1 /*
2 * Copyright (c) 2018-2020 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 */
24 #include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
25
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/CL/OpenCL.h"
30 #include "arm_compute/core/Helpers.h"
31 #include "arm_compute/core/TensorInfo.h"
32 #include "arm_compute/core/Utils.h"
33 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
34 #include "src/core/AccessWindowStatic.h"
35 #include "src/core/CL/CLValidate.h"
36 #include "src/core/CL/gemm/CLGEMMHelpers.h"
37 #include "src/core/helpers/AutoConfiguration.h"
38 #include "src/core/helpers/WindowHelpers.h"
39 #include "support/StringSupport.h"
40
41 namespace arm_compute
42 {
43 using namespace arm_compute::misc::shape_calculator;
44
45 namespace
46 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,const GEMMRHSMatrixInfo & rhs_info)47 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const GEMMRHSMatrixInfo &rhs_info)
48 {
49 ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.n0 == 0);
50 ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 == 0);
51 ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.h0 == 0);
52 ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0");
53 ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 & (rhs_info.k0 - 1)) && (rhs_info.k0 != 1) && (rhs_info.k0 != 3)), "Only 1,2,3,4,8,16 are supported for k0");
54 ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.n0 > 16);
55 ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16);
56 ARM_COMPUTE_RETURN_ERROR_ON((rhs_info.k0 == 1) && (rhs_info.transpose));
57
58 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
59 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
60
61 if(rhs_info.export_to_cl_image)
62 {
63 const TensorInfo tensor_reshaped_info(compute_rhs_reshaped_shape(*input, rhs_info), 1, input->data_type());
64 ARM_COMPUTE_RETURN_ON_ERROR(cl_gemm::validate_image2d_support_on_rhs(tensor_reshaped_info, rhs_info));
65 }
66
67 if(output->total_size() != 0)
68 {
69 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_rhs_reshaped_shape(*input, rhs_info));
70 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
71 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
72 }
73
74 return Status{};
75 }
76
validate_and_configure_window(ITensorInfo * input,ITensorInfo * output,const GEMMRHSMatrixInfo & rhs_info)77 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const GEMMRHSMatrixInfo &rhs_info)
78 {
79 const unsigned int num_elems_processed_per_iteration_x = rhs_info.n0;
80 const unsigned int num_elems_processed_per_iteration_y = rhs_info.k0;
81 bool window_changed = false;
82
83 // Output auto initialization if not yet initialized
84 auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*input, rhs_info)));
85
86 // Configure window
87 Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
88
89 AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
90 AccessWindowStatic output_access(output, 0, 0, output->dimension(0), output->dimension(1));
91
92 window_changed = update_window_and_padding(win, input_access);
93 output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
94
95 if(rhs_info.export_to_cl_image)
96 {
97 arm_compute::cl_gemm::update_padding_for_cl_image(output);
98 }
99
100 // Collapse along the Z direction
101 // This collapse needs to be here in order to tune the Z dimension of LWS
102 Window collapsed = win.collapse(win, Window::DimZ);
103
104 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
105 return std::make_pair(err, collapsed);
106 }
107 } // namespace
108
CLGEMMReshapeRHSMatrixKernel()109 CLGEMMReshapeRHSMatrixKernel::CLGEMMReshapeRHSMatrixKernel()
110 : _input(nullptr), _output(nullptr)
111 {
112 }
113
configure(const ICLTensor * input,ICLTensor * output,const GEMMRHSMatrixInfo & rhs_info)114 void CLGEMMReshapeRHSMatrixKernel::configure(const ICLTensor *input, ICLTensor *output, const GEMMRHSMatrixInfo &rhs_info)
115 {
116 configure(CLKernelLibrary::get().get_compile_context(), input, output, rhs_info);
117 }
118
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,const GEMMRHSMatrixInfo & rhs_info)119 void CLGEMMReshapeRHSMatrixKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const GEMMRHSMatrixInfo &rhs_info)
120 {
121 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
122
123 // Perform validate step
124 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), rhs_info));
125
126 _input = input;
127 _output = output;
128
129 // Create build options
130 CLBuildOptions build_opts;
131 build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
132 build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
133 build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
134 build_opts.add_option_if(rhs_info.transpose, "-DTRANSPOSE");
135 build_opts.add_option_if(rhs_info.interleave, "-DINTERLEAVE");
136 build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
137 build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
138
139 std::string kernel_name("gemm_reshape_rhs_matrix_");
140 kernel_name += rhs_info.transpose ? "t" : "nt";
141
142 // Create kernel
143 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
144
145 // Configure kernel window
146 auto win_config = validate_and_configure_window(input->info(), output->info(), rhs_info);
147 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
148 ICLKernel::configure_internal(win_config.second);
149 }
150
validate(const ITensorInfo * input,const ITensorInfo * output,const GEMMRHSMatrixInfo & rhs_info)151 Status CLGEMMReshapeRHSMatrixKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const GEMMRHSMatrixInfo &rhs_info)
152 {
153 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, rhs_info));
154 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), rhs_info).first);
155
156 return Status{};
157 }
158
run(const Window & window,cl::CommandQueue & queue)159 void CLGEMMReshapeRHSMatrixKernel::run(const Window &window, cl::CommandQueue &queue)
160 {
161 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
162 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
163
164 Window slice = window.first_slice_window_3D();
165
166 do
167 {
168 unsigned int idx = 0;
169 add_3D_tensor_argument(idx, _input, slice);
170 add_3D_tensor_argument(idx, _output, slice);
171 enqueue(queue, *this, slice, lws_hint());
172 }
173 while(window.slide_window_slice_3D(slice));
174 }
175 } // namespace arm_compute
176