1 /**
2 * Copyright 2022 Huawei Technologies Co., Ltd
3 *
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
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
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 gather_dress or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "nnacl/kernel/gather_d.h"
18 #include "nnacl/gather_parameter.h"
19 #include "nnacl/tensor_c_utils.h"
20 #include "nnacl/op_base.h"
21 #include "nnacl/base/gather_d_base.h"
22 #include "nnacl/kernel/default_kernel_base.h"
23
24 typedef struct GatherDStru {
25 KernelBase base;
26 } GatherDStru;
27
GatherDPrepare(struct KernelBase * self)28 int GatherDPrepare(struct KernelBase *self) {
29 GatherDStru *gather_d = (GatherDStru *)self;
30 NNACL_CHECK_NULL_RETURN_ERR(gather_d);
31 GatherParameter *param = (GatherParameter *)gather_d->base.param_;
32 NNACL_CHECK_NULL_RETURN_ERR(param);
33 NNACL_CHECK_FALSE(self->in_size_ < kInputSize2 || self->out_size_ < 1, NNACL_TENSOR_SIZE_INVALID);
34
35 param->axis_ = ((int *)(gather_d->base.in_[1]->data_))[0];
36 return NNACL_OK;
37 }
38
GatherDResize(struct KernelBase * self)39 int GatherDResize(struct KernelBase *self) {
40 GatherDStru *gather_d = (GatherDStru *)self;
41 NNACL_CHECK_NULL_RETURN_ERR(gather_d);
42 GatherParameter *param = (GatherParameter *)gather_d->base.param_;
43 NNACL_CHECK_NULL_RETURN_ERR(param);
44 int input_rank = (int)gather_d->base.in_[0]->shape_size_;
45 NNACL_CHECK_FALSE(param->axis_ >= input_rank || param->axis_ < -input_rank, NNACL_GATHER_D_AXIS_INVALID);
46
47 if (param->axis_ < 0) {
48 param->axis_ = param->axis_ + input_rank;
49 }
50 return NNACL_OK;
51 }
52
GatherDCompute(struct KernelBase * self)53 int GatherDCompute(struct KernelBase *self) {
54 GatherDStru *gather_d_stru = (GatherDStru *)self;
55 NNACL_CHECK_NULL_RETURN_ERR(gather_d_stru);
56 GatherParameter *param = (GatherParameter *)gather_d_stru->base.param_;
57 NNACL_CHECK_NULL_RETURN_ERR(param);
58
59 TensorC *input = gather_d_stru->base.in_[0];
60 NNACL_CHECK_NULL_RETURN_ERR(input);
61 TensorC *output = gather_d_stru->base.out_[0];
62 NNACL_CHECK_NULL_RETURN_ERR(output);
63 const void *input_data = input->data_;
64 NNACL_CHECK_NULL_RETURN_ERR(input_data);
65 const void *index_data = gather_d_stru->base.in_[2]->data_;
66 NNACL_CHECK_NULL_RETURN_ERR(index_data);
67 void *output_data = output->data_;
68 NNACL_CHECK_NULL_RETURN_ERR(output_data);
69
70 size_t input_shape[MAX_SHAPE_SIZE];
71 for (size_t i = 0; i < input->shape_size_; i++) {
72 input_shape[i] = input->shape_[i];
73 }
74 size_t output_shape[MAX_SHAPE_SIZE];
75 for (size_t i = 0; i < output->shape_size_; i++) {
76 output_shape[i] = output->shape_[i];
77 }
78
79 int input_dtype = input->data_type_;
80 int index_dtype = gather_d_stru->base.in_[THIRD_INPUT]->data_type_;
81 int status = NNACL_ERR;
82 if (index_dtype == kNumberTypeInt32) {
83 if (input_dtype == kNumberTypeFloat32) {
84 status = GATHER_D(float, int32_t, (float *)output_data, (float *)input_data, (int32_t *)index_data, input_shape,
85 input->shape_size_, output_shape, output->shape_size_, param->axis_);
86 } else if (input_dtype == kNumberTypeFloat16) {
87 #ifdef ENABLE_FP16
88 status = GATHER_D(float16_t, int32_t, (float16_t *)output_data, (float16_t *)input_data, (int32_t *)index_data,
89 input_shape, input->shape_size_, output_shape, output->shape_size_, param->axis_);
90 #endif
91 } else if (input_dtype == kNumberTypeInt32) {
92 status = GATHER_D(int32_t, int32_t, (int32_t *)output_data, (int32_t *)input_data, (int32_t *)index_data,
93 input_shape, input->shape_size_, output_shape, output->shape_size_, param->axis_);
94 }
95 } else if (index_dtype == kNumberTypeInt64) {
96 if (input_dtype == kNumberTypeFloat32) {
97 status = GATHER_D(float, int64_t, (float *)output_data, (float *)input_data, (int64_t *)index_data, input_shape,
98 input->shape_size_, output_shape, output->shape_size_, param->axis_);
99 } else if (input_dtype == kNumberTypeFloat16) {
100 #ifdef ENABLE_FP16
101 status = GATHER_D(float16_t, int64_t, (float16_t *)output_data, (float16_t *)input_data, (int64_t *)index_data,
102 input_shape, input->shape_size_, output_shape, output->shape_size_, param->axis_);
103 #endif
104 } else if (input_dtype == kNumberTypeInt32) {
105 status = GATHER_D(int32_t, int64_t, (int32_t *)output_data, (int32_t *)input_data, (int64_t *)index_data,
106 input_shape, input->shape_size_, output_shape, output->shape_size_, param->axis_);
107 }
108 }
109 return status;
110 }
111
CreateGatherD(OpParameter * param,int data_type)112 KernelBase *CreateGatherD(OpParameter *param, int data_type) {
113 GatherDStru *gather_d = (GatherDStru *)malloc(sizeof(GatherDStru));
114 NNACL_MALLOC_CHECK_NULL_RETURN_NULL(gather_d);
115 gather_d->base.Prepare = GatherDPrepare;
116 gather_d->base.Resize = GatherDResize;
117 gather_d->base.Release = DefaultRelease;
118 gather_d->base.Compute = GatherDCompute;
119 return (KernelBase *)gather_d;
120 }
121
122 REG_KERNEL_CREATOR(PrimType_GatherD, kNumberTypeFloat32, CreateGatherD);
123 REG_KERNEL_CREATOR(PrimType_GatherD, kNumberTypeInt32, CreateGatherD);
124 REG_KERNEL_CREATOR(PrimType_GatherD, kNumberTypeFloat16, CreateGatherD);
125