-
Notifications
You must be signed in to change notification settings - Fork 454
/
output.h
579 lines (489 loc) · 22.1 KB
/
output.h
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
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
// Copyright 2015 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.
// output.h: processing the 32-bit accumulators output by the unpack
// stage, obtaining the final result matrix entries and storing them into
// the destination matrix.
#ifndef GEMMLOWP_INTERNAL_OUTPUT_H_
#define GEMMLOWP_INTERNAL_OUTPUT_H_
#include <cmath>
#include <tuple>
#include <type_traits>
#include <typeinfo>
#include "../fixedpoint/fixedpoint.h"
#include "../public/output_stages.h"
#include "simd_wrappers.h"
namespace gemmlowp {
template <typename OutputStage, typename InputBufferType>
struct OutputStageEvalBufferImpl {
// This generic template body should never be hit.
static_assert(
std::is_same<InputBufferType, void>::value,
"Unimplemented: missing implementation of this output pipeline stage "
"for this data type. This would happen if some architecture-specific "
"SIMD back-end (output_$arch.h) were incomplete.");
};
template <typename OutputStage, typename InputType>
struct OutputStageEvalImpl {
static constexpr int kRows = InputType::kRows;
static constexpr int kCols = InputType::kCols;
using InputBufferType = typename InputType::BufferType;
using BufferEvalImplType =
OutputStageEvalBufferImpl<OutputStage, InputBufferType>;
using OutputBufferType = typename BufferEvalImplType::OutputType;
using OutputScalarType = typename OutputBufferType::ScalarType;
using OutputType = RegisterBlock<OutputScalarType, kRows, kCols>;
OutputStageEvalImpl(const OutputStage& s) : buffer_eval_impl(s) {}
OutputType Eval(InputType input, int, int) const {
OutputType output;
output.buf = buffer_eval_impl.Eval(input.buf);
return output;
}
const BufferEvalImplType buffer_eval_impl;
};
template <int Size>
struct OutputStageEvalBufferImpl<OutputStageQuantizeDownInt32ToUint8Scale,
RegisterBuffer<std::int32_t, Size>> {
using InputType = RegisterBuffer<std::int32_t, Size>;
using OutputType = RegisterBuffer<std::int32_t, Size>;
typedef OutputStageQuantizeDownInt32ToUint8Scale OutputStage;
OutputStageEvalBufferImpl(const OutputStage& s) : output_stage(s) {}
OutputType Eval(InputType input) const {
const int result_shift = output_stage.result_shift;
const std::int32_t result_mult_int = output_stage.result_mult_int;
using RegisterType = typename InputType::RegisterType;
const RegisterType result_offset =
Dup<RegisterType>(output_stage.result_offset);
OutputType output;
for (int i = 0; i < InputType::kRegisterCount; i++) {
output.reg[i] = RoundingDivideByPOT(
Mul(Add(input.reg[i], result_offset), result_mult_int), result_shift);
}
return output;
}
const OutputStage& output_stage;
};
template <int Rows, int Cols, VectorShape Shape>
struct OutputStageEvalImpl<OutputStageQuantizeDownInt32ToUint8ScalePC<Shape>,
RegisterBlock<std::int32_t, Rows, Cols>> {
typedef RegisterBlock<std::int32_t, Rows, Cols> InputType;
typedef RegisterBlock<std::int32_t, Rows, Cols> OutputType;
typedef OutputStageQuantizeDownInt32ToUint8ScalePC<Shape> OutputStage;
OutputStageEvalImpl(const OutputStage& s) : output_stage(s) {}
OutputType Eval(InputType input, int row, int col) const {
OutputType output;
const int result_shift = output_stage.result_shift;
const int pos = Shape == VectorShape::Col ? row : col;
const auto result_mult_int =
LoadForBroadcasting<InputType>(output_stage.result_mult_int, pos);
const auto result_offset =
LoadForBroadcasting<InputType>(output_stage.result_offset, pos);
const auto dividend = BroadcastMul<InputType>(
BroadcastAdd<InputType>(input, result_offset), result_mult_int);
for (int i = 0; i < InputType::kRegisterCount; i++) {
output.buf.reg[i] =
RoundingDivideByPOT(dividend.buf.reg[i], result_shift);
}
return output;
}
const OutputStage& output_stage;
};
template <int Size>
struct OutputStageEvalBufferImpl<
OutputStageQuantizeDownInt32ByFixedPoint,
RegisterBuffer<std::int32_t, Size>> {
typedef RegisterBuffer<std::int32_t, Size> InputType;
typedef RegisterBuffer<std::int32_t, Size> OutputType;
typedef OutputStageQuantizeDownInt32ByFixedPoint OutputStage;
OutputStageEvalBufferImpl(const OutputStage& s) : output_stage(s) {}
OutputType Eval(InputType input) const {
OutputType output;
using RegisterType = typename InputType::RegisterType;
const RegisterType result_offset_after_shift =
Dup<RegisterType>(output_stage.result_offset_after_shift);
for (int i = 0; i < InputType::kRegisterCount; i++) {
const RegisterType mulhigh_val = SaturatingRoundingDoublingHighMul(
input.reg[i], output_stage.result_fixedpoint_multiplier);
output.reg[i] =
Add(RoundingDivideByPOT(mulhigh_val, output_stage.result_shift),
result_offset_after_shift);
}
return output;
}
const OutputStage& output_stage;
};
template <int Size>
struct OutputStageEvalBufferImpl<OutputStageScaleInt32ByFixedPointAndExponent,
RegisterBuffer<std::int32_t, Size>> {
typedef RegisterBuffer<std::int32_t, Size> InputType;
typedef RegisterBuffer<std::int32_t, Size> OutputType;
typedef OutputStageScaleInt32ByFixedPointAndExponent OutputStage;
OutputStageEvalBufferImpl(const OutputStage& s) : output_stage(s) {
left_shift = std::max(0, output_stage.result_exponent);
right_shift = std::max(0, -output_stage.result_exponent);
}
OutputType Eval(InputType input) const {
OutputType output;
using RegisterType = typename InputType::RegisterType;
const RegisterType result_offset_after_shift =
Dup<RegisterType>(output_stage.result_offset_after_shift);
for (int i = 0; i < InputType::kRegisterCount; i++) {
const RegisterType mulhigh_val = SaturatingRoundingDoublingHighMul(
ShiftLeft(input.reg[i], left_shift),
output_stage.result_fixedpoint_multiplier);
output.reg[i] = Add(RoundingDivideByPOT(mulhigh_val, right_shift),
result_offset_after_shift);
}
return output;
}
const OutputStage& output_stage;
int left_shift;
int right_shift;
};
template <int Rows, int Cols, VectorShape Shape>
struct OutputStageEvalImpl<
OutputStageScaleInt32ByFixedPointAndExponentPC<Shape>,
RegisterBlock<std::int32_t, Rows, Cols>> {
typedef RegisterBlock<std::int32_t, Rows, Cols> InputType;
typedef RegisterBlock<std::int32_t, Rows, Cols> OutputType;
typedef OutputStageScaleInt32ByFixedPointAndExponentPC<Shape> OutputStage;
OutputStageEvalImpl(const OutputStage& s) : output_stage(s) {}
OutputType Eval(InputType input, int row, int col) const {
OutputType output;
const int pos = Shape == VectorShape::Row ? col : row;
using RegisterType = typename InputType::RegisterType;
const RegisterType result_offset_after_shift =
Dup<RegisterType>(output_stage.result_offset_after_shift);
auto left_shift =
LoadForBroadcasting<InputType>(output_stage.result_exponent, pos);
auto right_shift =
LoadForBroadcasting<InputType>(output_stage.result_exponent, pos);
const auto result_fixedpoint_multiplier = LoadForBroadcasting<InputType>(
output_stage.result_fixedpoint_multiplier, pos);
for (int i = 0; i < decltype(left_shift)::kRegisterCount; i++) {
left_shift.buf.reg[i] = Max(left_shift.buf.reg[i], 0);
right_shift.buf.reg[i] = Max(-right_shift.buf.reg[i], 0);
}
const auto mulhigh_val = BroadcastSaturatingRoundingDoublingHighMul(
BroadcastShiftLeft(input, left_shift), result_fixedpoint_multiplier);
const auto rdpot_val =
BroadcastRoundingDivideByPOT(mulhigh_val, right_shift);
for (int i = 0; i < InputType::kRegisterCount; i++) {
output.buf.reg[i] = Add(rdpot_val.buf.reg[i], result_offset_after_shift);
}
return output;
}
const OutputStage& output_stage;
};
// Implementation of OutputStageSaturatingCastToUint8 for scalar data.
template <int Size>
struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToUint8,
RegisterBuffer<std::int32_t, Size>> {
typedef RegisterBuffer<std::int32_t, Size> InputType;
typedef RegisterBuffer<std::uint8_t, Size> OutputType;
static_assert(InputType::kRegisterLanes == 1,
"This path is only for scalar values");
typedef OutputStageSaturatingCastToUint8 OutputStage;
OutputStageEvalBufferImpl(const OutputStage&) {}
OutputType Eval(InputType input) const {
OutputType output;
for (int i = 0; i < InputType::kRegisterCount; i++) {
std::int32_t data = input.reg[i];
output.reg[i] = data > 255 ? 255 : data < 0 ? 0 : data;
}
return output;
}
};
// Implementation of OutputStageSaturatingCastToInt8 for scalar data.
template <int Size>
struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToInt8,
RegisterBuffer<std::int32_t, Size>> {
typedef RegisterBuffer<std::int32_t, Size> InputType;
typedef RegisterBuffer<std::int8_t, Size> OutputType;
static_assert(InputType::kRegisterLanes == 1,
"This path is only for scalar values");
typedef OutputStageSaturatingCastToInt8 OutputStage;
OutputStageEvalBufferImpl(const OutputStage&) {}
OutputType Eval(InputType input) const {
OutputType output;
for (int i = 0; i < InputType::kRegisterCount; i++) {
std::int32_t data = input.reg[i];
output.reg[i] = data > 127 ? 127 : data < -128 ? -128 : data;
}
return output;
}
};
// Implementation of OutputStageSaturatingCastToInt16 for scalar data.
template <int Size>
struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToInt16,
RegisterBuffer<std::int32_t, Size>> {
typedef RegisterBuffer<std::int32_t, Size> InputType;
typedef RegisterBuffer<std::int16_t, Size> OutputType;
static_assert(InputType::kRegisterLanes == 1,
"This path is only for scalar values");
typedef OutputStageSaturatingCastToInt16 OutputStage;
OutputStageEvalBufferImpl(const OutputStage&) {}
OutputType Eval(InputType input) const {
OutputType output;
for (int i = 0; i < InputType::kRegisterCount; i++) {
std::int32_t data = input.reg[i];
output.reg[i] = data > 32767 ? 32767 : data < -32768 ? -32768 : data;
}
return output;
}
};
// Implementation of OutputStageTruncatingCastToUint8 for scalar data
template <int Size>
struct OutputStageEvalBufferImpl<OutputStageTruncatingCastToUint8,
RegisterBuffer<std::int32_t, Size>> {
typedef RegisterBuffer<std::int32_t, Size> InputType;
typedef RegisterBuffer<std::uint8_t, Size> OutputType;
static_assert(InputType::kRegisterLanes == 1,
"This path is only for scalar values");
typedef OutputStageTruncatingCastToUint8 OutputStage;
OutputStageEvalBufferImpl(const OutputStage&) {}
OutputType Eval(InputType input) const {
OutputType output;
for (int i = 0; i < InputType::kRegisterCount; i++) {
output.reg[i] = input.reg[i];
}
return output;
}
};
template <int Rows, int Cols, typename VectorType>
struct OutputStageEvalImpl<OutputStageBiasAddition<VectorType>,
RegisterBlock<std::int32_t, Rows, Cols>> {
typedef RegisterBlock<std::int32_t, Rows, Cols> InputType;
typedef RegisterBlock<std::int32_t, Rows, Cols> OutputType;
typedef OutputStageBiasAddition<VectorType> OutputStage;
OutputStageEvalImpl(const OutputStage& s) : output_stage(s) {}
OutputType Eval(InputType input, int row, int col) const {
const int pos = VectorType::kShape == VectorShape::Row ? col : row;
return BroadcastAdd<InputType>(
input, LoadForBroadcasting<InputType>(output_stage.bias_vector, pos));
}
const OutputStage& output_stage;
};
template <int Size>
struct OutputStageEvalBufferImpl<OutputStageClamp,
RegisterBuffer<std::int32_t, Size>> {
typedef RegisterBuffer<std::int32_t, Size> InputType;
typedef RegisterBuffer<std::int32_t, Size> OutputType;
typedef OutputStageClamp OutputStage;
OutputStageEvalBufferImpl(const OutputStage& s) : output_stage(s) {}
OutputType Eval(InputType input) const {
using RegisterType = typename InputType::RegisterType;
const RegisterType min = Dup<RegisterType>(output_stage.min);
const RegisterType max = Dup<RegisterType>(output_stage.max);
OutputType output;
for (int i = 0; i < InputType::kRegisterCount; i++) {
output.reg[i] = Min(Max(input.reg[i], min), max);
}
return output;
}
const OutputStage& output_stage;
};
template <int Size>
struct OutputStageEvalBufferImpl<OutputStageTanh,
RegisterBuffer<std::int32_t, Size>> {
typedef RegisterBuffer<std::int32_t, Size> InputType;
typedef RegisterBuffer<std::int32_t, Size> OutputType;
using RegisterType = typename InputType::RegisterType;
typedef RegisterType DataType;
typedef OutputStageTanh OutputStage;
OutputStageEvalBufferImpl(const OutputStage& s) : output_stage(s) {
const std::int32_t real_zero_as_int32 = output_stage.real_zero_as_int32;
const std::int32_t real_amplitude_as_int32 =
output_stage.real_amplitude_as_int32;
input_cutoff_min = real_zero_as_int32 - 8 * real_amplitude_as_int32;
input_cutoff_max = real_zero_as_int32 + 8 * real_amplitude_as_int32;
output_min = real_zero_as_int32 - real_amplitude_as_int32;
output_max = real_zero_as_int32 + real_amplitude_as_int32;
double inverse_amplitude_normalized_double = 1.0 / real_amplitude_as_int32;
inverse_amplitude_neg_exponent = 0;
while (inverse_amplitude_normalized_double < 0.5) {
inverse_amplitude_normalized_double *= 2;
inverse_amplitude_neg_exponent++;
}
inverse_amplitude_normalized = FixedPoint<DataType, 0>::FromDouble(
inverse_amplitude_normalized_double);
double amplitude_normalized_double = real_amplitude_as_int32;
amplitude_exponent = 0;
while (amplitude_normalized_double >= 1.0) {
amplitude_normalized_double *= 0.5;
amplitude_exponent++;
}
amplitude_normalized =
FixedPoint<DataType, 0>::FromDouble(amplitude_normalized_double);
}
OutputType Eval(InputType input) const {
const std::int32_t real_zero_as_int32 = output_stage.real_zero_as_int32;
typedef FixedPoint<DataType, 3> F3;
typedef FixedPoint<DataType, 0> F0;
OutputType output;
for (int i = 0; i < OutputType::kRegisterCount; i++) {
// fixed-point affine transformation
DataType input_centered =
Sub(input.reg[i], Dup<DataType>(real_zero_as_int32));
F3 fixedpoint_input =
F3::FromRaw(input_centered) * inverse_amplitude_normalized;
// left shift
fixedpoint_input.raw() = ShiftLeft(fixedpoint_input.raw(),
28 - inverse_amplitude_neg_exponent);
// fixed-point tanh and multiplication
F0 fixedpoint_output = tanh(fixedpoint_input) * amplitude_normalized;
// right shift
DataType int32_output =
Add(Dup<DataType>(real_zero_as_int32),
ShiftRight(fixedpoint_output.raw(), 31 - amplitude_exponent));
DataType mask_if_below_cutoff_min =
MaskIfLessThanOrEqual(input.reg[i], Dup<DataType>(input_cutoff_min));
DataType mask_if_above_cutoff_max = MaskIfGreaterThanOrEqual(
input.reg[i], Dup<DataType>(input_cutoff_max));
output.reg[i] = SelectUsingMask(
mask_if_below_cutoff_min, Dup<DataType>(output_min),
SelectUsingMask(mask_if_above_cutoff_max, Dup<DataType>(output_max),
int32_output));
}
return output;
}
const OutputStage& output_stage;
std::int32_t input_cutoff_min, input_cutoff_max;
std::int32_t output_min, output_max;
FixedPoint<DataType, 0> inverse_amplitude_normalized;
int inverse_amplitude_neg_exponent;
FixedPoint<DataType, 0> amplitude_normalized;
int amplitude_exponent;
};
// OutputPipelineOutputType is a helper to determine the output data type of a
// pipeline, for a
// given input data type. It is a recursive template; see the explanation on
// OutputPipelineEvalImpl below.
template <typename OutputPipelineType, int FirstStage, typename InputType,
bool StopRecursion =
FirstStage == std::tuple_size<OutputPipelineType>::value>
struct OutputPipelineOutputType {
typedef typename std::tuple_element<FirstStage, OutputPipelineType>::type
FirstStageType;
typedef typename OutputStageEvalImpl<FirstStageType, InputType>::OutputType
FirstStageOutputType;
typedef typename OutputPipelineOutputType<OutputPipelineType, FirstStage + 1,
FirstStageOutputType>::Type Type;
};
template <typename OutputPipelineType, int FirstStage, typename InputType>
struct OutputPipelineOutputType<OutputPipelineType, FirstStage, InputType,
true> {
typedef InputType Type;
};
// OutputPipelineEvalImpl is a helper to implement the evaluation of
// the whole pipeline. It is a recursive template to implement compile-time
// unrolling of the loop over all pipeline stages. The 'FirstStage' parameter
// is how we implement recursion: each specialization implements only
// evaluation starting at 'FirstStage'. The StopRecursion parameter is just a
// helper to implement the termination of the recursion as a partial
// specialization below.
template <typename OutputPipelineType, int FirstStage, typename InputType,
bool StopRecursion =
FirstStage == std::tuple_size<OutputPipelineType>::value>
struct OutputPipelineEvalImpl {
typedef typename std::tuple_element<FirstStage, OutputPipelineType>::type
FirstStageType;
typedef typename OutputStageEvalImpl<FirstStageType, InputType>::OutputType
FirstStageOutputType;
typedef typename OutputPipelineOutputType<OutputPipelineType, FirstStage,
InputType>::Type OutputType;
OutputPipelineEvalImpl(const OutputPipelineType& output_pipeline)
: head_impl(std::get<FirstStage>(output_pipeline)),
tail_impl(output_pipeline) {}
OutputType Eval(InputType input, int row, int col) const {
// Evaluate the first stage.
FirstStageOutputType first_stage_output = head_impl.Eval(input, row, col);
// Recurse into the remaining stages.
return tail_impl.Eval(first_stage_output, row, col);
}
const OutputStageEvalImpl<FirstStageType, InputType> head_impl;
const OutputPipelineEvalImpl<OutputPipelineType, FirstStage + 1,
FirstStageOutputType>
tail_impl;
};
// Specialization on 'StopRecursion' for terminating the recursion.
template <typename OutputPipelineType, int FirstStage, typename InputType>
struct OutputPipelineEvalImpl<OutputPipelineType, FirstStage, InputType, true> {
OutputPipelineEvalImpl(const OutputPipelineType&) {}
InputType Eval(InputType input, int, int) const {
// Terminating the recursion.
return input;
}
};
template <typename RegisterBlockType, typename DstType>
struct StoreFinalOutputImpl {
static_assert(std::is_same<RegisterBlockType, void>::value,
"This generic impl should never be hit");
};
template <typename ScalarType, int Rows, int Cols, typename DstType>
struct StoreFinalOutputImpl<RegisterBlock<ScalarType, Rows, Cols>, DstType> {
using RegisterBlockType = RegisterBlock<ScalarType, Rows, Cols>;
static void Run(const RegisterBlockType& src, DstType* dst, int row,
int col) {
for (int r = 0; r < Rows; r++) {
for (int c = 0; c < Cols; c++) {
*dst->data(row + r, col + c) = src.buf.reg[r + c * Rows];
}
}
}
};
// StoreFinalOutput takes the final value at the end of the output pipeline and
// stores it into the destination matrix. It can be specialized for different
// data types; the generic implementation here is typically used only for plain
// old scalar (not SIMD) types.
template <typename RegisterBlockType, typename DstType>
void StoreFinalOutput(RegisterBlockType src, DstType* dst, int row, int col) {
StoreFinalOutputImpl<RegisterBlockType, DstType>::Run(src, dst, row, col);
}
template <typename OutputPipelineType, typename InputType>
struct OutputPipelineExecutor {
OutputPipelineExecutor(const OutputPipelineType& output_pipeline)
: output_pipeline_eval_impl_(output_pipeline) {}
// Execute is the entry point into the output pipeline evaluation
// code. It should be the only thing that unpack code calls. It takes the
// result
// of the unpack stage and stores it into the destination matrix.
template <typename DstType>
void Execute(InputType input, DstType* dst, int src_global_row,
int src_global_col, int dst_row, int dst_col) const {
// Statically assert that the output pipeline matches the given destination
// matrix's scalar type.
typedef typename OutputPipelineOutputType<
OutputPipelineType, 0, InputType>::Type::BufferType::ScalarType
ScalarOutputType;
typedef typename DstType::Scalar ScalarDstType;
static_assert(std::is_same<ScalarOutputType, ScalarDstType>::value,
"mismatched destination scalar type and output pipeline");
// Evaluate the output pipeline.
auto output =
output_pipeline_eval_impl_.Eval(input, src_global_row, src_global_col);
// Store the result into the destination matrix.
StoreFinalOutput(output, dst, dst_row, dst_col);
}
const OutputPipelineEvalImpl<OutputPipelineType, 0, InputType>
output_pipeline_eval_impl_;
};
} // namespace gemmlowp
#ifdef GEMMLOWP_NEON
#include "output_neon.h"
#elif defined(GEMMLOWP_SSE4)
#include "output_sse.h"
#elif defined(GEMMLOWP_MSA)
#include "output_msa.h"
#endif
#endif // GEMMLOWP_INTERNAL_OUTPUT_H_