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
|
# mypy: allow-untyped-defs
import argparse
import sys
# Unroll loops when block_size is a multiple of vector length.
def unroll(num_unrolls, IndexType, InType, OutType, use_weights):
def compute(regid, InType, use_weights):
code = []
if InType == "float":
code.append(
f" vsum{regid} =\n"
" svmad_f32_x("
f"svAll, vwgt, svld1_f32(svAll, &ip[{regid} * vLen]),"
f" vsum{regid});"
)
elif InType == "at::Half":
code.append(
f" vsum{regid} = svmad_f32_x(\n"
" svAll,\n"
" vwgt,\n"
" svcvt_f32_f16_x(\n"
" svAll,\n"
" svreinterpret_f16_u32(svld1uh_u32(\n"
" svAll, reinterpret_cast<const uint16_t*>("
f"&ip[{regid} * vLen])))),\n" # noqa
f" vsum{regid});"
)
elif InType == "at::BFloat16":
code.append(
f" vsum{regid} = svmad_f32_x(\n"
" svAll,\n"
" vwgt,\n"
" svreinterpret_f32_u32(svlsl_n_u32_x(\n"
" svAll,\n"
" svld1uh_u32(\n"
" svAll, reinterpret_cast<const uint16_t*>("
f"&ip[{regid} * vLen])),\n"
" 16)),\n" # noqa
f" vsum{regid});"
)
elif InType == "uint8_t":
code.append(
f" vsum{regid} = svmad_f32_x(\n"
" svAll,\n"
" vwgt,\n"
" svcvt_f32_u32_x(svAll,"
f" svld1ub_u32(svAll, &ip[{regid} * vLen])),\n" # noqa
f" svadd_f32_x(svAll, vsum{regid}, vbio));"
)
else:
raise ValueError(f"Unknown datatype \"{InType}\"")
return code
code = []
code.append(f" // unrolling {num_unrolls} times")
code.append(" for (int64_t i = 0; i < output_size; ++i) {")
code.append(" " + OutType + "* const op = &out[i * block_size];")
code.append(
" if (pos != offsets[i] - offsets[0]) {\n"
+ " return false;\n"
+ " }"
)
# Initialise vector sum registers
for i in range(num_unrolls):
code.append(f" svfloat32_t vsum{i} = svdup_n_f32(0);")
# inner loop
code.append("""\
int64_t start_offset = offsets[i];
int64_t end_offset = offsets[i + 1];""")
code.append(
" for ("
+ "int64_t"
+ " j = start_offset; j < end_offset; ++j) {" # noqa
)
code.append(" const auto idx = indices[pos];")
code.append(
" if (idx < 0 || idx >= data_size) {\n"
+ " return false;\n"
+ " }"
)
if InType == "uint8_t":
code.append(" " + OutType + " wgt = 1.f;")
code.append(" " + OutType + " bio{};")
code.append(" if (weights) {")
code.append(
" wgt = weights[IS_WEIGHT_POSITIONAL ? (j - start_offset) : pos];" # noqa
)
code.append(" }")
code.append(" if (scale_bias) {")
code.append(" bio = wgt * scale_bias[2 * idx + 1];")
code.append(" wgt = wgt * scale_bias[2 * idx];")
code.append(" }")
code.append(" svfloat32_t vbio = svdup_n_f32(bio);")
else:
code.append(" " + OutType + " wgt = 1.f;")
code.append(" if (weights) {")
code.append(
" wgt = weights[IS_WEIGHT_POSITIONAL ? (j - start_offset) : pos];" # noqa
)
code.append(" }")
code.append(" const svfloat32_t vwgt = svdup_n_f32(wgt);")
code.append(f" const {InType}* const ip = &input[idx * block_size];")
code.append(" // weight * input + out")
for i in range(num_unrolls):
code.extend(compute(i, InType, use_weights))
code.append(" ++pos;")
code.append(" }")
code.append(" // Normalisation")
code.append(" const int64_t length = end_offset - start_offset;")
code.append(" if (normalize_by_lengths && length != 0) {")
code.append(" const float len_inv = 1.0f / length;")
code.append(" const svfloat32_t vlen_inv = svdup_n_f32(len_inv);")
for i in range(num_unrolls):
code.append(f" svst1_f32(svAll, &op[{i} * vLen],"
+ f" svmul_f32_x(svAll, vsum{i}, vlen_inv));")
code.append(" } else {")
# inv of length
for i in range(num_unrolls):
code.append(f" svst1_f32(svAll, &op[{i} * vLen], vsum{i});")
code.append(" }")
code.append(" }")
return code
# Handle the case where block_size is not a multiple of vector length.
def generic(IndexType, InType, OutType, use_weights):
def compute(InType, use_weights):
code = []
if InType == "float":
code.append(
" svst1_f32(\n"
" pg,\n"
" &op[k],\n"
" svmad_f32_x(\n"
" pg, vwgt, svld1_f32(pg, &ip[k]),"
" svld1_f32(pg, &op[k])));"
)
elif InType == "at::Half":
code.append(
" svst1_f32(\n"
" pg,\n"
" &op[k],\n"
" svmad_f32_x(\n"
" pg,\n"
" vwgt,\n"
" svcvt_f32_f16_x(\n"
" pg,\n"
" svreinterpret_f16_u32(svld1uh_u32(\n"
" pg,"
" reinterpret_cast<const uint16_t*>(&ip[k])))),\n"
" svld1_f32(pg, &op[k])));"
)
elif InType == "at::BFloat16":
code.append(
" svst1_f32(\n"
" pg,\n"
" &op[k],\n"
" svmad_f32_x(\n"
" pg,\n"
" vwgt,\n"
" svreinterpret_f32_u32(svlsl_n_u32_x(\n"
" pg,\n"
" svld1uh_u32(\n"
" pg,"
" reinterpret_cast<const uint16_t*>(&ip[k])),\n"
" 16)),\n"
" svld1_f32(pg, &op[k])));"
)
elif InType == "uint8_t":
code.append(
" svst1_f32(\n"
" pg,\n"
" &op[k],\n"
" svmad_f32_x(\n"
" pg,\n"
" vwgt,\n"
" svcvt_f32_u32_x(pg,"
" svld1ub_u32(pg, &ip[k])),\n" # noqa
" svadd_f32_x(pg,"
" svld1_f32(pg, &op[k]), vbio)));"
)
else:
raise ValueError(f"Unknown datatype \"{InType}\"")
return code
code = []
code.append(
" for (int64_t i = 0; i < output_size; ++i) {"
)
code.append(" " + OutType + "* const op = &out[i * block_size];")
# initialize to 0
code.append(" memset(op, 0, sizeof(float) * block_size);")
# inner loop
code.append(
" if (pos != offsets[i] - offsets[0]) {\n"
+ " return false;\n"
+ " }"
)
code.append(
" int64_t start_offset = offsets[i];\n"
+ " int64_t end_offset = offsets[i + 1];"
)
code.append(
" for ("
+ "int64_t"
+ " j = start_offset; j < end_offset; ++j) {" # noqa
)
code.append(" const auto idx = indices[pos];")
code.append(
" if (idx < 0 || idx >= data_size) {\n"
+ " return false;\n"
+ " }"
)
if InType == "uint8_t":
code.append(" // unimplemented")
code.append(" " + OutType + " wgt = 1.f;")
code.append(" " + OutType + " bio{};")
code.append(" if (weights) {")
code.append(
" wgt = weights[IS_WEIGHT_POSITIONAL ? (j - start_offset) : pos];" # noqa
)
code.append(" }")
code.append(" if (scale_bias) {")
code.append(" bio = wgt * scale_bias[2 * idx + 1];")
code.append(" wgt = wgt * scale_bias[2 * idx];")
code.append(" }")
code.append(" svfloat32_t vbio = svdup_n_f32(bio);")
else:
code.append(" " + OutType + " wgt = 1.f;")
code.append(" if (weights) {")
code.append(
" wgt = weights[IS_WEIGHT_POSITIONAL ? (j - start_offset) : pos];" # noqa
)
code.append(" }")
code.append(" const svfloat32_t vwgt = svdup_n_f32(wgt);")
code.append(f" const {InType}* ip = &input[idx * block_size];")
# compute and store main loop
code.append(" svbool_t pg;")
code.append(" for (int64_t k = 0;")
code.append(" svptest_first(svAll, pg = svwhilelt_b32_s64("
+ "k, block_size));")
code.append(" k += vLen) {")
code.extend(compute(InType, use_weights))
code.append(" }\n")
code.append(" ++pos;")
code.append(" }")
code.append(" const int64_t length = end_offset - start_offset;\n")
code.append(" if (normalize_by_lengths && length != 0) {")
code.append(" const float len_inv = 1.0f / length;")
code.append(" svfloat32_t vlen_inv = svdup_n_f32(len_inv);")
code.append(" svbool_t pg;")
code.append(" for (int64_t j = 0;\n"
" svptest_first(svAll, pg = svwhilelt_b32_s64("
"j, block_size));")
code.append(" j += vLen) {")
code.append(
" svst1_f32(\n"
" pg, &op[j], svmul_f32_x(pg, svld1_f32(pg, &op[j]), vlen_inv));"
)
code.append(" }")
code.append(" }")
code.append(" }")
return code
def main():
parser = argparse.ArgumentParser()
parser.add_argument("-f", "--filename", help="file name")
opts = parser.parse_args()
if opts.filename:
filename = opts.filename
else:
filename = "embedding_lookup_idx_sve.cc"
options = [
["int32_t", "int32_t", "float", "float", "float", "float"],
["int64_t", "int64_t", "float", "float", "float", "float"],
["int32_t", "int32_t", "half", "at::Half", "float", "float"],
["int64_t", "int64_t", "half", "at::Half", "float", "float"],
["int32_t", "int32_t", "bfloat16", "at::BFloat16", "float", "float"],
["int64_t", "int64_t", "bfloat16", "at::BFloat16", "float", "float"],
["int32_t", "int32_t", "uint8_t", "uint8_t", "float", "float"],
["int64_t", "int64_t", "uint8_t", "uint8_t", "float", "float"],
]
code = []
# includes
code.append("//// --------------------------")
code.append("//// ATTENTION:")
code.append("//// THIS CODE IS AUTOGENERATED")
code.append(f"//// BY {' '.join(sys.argv)}")
code.append("//// DO NOT MODIFY!!!")
code.append("//// --------------------------\n")
code.append("#include <arm_sve.h>")
code.append("#include <c10/util/BFloat16.h>")
code.append("#include <c10/util/Half.h>")
code.append("#include <cstdint>")
code.append("#include <cstring>")
code.append("namespace caffe2 {\n")
for o in options:
[IndexTypeName, IndexType, InTypeName, InType, OutTypeName, OutType] = o
code.append("template <bool IS_WEIGHT_POSITIONAL>")
fn_base = f"EmbeddingLookupIdx_{IndexTypeName}_{InTypeName}_{OutTypeName}"
suffix = "__sve"
fn = "static bool " + fn_base + suffix
code.append(fn + "(")
args = []
args.append(" const int64_t block_size,")
args.append(" const int64_t output_size,")
args.append(" const int64_t index_size,")
args.append(" const int64_t data_size,")
args.append(" const " + InType + "* input,")
args.append(" const " + IndexType + "* indices,")
args.append(" const " + IndexType + "* offsets,")
args.append(" const float* weights,")
args.append(" const float* scale_bias,")
args.append(" bool normalize_by_lengths,")
args.append(" " + OutType + "* out) {")
code += args
code.append(" const svbool_t svAll = svptrue_b32();")
code.append(" const auto vLen = static_cast<int64_t>(svcntw());")
code.append(" int64_t pos = 0;")
code.append(" if (block_size == 32 * vLen) {")
code += unroll(32, IndexType, InType, OutType, True)
code.append(" } else if (block_size == 16 * vLen) {")
code += unroll(16, IndexType, InType, OutType, True)
code.append(" } else if (block_size == 8 * vLen) {")
code += unroll(8, IndexType, InType, OutType, True)
code.append(" } else if (block_size == 4 * vLen) {")
code += unroll(4, IndexType, InType, OutType, True)
code.append(" } else if (block_size == 2 * vLen) {")
code += unroll(2, IndexType, InType, OutType, True)
code.append(" } else {")
code.append(" // generic code:")
code += generic(IndexType, InType, OutType, True)
code.append(" }")
code.append(" return pos == index_size;")
code.append("}")
for is_weight_positional in ["false", "true"]:
code.append("bool " + fn_base + "_" + is_weight_positional + suffix + "(")
code += args
# Resolve the Lint warnings: Limit of 80 characters in one line.
extra_space = "\n "
ret_string = " return " + fn_base + suffix \
+ "<" + is_weight_positional + ">("
if len(ret_string) <= 80:
code.append(ret_string)
else:
code.append(" return " + fn_base + suffix + "<" + extra_space + is_weight_positional + ">(")
code.append(" block_size,")
code.append(" output_size,")
code.append(" index_size,")
code.append(" data_size,")
code.append(" input,")
code.append(" indices,")
code.append(" offsets,")
code.append(" weights,")
code.append(" scale_bias,")
code.append(" normalize_by_lengths,")
code.append(" out);")
code.append("}")
code.append("")
code.append("} // namespace caffe2")
with open(filename, "w") as fout:
fout.write("\n".join(code) + "\n")
print("Created " + filename)
if __name__ == "__main__":
main()
|
