2 * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 #include <cker/operation/DepthwiseConv.h>
18 #include <misc/polymorphic_downcast.h>
20 #include "OperationUtil.h"
22 #include "interp/Registration.h"
23 #include "ir/operation/DepthwiseConv2D.h"
24 #include "util/Utils.h"
25 #include "util/ShapeInference.h"
35 void prepareDepthwiseConv(ExecEnv *env, const ir::Operation &node)
37 const auto in_index = node.getInputs().at(ir::operation::DepthwiseConv2D::INPUT);
38 const auto kernel_index = node.getInputs().at(ir::operation::DepthwiseConv2D::KERNEL);
39 const auto bias_index = node.getInputs().at(ir::operation::DepthwiseConv2D::BIAS);
40 const auto out_index = node.getOutputs().at(0);
42 const auto in_tensor = env->tensorAt(in_index);
43 const auto kernel_tensor = env->tensorAt(kernel_index);
44 const auto bias_tensor = env->tensorAt(bias_index);
46 assert(in_tensor->num_dimensions() == 4);
47 assert(kernel_tensor->num_dimensions() == 4);
48 assert(bias_tensor->num_dimensions() == 1);
50 UNUSED_RELEASE(in_tensor);
51 UNUSED_RELEASE(kernel_tensor);
52 UNUSED_RELEASE(bias_tensor);
54 // TODO handle unspecified output shape:
55 // calculate output shape using ifm shape, kernel shape, padding, stride
56 const auto output_info = env->graph().operands().at(out_index).info();
57 if (output_info.total_size() == 0)
59 // Handle unspecified output shape
60 const auto &depth_conv_node =
61 nnfw::misc::polymorphic_downcast<const ir::operation::DepthwiseConv2D &>(node);
62 const auto infered_output_shape = shape_inference::inferDepthwiseConv2DShape(
63 in_tensor->tensorInfo().shape(), kernel_tensor->tensorInfo().shape(),
64 depth_conv_node.param());
65 env->allocateIfNeeded(
66 out_index, ir::OperandInfo::createStaticInfo(infered_output_shape, output_info.typeInfo()));
70 env->allocateIfNeeded(out_index, output_info);
73 auto out_tensor = env->tensorAt(out_index);
74 UNUSED_RELEASE(out_tensor);
76 // Handle same ifm & ofm data type only
77 assert(in_tensor->data_type() == out_tensor->data_type());
78 assert(out_tensor->num_dimensions() == 4);
81 void invoke(const ITensor *ifm_tensor, const ITensor *ker_tensor, const ITensor *bias_tensor,
82 const ITensor *ofm_tensor, const ir::operation::DepthwiseConv2D::Param ¶m)
84 // TODO Support NCHW frontend
85 const auto ifm_shape = ifm_tensor->tensorInfo().shape().asFeature(ir::Layout::NHWC);
86 const auto ofm_shape = ofm_tensor->tensorInfo().shape().asFeature(ir::Layout::NHWC);
87 // Kernel format is [1, kernel_height, kernel_width, depth_out].
88 const auto &ker_shape = ker_tensor->tensorInfo().shape();
89 const auto ker_height = ker_shape.dim(1);
90 const auto ker_width = ker_shape.dim(2);
91 const auto padding = ir::calculatePadding(param.padding, ifm_shape, ofm_shape, param.stride,
92 ker_width, ker_height);
95 float activation_min, activation_max;
96 calculateActivationRange(param.activation, &activation_min, &activation_max);
98 nnfw::cker::DepthwiseConvParams cker_param;
99 cker_param.padding_values.width = padding.left;
100 cker_param.padding_values.height = padding.top;
101 cker_param.depth_multiplier = param.multiplier;
102 cker_param.stride_width = param.stride.horizontal;
103 cker_param.stride_height = param.stride.vertical;
104 cker_param.dilation_width_factor = 1;
105 cker_param.dilation_height_factor = 1;
106 cker_param.float_activation_min = activation_min;
107 cker_param.float_activation_max = activation_max;
109 const auto cker_ifm_shape = convertShape(ifm_tensor->tensorInfo().shape());
110 const auto cker_ker_shape = convertShape(ker_tensor->tensorInfo().shape());
111 const auto cker_bias_shape = convertShape(bias_tensor->tensorInfo().shape());
112 const auto cker_ofm_shape = convertShape(ofm_tensor->tensorInfo().shape());
113 const float *ifm_ptr = reinterpret_cast<const float *>(ifm_tensor->bufferRO());
114 const float *ker_ptr = reinterpret_cast<const float *>(ker_tensor->bufferRO());
115 const float *bias_ptr = reinterpret_cast<const float *>(bias_tensor->bufferRO());
116 float *ofm_ptr = reinterpret_cast<float *>(ofm_tensor->buffer());
118 nnfw::cker::DepthwiseConv(cker_param, cker_ifm_shape, ifm_ptr, cker_ker_shape, ker_ptr,
119 cker_bias_shape, bias_ptr, cker_ofm_shape, ofm_ptr);
122 void invokeDepthwiseConv(const ExecEnv *env, const ir::Operation &node)
124 const auto &conv_node = static_cast<const ir::operation::DepthwiseConv2D &>(node);
126 const auto ifm_index = node.getInputs().at(ir::operation::DepthwiseConv2D::INPUT);
127 const auto ker_index = node.getInputs().at(ir::operation::DepthwiseConv2D::KERNEL);
128 const auto bias_index = node.getInputs().at(ir::operation::DepthwiseConv2D::BIAS);
129 const auto ofm_index = node.getOutputs().at(0);
131 const auto ifm_tensor = env->tensorAt(ifm_index);
132 const auto ker_tensor = env->tensorAt(ker_index);
133 const auto bias_tensor = env->tensorAt(bias_index);
134 const auto ofm_tensor = env->tensorAt(ofm_index);
136 const auto data_type = ifm_tensor->data_type();
137 if (data_type == ir::DataType::FLOAT32)
139 invoke(ifm_tensor, ker_tensor, bias_tensor, ofm_tensor, conv_node.param());
143 throw std::runtime_error{"NYI: Support float32 only"};
149 OpKernel *getDepthwiseConv2D()
151 static OpKernel kernel = {prepareDepthwiseConv, invokeDepthwiseConv};
155 } // namespace interp