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
|
/*
* Copyright (c) 2018-2020 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_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H
#define ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H
#include "arm_compute/core/CL/ICLKernel.h"
#include "arm_compute/core/KernelDescriptors.h"
namespace arm_compute
{
class ICLTensor;
/** OpenCL kernel to multiply matrices when both the input matrices LHS (input0) and RHS (input1) have been reshaped
*
* @note The input matrices @p input0 and @p input1 must be reshaped through @ref CLGEMMReshapeLHSMatrixKernel and @ref CLGEMMReshapeRHSMatrixKernel
*/
class CLGEMMMatrixMultiplyReshapedKernel : public ICLKernel
{
public:
/** Default Constructor */
CLGEMMMatrixMultiplyReshapedKernel();
/** Prevent instances of this class from being copied (As this class contains pointers) */
CLGEMMMatrixMultiplyReshapedKernel(const CLGEMMMatrixMultiplyReshapedKernel &) = delete;
/** Prevent instances of this class from being copied (As this class contains pointers) */
CLGEMMMatrixMultiplyReshapedKernel &operator=(const CLGEMMMatrixMultiplyReshapedKernel &) = delete;
/** Allow instances of this class to be moved */
CLGEMMMatrixMultiplyReshapedKernel(CLGEMMMatrixMultiplyReshapedKernel &&) = default;
/** Allow instances of this class to be moved */
CLGEMMMatrixMultiplyReshapedKernel &operator=(CLGEMMMatrixMultiplyReshapedKernel &&) = default;
/** Initialise the kernel's input and output.
*
* @note The F16 computation also supports mixed precision through the gemm_info.fp_mixed_precision flag.
* Mixed precision combines different floating precisions during the computation, in particular, F32 for the accumulations and F16 for the
* multiplications. i.e. float c = (half)a * (half)b
*
* @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will fetch the RHS data using the OpenCL read_image built-in function.
* Reading from the OpenCL image object can increase the performance. However, since the OpenCL image object is created importing the OpenCL buffer,
* the following conditions are required:
* -# rhs_info.n0 can only be 4, 8 and 16
* -# rhs_info.k0 can only be 4, 8 and 16
* -# Data type can only be F32
* -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension
* -# The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement
* -# input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
* -# input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
*
* @param[in] input0 Input tensor containing the LHS reshaped matrix. Data type supported: F16/F32 (only F32 if rhs_info.export_to_cl_image = true). The number of dimensions for the LHS matrix must be less or equal than 4
* @param[in] input1 Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0. The number of dimensions for the RHS matrix must be less or equal than 3
* @param[in] input2 Input tensor containing the bias matrix. Data type supported: same as @p input0.
* @param[out] output Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
* @param[in] alpha Weight of the matrix product
* @param[in] beta Weight of the matrix bias
* @param[in] lhs_info LHS matrix information used for reshaping the input0 tensor. Only the following values are supported:
* lhs_info.m0: 2,3,4,5,6,7,8
* lhs_info.k0: 2,3,4,8,16
* lhs_info.transpose: false
* @param[in] rhs_info RHS matrix information used for reshaping the input1 tensor. Only the following values are supported:
* rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
* rhs_info.k0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
* rhs_info.transpose: true
* @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices
*
* @note lhs_info.k0 must be equal to rhs_info.k0
*/
void configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
const GEMMRHSMatrixInfo &rhs_info,
const GEMMKernelInfo &gemm_info);
/** Initialise the kernel's input and output.
*
* @note The F16 computation also supports mixed precision through the gemm_info.fp_mixed_precision flag.
* Mixed precision combines different floating precisions during the computation, in particular, F32 for the accumulations and F16 for the
* multiplications. i.e. float c = (half)a * (half)b
*
* @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will fetch the RHS data using the OpenCL read_image built-in function.
* Reading from the OpenCL image object can increase the performance. However, since the OpenCL image object is created importing the OpenCL buffer,
* the following conditions are required:
* -# rhs_info.n0 can only be 4, 8 and 16
* -# rhs_info.k0 can only be 4, 8 and 16
* -# Data type can only be F32
* -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension
* -# The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement
* -# input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
* -# input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
*
* @param[in] compile_context The compile context to be used.
* @param[in] input0 Input tensor containing the LHS reshaped matrix. Data type supported: F16/F32 (only F32 if rhs_info.export_to_cl_image = true). The number of dimensions for the LHS matrix must be less or equal than 4
* @param[in] input1 Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0. The number of dimensions for the RHS matrix must be less or equal than 3
* @param[in] input2 Input tensor containing the bias matrix. Data type supported: same as @p input0.
* @param[out] output Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
* @param[in] alpha Weight of the matrix product
* @param[in] beta Weight of the matrix bias
* @param[in] lhs_info LHS matrix information used for reshaping the input0 tensor. Only the following values are supported:
* lhs_info.m0: 2,3,4,5,6,7,8
* lhs_info.k0: 2,3,4,8,16
* lhs_info.transpose: false
* @param[in] rhs_info RHS matrix information used for reshaping the input1 tensor. Only the following values are supported:
* rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
* rhs_info.k0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
* rhs_info.transpose: true
* @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices
*
* @note lhs_info.k0 must be equal to rhs_info.k0
*/
void configure(const CLCompileContext &compile_context, const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta,
const GEMMLHSMatrixInfo &lhs_info,
const GEMMRHSMatrixInfo &rhs_info,
const GEMMKernelInfo &gemm_info);
/** Static function to check if given info will lead to a valid configuration of @ref CLGEMMMatrixMultiplyReshapedKernel
*
* @note The F16 computation also supports mixed precision through the gemm_info.fp_mixed_precision flag.
* Mixed precision combines different floating precisions during the computation, in particular, F32 for the accumulations and F16 for the
* multiplications. i.e. float c = (half)a * (half)b
*
* @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will fetch the RHS data using the OpenCL read_image built-in function.
* Reading from the OpenCL image object can increase the performance. However, since the OpenCL image object is created importing the OpenCL buffer,
* the following conditions are required:
* -# rhs_info.n0 can only be 4, 8 and 16
* -# rhs_info.k0 can only be 4, 8 and 16
* -# Data type can only be F32
* -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension
* -# The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement
* -# input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
* -# input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
*
* @param[in] input0 Input tensor containing the LHS reshaped matrix. Data type supported: F16/F32 (only F32 if rhs_info.export_to_cl_image = true). The number of dimensions for the LHS matrix must be less or equal than 4
* @param[in] input1 Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0. The number of dimensions for the RHS matrix must be less or equal than 3
* @param[in] input2 Input tensor info containing the bias matrix. Data type supported: same as @p input0.
* @param[in] output Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
* @param[in] alpha Weight of the matrix product
* @param[in] beta Weight of the matrix bias
* @param[in] lhs_info LHS matrix information used for reshaping the input0 tensor. Only the following values are supported:
* lhs_info.m0: 2,3,4,5,6,7,8
* lhs_info.k0: 2,3,4,8,16
* lhs_info.transpose: false
* @param[in] rhs_info RHS matrix information used for reshaping the input1 tensor. Only the following values are supported:
* rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
* rhs_info.k0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
* rhs_info.transpose: true
* @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices
*
* @note lhs_info.k0 must be equal to rhs_info.k0
*
* @return a status
*/
static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
const GEMMRHSMatrixInfo &rhs_info,
const GEMMKernelInfo &gemm_info);
// Inherited methods overridden:
void run(const Window &window, cl::CommandQueue &queue) override;
private:
const ICLTensor *_input0;
const ICLTensor *_input1;
const ICLTensor *_input2;
ICLTensor *_output;
bool _slide_matrix_b;
bool _reinterpret_output_as_3d;
bool _use_dummy_work_items;
bool _add_bias;
bool _broadcast_bias;
bool _export_to_cl_image;
unsigned int _k;
};
} // namespace arm_compute
#endif /*ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H*/
|
