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
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
|
// Copyright 2018 The Gemmlowp Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// pack_msa.h: optimized MSA specializations of the templates in pack.h.
#ifndef GEMMLOWP_INTERNAL_PACK_MSA_H_
#define GEMMLOWP_INTERNAL_PACK_MSA_H_
#include "pack.h"
#include <msa.h>
namespace gemmlowp {
typedef SideMap<const std::uint8_t, SideMapOrder::WidthMajor>
WidthMajorUint8SideMap;
template <int Cells>
using DepthMajorSideFormatNCells4x2 = KernelSideFormat<CellFormat<4, 2>, Cells>;
template <int Cells>
class PackingRegisterBlock<
WidthMajorUint8SideMap,
PackedSideBlock<DepthMajorSideFormatNCells4x2<Cells>>>
: public PackingRegisterBlockBase<
WidthMajorUint8SideMap,
PackedSideBlock<DepthMajorSideFormatNCells4x2<Cells>>> {
public:
typedef DepthMajorSideFormatNCells4x2<Cells> KernelSideFormat;
typedef typename KernelSideFormat::Cell CellFormat;
static constexpr int kCells = KernelSideFormat::kCells;
static const int kCellWidth = CellFormat::kWidth;
static const int kKernelWidth = CellFormat::kWidth * kCells;
static const int kCellDepth = CellFormat::kDepth;
static const int kCellSize = CellFormat::kSize;
void Pack(PackedSideBlock<KernelSideFormat>* dst, int start_width) {
std::uint8_t* dst_ptr = dst->current_data();
const std::uint8_t* const src_ptr = this->complete_src_.data();
const int stride = this->complete_src_.stride();
// Load source WidthMajor data
v16i8 src_lines[4 * kCells];
for (int i = 0; i < 4 * kCells; i++) {
src_lines[i] = __builtin_msa_ld_b(
const_cast<std::uint8_t*>(src_ptr + i * stride), 0);
}
// Reorder the data within registers to make DepthMajor 4x2 cells
v16i8 src_lines_intertwined_2x[2 * kCells][2];
for (int i = 0; i < kCells; i++) {
src_lines_intertwined_2x[2 * i][0] =
__builtin_msa_ilvr_b(src_lines[4 * i + 2], src_lines[4 * i]);
src_lines_intertwined_2x[2 * i][1] =
__builtin_msa_ilvl_b(src_lines[4 * i + 2], src_lines[4 * i]);
src_lines_intertwined_2x[2 * i + 1][0] =
__builtin_msa_ilvr_b(src_lines[4 * i + 3], src_lines[4 * i + 1]);
src_lines_intertwined_2x[2 * i + 1][1] =
__builtin_msa_ilvl_b(src_lines[4 * i + 3], src_lines[4 * i + 1]);
}
v16i8 src_lines_intertwined_4x[2 * kCells][2];
for (int i = 0; i < kCells; i++) {
src_lines_intertwined_4x[2 * i][0] =
__builtin_msa_ilvr_b(src_lines_intertwined_2x[2 * i + 1][0],
src_lines_intertwined_2x[2 * i][0]);
src_lines_intertwined_4x[2 * i][1] =
__builtin_msa_ilvl_b(src_lines_intertwined_2x[2 * i + 1][0],
src_lines_intertwined_2x[2 * i][0]);
src_lines_intertwined_4x[2 * i + 1][0] =
__builtin_msa_ilvr_b(src_lines_intertwined_2x[2 * i + 1][1],
src_lines_intertwined_2x[2 * i][1]);
src_lines_intertwined_4x[2 * i + 1][1] =
__builtin_msa_ilvl_b(src_lines_intertwined_2x[2 * i + 1][1],
src_lines_intertwined_2x[2 * i][1]);
}
// Store the resulting DepthMajor 4x2 cells in the destination packed block
for (int outer = 0; outer < 2; outer++) {
for (int inner = 0; inner < 2; inner++) {
if (kCells % 2 == 0) {
for (int cell = 0; cell < kCells; cell += 2) {
v2i64 tmp = __builtin_msa_ilvr_d(
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * (cell + 1) + outer][inner]),
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * cell + outer][inner]));
__builtin_msa_st_b(reinterpret_cast<v16i8>(tmp), dst_ptr, 0);
dst_ptr += 16;
}
for (int cell = 0; cell < kCells; cell += 2) {
v2i64 tmp = __builtin_msa_ilvl_d(
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * (cell + 1) + outer][inner]),
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * cell + outer][inner]));
__builtin_msa_st_b(reinterpret_cast<v16i8>(tmp), dst_ptr, 0);
dst_ptr += 16;
}
} else {
// Store even number of low vector halves.
for (int cell = 0; cell < kCells - 1; cell += 2) {
v2i64 tmp = __builtin_msa_ilvr_d(
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * (cell + 1) + outer][inner]),
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * cell + outer][inner]));
__builtin_msa_st_b(reinterpret_cast<v16i8>(tmp), dst_ptr, 0);
dst_ptr += 16;
}
// Store last low half and first high half.
v2i64 tmp = reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * 0 + outer][inner]);
tmp = __builtin_msa_insve_d(
tmp, 0,
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * (kCells - 1) + outer][inner]));
__builtin_msa_st_b(reinterpret_cast<v16i8>(tmp), dst_ptr, 0);
dst_ptr += 16;
// Store even number of high vector halves.
for (int cell = 1; cell < kCells; cell += 2) {
v2i64 tmp = __builtin_msa_ilvl_d(
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * (cell + 1) + outer][inner]),
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * cell + outer][inner]));
__builtin_msa_st_b(reinterpret_cast<v16i8>(tmp), dst_ptr, 0);
dst_ptr += 16;
}
}
}
}
// Compute sums across the depth dimension
v8i16 sums_of_2_cells[kCells][4];
const v16i8 zeroes = __builtin_msa_ldi_b(0);
for (int outer = 0; outer < 2; outer++) {
for (int inner = 0; inner < 2; inner++) {
int i = 2 * outer + inner;
for (int cell = 0; cell < kCells; cell++) {
v8i16 tmp0 = reinterpret_cast<v8i16>(__builtin_msa_ilvr_b(
zeroes, src_lines_intertwined_4x[2 * cell + outer][inner]));
v8i16 tmp1 = reinterpret_cast<v8i16>(__builtin_msa_ilvl_b(
zeroes, src_lines_intertwined_4x[2 * cell + outer][inner]));
sums_of_2_cells[cell][i] = __builtin_msa_addv_h(tmp0, tmp1);
}
}
}
v4i32 sums_of_4_cells[kCells][4];
for (int i = 0; i < 4; i++) {
for (int cell = 0; cell < kCells; cell++) {
v4i32 tmp0 = reinterpret_cast<v4i32>(__builtin_msa_ilvr_h(
reinterpret_cast<v8i16>(zeroes), sums_of_2_cells[cell][i]));
v4i32 tmp1 = reinterpret_cast<v4i32>(__builtin_msa_ilvl_h(
reinterpret_cast<v8i16>(zeroes), sums_of_2_cells[cell][i]));
sums_of_4_cells[cell][i] = __builtin_msa_addv_w(tmp0, tmp1);
}
}
// Update the sums_of_each_slice vector
for (int cell = 0; cell < kCells; cell++) {
v4i32 s01 = __builtin_msa_addv_w(sums_of_4_cells[cell][0],
sums_of_4_cells[cell][1]);
v4i32 s23 = __builtin_msa_addv_w(sums_of_4_cells[cell][2],
sums_of_4_cells[cell][3]);
v4i32 s = __builtin_msa_addv_w(s01, s23);
std::int32_t* sums_of_each_slice_ptr =
dst->sums_of_each_slice() + start_width + 4 * cell;
v4i32 tmp = __builtin_msa_ld_w(sums_of_each_slice_ptr, 0);
tmp = __builtin_msa_addv_w(tmp, s);
__builtin_msa_st_w(tmp, sums_of_each_slice_ptr, 0);
}
dst->seek_forward_n_cells(kCells * kRegisterSize / kCellDepth);
}
};
template <int Cells>
using WidthMajorSideFormatNCells4x2 =
KernelSideFormat<CellFormat<4, 2, CellOrder::WidthMajor>, Cells>;
template <int Cells>
class PackingRegisterBlock<
WidthMajorUint8SideMap,
PackedSideBlock<WidthMajorSideFormatNCells4x2<Cells>>>
: public PackingRegisterBlockBase<
WidthMajorUint8SideMap,
PackedSideBlock<WidthMajorSideFormatNCells4x2<Cells>>> {
public:
typedef WidthMajorSideFormatNCells4x2<Cells> KernelSideFormat;
typedef typename KernelSideFormat::Cell CellFormat;
static constexpr int kCells = KernelSideFormat::kCells;
static const int kCellWidth = CellFormat::kWidth;
static const int kKernelWidth = CellFormat::kWidth * kCells;
static const int kCellDepth = CellFormat::kDepth;
static const int kCellSize = CellFormat::kSize;
void Pack(PackedSideBlock<KernelSideFormat>* dst, int start_width) {
std::uint8_t* dst_ptr = dst->current_data();
const std::uint8_t* src_ptr = this->complete_src_.data();
const int stride = this->complete_src_.stride();
// Load source WidthMajor data
v8i16 src_lines[kCells * 4];
for (int i = 0; i < kCells; i++) {
#define GEMMLOWP_UNROLLED_LOOP_ITER(k) \
src_lines[4 * i + k] = \
__builtin_msa_ld_h(const_cast<std::uint8_t*>(src_ptr), 0); \
src_ptr += stride;
GEMMLOWP_UNROLLED_LOOP_ITER(0)
GEMMLOWP_UNROLLED_LOOP_ITER(1)
GEMMLOWP_UNROLLED_LOOP_ITER(2)
GEMMLOWP_UNROLLED_LOOP_ITER(3)
#undef GEMMLOWP_UNROLLED_LOOP_ITER
}
// Reorder the data within registers to make WidthMajor 4x2 cells
v8i16 src_lines_intertwined_2x[2 * kCells][2];
for (int i = 0; i < kCells; i++) {
src_lines_intertwined_2x[2 * i][0] =
__builtin_msa_ilvr_h(src_lines[4 * i + 2], src_lines[4 * i]);
src_lines_intertwined_2x[2 * i][1] =
__builtin_msa_ilvl_h(src_lines[4 * i + 2], src_lines[4 * i]);
src_lines_intertwined_2x[2 * i + 1][0] =
__builtin_msa_ilvr_h(src_lines[4 * i + 3], src_lines[4 * i + 1]);
src_lines_intertwined_2x[2 * i + 1][1] =
__builtin_msa_ilvl_h(src_lines[4 * i + 3], src_lines[4 * i + 1]);
}
v8i16 src_lines_intertwined_4x[2 * kCells][2];
for (int i = 0; i < kCells; i++) {
src_lines_intertwined_4x[2 * i][0] =
__builtin_msa_ilvr_h(src_lines_intertwined_2x[2 * i + 1][0],
src_lines_intertwined_2x[2 * i][0]);
src_lines_intertwined_4x[2 * i][1] =
__builtin_msa_ilvl_h(src_lines_intertwined_2x[2 * i + 1][0],
src_lines_intertwined_2x[2 * i][0]);
src_lines_intertwined_4x[2 * i + 1][0] =
__builtin_msa_ilvr_h(src_lines_intertwined_2x[2 * i + 1][1],
src_lines_intertwined_2x[2 * i][1]);
src_lines_intertwined_4x[2 * i + 1][1] =
__builtin_msa_ilvl_h(src_lines_intertwined_2x[2 * i + 1][1],
src_lines_intertwined_2x[2 * i][1]);
}
// Store the resulting WidthMajor 4x2 cells in the destination packed block
for (int outer = 0; outer < 2; outer++) {
for (int inner = 0; inner < 2; inner++) {
if (kCells % 2 == 0) {
for (int cell = 0; cell < kCells; cell += 2) {
v2i64 tmp = __builtin_msa_ilvr_d(
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * (cell + 1) + outer][inner]),
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * cell + outer][inner]));
__builtin_msa_st_b(reinterpret_cast<v16i8>(tmp), dst_ptr, 0);
dst_ptr += 16;
}
for (int cell = 0; cell < kCells; cell += 2) {
v2i64 tmp = __builtin_msa_ilvl_d(
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * (cell + 1) + outer][inner]),
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * cell + outer][inner]));
__builtin_msa_st_b(reinterpret_cast<v16i8>(tmp), dst_ptr, 0);
dst_ptr += 16;
}
} else {
// Store even number of low vector halves.
for (int cell = 0; cell < kCells - 1; cell += 2) {
v2i64 tmp = __builtin_msa_ilvr_d(
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * (cell + 1) + outer][inner]),
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * cell + outer][inner]));
__builtin_msa_st_b(reinterpret_cast<v16i8>(tmp), dst_ptr, 0);
dst_ptr += 16;
}
// Store last low half and first high half.
v2i64 tmp = reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * 0 + outer][inner]);
tmp = __builtin_msa_insve_d(
tmp, 0,
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * (kCells - 1) + outer][inner]));
__builtin_msa_st_b(reinterpret_cast<v16i8>(tmp), dst_ptr, 0);
dst_ptr += 16;
// Store even number of high vector halves.
for (int cell = 1; cell < kCells; cell += 2) {
v2i64 tmp = __builtin_msa_ilvl_d(
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * (cell + 1) + outer][inner]),
reinterpret_cast<v2i64>(
src_lines_intertwined_4x[2 * cell + outer][inner]));
__builtin_msa_st_b(reinterpret_cast<v16i8>(tmp), dst_ptr, 0);
dst_ptr += 16;
}
}
}
}
// Compute sums across the depth dimension
v8i16 sums_of_2[kCells][4];
for (int outer = 0; outer < 2; outer++) {
for (int inner = 0; inner < 2; inner++) {
int i = 2 * outer + inner;
for (int cell = 0; cell < kCells; cell++) {
sums_of_2[cell][i] = reinterpret_cast<v8i16>(__builtin_msa_hadd_u_h(
reinterpret_cast<v16u8>(
src_lines_intertwined_4x[2 * cell + outer][inner]),
reinterpret_cast<v16u8>(
src_lines_intertwined_4x[2 * cell + outer][inner])));
}
}
}
v8i16 sums_of_4[kCells][2];
for (int i = 0; i < 2; i++) {
for (int cell = 0; cell < kCells; cell++) {
sums_of_4[cell][i] = __builtin_msa_addv_h(sums_of_2[cell][2 * i],
sums_of_2[cell][2 * i + 1]);
}
}
v8i16 sums_of_8[kCells];
for (int cell = 0; cell < kCells; cell++) {
sums_of_8[cell] =
__builtin_msa_addv_h(sums_of_4[cell][0], sums_of_4[cell][1]);
}
v4i32 sums_of_16[kCells];
const v8i16 zeroes = __builtin_msa_ldi_h(0);
for (int cell = 0; cell < kCells; cell++) {
sums_of_16[cell] = reinterpret_cast<v4i32>(
__builtin_msa_ilvr_h(zeroes, sums_of_8[cell]));
v8i16 tmp = __builtin_msa_ilvl_h(zeroes, sums_of_8[cell]);
sums_of_16[cell] =
__builtin_msa_addv_w(sums_of_16[cell], reinterpret_cast<v4i32>(tmp));
}
// Update the sums_of_each_slice vector
for (int cell = 0; cell < kCells; cell++) {
std::int32_t* sums_of_each_slice_ptr =
dst->sums_of_each_slice() + start_width + 4 * cell;
v4i32 tmp = __builtin_msa_ld_w(sums_of_each_slice_ptr, 0);
tmp = __builtin_msa_addv_w(tmp, sums_of_16[cell]);
__builtin_msa_st_w(tmp, sums_of_each_slice_ptr, 0);
}
dst->seek_forward_n_cells(kCells * kRegisterSize / kCellDepth);
}
};
template <int Width>
using Int8FastKernelFormat =
KernelSideFormatInt8<CellFormat<Width, 16, CellOrder::WidthMajor>, 1>;
template <int Width>
class PackingRegisterBlock<WidthMajorUint8SideMap,
PackedSideBlock<Int8FastKernelFormat<Width>>>
: public PackingRegisterBlockBase<
WidthMajorUint8SideMap,
PackedSideBlock<Int8FastKernelFormat<Width>>> {
public:
static_assert(Width == 2 || Width == 4, "");
typedef Int8FastKernelFormat<Width> KernelSideFormat;
typedef typename KernelSideFormat::Cell CellFormat;
static const int kCells = KernelSideFormat::kCells;
static const int kCellWidth = CellFormat::kWidth;
static const int kKernelWidth = CellFormat::kWidth * kCells;
static const int kCellDepth = CellFormat::kDepth;
static const int kCellSize = CellFormat::kSize;
void Pack(PackedSideBlock<KernelSideFormat>* dst, int start_width) {
std::int32_t* sums_ptr = dst->sums_of_each_slice() + start_width;
std::uint8_t* dst_ptr = dst->current_data();
const std::uint8_t* const src_ptr = this->complete_src_.data();
const int stride = this->complete_src_.stride();
// Load source WidthMajor data.
v16i8 src_lines[Width];
for (int i = 0; i < Width; i++) {
src_lines[i] = __builtin_msa_ld_b(
const_cast<std::uint8_t*>(src_ptr + i * stride), 0);
}
for (int i = 0; i < Width; i++) {
// Subtract 128 by inverting bit 7.
src_lines[i] = reinterpret_cast<v16i8>(
__builtin_msa_bnegi_b(reinterpret_cast<v16u8>(src_lines[i]), 7));
}
for (int i = 0; i < Width; i++) {
__builtin_msa_st_b(src_lines[i], dst_ptr + 16 * i, 0);
}
v8i16 sums2[Width];
for (int i = 0; i < Width; i++) {
sums2[i] = __builtin_msa_hadd_s_h(src_lines[i], src_lines[i]);
}
v4i32 sums4_wide[Width];
for (int i = 0; i < Width; i++) {
sums4_wide[i] = __builtin_msa_hadd_s_w(sums2[i], sums2[i]);
}
v8i16 sums4[Width / 2];
for (int i = 0; i < Width / 2; i++) {
sums4[i] = __builtin_msa_pckev_h(
reinterpret_cast<v8i16>(sums4_wide[2 * i + 1]),
reinterpret_cast<v8i16>(sums4_wide[2 * i]));
}
v4i32 sums8_wide[Width / 2];
for (int i = 0; i < Width / 2; i++) {
sums8_wide[i] = __builtin_msa_hadd_s_w(sums4[i], sums4[i]);
}
if (Width == 4) {
v4i32 sum = __builtin_msa_ld_w(const_cast<std::int32_t*>(sums_ptr), 0);
v8i16 sums8 = __builtin_msa_pckev_h(
reinterpret_cast<v8i16>(sums8_wide[1]),
reinterpret_cast<v8i16>(sums8_wide[0]));
v4i32 sums16 = __builtin_msa_hadd_s_w(sums8, sums8);
sum = __builtin_msa_addv_w(sum, sums16);
__builtin_msa_st_w(sum, sums_ptr, 0);
} else {
assert(Width == 2);
std::int32_t sum[2] = { sums_ptr[0], sums_ptr[1] };
v2i64 sums16 = __builtin_msa_hadd_s_d(sums8_wide[0], sums8_wide[0]);
sum[0] += __builtin_msa_copy_s_w(reinterpret_cast<v4i32>(sums16), 0);
sum[1] += __builtin_msa_copy_s_w(reinterpret_cast<v4i32>(sums16), 2);
sums_ptr[0] = sum[0];
sums_ptr[1] = sum[1];
}
dst->seek_forward_n_cells(1);
}
};
} // namespace gemmlowp
#endif // GEMMLOWP_INTERNAL_PACK_MSA_H_
|
