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
|
#include <metal_stdlib>
using namespace metal;
METAL_FUNC uint get_strided_index(
uint idx,
constant size_t &num_dims,
constant size_t *dims,
constant size_t *strides
) {
uint strided_i = 0;
for (uint d = 0; d < num_dims; d++) {
uint dim_idx = num_dims - 1 - d;
strided_i += (idx % dims[dim_idx]) * strides[dim_idx];
idx /= dims[dim_idx];
}
return strided_i;
}
template<typename TYPENAME, typename INDEX_TYPENAME>
METAL_FUNC void index(
constant size_t &dst_size,
constant size_t &left_size,
constant size_t &src_dim_size,
constant size_t &right_size,
constant size_t &ids_size,
constant bool &contiguous,
constant size_t *src_dims,
constant size_t *src_strides,
const device TYPENAME *input,
const device INDEX_TYPENAME *input_ids,
device TYPENAME *output,
uint tid [[ thread_position_in_grid ]]
) {
if (tid >= dst_size) {
return;
}
const size_t id_i = (tid / right_size) % ids_size;
const INDEX_TYPENAME input_i = min(input_ids[id_i], (INDEX_TYPENAME)(src_dim_size - 1));
const size_t right_rank_i = tid % right_size;
const size_t left_rank_i = tid / right_size / ids_size;
/*
// Force prevent out of bounds indexing
// since there doesn't seem to be a good way to force crash
// No need to check for zero we're only allowing unsized.
*/
const size_t src_i = left_rank_i * src_dim_size * right_size + input_i * right_size + right_rank_i;
const size_t strided_src_i = contiguous ? src_i : get_strided_index(src_i, src_dim_size, src_dims, src_strides);
output[tid] = input[strided_src_i];
}
# define INDEX_OP(NAME, INDEX_TYPENAME, TYPENAME) \
kernel void NAME( \
constant size_t &dst_size, \
constant size_t &left_size, \
constant size_t &src_dim_size, \
constant size_t &right_size, \
constant size_t &ids_size, \
constant bool &contiguous, \
constant size_t *src_dims, \
constant size_t *src_strides, \
const device TYPENAME *input, \
const device INDEX_TYPENAME *input_ids, \
device TYPENAME *output, \
uint tid [[ thread_position_in_grid ]] \
) { \
index<TYPENAME, INDEX_TYPENAME>(dst_size, left_size, src_dim_size, right_size, ids_size, contiguous, src_dims, src_strides, input, input_ids, output, tid); \
}
template<typename TYPENAME, typename INDEX_TYPENAME>
METAL_FUNC void gather(
constant size_t &dst_size,
constant size_t &left_size,
constant size_t &src_dim_size,
constant size_t &right_size,
constant size_t &ids_size,
const device TYPENAME *input,
const device INDEX_TYPENAME *input_ids,
device TYPENAME *output,
uint tid [[ thread_position_in_grid ]]
) {
if (tid >= dst_size) {
return;
}
const INDEX_TYPENAME input_i = input_ids[tid];
const size_t right_rank_i = tid % right_size;
const size_t left_rank_i = tid / right_size / ids_size;
const size_t src_i = (left_rank_i * src_dim_size + input_i) * right_size + right_rank_i;
output[tid] = input[src_i];
}
# define GATHER_OP(NAME, INDEX_TYPENAME, TYPENAME) \
kernel void NAME( \
constant size_t &dst_size, \
constant size_t &left_size, \
constant size_t &src_dim_size, \
constant size_t &right_size, \
constant size_t &ids_size, \
const device TYPENAME *input, \
const device INDEX_TYPENAME *input_ids, \
device TYPENAME *output, \
uint tid [[ thread_position_in_grid ]] \
) { \
gather<TYPENAME, INDEX_TYPENAME>(dst_size, left_size, src_dim_size, right_size, ids_size, input, input_ids, output, tid); \
}
template<typename TYPENAME, typename INDEX_TYPENAME>
METAL_FUNC void scatter_add(
constant size_t &dst_size,
constant size_t &left_size,
constant size_t &src_dim_size,
constant size_t &right_size,
constant size_t &dst_dim_size,
const device TYPENAME *input,
const device INDEX_TYPENAME *input_ids,
device TYPENAME *output,
uint tid [[ thread_position_in_grid ]]
) {
if (tid >= dst_size) {
return;
}
const size_t right_rank_i = tid % right_size;
const size_t left_rank_i = tid / right_size;
for (unsigned int j = 0; j < src_dim_size; ++j) {
const size_t src_i = (left_rank_i * src_dim_size + j) * right_size + right_rank_i;
const INDEX_TYPENAME idx = input_ids[src_i];
const size_t dst_i = (left_rank_i * dst_dim_size + idx) * right_size + right_rank_i;
output[dst_i] += input[src_i];
}
}
# define SCATTER_ADD_OP(NAME, INDEX_TYPENAME, TYPENAME) \
kernel void NAME( \
constant size_t &dst_size, \
constant size_t &left_size, \
constant size_t &src_dim_size, \
constant size_t &right_size, \
constant size_t &dst_dim_size, \
const device TYPENAME *input, \
const device INDEX_TYPENAME *input_ids, \
device TYPENAME *output, \
uint tid [[ thread_position_in_grid ]] \
) { \
scatter_add<TYPENAME, INDEX_TYPENAME>(dst_size, left_size, src_dim_size, right_size, dst_dim_size, input, input_ids, output, tid); \
}
template<typename TYPENAME, typename INDEX_TYPENAME>
METAL_FUNC void index_add(
constant size_t &dst_size,
constant size_t &left_size,
constant size_t &src_dim_size,
constant size_t &right_size,
constant size_t &dst_dim_size,
constant size_t &ids_dim_size,
const device TYPENAME *input,
const device INDEX_TYPENAME *input_ids,
device TYPENAME *output,
uint tid [[ thread_position_in_grid ]]
) {
if (tid >= dst_size) {
return;
}
const size_t right_rank_i = tid % right_size;
const size_t left_rank_i = tid / right_size;
for (unsigned int j = 0; j < ids_dim_size; ++j) {
const INDEX_TYPENAME idx = input_ids[j];
const size_t src_i = (left_rank_i * src_dim_size + j) * right_size + right_rank_i;
const size_t dst_i = (left_rank_i * dst_dim_size + idx) * right_size + right_rank_i;
output[dst_i] += input[src_i];
}
}
# define INDEX_ADD_OP(NAME, INDEX_TYPENAME, TYPENAME) \
kernel void NAME( \
constant size_t &dst_size, \
constant size_t &left_size, \
constant size_t &src_dim_size, \
constant size_t &right_size, \
constant size_t &dst_dim_size, \
constant size_t &ids_dim_size, \
const device TYPENAME *input, \
const device INDEX_TYPENAME *input_ids, \
device TYPENAME *output, \
uint tid [[ thread_position_in_grid ]] \
) { \
index_add<TYPENAME, INDEX_TYPENAME>(dst_size, left_size, src_dim_size, right_size, dst_dim_size, ids_dim_size, input, input_ids, output, tid); \
}
INDEX_OP(is_i64_f32, int64_t, float)
INDEX_OP(is_i64_f16, int64_t, half)
#if defined(__HAVE_BFLOAT__)
INDEX_OP(is_i64_bf16, int64_t, bfloat)
#endif
INDEX_OP(is_u32_u8, uint32_t, uint8_t)
INDEX_OP(is_u32_u32, uint32_t, uint32_t)
INDEX_OP(is_u32_f32, uint32_t, float)
INDEX_OP(is_u32_f16, uint32_t, half)
#if defined(__HAVE_BFLOAT__)
INDEX_OP(is_u32_bf16, uint32_t, bfloat)
#endif
INDEX_OP(is_u8_u8, uint8_t, uint8_t)
INDEX_OP(is_u8_u32, uint8_t, uint32_t)
INDEX_OP(is_u8_f32, uint8_t, float)
INDEX_OP(is_u8_f16, uint8_t, half)
#if defined(__HAVE_BFLOAT__)
INDEX_OP(is_u8_bf16, uint8_t, bfloat)
#endif
GATHER_OP(gather_u32_f32, uint, float)
GATHER_OP(gather_u32_f16, uint, half)
#if defined(__HAVE_BFLOAT__)
GATHER_OP(gather_u32_bf16, uint, bfloat)
#endif
GATHER_OP(gather_u32_u32, uint, uint)
SCATTER_ADD_OP(sa_u32_f32, uint32_t, float)
SCATTER_ADD_OP(sa_u8_f32, uint8_t, float)
SCATTER_ADD_OP(sa_i64_f32, int64_t, float)
SCATTER_ADD_OP(sa_u32_u32, uint32_t, uint32_t)
SCATTER_ADD_OP(sa_u32_f16, uint32_t, half)
SCATTER_ADD_OP(sa_u8_f16, uint8_t, half)
SCATTER_ADD_OP(sa_i64_f16, int64_t, half)
#if defined(__HAVE_BFLOAT__)
SCATTER_ADD_OP(sa_u32_bf16, uint32_t, bfloat)
SCATTER_ADD_OP(sa_u8_bf16, uint8_t, bfloat)
SCATTER_ADD_OP(sa_i64_bf16, int64_t, bfloat)
#endif
// i64
INDEX_ADD_OP(ia_i64_f16, int64_t, half)
INDEX_ADD_OP(ia_i64_f32, int64_t, float)
INDEX_ADD_OP(ia_i64_i64, int64_t, int64_t)
INDEX_ADD_OP(ia_i64_u32, int64_t, uint32_t)
INDEX_ADD_OP(ia_i64_u8, int64_t, uint8_t)
#if defined(__HAVE_BFLOAT__)
INDEX_ADD_OP(ia_i64_bf16, int64_t, bfloat)
#endif
// u32
INDEX_ADD_OP(ia_u32_f16, uint32_t, half)
INDEX_ADD_OP(ia_u32_f32, uint32_t, float)
INDEX_ADD_OP(ia_u32_i64, uint32_t, int64_t)
INDEX_ADD_OP(ia_u32_u32, uint32_t, uint32_t)
INDEX_ADD_OP(ia_u32_u8, uint32_t, uint8_t)
#if defined(__HAVE_BFLOAT__)
INDEX_ADD_OP(ia_u32_bf16, uint32_t, bfloat)
#endif
// u8
INDEX_ADD_OP(ia_u8_f16, uint8_t, half)
INDEX_ADD_OP(ia_u8_f32, uint8_t, float)
INDEX_ADD_OP(ia_u8_i64, uint8_t, int64_t)
INDEX_ADD_OP(ia_u8_u32, uint8_t, uint32_t)
INDEX_ADD_OP(ia_u8_u8, uint8_t, uint8_t)
#if defined(__HAVE_BFLOAT__)
INDEX_ADD_OP(ia_u8_bf16, uint8_t, bfloat)
#endif
|
