1 /*
2 * Copyright (c) 2018-2021 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/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.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/helpers/AutoConfiguration.h"
37 #include "src/core/helpers/WindowHelpers.h"
38 #include "support/Cast.h"
39 #include "support/StringSupport.h"
40
41 namespace arm_compute
42 {
43 namespace opencl
44 {
45 namespace kernels
46 {
47 namespace
48 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,bool reinterpret_input_as_3d)49 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info, bool reinterpret_input_as_3d)
50 {
51 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
52 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 == 0);
53 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 == 0);
54 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.v0 == 0);
55 ARM_COMPUTE_RETURN_ERROR_ON_MSG(((lhs_info.k0 & (lhs_info.k0 - 1)) && lhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0");
56 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16);
57 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8);
58 ARM_COMPUTE_RETURN_ERROR_ON((lhs_info.m0 > 4 && lhs_info.m0 < 8) && lhs_info.transpose);
59
60 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
61 ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN);
62
63 if(dst->total_size() != 0)
64 {
65 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(),
66 misc::shape_calculator::compute_lhs_reshaped_shape(*src, lhs_info, reinterpret_input_as_3d));
67 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
68 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst);
69 }
70
71 return Status{};
72 }
73
configure_window(ITensorInfo * src,ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,bool reinterpret_input_as_3d)74 Window configure_window(ITensorInfo *src, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info, bool reinterpret_input_as_3d)
75 {
76 const unsigned int num_elems_processed_per_iteration_x = lhs_info.k0;
77 const unsigned int num_elems_processed_per_iteration_y = lhs_info.m0;
78
79 TensorInfo tmp_info(*src);
80
81 if(reinterpret_input_as_3d)
82 {
83 // Since the src tensor has to be reinterpreted as 3D and the execute window is based on a 2D interleave,
84 // the window needs to be constructed on the 2D collapsed version of the tensor
85 TensorShape tmp_shape(src->tensor_shape());
86 tmp_shape.collapse(2U, 1U);
87 tmp_info.set_tensor_shape(tmp_shape);
88 }
89
90 // dst auto inizialitation if not yet initialized
91 auto_init_if_empty(*dst, src->clone()->set_tensor_shape(misc::shape_calculator::compute_lhs_reshaped_shape(*src, lhs_info, reinterpret_input_as_3d)));
92
93 // Configure window
94 Window win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
95
96 // Collapse along the Z direction
97 // This collapse needs to be here in order to tune the Z dimension of LWS
98 Window collapsed = win.collapse(win, Window::DimZ);
99
100 return collapsed;
101 }
102 } // namespace
103
ClGemmReshapeLhsMatrixKernel()104 ClGemmReshapeLhsMatrixKernel::ClGemmReshapeLhsMatrixKernel()
105 {
106 _type = CLKernelType::ELEMENTWISE;
107 }
108
configure(const CLCompileContext & compile_context,ITensorInfo * src,ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,bool reinterpret_input_as_3d)109 void ClGemmReshapeLhsMatrixKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info, bool reinterpret_input_as_3d)
110 {
111 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
112
113 // Perform validate step
114 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, lhs_info, reinterpret_input_as_3d));
115
116 auto padding_info = get_padding_info({ src });
117
118 const unsigned int src_w = src->dimension(0);
119 const unsigned int m = reinterpret_input_as_3d ? src->dimension(1) * src->dimension(2) : src->dimension(1);
120 const unsigned int partial_m0 = m % lhs_info.m0;
121 const unsigned int partial_k0 = src_w % lhs_info.k0;
122
123 // Create build options
124 CLBuildOptions build_opts;
125 build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
126 build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0));
127 build_opts.add_option_if(lhs_info.interleave, "-DINTERLEAVE");
128 build_opts.add_option_if_else(lhs_info.transpose, "-DRESHAPE_LHS_T", "-DRESHAPE_LHS_NT");
129 build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(src->element_size()));
130 build_opts.add_option("-DPARTIAL_M0=" + support::cpp11::to_string(partial_m0));
131 build_opts.add_option("-DPARTIAL_K0=" + support::cpp11::to_string(partial_k0));
132
133 std::string kernel_name("gemm_reshape_lhs_matrix_");
134 kernel_name += lhs_info.transpose ? "t" : "nt";
135
136 // Create kernel
137 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
138
139 // Configure kernel window
140 auto win_config = configure_window(src, dst, lhs_info, reinterpret_input_as_3d);
141 ICLKernel::configure_internal(win_config);
142
143 unsigned int idx = 2 * num_arguments_per_3d_tensor_nhw();
144 _kernel.setArg<cl_int>(idx++, m);
145 _kernel.setArg<cl_int>(idx++, lhs_info.v0);
146
147 // Set config_id for enabling LWS tuning
148 _config_id = "gemm_reshape_lhs_matrix_";
149 _config_id += (reinterpret_input_as_3d ? "3d_" : "");
150 _config_id += lower_string(string_from_data_type(src->data_type()));
151 _config_id += "_";
152 _config_id += support::cpp11::to_string(dst->dimension(0));
153 _config_id += "_";
154 _config_id += support::cpp11::to_string(dst->dimension(1));
155 _config_id += "_";
156 _config_id += support::cpp11::to_string(dst->dimension(2));
157 _config_id += "_";
158 _config_id += support::cpp11::to_string(lhs_info.m0);
159 _config_id += "_";
160 _config_id += support::cpp11::to_string(lhs_info.k0);
161 _config_id += "_";
162 _config_id += support::cpp11::to_string(lhs_info.v0);
163 _config_id += "_";
164 _config_id += support::cpp11::to_string(lhs_info.interleave);
165 _config_id += "_";
166 _config_id += support::cpp11::to_string(lhs_info.transpose);
167
168 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
169 }
170
validate(const ITensorInfo * src,const ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,bool reinterpret_input_as_3d)171 Status ClGemmReshapeLhsMatrixKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info, bool reinterpret_input_as_3d)
172 {
173 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, lhs_info, reinterpret_input_as_3d));
174 return Status{};
175 }
176
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)177 void ClGemmReshapeLhsMatrixKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
178 {
179 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
180 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
181
182 const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
183 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
184
185 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
186
187 Window slice = window.first_slice_window_3D();
188
189 do
190 {
191 unsigned int idx = 0;
192 add_3d_tensor_nhw_argument(idx, src);
193 add_3d_tensor_nhw_argument(idx, dst);
194 enqueue(queue, *this, slice, lws_hint());
195 }
196 while(window.slide_window_slice_3D(slice));
197 }
198 } // namespace kernels
199 } // namespace opencl
200 } // namespace arm_compute
201