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/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/helpers/AutoConfiguration.h"
37 #include "src/core/helpers/WindowHelpers.h"
38 #include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h"
39 #include "support/Cast.h"
40 #include "support/StringSupport.h"
41
42 namespace arm_compute
43 {
44 namespace opencl
45 {
46 namespace kernels
47 {
48 namespace
49 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst,const GEMMRHSMatrixInfo & rhs_info)50 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const GEMMRHSMatrixInfo &rhs_info)
51 {
52 ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.n0 == 0);
53 ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 == 0);
54 ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.h0 == 0);
55 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");
56 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");
57 ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.n0 > 16);
58 ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16);
59 ARM_COMPUTE_RETURN_ERROR_ON((rhs_info.k0 == 1) && (rhs_info.transpose));
60
61 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
62 ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN);
63
64 if(rhs_info.export_to_cl_image)
65 {
66 const TensorInfo tensor_reshaped_info(misc::shape_calculator::compute_rhs_reshaped_shape(*src, rhs_info), 1, src->data_type());
67 ARM_COMPUTE_RETURN_ON_ERROR(gemm::validate_image2d_support_on_rhs(tensor_reshaped_info, rhs_info));
68 }
69
70 if(dst->total_size() != 0)
71 {
72 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), misc::shape_calculator::compute_rhs_reshaped_shape(*src, rhs_info));
73 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
74 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst);
75 }
76
77 return Status{};
78 }
79
validate_and_configure_window(ITensorInfo * src,ITensorInfo * dst,const GEMMRHSMatrixInfo & rhs_info)80 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITensorInfo *dst, const GEMMRHSMatrixInfo &rhs_info)
81 {
82 const unsigned int num_elems_processed_per_iteration_x = rhs_info.n0;
83 const unsigned int num_elems_processed_per_iteration_y = rhs_info.k0;
84 bool window_changed = false;
85
86 // dst auto initialization if not yet initialized
87 auto_init_if_empty(*dst, src->clone()->set_tensor_shape(misc::shape_calculator::compute_rhs_reshaped_shape(*src, rhs_info)));
88
89 // Configure window
90 Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
91
92 AccessWindowRectangle src_access(src, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
93
94 window_changed = update_window_and_padding(win, src_access);
95
96 if(rhs_info.export_to_cl_image)
97 {
98 gemm::update_padding_for_cl_image(dst);
99 }
100
101 // Collapse along the Z direction
102 // This collapse needs to be here in order to tune the Z dimension of LWS
103 Window collapsed = win.collapse(win, Window::DimZ);
104
105 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
106 return std::make_pair(err, collapsed);
107 }
108 } // namespace
109
ClGemmReshapeRhsMatrixKernel()110 ClGemmReshapeRhsMatrixKernel::ClGemmReshapeRhsMatrixKernel()
111 {
112 _type = CLKernelType::ELEMENTWISE;
113 }
114
configure(const CLCompileContext & compile_context,ITensorInfo * src,ITensorInfo * dst,const GEMMRHSMatrixInfo & rhs_info)115 void ClGemmReshapeRhsMatrixKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const GEMMRHSMatrixInfo &rhs_info)
116 {
117 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
118
119 // Perform validate step
120 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, rhs_info));
121
122 // Create build options
123 CLBuildOptions build_opts;
124 build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
125 build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
126 build_opts.add_option_if(rhs_info.interleave, "-DINTERLEAVE");
127 build_opts.add_option_if(rhs_info.transpose, "-DRESHAPE_RHS_T");
128 build_opts.add_option_if(!rhs_info.transpose, "-DRESHAPE_RHS_NT");
129 build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(src->element_size()));
130
131 std::string kernel_name("gemm_reshape_rhs_matrix_");
132 kernel_name += rhs_info.transpose ? "t" : "nt";
133
134 // Create kernel
135 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
136
137 // Configure kernel window
138 auto win_config = validate_and_configure_window(src, dst, rhs_info);
139 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
140 ICLKernel::configure_internal(win_config.second);
141
142 unsigned int idx = 2 * num_arguments_per_3d_tensor_nhw();
143 _kernel.setArg<cl_int>(idx++, rhs_info.h0);
144 }
145
validate(const ITensorInfo * src,const ITensorInfo * dst,const GEMMRHSMatrixInfo & rhs_info)146 Status ClGemmReshapeRhsMatrixKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const GEMMRHSMatrixInfo &rhs_info)
147 {
148 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, rhs_info));
149 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get(), rhs_info).first);
150
151 return Status{};
152 }
153
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)154 void ClGemmReshapeRhsMatrixKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
155 {
156 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
157 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
158
159 const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
160 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
161
162 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
163
164 Window slice = window.first_slice_window_3D();
165
166 do
167 {
168 unsigned int idx = 0;
169 add_3d_tensor_nhw_argument(idx, src);
170 add_3d_tensor_nhw_argument(idx, dst);
171 enqueue(queue, *this, slice, lws_hint());
172 }
173 while(window.slide_window_slice_3D(slice));
174 }
175 } // namespace kernels
176 } // namespace opencl
177 } // namespace arm_compute