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
|
/*
* Copyright (c) 2017-2018 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 "arm_compute/runtime/CL/functions/CLGEMM.h"
#include "arm_compute/core/CL/ICLTensor.h"
#include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h"
#include "arm_compute/core/CL/kernels/CLGEMMMatrixAdditionKernel.h"
#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h"
#include "arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/runtime/CL/CLScheduler.h"
#include "arm_compute/runtime/ITensorAllocator.h"
using namespace arm_compute;
namespace
{
inline bool is_interleaved_transposed(int m, int n, int k, DataType data_type, bool reshape_b_only_on_first_run, GPUTarget gpu_target)
{
bool flag = true;
if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX))
{
// COMPMID-852
if(k > 256 && m > 4 && is_data_type_float(data_type) && reshape_b_only_on_first_run)
{
const float scale = k < 1024 ? 2.0f : 2.5f;
flag = (scale * n) > ((1.66f * n) + 38.4f);
}
else
{
flag = false;
}
}
else
{
// We reshape the matrices only if we do not have the vector-by-matrix case and we reshape the matrix B only once
flag = m != 1 && reshape_b_only_on_first_run;
}
return flag;
}
Status validate_arguments(const ITensorInfo *a, const ITensorInfo *b, const ICLTensor *c, const ITensorInfo *output, const float alpha, const float beta, const GEMMInfo &gemm_info = GEMMInfo())
{
ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b);
ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
if(c != nullptr)
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, c->info());
ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != c->info()->dimension(1), "The matrix C must have the same number of rows as the matrix A");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != c->info()->dimension(0), "The matrix C must have the same number of columns as the matrix B");
}
if(output->total_size() != 0)
{
ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != output->dimension(0), "The output matrix must have the same number of columns as the matrix B");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != output->dimension(1), "The output matrix must have the same number of rows as the matrix A");
}
ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(0) != b->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
ARM_COMPUTE_UNUSED(alpha);
ARM_COMPUTE_UNUSED(beta);
return Status{};
}
} // namespace
CLGEMM::CLGEMM(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _original_b(nullptr), _is_interleaved_transposed(false),
_run_addition(false), _reshape_b_only_on_first_run(false), _is_prepared(false)
{
}
void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
// Perform validation step
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(a->info(), b->info(), c, output->info(), alpha, beta, gemm_info));
// Store original b matrix
_original_b = b;
// Check if we need to reshape the matrix B only on the first run
_reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run();
_is_prepared = false;
const ICLTensor *matrix_a = a;
const ICLTensor *matrix_b = b;
// Get the GPU target
const GPUTarget gpu_target = CLScheduler::get().target();
// Set the target for the kernels
_interleave_kernel.set_target(gpu_target);
_mm_kernel.set_target(gpu_target);
// Arguments used by GEMMReshapeInfo
// If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo
// in order to know how the matrices have been reshaped
const int m = a->info()->dimension(1);
const int n = b->info()->dimension(0);
const int k = a->info()->dimension(0);
int mult_transpose1xW_width = 1;
int mult_interleave4x4_height = 1;
if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX))
{
mult_transpose1xW_width = 4;
mult_interleave4x4_height = 2;
}
// Check if we need to reshape the matrix A and matrix B
_is_interleaved_transposed = is_interleaved_transposed(m, n, k, a->info()->data_type(), _reshape_b_only_on_first_run, gpu_target);
if(_is_interleaved_transposed)
{
matrix_a = &_tmp_a;
matrix_b = &_tmp_b;
// Manage intermediate buffers
_memory_group.manage(&_tmp_a);
if(!_reshape_b_only_on_first_run)
{
_memory_group.manage(&_tmp_b);
}
// _tmp_a and _tmp_b will be auto configured in _interleave_kernel and in _transpose_kernel
// Configure interleave kernel
_interleave_kernel.configure(a, &_tmp_a, mult_interleave4x4_height);
// Configure transpose kernel
_transpose_kernel.configure(b, &_tmp_b, mult_transpose1xW_width);
}
_mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height));
if(_is_interleaved_transposed)
{
// Allocate intermediate tensors
_tmp_a.allocator()->allocate();
if(!_reshape_b_only_on_first_run)
{
_tmp_b.allocator()->allocate();
}
}
// Configure matrix addition kernel
if(beta != 0 && c != nullptr)
{
_ma_kernel.configure(c, output, beta);
_run_addition = true;
}
}
Status CLGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ICLTensor *c, const ITensorInfo *output, const float alpha, const float beta, const GEMMInfo &gemm_info)
{
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(a, b, c, output, alpha, beta, gemm_info));
return Status{};
}
void CLGEMM::run()
{
prepare();
_memory_group.acquire();
if(_is_interleaved_transposed)
{
// Run interleave kernel
CLScheduler::get().enqueue(_interleave_kernel, false);
if(!_reshape_b_only_on_first_run)
{
// Run transpose kernel
CLScheduler::get().enqueue(_transpose_kernel, false);
}
}
// Run matrix multiply kernel
CLScheduler::get().enqueue(_mm_kernel, !_run_addition);
// Run matrix addition kernel
if(_run_addition)
{
CLScheduler::get().enqueue(_ma_kernel);
}
_memory_group.release();
}
void CLGEMM::prepare()
{
if(!_is_prepared)
{
if(_is_interleaved_transposed && _reshape_b_only_on_first_run)
{
// Run transpose kernel
_tmp_b.allocator()->allocate();
CLScheduler::get().enqueue(_transpose_kernel, false);
_original_b->mark_as_unused();
}
CLScheduler::get().queue().finish();
_is_prepared = true;
}
}
|
