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
|
/*
* Copyright (c) 2018-2019 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.
*/
#ifndef __ARM_COMPUTE_NEGEMMINTERLEAVEDWRAPPER_H__
#define __ARM_COMPUTE_NEGEMMINTERLEAVEDWRAPPER_H__
#include "arm_compute/core/NEON/kernels/assembly/Helpers.h"
#include "arm_compute/core/NEON/kernels/assembly/INEGEMMWrapperKernel.h"
#include "arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedMatrixMultiplyWrapper.h"
#include "arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedPrepareBWrapperKernel.h"
#include "arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedTransformAWrapper.h"
#include "arm_compute/runtime/IFunction.h"
#include "arm_compute/runtime/IMemoryManager.h"
#include "arm_compute/runtime/IScheduler.h"
#include "arm_compute/runtime/IWeightsManager.h"
#include "arm_compute/runtime/MemoryGroup.h"
#include "arm_compute/runtime/Tensor.h"
#include <memory>
namespace arm_compute
{
// Forward declarations
class ITensor;
/** Buffer manager used when reshaping B on the fly
*
* The typical workflow is:
* - lock_to_reshape_if_needed()
* - If the previous lock was successful: mark_as_reshaped()
* - wait_for_reshaping() wait for the reshaping to be complete
* - mark_as_unused() once the thread is done using this given buffer.
*
* Calls for different indices might be interleaved, however the calls for a given index must always be in that order.
*/
class IBufferManager
{
public:
/** Lock a buffer for the given index if it's available else return
*
* @param[in] index Index of the buffer to lock
*
* @return True if the buffer has been successfully locked, false if it's already reshaped / being reshaped.
*/
virtual bool lock_to_reshape_if_needed(unsigned int index) = 0;
/** Mark a buffer previously locked as reshaped
*
* @pre The thread calling this function must have locked the given buffer through lock_to_reshape_if_needed()
*
* @param[in] index Index of the buffer to mark as reshaped
*/
virtual void mark_as_reshaped(unsigned int index) = 0;
/** Block until the given buffer is marked as reshaped
*
* @param[in] index Index of the buffer
*/
virtual void wait_for_reshaping(unsigned int index) = 0;
/** Mark a reshaped buffer as unused
*
* Once all the users have marked a buffer as unused then it goes back to being free
*/
virtual void mark_as_unused(unsigned int index) = 0;
/** Number of buffers used internally
*
* @return The number of buffers used by the manager.
*/
virtual unsigned int num_buffers() const = 0;
/** Default destructor */
virtual ~IBufferManager() = default;
};
/** Equivalent to arm_gemm::GemmInterleaved but using Compute Library types.
*/
class NEGEMMInterleavedWrapper : public IFunction
{
public:
NEGEMMInterleavedWrapper(std::shared_ptr<IMemoryManager> memory_manager = nullptr, IWeightsManager *weights_manager = nullptr);
~NEGEMMInterleavedWrapper() = default;
NEGEMMInterleavedWrapper(const NEGEMMInterleavedWrapper &) = delete;
NEGEMMInterleavedWrapper &operator=(const NEGEMMInterleavedWrapper &) = delete;
/** Initialise the kernel's input and output.
*
* @note The input and output tensor must have the same dimensions
*
* @param[in] a Input tensor (Matrix A)
* @param[in] b Input tensor (Matrix B)
* @param[out] c Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0.
* @param[in] alpha Scalar multiplier to apply to AB matrix product.
* @param[in] beta Scalar multiplier to apply to input C matrix before adding product.
* @param[in] gemm_info GEMM meta-data
*/
void configure(const ITensor *a, const ITensor *b, ITensor *c, float alpha, float beta, const GEMMInfo &gemm_info);
// Inherited methods overridden:
void run() override;
void prepare() override;
private:
MemoryGroup _memory_group;
bool _is_prepared{ false };
bool _pretranspose_b{ false };
Window _block_walker{};
Window _batch_window{};
const ITensor *_a{ nullptr };
const ITensor *_b{ nullptr };
ITensor *_c{ nullptr };
Tensor _transformed_b{};
Tensor _transformed_a{};
Tensor _tmp_c{};
INEGEMMWrapperKernel::Params _params{};
BlockSizes _block_sizes{};
std::unique_ptr<NEGEMMInterleavedPrepareBWrapperKernel> _prepare_b{ nullptr };
std::unique_ptr<NEGEMMInterleavedTransformAWrapper> _transform_a{ nullptr };
std::unique_ptr<NEGEMMInterleavedMatrixMultiplyWrapper> _matrix_multiply{ nullptr };
std::unique_ptr<IBufferManager> _buffer_manager{ nullptr };
std::vector<TransformAWorkload> _a_workloads{};
std::vector<PrepareBWorkload> _b_workloads{};
std::vector<MatrixMultiplyWorkload> _mm_workloads{};
std::vector<IScheduler::Workload> _workloads{};
std::string _tag{};
};
} // namespace arm_compute
#endif /* __ARM_COMPUTE_NEGEMMINTERLEAVEDWRAPPER_H__ */
|
