1 /*
2 * Copyright (c) 2021-2022 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 */
24 #include "src/gpu/cl/kernels/ClDirectConv3dKernel.h"
25
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
28 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
29 #include "src/core/CL/CLValidate.h"
30 #include "src/core/helpers/WindowHelpers.h"
31 #include "support/Cast.h"
32
33 namespace arm_compute
34 {
35 namespace opencl
36 {
37 namespace kernels
38 {
39 namespace
40 {
validate_arguments(const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * src2,const ITensorInfo * dst,const Conv3dInfo & conv3d_info)41 Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const Conv3dInfo &conv3d_info)
42 {
43 ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_LAYOUT(src0, src1, dst);
44 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->data_layout() != DataLayout::NDHWC, "Only NDHWC layout supported");
45
46 // When fusing activation, same workaround introduced for COMPMID-5324 may be necessary
47 ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv3d_info.act_info.enabled(), "Fused activation not supported");
48
49 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src0);
50 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::F16, DataType::F32, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
51 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
52 ARM_COMPUTE_RETURN_ERROR_ON(conv3d_info.dilation != Size3D(1U, 1U, 1U));
53
54 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->dimension(1) != src0->dimension(0), "Weights feature map dimension should match the respective src's one");
55 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 5, "Weights can be at most 5 dimensional");
56
57 ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(2) > (src0->dimension(1) + conv3d_info.padding.left + conv3d_info.padding.right));
58 ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(3) > (src0->dimension(2) + conv3d_info.padding.top + conv3d_info.padding.bottom));
59 ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(4) > (src0->dimension(3) + conv3d_info.padding.front + conv3d_info.padding.back));
60
61 if(src2 != nullptr)
62 {
63 if(is_data_type_quantized(src0->data_type()))
64 {
65 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src2, 1, DataType::S32);
66 }
67 else
68 {
69 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src1, src2);
70 }
71 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src2->dimension(0) != src1->dimension(0), "Biases size and number of dst feature maps should match");
72 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src2->num_dimensions() > 1, "Biases should be one dimensional");
73 }
74
75 // Checks performed when dst is configured
76 if(dst->total_size() != 0)
77 {
78 ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst->dimension(0) != src1->dimension(0), "Weights and dst OFMs should match");
79 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), misc::shape_calculator::compute_conv3d_shape(src0->tensor_shape(), src1->tensor_shape(), conv3d_info));
80 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst);
81 }
82
83 return Status{};
84 }
85 } // namespace
86
ClDirectConv3dKernel()87 ClDirectConv3dKernel::ClDirectConv3dKernel()
88 {
89 _type = CLKernelType::DIRECT;
90 }
91
configure(const CLCompileContext & compile_context,const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * src2,ITensorInfo * dst,const Conv3dInfo & conv3d_info)92 void ClDirectConv3dKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, ITensorInfo *dst,
93 const Conv3dInfo &conv3d_info)
94 {
95 ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
96
97 // Perform validation
98 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, conv3d_info));
99
100 // Create window and update padding
101 const DataType data_type = src0->data_type();
102 const size_t src_width = src0->dimension(1);
103 const size_t src_height = src0->dimension(2);
104 const size_t src_depth = src0->dimension(3);
105 const size_t src_channels = src0->dimension(0);
106 const size_t dst_width = dst->dimension(1);
107 const size_t dst_height = dst->dimension(2);
108 const size_t dst_depth = dst->dimension(3);
109 const size_t dst_channels = dst->dimension(0);
110 const size_t weights_width = src1->dimension(2);
111 const size_t weights_height = src1->dimension(3);
112 const size_t weights_depth = src1->dimension(4);
113 const size_t pad_left = conv3d_info.padding.left;
114 const size_t pad_top = conv3d_info.padding.top;
115 const size_t pad_front = conv3d_info.padding.front;
116 const size_t conv_stride_x = conv3d_info.stride.x();
117 const size_t conv_stride_y = conv3d_info.stride.y();
118 const size_t conv_stride_z = conv3d_info.stride.z();
119
120 const size_t n0 = std::min(dst->dimension(0), static_cast<size_t>(4u));
121 const size_t m0 = (dst->tensor_shape()[0] > 16) ? ((data_type == DataType::F32) ? 2U : 4U) : 1U;
122 const size_t k0 = adjust_vec_size(8u, src0->dimension(0));
123 const size_t partial_store_n0 = dst->dimension(0) % n0;
124
125 CLBuildOptions build_options;
126 build_options.add_option("-cl-fast-relaxed-math");
127 build_options.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
128 build_options.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src_width));
129 build_options.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src_height));
130 build_options.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(src_depth));
131 build_options.add_option("-DSRC_CHANNELS=" + support::cpp11::to_string(src_channels));
132 build_options.add_option("-DDST_WIDTH=" + support::cpp11::to_string(dst_width));
133 build_options.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst_height));
134 build_options.add_option("-DDST_DEPTH=" + support::cpp11::to_string(dst_depth));
135 build_options.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(dst_channels));
136 build_options.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(weights_width));
137 build_options.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(weights_height));
138 build_options.add_option("-DWEI_DEPTH=" + support::cpp11::to_string(weights_depth));
139 build_options.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_stride_x));
140 build_options.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_stride_y));
141 build_options.add_option("-DSTRIDE_Z=" + support::cpp11::to_string(conv_stride_z));
142 build_options.add_option("-DPAD_LEFT=" + support::cpp11::to_string(pad_left));
143 build_options.add_option("-DPAD_TOP=" + support::cpp11::to_string(pad_top));
144 build_options.add_option("-DPAD_FRONT=" + support::cpp11::to_string(pad_front));
145 build_options.add_option("-DN0=" + support::cpp11::to_string(n0));
146 build_options.add_option("-DM0=" + support::cpp11::to_string(m0));
147 build_options.add_option("-DK0=" + support::cpp11::to_string(k0));
148 build_options.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0));
149
150 if(src2 != nullptr)
151 {
152 build_options.add_option(std::string("-DHAS_BIAS"));
153 build_options.add_option(std::string("-DBIA_DATA_TYPE=" + get_cl_type_from_data_type(src2->data_type())));
154 }
155
156 if(is_data_type_quantized(data_type))
157 {
158 const UniformQuantizationInfo iqinfo = src0->quantization_info().uniform();
159 const UniformQuantizationInfo wqinfo = src1->quantization_info().uniform();
160 const UniformQuantizationInfo oqinfo = dst->quantization_info().uniform();
161
162 PixelValue zero_value = PixelValue(0, src0->data_type(), src0->quantization_info());
163 int zero_value_s32;
164 zero_value.get(zero_value_s32);
165
166 float multiplier = iqinfo.scale * wqinfo.scale / oqinfo.scale;
167 int output_multiplier = 0;
168 int output_shift = 0;
169 quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
170 build_options.add_option("-DIS_QUANTIZED");
171 build_options.add_option("-DDST_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
172 build_options.add_option("-DDST_SHIFT=" + support::cpp11::to_string(output_shift));
173 build_options.add_option("-DSRC_OFFSET=" + support::cpp11::to_string(-iqinfo.offset));
174 build_options.add_option("-DWEI_OFFSET=" + support::cpp11::to_string(-wqinfo.offset));
175 build_options.add_option("-DDST_OFFSET=" + support::cpp11::to_string(oqinfo.offset));
176 build_options.add_option("-DZERO_VALUE=" + support::cpp11::to_string(zero_value_s32));
177 build_options.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(DataType::S32));
178 }
179 else
180 {
181 build_options.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(DataType::F32));
182 build_options.add_option("-DZERO_VALUE=" + support::cpp11::to_string(0));
183 build_options.add_option("-DSRC_OFFSET=" + support::cpp11::to_string(0));
184 build_options.add_option("-DWEI_OFFSET=" + support::cpp11::to_string(0));
185 build_options.add_option("-DDST_OFFSET=" + support::cpp11::to_string(0));
186 }
187
188 std::string kernel_name = "direct_convolution3d_ndhwc";
189 _kernel = create_kernel(compile_context, kernel_name, build_options.options());
190
191 // Configure kernel window
192 Window win = calculate_max_window(*dst, Steps(n0, m0));
193 ICLKernel::configure_internal(win);
194
195 // Set config_id for enabling LWS tuning
196 _config_id = kernel_name;
197 _config_id += "_";
198 _config_id += lower_string(string_from_data_type(data_type));
199 _config_id += "_";
200 _config_id += support::cpp11::to_string(weights_width);
201 _config_id += "_";
202 _config_id += support::cpp11::to_string(weights_height);
203 _config_id += "_";
204 _config_id += support::cpp11::to_string(weights_depth);
205 _config_id += "_";
206 _config_id += support::cpp11::to_string(conv_stride_x);
207 _config_id += "_";
208 _config_id += support::cpp11::to_string(conv_stride_y);
209 _config_id += "_";
210 _config_id += support::cpp11::to_string(conv_stride_z);
211 _config_id += "_";
212 _config_id += support::cpp11::to_string(dst_width);
213 _config_id += "_";
214 _config_id += support::cpp11::to_string(dst_height);
215 _config_id += "_";
216 _config_id += support::cpp11::to_string(dst_channels);
217 }
218
validate(const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * src2,const ITensorInfo * dst,const Conv3dInfo & conv3d_info)219 Status ClDirectConv3dKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const Conv3dInfo &conv3d_info)
220 {
221 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, src2, dst, conv3d_info));
222 return Status{};
223 }
224
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)225 void ClDirectConv3dKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
226 {
227 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
228 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
229
230 const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
231 const auto weights = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
232 const auto biases = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
233 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
234
235 // Get initial windows
236 Window slice = window.first_slice_window_3D();
237 slice.set(Window::DimY, Window::Dimension(0, ceil_to_multiple(dst->info()->dimension(1) * dst->info()->dimension(2) * dst->info()->dimension(3), slice.y().step()), slice.y().step()));
238 slice.set(Window::DimZ, Window::Dimension(0, dst->info()->dimension(4), 1));
239
240 unsigned int idx = 0;
241 add_4D_tensor_argument(idx, src, slice);
242 add_4D_tensor_argument(idx, dst, slice);
243 add_4D_tensor_argument(idx, weights, slice);
244 if(biases != nullptr)
245 {
246 add_1D_tensor_argument(idx, biases, slice);
247 }
248 enqueue(queue, *this, slice, lws_hint());
249 }
250 } // namespace kernels
251 } // namespace opencl
252 } // namespace arm_compute
253