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
|
/*
* Copyright (c) 2017-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "src/gpu/cl/operators/ClFullyConnected.h"
#include "arm_compute/core/Size2D.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
#include "arm_compute/runtime/CL/CLScheduler.h"
#include "src/core/CL/kernels/CLFillBorderKernel.h"
#include "src/core/helpers/MemoryHelpers.h"
#include "src/gpu/cl/operators/ClConvertFullyConnectedWeights.h"
#include "src/gpu/cl/operators/ClFlatten.h"
#include "src/gpu/cl/operators/ClGemm.h"
#include "src/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.h"
#include "src/gpu/cl/operators/ClTranspose.h"
#include "src/gpu/cl/utils/ClAuxTensorHandler.h"
#include "support/Cast.h"
#include <algorithm>
namespace arm_compute
{
namespace opencl
{
using namespace arm_compute::experimental;
using namespace arm_compute::misc::shape_calculator;
namespace
{
Status construct_gemmlowp_output_stage(const ITensorInfo &src, const ITensorInfo &weights, const ITensorInfo &dst,
GEMMLowpOutputStageInfo &gemmlowp_output_stage, ActivationLayerInfo activation_info)
{
gemmlowp_output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
gemmlowp_output_stage.gemmlowp_offset = 0;
gemmlowp_output_stage.gemmlowp_multiplier = 0;
gemmlowp_output_stage.gemmlowp_shift = 0;
const auto data_type = src.data_type();
// Configure output stage for quantized case
if(is_data_type_quantized_asymmetric(data_type))
{
const QuantizationInfo oq_info = dst.quantization_info();
const UniformQuantizationInfo iq_unif = src.quantization_info().uniform();
const UniformQuantizationInfo wq_unif = weights.quantization_info().uniform();
const UniformQuantizationInfo oq_unif = oq_info.uniform();
const auto output_quant_info = (dst.total_size() == 0) ? iq_unif : oq_unif;
const float multiplier = (iq_unif.scale * wq_unif.scale) / output_quant_info.scale;
int output_multiplier = 0;
int output_shift = 0;
ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
PixelValue type_min{};
PixelValue type_max{};
std::tie(type_min, type_max) = get_min_max(data_type);
if(activation_info.enabled())
{
std::tie(type_min, type_max) = get_quantized_activation_min_max(activation_info, data_type, output_quant_info);
}
// Set the GEMMLowp output stage info
gemmlowp_output_stage.gemmlowp_offset = output_quant_info.offset;
gemmlowp_output_stage.gemmlowp_multiplier = output_multiplier;
gemmlowp_output_stage.gemmlowp_shift = output_shift;
gemmlowp_output_stage.gemmlowp_multipliers.push_back(output_multiplier);
gemmlowp_output_stage.gemmlowp_shifts.push_back(output_shift);
type_min.get(gemmlowp_output_stage.gemmlowp_min_bound);
type_max.get(gemmlowp_output_stage.gemmlowp_max_bound);
}
return Status{};
}
Status validate_mm(const ITensorInfo &src, const ITensorInfo &weights, const ITensorInfo *bias, const ITensorInfo &dst, const FullyConnectedLayerInfo &fc_info)
{
GEMMLowpOutputStageInfo gemmlowp_output_stage;
ARM_COMPUTE_RETURN_ON_ERROR(construct_gemmlowp_output_stage(src, weights, dst, gemmlowp_output_stage, fc_info.activation_info));
const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped
false, // is_b_reshaped
true, // reshape_b_only_on_first_run
0, // depth_output_gemm3d
false, // reinterpret_input_as_3d
fc_info.retain_internal_weights, // retain_internal_weights
gemmlowp_output_stage, // gemmlowp_output_stage
fc_info.fp_mixed_precision, // fp_mixed_precision
false, // fast_math
true, // broadcast_bias
ActivationLayerInfo()); // activation_info
if(is_data_type_quantized_asymmetric(src.data_type()))
{
const UniformQuantizationInfo iq_info = src.quantization_info().uniform();
const UniformQuantizationInfo wq_info = weights.quantization_info().uniform();
// Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
// Extract and negate src and weights offset
const QuantizationInfo src_quantization_info(iq_info.scale, -iq_info.offset);
const QuantizationInfo weights_quantization_info(wq_info.scale, -wq_info.offset);
// Validate gemmlowp function
ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixMultiplyCore::validate(&src.clone()->set_quantization_info(src_quantization_info),
&weights.clone()->set_quantization_info(weights_quantization_info),
bias,
&dst,
gemm_info));
}
else
{
ARM_COMPUTE_RETURN_ON_ERROR(ClGemm::validate(&src, &weights, bias, &dst, 1.f, 1.f, gemm_info));
}
return Status{};
}
} // namespace
ClFullyConnected::ClFullyConnected()
: _convert_weights(nullptr),
_flatten(nullptr),
_reshape_weights(nullptr),
_mm_gemm(nullptr),
_mm_gemmlowp(nullptr),
_aux_mem(Count)
{
}
ClFullyConnected::~ClFullyConnected() = default;
void ClFullyConnected::configure_mm(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *dst,
const FullyConnectedLayerInfo &fc_info)
{
GEMMLowpOutputStageInfo gemmlowp_output_stage;
construct_gemmlowp_output_stage(*src, *weights, *dst, gemmlowp_output_stage, fc_info.activation_info);
const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped
false, // is_b_reshaped
true, // reshape_b_only_on_first_run
0, // depth_output_gemm3d
false, // reinterpret_input_as_3d
fc_info.retain_internal_weights, // retain_internal_weights
gemmlowp_output_stage, // gemmlowp_output_stage
fc_info.fp_mixed_precision, // fp_mixed_precision
false, // fast_math
true, // broadcast_bias
fc_info.activation_info, // activation_info
fc_info.constant_weights); // constant_weights
if(_is_quantized)
{
// Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
// Extract and negate input and weights offset
const QuantizationInfo src_quantization_info = src->quantization_info();
const QuantizationInfo weights_quantization_info = weights->quantization_info();
TensorInfo src_info = src->clone()->set_quantization_info(src_quantization_info);
TensorInfo weights_info = weights->clone()->set_quantization_info(weights_quantization_info);
src_info.set_quantization_info(QuantizationInfo(src_quantization_info.uniform().scale, -src_quantization_info.uniform().offset));
weights_info.set_quantization_info(QuantizationInfo(weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset));
// Configure gemmlowp function
_mm_gemmlowp = std::make_unique<ClGemmLowpMatrixMultiplyCore>();
_mm_gemmlowp->configure(compile_context, &src_info, &weights_info, bias, dst, gemm_info);
}
else
{
// Configure matrix multiply kernel
_mm_gemm = std::make_unique<ClGemm>();
_mm_gemm->configure(compile_context, src, weights, bias, dst, 1.f, 1.f, gemm_info);
}
}
void ClFullyConnected::configure_conv_fc(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *dst,
const FullyConnectedLayerInfo &fc_info)
{
ARM_COMPUTE_ERROR_ON((weights->dimension(1) != (src->dimension(0) * src->dimension(1) * src->dimension(2))));
// If the fully connected layer is called after a convolution layer, the input tensor must be linearized
// Initialize output tensor for flatten
_flattened_src = src->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_flatten_shape(src)).set_data_layout(DataLayout::NCHW);
// Configure flatten kernel
_flatten = std::make_unique<ClFlatten>();
_flatten->configure(compile_context, src, &_flattened_src);
// Configure matrix multiply kernel
configure_mm(compile_context, &_flattened_src, weights, bias, dst, fc_info);
}
void ClFullyConnected::configure_fc_fc(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *dst,
const FullyConnectedLayerInfo &fc_info)
{
ARM_COMPUTE_ERROR_ON(src->dimension(0) != weights->dimension(1));
// Configure matrix multiply kernel
configure_mm(compile_context, src, weights, bias, dst, fc_info);
}
void ClFullyConnected::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst,
FullyConnectedLayerInfo fc_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst);
// Perform validate step
ARM_COMPUTE_ERROR_THROW_ON(ClFullyConnected::validate(src, weights, biases, dst, fc_info));
_are_weights_converted = true;
_are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
_is_fc_after_conv = true;
_is_quantized = is_data_type_quantized_asymmetric(src->data_type());
_is_prepared = fc_info.retain_internal_weights;
_weights_to_use = TensorInfo(*weights);
_weights_to_use_idx = ACL_SRC_1;
// With the Fully Connected layer we can have 4 different cases:
// 1) Convolution layer -> Fully Connected layer without batches
// 2) Fully Connected layer -> Fully Connected layer without batches
// 3) Convolution layer -> Fully Connected layer with batches
// 4) Fully Connected layer -> Fully Connected layer with batches
// Check if we have a fully connected layer with batches
const bool is_batched_fc_layer = dst->dimension(1) > 1;
if(is_batched_fc_layer)
{
_is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(src->tensor_shape().cbegin() + 3,
src->tensor_shape().cend(),
dst->tensor_shape().cbegin() + 1));
}
else
{
_is_fc_after_conv = src->num_dimensions() > 1;
}
ITensorInfo *weights_used = weights;
// Reshape weights if needed
if(!_are_weights_reshaped)
{
// Reshape the weights
_reshape_weights = std::make_unique<ClTranspose>();
_reshape_weights->configure(compile_context, weights, &_reshaped_weights);
weights_used = &_reshaped_weights;
_weights_to_use_idx = offset_int_vec(TransposedWeights);
}
// Convert weights if needed
if(_is_fc_after_conv && (src->data_layout() != fc_info.weights_trained_layout))
{
// Convert weights
_convert_weights = std::make_unique<ClConvertFullyConnectedWeights>();
_convert_weights->configure(compile_context,
weights_used,
&_converted_weights,
src->tensor_shape(),
fc_info.weights_trained_layout);
weights_used = &_converted_weights;
_weights_to_use_idx = offset_int_vec(ConvertedWeights);
_are_weights_converted = false;
}
if(_is_fc_after_conv)
{
// Fully Connected layer after a Convolution Layer without batches
configure_conv_fc(compile_context, src, weights_used, biases, dst, fc_info);
}
else
{
// Fully Connected layer after a Fully Connected Layer without batches
configure_fc_fc(compile_context, src, weights_used, biases, dst, fc_info);
}
// Update TensorInfo of final weights used (Need to be done in the end due to padding expansion)
_weights_to_use = *weights_used;
// Set auxiliary memory requirements
auto gemm_mem_req = (_is_quantized) ? _mm_gemmlowp->workspace() : _mm_gemm->workspace();
for(unsigned int i = 0; i < gemm_mem_req.size(); ++i)
{
_aux_mem[i] = gemm_mem_req[i];
}
if(_aux_mem[1].size > 0 || _aux_mem[2].size > 0) // Persistent weights memory on GEMMs
{
// Release permuted weights at the of prepare as they are further transposed by the assembly dispatch
_aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), MemoryLifetime::Prepare, _reshaped_weights.total_size());
_aux_mem[ConvertedWeights] = MemoryInfo(offset_int_vec(ConvertedWeights), MemoryLifetime::Prepare, _converted_weights.total_size());
}
else
{
// Release permuted weights at the of prepare as they are further transposed by the assembly dispatch
const auto transposed_wei_lft = (_weights_to_use_idx == offset_int_vec(TransposedWeights)) ? MemoryLifetime::Persistent : MemoryLifetime::Prepare;
const auto converted_wei_lft = (_weights_to_use_idx == offset_int_vec(ConvertedWeights)) ? MemoryLifetime::Persistent : MemoryLifetime::Prepare;
_aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), transposed_wei_lft, _reshaped_weights.total_size());
_aux_mem[ConvertedWeights] = MemoryInfo(offset_int_vec(ConvertedWeights), converted_wei_lft, _converted_weights.total_size());
}
_aux_mem[FlattenedSrc] = MemoryInfo(offset_int_vec(FlattenedSrc), MemoryLifetime::Temporary, _flattened_src.total_size());
}
Status ClFullyConnected::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst,
FullyConnectedLayerInfo fc_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights, dst);
ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
ARM_COMPUTE_RETURN_ERROR_ON(fc_info.activation_info.enabled() && is_data_type_quantized(src->data_type()) && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::RELU
&& fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU);
ARM_COMPUTE_RETURN_ERROR_ON(!fc_info.constant_weights && (!fc_info.are_weights_reshaped || fc_info.transpose_weights));
bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
bool is_fc_after_conv = true;
const ITensorInfo &flatten_src = TensorInfo(src->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_flatten_shape(src)).set_data_layout(DataLayout::NCHW));
const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights)));
const ITensorInfo &converted_weights = weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding()) : TensorInfo(*reshaped_weights.clone());
// With the Fully Connected layer we can have 4 different cases:
// 1) Convolution layer -> Fully Connected layer without batches
// 2) Fully Connected layer -> Fully Connected layer without batches
// 3) Convolution layer -> Fully Connected layer with batches
// 4) Fully Connected layer -> Fully Connected layer with batches
const ITensorInfo *src_to_use = src;
const ITensorInfo *weights_to_use = weights;
// Check if we have a fully connected layer with batches
const bool is_batched_fc_layer = dst->dimension(1) > 1;
if(is_batched_fc_layer)
{
is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(src->tensor_shape().cbegin() + 3,
src->tensor_shape().cend(),
dst->tensor_shape().cbegin() + 1));
}
else
{
is_fc_after_conv = src->num_dimensions() > 1;
}
if(!weights_reshaped)
{
// Validate reshape weights kernel
ARM_COMPUTE_RETURN_ON_ERROR(ClTranspose::validate(weights, &reshaped_weights));
weights_to_use = &reshaped_weights;
}
if(is_fc_after_conv && (src->data_layout() != fc_info.weights_trained_layout))
{
// Validate convert weights kernel
ARM_COMPUTE_RETURN_ON_ERROR(ClConvertFullyConnectedWeights::validate(weights_to_use,
&converted_weights,
src->tensor_shape(),
fc_info.weights_trained_layout));
weights_to_use = &converted_weights;
}
if(is_fc_after_conv)
{
// Fully Connected layer after a Convolution Layer without batches
ARM_COMPUTE_RETURN_ERROR_ON((weights_to_use->dimension(1) != (src->dimension(0) * src->dimension(1) * src->dimension(2))));
// Validate flatten kernel
ARM_COMPUTE_RETURN_ON_ERROR(ClFlatten::validate(src, &flatten_src));
src_to_use = &flatten_src;
}
else
{
// Fully Connected layer after a Fully Connected Layer without batches
ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != weights_to_use->dimension(1));
}
// Validate matrix multiply kernel
ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(*src_to_use, *weights_to_use, biases, *dst, fc_info));
return Status{};
}
void ClFullyConnected::run(ITensorPack &tensors)
{
prepare(tensors);
auto src = tensors.get_const_tensor(ACL_SRC_0);
CLAuxTensorHandler flattened_src(offset_int_vec(FlattenedSrc), _flattened_src, tensors, false);
CLAuxTensorHandler weights(_weights_to_use_idx, _weights_to_use, tensors, false);
// Linearize input if it comes from a convolutional layer
if(_is_fc_after_conv)
{
ITensorPack flatten_pack{ { ACL_SRC, src }, { ACL_DST, flattened_src.get() } };
_flatten->run(flatten_pack);
}
ITensorPack gemm_pack = tensors;
gemm_pack.add_const_tensor(ACL_SRC_0, (_is_fc_after_conv) ? flattened_src.get() : src);
if(_weights_to_use_idx != ACL_SRC_1)
{
gemm_pack.add_const_tensor(ACL_SRC_1, weights.get());
}
// Run matrix multiply
if(_is_quantized)
{
_mm_gemmlowp->run(gemm_pack);
}
else
{
_mm_gemm->run(gemm_pack);
}
}
void ClFullyConnected::prepare(ITensorPack &tensors)
{
if(!_is_prepared)
{
auto weights = tensors.get_const_tensor(ACL_SRC_1);
CLAuxTensorHandler reshaped_weights(offset_int_vec(TransposedWeights), _reshaped_weights, tensors, false);
CLAuxTensorHandler converted_weights(offset_int_vec(ConvertedWeights), _converted_weights, tensors, false);
// Pointer to current weights
const ITensor *cur_weights = weights;
// Reshape of the weights if needed (happens only once)
if(!_are_weights_reshaped)
{
// Run reshape weights kernel and mark weights as unused
ITensorPack transpose_pack{ { ACL_SRC, weights }, { ACL_DST, reshaped_weights.get() } };
_reshape_weights->run(transpose_pack);
cur_weights->mark_as_unused();
cur_weights = reshaped_weights.get();
_are_weights_reshaped = true;
}
// Convert weights if needed (happens only once)
if(!_are_weights_converted)
{
ITensorPack convert_pack{ { ACL_SRC, cur_weights }, { ACL_DST, converted_weights.get() } };
_convert_weights->run(convert_pack);
cur_weights->mark_as_unused();
cur_weights = converted_weights.get();
_are_weights_converted = true;
}
tensors.add_const_tensor(ACL_SRC_1, cur_weights);
// Prepare GEMM prepare and release unused weights
if(!_is_quantized)
{
_mm_gemm->prepare(tensors);
}
else
{
_mm_gemmlowp->prepare(tensors);
}
_is_prepared = true;
}
}
experimental::MemoryRequirements ClFullyConnected::workspace() const
{
return _aux_mem;
}
} // namespace opencl
} // namespace arm_compute
|