // use_32_bit_indices
true); // prefetch
// running with normalize by lengths
// run_benchmark(batch_size, num_rows, embedding_dim, average_len,
// true); run_benchmark(
// batch_size, num_rows, embedding_dim, average_len, true,
// true);
// run_benchmark(
// batch_size,
// num_rows,
// embedding_dim,
// average_len,
// false,
// true,
// true);
}
}
return 0;
}