File size: 12,524 Bytes
d1d4335
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
/*

 * 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.

 */

#pragma once
#include <cstdint>
#include <functional>

#include "fbgemm/FbgemmBuild.h"

namespace fbgemm {

template <
    typename InType,
    typename IndexType,
    typename OffsetType = std::int32_t,
    typename OutType = float>
class EmbeddingSpMDMKernelSignature {
 public:
  /**

   * Behavior is as the follow pseudocode

   * (when use_offsets == true, lengths[i] == offsets[i + 1] - offsets[i])

   * (when is_weight_positional == true, use weights[j - offsets[i]] instead of

   *  weights[j])

   *

   * for i in range(output_size):

   *  out[i * block_size : (i + 1) * block_size] = 0

   *  for j in range(offsets[i], offsets[i + 1]):

   *   for k in range(block_size):

   *    out[i * block_size + k] += input[indices[j] * block_size + k] *

   *                               weights ? weights[j] : 1;

   *  if normalize_weights and lengths[i] > 0:

   *   out[i * block_size : (i + 1) * block_size] /= lengths[i]

   *

   * @param data_size the number of rows in embedding table

   */
  using Type = std::function<bool(
      std::int64_t output_size,
      std::int64_t index_size,
      std::int64_t data_size,
      const InType* input,
      const IndexType* indices,
      const OffsetType* offsets_or_lengths,
      const float* weights, // optional, can be null for non-weighted sum
      OutType* out)>;
};

/**

 * @tparam InType can be float, float16, or uint8_t

 * @tparam IndexType can be int32_t or int64_t

 * @tparam IndexType can be int32_t or int64_t

 *

 * @param use_offsets If true, the generated code assumes we will pass offsets

 *                    instead of lengths that confirms PyTorch EmbeddingBag

 *                    interface. In this case, the length of offsets array

 *                    should be output_size + 1 and offsets[output_size] should

 *                    be index_size.

 *                    If false, the generate code assumes we will pass lengths

 *                    that confirms Caffe2 SparseLengthsSum interface.

 */
template <
    typename InType,
    typename IndexType,
    typename OffsetType = std::int32_t,
    typename OutType = float,
    bool THREAD_LOCAL = false>
FBGEMM_API typename EmbeddingSpMDMKernelSignature<
    InType,
    IndexType,
    OffsetType,
    OutType>::Type
GenerateEmbeddingSpMDM(
    const std::int64_t block_size,
    bool has_weight,
    bool normalize_by_lengths,
    int prefetch = 16,
    bool is_weight_positional = false,
    bool use_offsets = true,
    bool is_bf16_out = false,
    bool is_bf16_in = false);

/**

 * @param output_stride If -1, output_stride is same as block_size

 * @param input_stride If -1, input_stride is same as block_size

 * @param scale_bias_last if false, scale and bias appear at the beginning

 *        of each row and are in fp16 for table batched embedding (TBE)

 *        in FBGEMM_GPU. If false, it can also take -1 indices (output from

 *        pruned embedding id mapping)

 */
template <
    typename InType,
    typename IndexType,
    typename OffsetType = std::int32_t,
    typename OutType = float,
    bool THREAD_LOCAL = false>
FBGEMM_API typename EmbeddingSpMDMKernelSignature<
    InType,
    IndexType,
    OffsetType,
    OutType>::Type
GenerateEmbeddingSpMDMWithStrides(
    const std::int64_t block_size,
    bool has_weight,
    bool normalize_by_lengths,
    int prefetch = 16,
    bool is_weight_positional = false,
    bool use_offsets = true,
    std::int64_t output_stride = -1,
    std::int64_t input_stride = -1,
    bool scale_bias_last = true,
    bool no_bag = false,
    bool is_bf16_out = false,
    bool is_bf16_in = false);

/**

 * @tparam IndexType can be int32_t or int64_t

 * @tparam OffsetType can be int32_t or int64_t

 * @param bit_rate can be 2 or 4

 */
template <
    typename IndexType,
    typename OffsetType = std::int32_t,
    typename OutType = float>
FBGEMM_API typename EmbeddingSpMDMKernelSignature<
    std::uint8_t,
    IndexType,
    OffsetType,
    OutType>::Type
GenerateEmbeddingSpMDMNBit(
    int bit_rate,
    const std::int64_t block_size,
    bool has_weight,
    bool normalize_by_lengths,
    int prefetch = 16,
    bool is_weight_positional = false,
    bool use_offsets = true);

/**

 * @param output_stride If -1, output_stride is same as block_size

 * @param input_stride in Bytes. If -1, input_stride is same as

 *                     block_size / num_elem_per_byte + 2 * sizeof(float16)

 * @param scale_bias_last if false, scale and bias appear at the beginning

 *        of each row and are in fp16 for table batched embedding (TBE)

 *        in FBGEMM_GPU. If false, it can also take -1 indices (output from

 *        pruned embedding id mapping)

 */
template <
    typename IndexType,
    typename OffsetType = std::int32_t,
    typename OutType = float,
    bool THREAD_LOCAL = false>
FBGEMM_API typename EmbeddingSpMDMKernelSignature<
    std::uint8_t,
    IndexType,
    OffsetType,
    OutType>::Type
GenerateEmbeddingSpMDMNBitWithStrides(
    const int input_bit_rate,
    const std::int64_t block_size,
    bool has_weight,
    bool normalize_by_lengths,
    int prefetch = 16,
    bool is_weight_positional = false,
    bool use_offsets = true,
    std::int64_t output_stride = -1,
    std::int64_t input_stride = -1,
    bool scale_bias_last = true,
    const bool is_bf16_out = false,
    const bool no_bag = false,
    int output_bit_rate = -1);

/**

 * @param output_stride If -1, output_stride is same as block_size

 * @param input_stride in Bytes. If -1, input_stride is same as

 *                     block_size / num_elem_per_byte + 2 * sizeof(float16)

 * @param exponent_bits is the number of exponent bits in the FP8 encode

 *                      (normally 4 or 5)

 * @param exponent_bias is subtracted from the exponent to obtain the actual

 *                      exponent for the floating-point number

 */
template <
    typename IndexType,
    typename OffsetType = std::int32_t,
    typename OutType = float>
FBGEMM_API typename EmbeddingSpMDMKernelSignature<
    std::uint8_t,
    IndexType,
    OffsetType,
    OutType>::Type
GenerateEmbeddingSpMDMFP8WithStrides(
    const std::int64_t block_size,
    bool normalize_by_lengths,
    bool is_weight_positional = false,
    bool use_offsets = true,
    std::int64_t output_stride = -1,
    std::int64_t input_stride = -1,
    int exponent_bits = 4,
    int exponent_bias = 7,
    bool is_bf16_out = false);

template <
    typename InType,
    typename IndexType,
    typename OffsetType = std::int32_t>
class EmbeddingSpMDMRowWiseSparseKernelSignature {
 public:
  using Type = std::function<bool(
      std::int64_t output_size,
      std::int64_t index_size,
      std::int64_t uncompressed_data_size,
      // TODO: add compressed_data_size and check array bound
      const InType* input,
      const IndexType* indices,
      const OffsetType* offsets_or_lengths,
      const float* weights, // optional, can be null for non-weighted sum
      float* out,
      const std::int32_t* compressed_indices_table)>;
};

/**

 * @tparam InType can be float, float16, or uint8_t

 * @tparam IndexType can be int32_t or int64_t

 * @tparam OffsetType can be int32_t or int64_t

 */
template <
    typename InType,
    typename IndexType,
    typename OffsetType = std::int32_t>
FBGEMM_API typename EmbeddingSpMDMRowWiseSparseKernelSignature<
    InType,
    IndexType,
    OffsetType>::Type
GenerateEmbeddingSpMDMRowWiseSparse(
    const std::int64_t block_size,
    bool has_weight,
    bool normalize_by_lengths,
    int prefetch = 16,
    bool is_weight_positional = false,
    bool use_offsets = true);

/**

 * @tparam IndexType can be int32_t or int64_t

 * @tparam OffsetType can be int32_t or int64_t

 * @param bit_rate can be 2 or 4

 */
template <typename IndexType, typename OffsetType = std::int32_t>
FBGEMM_API typename EmbeddingSpMDMRowWiseSparseKernelSignature<
    std::uint8_t,
    IndexType,
    OffsetType>::Type
GenerateEmbeddingSpMDMNBitRowWiseSparse(
    int bit_rate,
    const std::int64_t block_size,
    bool has_weight,
    bool normalize_by_lengths,
    int prefetch = 16,
    bool is_weight_positional = false,
    bool use_offsets = true);

/**

 * @return The number of rows processed. If smaller than num_rows, an error

 *         must have happened at the last row processed.

 */
template <typename IndexType>
class SparseAdaGradSignature {
 public:
  using Type = std::function<int(
      int num_rows, // number of rows reading
      std::uint64_t param_size, // total number of parameters
      float* w, // input/output parameters
      const float* g, // input gradients
      float* h, // input/output momentums
      const IndexType* indices, // indices of each row
      float epsilon,
      float lr,
      float weight_decay,
      const double* counter, // used for weight_decay adjusted for frequency
                             // nullptr when frequency adjustment is not used.
                             // ignored when the kernel is generated with
                             // use_weight_decay = false.
      std::int64_t counter_halflife)>; // frequency adjust happens only after
};

template <typename IndexType>
FBGEMM_API typename SparseAdaGradSignature<IndexType>::Type
GenerateSparseAdaGrad(

    int block_size, // number of parameters per row

    bool rowwise = false,

    int prefetch = 16,

    bool use_weight_decay = false);

// RowWiseSparseAdaGrad fused with SLS gradient
// Weights can be either float or float16
template <
    typename IndexType,
    typename OffsetType = std::int32_t,
    typename DataType = float>
class RowWiseSparseAdaGradFusedSignature {
 public:
  using Type = std::function<bool(
      std::int64_t output_size,
      std::int64_t index_size,
      std::int64_t data_size, // number of rows in w
      DataType* w, // input/output parameters
      const float* g, // input gradients
      float* h, // input/output momentums
      const IndexType* indices, // indices of each row
      const OffsetType* offsets_or_lengths,
      float epsilon,
      float lr)>;
};

/**

 * @param grad_stride If -1, grad_stride is same as block size

 */
template <
    typename IndexType,
    typename OffsetType = std::int32_t,
    typename DataType = float>
FBGEMM_API typename RowWiseSparseAdaGradFusedSignature<
    IndexType,
    OffsetType,
    DataType>::Type
GenerateRowWiseSparseAdaGradFused(
    int block_size, // number of parameters per row
    int prefetch = 16,
    bool use_offsets = true,
    bool use_stochastic_rounding = true,
    int grad_stride = -1);

namespace internal {
// Specialization for block size 1 internally called by GenerateEmbeddingSpMDM
template <typename InType, typename IndexType, typename OffsetType>
FBGEMM_API bool EmbeddingSpMDMBlockSize1_(

    const std::int64_t output_size,

    const std::int64_t index_size,

    const std::int64_t data_size, // the number of rows in input

    const InType* input,

    const IndexType* indices,

    const OffsetType* offsets_or_lengths,

    const float* weights, // optional, can be null for non-weighted sum

    bool normalize_by_lengths,

    float* out,

    bool is_weight_positional = false,

    bool use_offsets = true,

    bool is_bf16 = false);

template <typename IndexType, bool HAS_WEIGHTS>
void compressed_indices_remap_avx512(

    std::int32_t offsets_numel,

    const IndexType* indices,

    const int32_t* compressed_indices_mapping,

    const IndexType* offsets,

    const float* weights, // optional, can be null,

    IndexType* out_indices,

    IndexType* out_offsets,

    float* out_weights);

} // namespace internal

template <typename IndexType>
FBGEMM_API void compressed_indices_remap(

    std::int32_t offsets_numel,

    const IndexType* indices,

    const int32_t* compressed_indices_mapping,

    const IndexType* offsets,

    const float* weights, // optional, can be null,

    IndexType* out_indices,

    IndexType* out_offsets,

    float* out_weights);

} // namespace fbgemm