aboutsummaryrefslogtreecommitdiff
path: root/src/runtime/NEON/functions/NEGEMM.cpp
blob: 914f088bf55bb3f94afdffc0490fd1ee9cc3ef31 (plain)
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
/*
 * Copyright (c) 2017-2019 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/NEON/functions/NEGEMM.h"

#include "arm_compute/core/CPP/Validate.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/ITensor.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/runtime/NEON/NEScheduler.h"
#include "arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
#include "arm_compute/runtime/TensorAllocator.h"
#include "support/ToolchainSupport.h"

#include <cmath>

using namespace arm_compute::misc::shape_calculator;

namespace arm_compute
{
NEGEMM::NEGEMM(std::shared_ptr<IMemoryManager> memory_manager)
    : _memory_group(memory_manager), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _asm_glue(memory_manager), _ma_kernel(), _tmp_a(), _tmp_b(), _original_b(nullptr),
      _run_vector_matrix_multiplication(false), _run_addition(false), _reshape_b_only_on_first_run(false), _is_prepared(false)
{
}

void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta, const GEMMInfo &gemm_info)
{
    ARM_COMPUTE_ERROR_THROW_ON(NEGEMM::validate(a->info(), b->info(), (c != nullptr) ? c->info() : nullptr, d->info(), alpha, beta, gemm_info));

    // Check if we need to reshape the matrix B only on the first run
    _is_prepared                      = false;
    _reshape_b_only_on_first_run      = gemm_info.reshape_b_only_on_first_run();
    _run_vector_matrix_multiplication = a->info()->dimension(1) < 2;
    _original_b                       = b;

    bool run_optimised = c == nullptr && bool(NEGEMMAssemblyDispatch::validate(a->info(), b->info(), d->info(), alpha, beta, _reshape_b_only_on_first_run));

    if(run_optimised)
    {
        if(MEMInfo::get_policy() == MemoryPolicy::MINIMIZE)
        {
            _asm_glue.configure(a, b, d, alpha, beta, false);
        }
        else
        {
            _asm_glue.configure(a, b, d, alpha, beta, _reshape_b_only_on_first_run);
        }
        ARM_COMPUTE_ERROR_ON(!_asm_glue.is_configured());
    }
    else
    {
        if(_run_vector_matrix_multiplication)
        {
            // Configure the matrix multiply kernel
            _mm_kernel.configure(a, b, d, alpha, false);
        }
        else
        {
            TensorShape shape_tmp_a = a->info()->tensor_shape();
            TensorShape shape_tmp_b = b->info()->tensor_shape();

            shape_tmp_a.set(0, a->info()->dimension(0) * 4);
            shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.0f));

            const unsigned int transpose_w = 16 / data_size_from_type(b->info()->data_type());
            shape_tmp_b.set(0, b->info()->dimension(1) * transpose_w);
            shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / static_cast<float>(transpose_w)));

            TensorInfo info_a = a->info()->clone()->set_tensor_shape(shape_tmp_a).set_is_resizable(true);
            TensorInfo info_b = b->info()->clone()->set_tensor_shape(shape_tmp_b).set_is_resizable(true);

            _tmp_a.allocator()->init(info_a);
            _tmp_b.allocator()->init(info_b);

            // Manage intermediate buffers
            _memory_group.manage(&_tmp_a);
            if(!_reshape_b_only_on_first_run)
            {
                _memory_group.manage(&_tmp_b);
            }

            int m = a->info()->dimension(1);
            int n = b->info()->dimension(0);
            int k = a->info()->dimension(0);

            // Configure interleave kernel
            _interleave_kernel.configure(a, &_tmp_a);

            // Configure transpose kernel
            _transpose_kernel.configure(b, &_tmp_b);

            // Configure matrix multiplication kernel
            _mm_kernel.configure(&_tmp_a, &_tmp_b, d, alpha, true, GEMMReshapeInfo(m, n, k));

            // Allocate once the all configure methods have been called
            _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, d, beta);
            _run_addition = true;
        }
    }
}

Status NEGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info)
{
    ARM_COMPUTE_UNUSED(alpha);

    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(a);
    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F16, DataType::F32);
    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output);
    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_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(gemm_info.depth_output_gemm3d() != 0);
        ARM_COMPUTE_RETURN_ERROR_ON(gemm_info.reinterpret_input_as_3d());
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, c);
        ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != c->dimension(1), "The C matrix must have the same number of rows as the matrix A");
        ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != c->dimension(0), "The C matrix must have the same number of columns as the matrix B");
    }

    if(output->total_size() != 0)
    {
        ARM_COMPUTE_RETURN_ERROR_ON(b->dimension(0) != output->dimension(0));
        if(gemm_info.depth_output_gemm3d() != 0)
        {
            if(gemm_info.reinterpret_input_as_3d())
            {
                ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1));
                ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(2) != output->dimension(2));
            }
            else
            {
                ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1) * output->dimension(2));
            }
        }
        else
        {
            ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1));
        }
    }

    // Check if we need to run the optimized assembly kernel
    const bool run_optimised = c == nullptr && bool(NEGEMMAssemblyDispatch::validate(a, b, output, alpha, beta, true));

    if(!run_optimised)
    {
        ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.reinterpret_input_as_3d(), "NEGEMM cannot reinterpret the input tensor as 3D");
        ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.depth_output_gemm3d() != 0, "NEGEMM cannot reinterpret the output tensor as 3D");

        // Check if the first input tensor is a vector.
        const bool run_vector_matrix_multiplication = a->dimension(1) < 2;
        // Check if we need to reshape the matrix A and matrix B
        const bool run_interleave_transpose = !run_vector_matrix_multiplication && !(gemm_info.reshape_b_only_on_first_run());

        // Arguments used by GEMMReshapeInfo
        // If we pass the matrix A and matrix B reshaped to NEGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to NEGEMMReshapeInfo
        // in order to know how the matrices have been reshaped
        const int m                         = a->dimension(1);
        const int n                         = b->dimension(0);
        const int k                         = a->dimension(0);
        int       mult_transpose1xW_width   = 1;
        int       mult_interleave4x4_height = 1;

        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, gemm_info.depth_output_gemm3d());

        const ITensorInfo *matrix_a_info = a;
        const ITensorInfo *matrix_b_info = b;

        TensorInfo tmp_a_info{};
        TensorInfo tmp_b_info{};
        TensorInfo tmp_output_info = *output->clone();

        if(run_interleave_transpose)
        {
            matrix_a_info = &tmp_a_info;
            matrix_b_info = &tmp_b_info;

            // Validate interleave kernel
            auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d())));
            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(a, &tmp_a_info));

            // Validate transpose kernel
            auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width)));
            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(b, &tmp_b_info));
        }

        // Validate matrix multiply
        auto_init_if_empty(tmp_output_info, matrix_a_info->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, run_interleave_transpose, reshape_info)));
        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, &tmp_output_info, alpha, run_interleave_transpose, reshape_info));
    }

    // Validate matrix addition kernel
    if(beta != 0 && c != nullptr)
    {
        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixAdditionKernel::validate(c, output, beta));
    }

    return Status{};
}

void NEGEMM::run()
{
    prepare();

    if(_asm_glue.is_configured())
    {
        _memory_group.acquire();
        _asm_glue.run();
        _memory_group.release();
    }
    else
    {
        _memory_group.acquire();

        if(!_run_vector_matrix_multiplication)
        {
            // Run interleave kernel
            NEScheduler::get().schedule(&_interleave_kernel, Window::DimY);

            if(!_reshape_b_only_on_first_run)
            {
                // Run transpose kernel
                NEScheduler::get().schedule(&_transpose_kernel, Window::DimY);
            }
        }

        NEScheduler::get().schedule(&_mm_kernel, _run_vector_matrix_multiplication ? Window::DimX : Window::DimY);

        _memory_group.release();

        // Run matrix addition kernel
        if(_run_addition)
        {
            NEScheduler::get().schedule(&_ma_kernel, Window::DimY);
        }
    }
}

void NEGEMM::prepare()
{
    if(!_is_prepared)
    {
        if(_asm_glue.is_configured())
        {
            ARM_COMPUTE_ERROR_ON(!_original_b->is_used());

            _asm_glue.prepare();
        }
        else if(_reshape_b_only_on_first_run && !_run_vector_matrix_multiplication && !_asm_glue.is_configured())
        {
            ARM_COMPUTE_ERROR_ON(!_original_b->is_used());

            _tmp_b.allocator()->allocate();
            NEScheduler::get().schedule(&_transpose_kernel, Window::DimY);
            _original_b->mark_as_unused();
        }

        _is_prepared = true;
    }
}
} // namespace arm_compute