xref: /aosp_15_r20/external/ComputeLibrary/src/gpu/cl/kernels/ClIndirectConv2dKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2022-2023 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/ClIndirectConv2dKernel.h"
25 
26 #include "arm_compute/core/CL/CLKernelLibrary.h"
27 #include "arm_compute/core/CL/ICLTensor.h"
28 #include "arm_compute/core/KernelDescriptors.h"
29 #include "arm_compute/core/Utils.h"
30 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
31 #include "src/core/CL/CLUtils.h"
32 #include "src/core/CL/CLValidate.h"
33 #include "src/core/helpers/AutoConfiguration.h"
34 #include "src/core/helpers/WindowHelpers.h"
35 #include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h"
36 #include "support/Cast.h"
37 #include "support/StringSupport.h"
38 
39 namespace arm_compute
40 {
41 namespace opencl
42 {
43 namespace kernels
44 {
45 namespace
46 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * indirect_buffer,const ITensorInfo * dst,const PadStrideInfo & conv_info,const ActivationLayerInfo & act_info,const DirectConvComputeKernelInfo & desc)47 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *indirect_buffer, const ITensorInfo *dst,
48                           const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, const DirectConvComputeKernelInfo &desc)
49 {
50     ARM_COMPUTE_UNUSED(act_info);
51     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
52     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32);
53     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(indirect_buffer, 1, DataType::S32);
54     ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(src, DataLayout::NHWC);
55     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights);
56     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(indirect_buffer->tensor_shape(),
57                                                        misc::shape_calculator::compute_indirect_buffer_shape(src->tensor_shape(),
58                                                                                                              src->data_layout(),
59                                                                                                              weights->tensor_shape(),
60                                                                                                              conv_info,
61                                                                                                              desc));
62 
63     constexpr int channel_idx = 0;
64     constexpr int batch_idx   = 3;
65 
66     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(channel_idx) != src->dimension(channel_idx), "Weights feature map dimension should match the respective src's one");
67     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4, "Weights can be at most 4 dimensional");
68 
69     ARM_COMPUTE_RETURN_ERROR_ON_MSG(desc.m0 <= 0 || desc.m0 > 8, "M0 can only be greater than 0 and less than or equal to 8");
70 
71     ARM_COMPUTE_RETURN_ERROR_ON_MSG(desc.n0 != 1 && desc.n0 != 2 && desc.n0 != 3 && desc.n0 != 4 && desc.n0 != 8 && desc.n0 != 16,
72                                     "N0 can only be: 1, 2, 3, 4, 8, and 16");
73     ARM_COMPUTE_RETURN_ERROR_ON_MSG(desc.k0 != 1 && desc.k0 != 2 && desc.k0 != 3 && desc.k0 != 4 && desc.k0 != 8 && desc.k0 != 16,
74                                     "K0 can only be: 1, 2, 3, 4, 8, and 16");
75 
76     if(desc.export_weights_to_cl_image)
77     {
78         ARM_COMPUTE_RETURN_ERROR_ON_MSG(desc.k0 != 4 && desc.k0 != 8 && desc.k0 != 16,
79                                         "K0 can only be: 4, 8, and 16");
80         ARM_COMPUTE_RETURN_ERROR_ON_MSG(!export_to_cl_image(weights),
81                                         "Export to CLImage is not supported for this weight configuration");
82     }
83 
84     if(biases != nullptr)
85     {
86         if(is_data_type_quantized_asymmetric(src->data_type()))
87         {
88             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
89         }
90         else
91         {
92             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
93         }
94         ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->dimension(channel_idx) != weights->dimension(batch_idx),
95                                         "Biases size and number of dst feature maps should match");
96         ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->num_dimensions() > 1,
97                                         "Biases should be one dimensional");
98     }
99 
100     // Checks performed when dst is configured
101     if(dst->total_size() != 0)
102     {
103         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(),
104                                                            misc::shape_calculator::compute_deep_convolution_shape(*src, *weights, conv_info));
105         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
106     }
107 
108     return Status{};
109 }
110 } // namespace
111 
ClIndirectConv2dKernel()112 ClIndirectConv2dKernel::ClIndirectConv2dKernel()
113 {
114     _type = CLKernelType::DIRECT;
115 }
116 
configure(const CLCompileContext & compile_context,ITensorInfo * src,ITensorInfo * weights,ITensorInfo * biases,ITensorInfo * indirect_buffer,ITensorInfo * dst,const PadStrideInfo & conv_info,const ActivationLayerInfo & act_info,const DirectConvComputeKernelInfo & desc)117 void ClIndirectConv2dKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *indirect_buffer, ITensorInfo *dst,
118                                        const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, const DirectConvComputeKernelInfo &desc)
119 {
120     ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, indirect_buffer, dst);
121 
122     // Perform validation
123     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, weights, biases, indirect_buffer, dst, conv_info, act_info, desc));
124 
125     constexpr unsigned int channel_idx   = 0;
126     constexpr unsigned int width_idx     = 1;
127     constexpr unsigned int height_idx    = 2;
128     const unsigned int     kernel_width  = weights->dimension(width_idx);
129     const unsigned int     kernel_height = weights->dimension(height_idx);
130     const DataType         data_type     = src->data_type();
131 
132     const GPUTarget gpu_target = get_target();
133 
134     // Get dst shape
135     TensorShape output_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *weights, conv_info);
136 
137     // Output auto inizialitation if not yet initialized
138     auto_init_if_empty(*dst, output_shape,
139                        1,
140                        src->data_type(),
141                        src->quantization_info());
142 
143     // Configure kernel window
144     Window win;
145     output_shape.collapse(2U, 1U);
146     const unsigned int n0 = adjust_vec_size(desc.n0, output_shape[0]);
147     const unsigned int m0 = adjust_vec_size(desc.m0, output_shape[1]);
148     const unsigned int k0 = adjust_vec_size(desc.k0, src->dimension(channel_idx));
149 
150     const unsigned int partial_store_n0 = dst->dimension(channel_idx) % n0;
151 
152     // Create window and update padding
153     win = calculate_max_window(output_shape, Steps(n0, m0));
154 
155     ICLKernel::configure_internal(win);
156 
157     std::stringstream kernel_name;
158     CLBuildOptions    build_options;
159 
160     kernel_name << "indirect_convolution_nhwc";
161 
162     _export_to_cl_image = desc.export_weights_to_cl_image;
163 
164     // Update the padding for the weights tensor if we can export to cl_image
165     if(_export_to_cl_image)
166     {
167         gemm::update_padding_for_cl_image(weights);
168     }
169 
170     // Add padding to indirect buffer to avoid out-of-bound reads
171     // When M0 is 5, 6, and 7, we use vload8 to fetch the data from the buffer
172     const unsigned int load_indirect_buf_size = m0 > 4 ? 8 : m0;
173     const unsigned int indirect_buf_width     = indirect_buffer->tensor_shape()[0];
174     const unsigned int round_up_width         = ((indirect_buf_width + load_indirect_buf_size - 1) / load_indirect_buf_size) * load_indirect_buf_size;
175     const unsigned int padding                = round_up_width - indirect_buf_width;
176     indirect_buffer->extend_padding(PaddingSize(0, indirect_buffer->padding().right + padding, 0, 0));
177 
178     if(biases != nullptr)
179     {
180         build_options.add_option(std::string("-DHAS_BIAS"));
181         build_options.add_option(std::string("-DBIA_DATA_TYPE=" + get_cl_type_from_data_type(biases->data_type())));
182     }
183 
184     // Conditions of -cl-fast-relaxed-math causing accuracy issues can be traced from COMPMID-5324
185     const auto act_function = act_info.activation();
186 
187     if((gpu_target != GPUTarget::G71 && (gpu_target & GPUTarget::GPU_ARCH_MASK) == GPUTarget::BIFROST)
188        && (act_function == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU || act_function == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
189        && (data_type == DataType::F32 || data_type == DataType::F16))
190     {
191         // -cl-fast-relaxed-math also sets -cl-finite-math-only and -cl-unsafe-math-optimizations
192         // to disable -cl-finite-math-only, we only include -cl-unsafe-math-optimizations
193         build_options.add_option("-cl-unsafe-math-optimizations");
194     }
195     else
196     {
197         build_options.add_option("-cl-fast-relaxed-math");
198     }
199 
200     build_options.add_option("-DSRC_TENSOR_TYPE=BUFFER");
201     build_options.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(data_type));
202     build_options.add_option("-DSRC_CHANNELS=" + support::cpp11::to_string(src->dimension(channel_idx)));
203     build_options.add_option("-DOFF_TENSOR_TYPE=BUFFER");
204     build_options.add_option("-DDST_WIDTH=" + support::cpp11::to_string(dst->dimension(width_idx)));
205     build_options.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst->dimension(height_idx)));
206     build_options.add_option("-DDST_TENSOR_TYPE=BUFFER");
207     build_options.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(data_type));
208     build_options.add_option_if_else(_export_to_cl_image, "-DWEI_TENSOR_TYPE=IMAGE", "-DWEI_TENSOR_TYPE=BUFFER");
209     build_options.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(kernel_width));
210     build_options.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(kernel_height));
211     build_options.add_option("-DWEI_DATA_TYPE=" + get_cl_type_from_data_type(data_type));
212     build_options.add_option("-DN0=" + support::cpp11::to_string(n0));
213     build_options.add_option("-DM0=" + support::cpp11::to_string(m0));
214     build_options.add_option("-DK0=" + support::cpp11::to_string(k0));
215     build_options.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0));
216     build_options.add_option("-DIND_BUFF_VEC_SIZE=" + support::cpp11::to_string(load_indirect_buf_size));
217     build_options.add_option_if((src->dimension(channel_idx) % k0) != 0, "-DLEFTOVER_LOOP");
218     build_options.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_function)));
219     build_options.add_option_if(act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(act_info.a()));
220     build_options.add_option_if(act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(act_info.b()));
221 
222     // A macro guard to compile ONLY the kernel of interest
223     build_options.add_option("-D" + upper_string(kernel_name.str()));
224 
225     if(compile_context.get_ddk_version() >= 30)
226     {
227         build_options.add_option("-fregister-allocation=64");
228     }
229 
230     _kernel = create_kernel(compile_context, kernel_name.str(), build_options.options());
231 
232     // Set config_id for enabling LWS tuning
233     _config_id = kernel_name.str();
234     _config_id += "_";
235     _config_id += lower_string(string_from_data_type(data_type));
236     _config_id += "_";
237     _config_id += support::cpp11::to_string(kernel_width);
238     _config_id += "_";
239     _config_id += support::cpp11::to_string(kernel_height);
240     _config_id += "_";
241     _config_id += support::cpp11::to_string(src->dimension(width_idx));
242     _config_id += "_";
243     _config_id += support::cpp11::to_string(src->dimension(height_idx));
244     _config_id += "_";
245     _config_id += support::cpp11::to_string(src->dimension(channel_idx));
246     _config_id += "_";
247     _config_id += support::cpp11::to_string(dst->dimension(width_idx));
248     _config_id += "_";
249     _config_id += support::cpp11::to_string(dst->dimension(height_idx));
250     _config_id += "_";
251     _config_id += support::cpp11::to_string(dst->dimension(channel_idx));
252 }
253 
validate(const ITensorInfo * src,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * indirect_buffer,const ITensorInfo * dst,const PadStrideInfo & conv_info,const ActivationLayerInfo & act_info,const DirectConvComputeKernelInfo & desc)254 Status ClIndirectConv2dKernel::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *indirect_buffer, const ITensorInfo *dst,
255                                         const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, const DirectConvComputeKernelInfo &desc)
256 {
257     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, biases, indirect_buffer, dst, conv_info, act_info, desc));
258     return Status{};
259 }
260 
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)261 void ClIndirectConv2dKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
262 {
263     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
264     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
265 
266     // Get initial windows
267     Window slice = window.first_slice_window_3D();
268 
269     const auto src             = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
270     const auto weights         = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
271     const auto biases          = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
272     const auto indirect_buffer = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_3));
273     auto       dst             = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
274 
275     cl::Image2D weights_cl_image;
276 
277     if(_export_to_cl_image)
278     {
279         const size_t      image_w = weights->info()->dimension(0) / 4;
280         const size_t      image_h = weights->info()->dimension(1) * weights->info()->dimension(2) * weights->info()->dimension(3);
281         const TensorShape shape2d(image_w, image_h);
282         const size_t      image_row_pitch = weights->info()->strides_in_bytes()[1];
283 
284         // Export cl_buffer to cl_image
285         weights_cl_image = create_image2d_from_buffer(CLKernelLibrary::get().context(), weights->cl_buffer(), shape2d, weights->info()->data_type(), image_row_pitch, CLImage2DType::ReadOnly);
286     }
287 
288     unsigned int idx = 0;
289     add_4d_tensor_nhwc_argument(idx, src);
290     add_4d_tensor_nhwc_argument(idx, indirect_buffer);
291     add_4d_tensor_nhwc_argument(idx, dst);
292     if(_export_to_cl_image)
293     {
294         _kernel.setArg(idx++, weights_cl_image);
295     }
296     add_4d_tensor_nhwc_argument(idx, weights);
297     if(biases != nullptr)
298     {
299         add_1D_tensor_argument(idx, biases, slice);
300     }
301     enqueue(queue, *this, slice, lws_hint());
302 }
303 } // namespace kernels
304 } // namespace opencl
305 } // namespace arm_compute
306