Compute Library
 22.11
CpuGemmMatrixMultiplyKernel.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2017-2022 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
25 
28 #include "arm_compute/core/Types.h"
31 #include "src/core/CPP/Validate.h"
36 
37 namespace arm_compute
38 {
39 namespace cpu
40 {
41 namespace kernels
42 {
43 namespace
44 {
45 static const std::vector<CpuGemmMatrixMultiplyKernel::GemmMatrixMulKernel> available_kernels =
46 {
47  {
48  "neon_fp32_gemm_matrix_mul",
49  [](const DataTypeISASelectorData & data)
50  {
51  return (data.dt == DataType::F32);
52  },
54  },
55  {
56  "neon_fp16_gemm_matrix_mul",
57  [](const DataTypeISASelectorData & data)
58  {
59  return (data.dt == DataType::F16) && data.isa.fp16;
60  },
62  },
63 };
64 
65 inline Status validate_arguments(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *dst, float alpha, bool is_interleaved, const GEMMReshapeInfo &reshape_info)
66 {
67  ARM_COMPUTE_UNUSED(alpha);
68 
72 
73  if(!is_interleaved)
74  {
75  ARM_COMPUTE_RETURN_ERROR_ON(lhs->dimension(0) != rhs->dimension(1));
76 
77  if(dst->total_size() != 0)
78  {
79  ARM_COMPUTE_RETURN_ERROR_ON(rhs->dimension(0) != dst->dimension(0));
80  ARM_COMPUTE_RETURN_ERROR_ON(lhs->dimension(1) != dst->dimension(1));
82  }
83  }
84  else
85  {
86  const int m = reshape_info.m();
87  const int n = reshape_info.n();
88  const int k = reshape_info.k();
89  const int mult_transpose1xW_width = reshape_info.mult_transpose1xW_width();
90  const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height();
91 
92  /* Interleave */
93  TensorShape tensor_shape0{ lhs->tensor_shape() };
94  tensor_shape0.set(0, k);
95  tensor_shape0.set(1, m);
96 
97  const TensorInfo tensor_info0 = lhs->clone()->set_tensor_shape(tensor_shape0);
98  const TensorInfo tensor_info_reshaped0 = lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_interleaved_shape(tensor_info0, mult_interleave4x4_height));
99  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(lhs, &tensor_info_reshaped0);
100 
101  if(n != 0) /* Transpose */
102  {
103  TensorShape tensor_shape1{ rhs->tensor_shape() };
104  tensor_shape1.set(0, n);
105  tensor_shape1.set(1, k);
106 
107  const TensorInfo tensor_info1 = rhs->clone()->set_tensor_shape(tensor_shape1);
108  const TensorInfo tensor_info_reshaped1 = rhs->clone()->set_tensor_shape(misc::shape_calculator::compute_transpose1xW_with_element_size_shape(tensor_info1, mult_transpose1xW_width));
109  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(rhs, &tensor_info_reshaped1);
110  }
111 
112  if(dst->total_size() != 0)
113  {
114  if(n != 0)
115  {
116  ARM_COMPUTE_RETURN_ERROR_ON(dst->dimension(0) != static_cast<size_t>(n));
117  }
118  ARM_COMPUTE_RETURN_ERROR_ON(dst->dimension(1) != static_cast<size_t>(m));
120  }
121  }
122 
123  return Status{};
124 }
125 
126 } // namespace
127 
128 void CpuGemmMatrixMultiplyKernel::configure(const ITensorInfo *lhs, const ITensorInfo *rhs, ITensorInfo *dst, float alpha, bool is_interleaved, const GEMMReshapeInfo &reshape_info)
129 {
130  ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst);
131 
132  // dst tensor auto inizialitation if not yet initialized
133  TensorShape tensor_shape{ lhs->tensor_shape() };
134  tensor_shape.set(0, is_interleaved ? reshape_info.n() : rhs->dimension(0));
135  tensor_shape.set(1, is_interleaved ? reshape_info.m() : lhs->dimension(1));
136 
137  auto_init_if_empty(*dst, lhs->clone()->set_tensor_shape(tensor_shape));
138 
139  // Perform validate step
140  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(lhs, rhs, dst, alpha, is_interleaved, reshape_info));
141 
142  _alpha = alpha;
143 
144  // Configure kernel window
145  Window win{};
146 
147  // Check if the dst tensor is a vector. If so,the kernel runs the vector-matrix multiplication
148  const bool is_dst_vector = (dst->dimension(1) == 1);
149  if(is_dst_vector)
150  {
151  const unsigned int num_elems_processed_per_iteration_x = (lhs->data_type() == DataType::F32) ? 16 : 32;
152 
153  win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x));
154  }
155  else
156  {
157  constexpr unsigned int num_elems_processed_per_iteration_x = 8;
158  constexpr unsigned int num_elems_processed_per_iteration_y = 4;
159 
160  win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
161  }
162 
165  _func = uk->ukernel;
166 
167  ICPPKernel::configure(win);
168 }
169 
170 Status CpuGemmMatrixMultiplyKernel::validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *dst, float alpha, bool is_interleaved,
171  const GEMMReshapeInfo &reshape_info)
172 {
173  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(lhs, rhs, dst, alpha, is_interleaved, reshape_info));
174 
175  return Status{};
176 }
177 
179 {
182  ARM_COMPUTE_ERROR_ON(tensors.empty());
183  ARM_COMPUTE_ERROR_ON(_func == nullptr);
184 
185  const ITensor *lhs = tensors.get_const_tensor(TensorType::ACL_SRC_0);
186  const ITensor *rhs = tensors.get_const_tensor(TensorType::ACL_SRC_1);
187  ITensor *dst = tensors.get_tensor(TensorType::ACL_DST);
188 
189  const bool is_dst_vector = (dst->info()->dimension(1) == 1);
190  (*_func)(lhs, rhs, dst, window, info, _alpha, is_dst_vector);
191 }
192 
194 {
195  return "CpuGemmMatrixMultiplyKernel";
196 }
197 
198 const std::vector<CpuGemmMatrixMultiplyKernel::GemmMatrixMulKernel> &CpuGemmMatrixMultiplyKernel::get_available_kernels()
199 {
200  return available_kernels;
201 }
202 } // namespace kernels
203 } // namespace cpu
204 } // namespace arm_compute
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
Shape of a tensor.
Definition: TensorShape.h:39
static const auto * get_implementation(const SelectorType &selector, KernelSelectionType selection_type=KernelSelectionType::Supported)
Micro-kernel selector.
Definition: ICpuKernel.h:53
TensorShape compute_transpose1xW_with_element_size_shape(const ITensorInfo &b, int mult_transpose1xW_width=1)
Calculate the transposed 1xW width element shape.
#define ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(tensor)
Definition: Validate.h:115
#define REGISTER_FP16_NEON(func_name)
Definition: Registrars.h:48
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
bool empty() const
Checks if pack is empty.
Definition: ITensorPack.cpp:80
void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override
Execute the kernel on the passed window.
GEMM reshape information class.
Definition: Types.h:2159
#define REGISTER_FP32_NEON(func_name)
Definition: Registrars.h:74
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
virtual DataType data_type() const =0
Data type used for each element of the tensor.
1 channel, 1 F32 per channel
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Status class.
Definition: Error.h:52
Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const PadStrideInfo &conv_info)
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
static Status validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *dst, float alpha, bool is_interleaved, const GEMMReshapeInfo &reshape_info)
Static function to check if given info will lead to a valid configuration of CpuGemmMatrixMultiplyKer...
Interface for CPU tensor.
Definition: ITensor.h:36
TensorShape compute_interleaved_shape(const ITensorInfo &a, int mult_interleave4x4_height=1, bool reinterpret_input_as_3d=false)
Calculate the interleaved shape of an input tensor.
Copyright (c) 2017-2022 Arm Limited.
1 channel, 1 F16 per channel
static const std::vector< GemmMatrixMulKernel > & get_available_kernels()
int n() const
Number of matrix B columns.
Definition: Types.h:2197
const ITensor * get_const_tensor(int id) const
Get constant tensor of a given id.
Definition: ITensorPack.cpp:54
void neon_fp16_gemm_matrix_mul(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha, const bool is_dst_vector)
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
void neon_fp32_gemm_matrix_mul(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha, const bool is_dst_vector)
Definition: fp32.cpp:31
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
const char * name() const override
Name of the kernel.
void configure(const ITensorInfo *lhs, const ITensorInfo *rhs, ITensorInfo *dst, float alpha, bool is_interleaved, const GEMMReshapeInfo &reshape_info=GEMMReshapeInfo())
Initialise the kernel&#39;s input and output.
bool auto_init_if_empty(ITensorInfo &info, const TensorShape &shape, int num_channels, DataType data_type, QuantizationInfo quantization_info=QuantizationInfo())
Auto initialize the tensor info (shape, number of channels and data type) if the current assignment i...
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:915
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
ITensor * get_tensor(int id)
Get tensor of a given id from the pac.
Definition: ITensorPack.cpp:64
Information about executing thread and CPU.
Definition: CPPTypes.h:179
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
Definition: Validate.h:439
int m() const
Number of matrix A rows.
Definition: Types.h:2189
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:541
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:788
Tensor packing service.
Definition: ITensorPack.h:39
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:157
static CPUInfo & get()
Access the KernelLibrary singleton.
Definition: CPPTypes.cpp:40
Describe a multidimensional execution window.
Definition: Window.h:39
TensorShape & set(size_t dimension, size_t value, bool apply_dim_correction=true, bool increase_dim_unit=true)
Accessor to set the value of one of the dimensions.
Definition: TensorShape.h:79
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:201
cpuinfo::CpuIsaInfo get_isa() const
Gets the current cpu&#39;s ISA information.
Definition: CPPTypes.cpp:124