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
|
#include "caffe2/perfkernels/fused_8bit_rowwise_embedding_lookup_idx.h"
#include "caffe2/core/types.h"
#include "caffe2/perfkernels/common.h"
#include "caffe2/utils/cpuid.h"
#include <c10/util/irange.h>
namespace caffe2 {
/**
* Base implementation does runtime dispatch for each segment of reduction
* @return false if there is an out-of-bound error
*/
template <
typename IndexType,
typename InType,
typename OutType,
bool IS_WEIGHT_POSITIONAL = false>
static bool Fused8BitRowwiseEmbeddingLookupGenericSlowIdx(
const int64_t block_size,
const int64_t output_size,
const int64_t index_size,
const int64_t data_size,
const InType* input,
const IndexType* indices,
const IndexType* offsets,
const float* weights, // optional, can be null for sum reducer
bool normalize_by_lengths,
OutType* out) {
// block_size is the number of elements and fused_block_size is the size of
// an entire row, including scale and bias.
const auto scale_bias_offset = 8 / sizeof(InType);
const int64_t fused_block_size = block_size + scale_bias_offset;
int64_t current = 0;
for (const auto m : c10::irange(output_size)) {
memset(out, 0, sizeof(OutType) * block_size);
if (current != offsets[m] - offsets[0]) {
return false;
}
int64_t start_offset = offsets[m];
int64_t end_offset = offsets[m + 1];
int64_t length = end_offset - start_offset;
for (const auto i : c10::irange(start_offset, end_offset)) {
int64_t idx = indices[current];
if (idx < 0 || idx >= data_size) {
return false;
}
#ifdef __GNUC__
if (current + 1 < index_size) {
__builtin_prefetch(
input + fused_block_size * indices[current + 1], 0, 1);
}
#endif // __GNUC__
const float* scale_bias = reinterpret_cast<const float*>(
input + fused_block_size * indices[current] + block_size);
float weight = 1.0f;
if (weights) {
weight = weights[IS_WEIGHT_POSITIONAL ? i : current];
}
const float scale = weight * scale_bias[0];
const float bias = weight * scale_bias[1];
for (const auto j : c10::irange(block_size)) {
out[j] += scale * input[fused_block_size * indices[current] + j] + bias;
}
++current;
}
if (normalize_by_lengths && length) {
float scale = 1.f / length;
for (const auto j : c10::irange(block_size)) {
out[j] *= scale;
}
}
out += block_size;
}
return current == index_size;
}
// clang-format off
// Proxy back to generic implementation
#define FUSED_8BIT_ROWWISE_EMBEDDING_IDX_SPECIALIZATION(IndexType, OutType) \
bool \
Fused8BitRowwiseEmbeddingLookupIdx_##IndexType##_uint8_t_##OutType##_false__base( \
const int64_t block_size, \
const int64_t output_size, \
const int64_t index_size, \
const int64_t data_size, \
const uint8_t* input, \
const IndexType* indices, \
const IndexType* offsets, \
const float* weights, \
bool normalize_by_lengths, \
OutType* out) { \
return Fused8BitRowwiseEmbeddingLookupGenericSlowIdx< \
IndexType, \
uint8_t, \
OutType, \
false>( \
block_size, \
output_size, \
index_size, \
data_size, \
input, \
indices, \
offsets, \
weights, \
normalize_by_lengths, \
out); \
} \
decltype( \
Fused8BitRowwiseEmbeddingLookupIdx_##IndexType##_uint8_t_##OutType##_false__base) \
Fused8BitRowwiseEmbeddingLookupIdx_##IndexType##_uint8_t_##OutType##_false__avx2_fma; \
bool Fused8BitRowwiseEmbeddingLookupIdx_##IndexType##_uint8_t_##OutType( \
const int64_t block_size, \
const int64_t output_size, \
const int64_t index_size, \
const int64_t data_size, \
const uint8_t* input, \
const IndexType* indices, \
const IndexType* offsets, \
const float* weights, \
bool normalize_by_lengths, \
OutType* out) { \
const int32_t one = 1; \
CAFFE_ENFORCE_EQ( \
reinterpret_cast<const uint8_t*>(&one)[0], \
1, \
"Fused8BitRowwiseEmbeddingLookup is not supported on this platform"); \
AVX2_FMA_DO( \
Fused8BitRowwiseEmbeddingLookupIdx_##IndexType##_uint8_t_##OutType##_false, \
block_size, \
output_size, \
index_size, \
data_size, \
input, \
indices, \
offsets, \
weights, \
normalize_by_lengths, \
out); \
BASE_DO( \
Fused8BitRowwiseEmbeddingLookupIdx_##IndexType##_uint8_t_##OutType##_false, \
block_size, \
output_size, \
index_size, \
data_size, \
input, \
indices, \
offsets, \
weights, \
normalize_by_lengths, \
out); \
} \
template <> \
void Fused8BitRowwiseEmbeddingLookupIdx<IndexType, uint8_t, OutType, false>( \
const int64_t block_size, \
const int64_t output_size, \
const int64_t index_size, \
const int64_t data_size, \
const uint8_t* input, \
const IndexType* indices, \
const IndexType* offsets, \
const float* weights, \
bool normalize_by_lengths, \
OutType* out) { \
bool success = \
Fused8BitRowwiseEmbeddingLookupIdx_##IndexType##_uint8_t_##OutType( \
block_size, \
output_size, \
index_size, \
data_size, \
input, \
indices, \
offsets, \
weights, \
normalize_by_lengths, \
out); \
if (success) { \
return; \
} \
int64_t current = 0; \
for (int m = 0; m < output_size; ++m) { \
for (int64_t i = offsets[m]; i < offsets[m + 1]; ++i) { \
CAFFE_ENFORCE_LT(current, index_size); \
IndexType idx = indices[current]; \
CAFFE_ENFORCE( \
0 <= idx && idx < data_size, \
"Index ", \
current, \
" is out of bounds: ", \
idx, \
", range 0 to ", \
data_size); \
++current; \
} \
} \
CAFFE_ENFORCE_EQ( \
current, \
index_size, \
"Your input seems to be incorrect: the sum of lengths values should be " \
"the size of the indices tensor, but it appears not."); \
}
// clang-format on
FUSED_8BIT_ROWWISE_EMBEDDING_IDX_SPECIALIZATION(int32_t, float);
FUSED_8BIT_ROWWISE_EMBEDDING_IDX_SPECIALIZATION(int64_t, float);
#undef FUSED_8BIT_ROWWISE_EMBEDDING_IDX_SPECIALIZATION
} // namespace caffe2
|
