• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /**
2  * Copyright 2021-2023 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 express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 #include "nnacl/infer/matmul_infer.h"
18 #include <math.h>
19 #include "nnacl/infer/infer_register.h"
20 
21 #define MIN_SHAPE_SIZE 2
22 
CheckMatmulInputShape(int * a_shape,size_t a_shape_size,int * b_shape,size_t b_shape_size,const int * bias_shape,size_t bias_shape_size,const MatMulParameter * param)23 int CheckMatmulInputShape(int *a_shape, size_t a_shape_size, int *b_shape, size_t b_shape_size, const int *bias_shape,
24                           size_t bias_shape_size, const MatMulParameter *param) {
25   if (a_shape_size < MIN_SHAPE_SIZE || b_shape_size < MIN_SHAPE_SIZE) {
26     return NNACL_PARAM_INVALID;
27   }
28   for (size_t i = 0; i < (a_shape_size - 2) && i < (b_shape_size - 2); ++i) {
29     int min_value = MSMIN(a_shape[i], b_shape[i]);
30     int max_value = MSMAX(a_shape[i], b_shape[i]);
31     if (min_value != 0 && max_value % min_value != 0) {
32       NNACL_LOG_ERROR("min_value != 0 && max_value is not multiple of min_value");
33       return NNACL_INPUT_TENSOR_ERROR;
34     }
35   }
36   if (param->a_transpose_) {
37     iswap(&a_shape[a_shape_size - 1], &a_shape[a_shape_size - DIMENSION_2D]);
38   }
39   if (param->b_transpose_) {
40     iswap(&b_shape[b_shape_size - 1], &b_shape[b_shape_size - 2]);
41   }
42   if (bias_shape_size == DIMENSION_1D && bias_shape[0] != b_shape[b_shape_size - 1]) {
43     NNACL_LOG_ERROR("bias_shape_size == DIMENSION_1D && bias_shape[0] != b_shape[b_shape_size - 1]");
44     return NNACL_ERR;
45   }
46   if (a_shape[a_shape_size - 1] != b_shape[b_shape_size - 2]) {
47     NNACL_LOG_ERROR("a_shape[a_shape_size - 1] != b_shape[b_shape_size - 2]");
48     return NNACL_ERR;
49   }
50   return NNACL_OK;
51 }
52 
CheckMatMulBias(int * shape,size_t dim_size)53 int CheckMatMulBias(int *shape, size_t dim_size) {
54   if (dim_size > 1) {
55     for (size_t i = 0; i < dim_size - 1; i++) {
56       if (shape[i] != DIMENSION_1D) {
57         NNACL_LOG_ERROR("shape[%zu] != DIMENSION_1D", i);
58         return NNACL_ERR;
59       }
60     }
61   }
62   return NNACL_OK;
63 }
64 
SetShape(const TensorC * const * inputs,size_t inputs_size,TensorC ** outputs,size_t outputs_size,OpParameter * parameter)65 int SetShape(const TensorC *const *inputs, size_t inputs_size, TensorC **outputs, size_t outputs_size,
66              OpParameter *parameter) {
67   TensorC *input0 = (TensorC *)inputs[0];
68   TensorC *input1 = (TensorC *)inputs[1];
69   TensorC *output = outputs[0];
70   MatMulParameter *param = (MatMulParameter *)parameter;
71   int a_shape[MAX_SHAPE_SIZE] = {0};
72   size_t a_shape_size = 0;
73   ShapeSet(a_shape, &a_shape_size, input0->shape_, input0->shape_size_);
74   int b_shape[MAX_SHAPE_SIZE] = {0};
75   size_t b_shape_size = 0;
76   ShapeSet(b_shape, &b_shape_size, input1->shape_, input1->shape_size_);
77   int *shape_align = a_shape_size > b_shape_size ? b_shape : a_shape;
78   size_t *shape_size_align = a_shape_size > b_shape_size ? &b_shape_size : &a_shape_size;
79   int diff = abs((int)a_shape_size - (int)b_shape_size);
80   for (int i = 0; i < diff; ++i) {
81     ShapeInsert(shape_align, shape_size_align, 0, 1);
82   }
83   int bias_shape[MAX_AXIS_SIZE] = {0};
84   size_t bias_shape_size = 0;
85   if (inputs_size == kInputSize2) {
86     TensorC *bias = (TensorC *)inputs[2];
87     ShapeSet(bias_shape, &bias_shape_size, bias->shape_, bias->shape_size_);
88     NNACL_CHECK_TRUE_RET(CheckMatMulBias(bias_shape, bias_shape_size) == NNACL_OK, NNACL_ERR);
89   }
90 
91   bool del_start = false;
92   bool del_end = false;
93   if (a_shape_size == 1) {
94     int insert_ret = ShapeInsert(a_shape, &a_shape_size, 0, 1);
95     if (insert_ret != NNACL_OK) {
96       NNACL_LOG_ERROR("ShapeInsert failed,insert_ret != NNACL_OK");
97       return NNACL_ERR;
98     }
99     del_start = true;
100   }
101   if (b_shape_size == 1) {
102     ShapePush(b_shape, &b_shape_size, 1);
103     del_end = true;
104   }
105   int ret = CheckMatmulInputShape(a_shape, a_shape_size, b_shape, b_shape_size, bias_shape, bias_shape_size, param);
106   if (ret != NNACL_OK) {
107     NNACL_LOG_ERROR("CheckMatmulInputShape failed,ret != NNACL_OK");
108     return NNACL_ERR;
109   }
110   int c_shape[MAX_SHAPE_SIZE];
111   size_t c_shape_size = 0;
112   ShapeSet(c_shape, &c_shape_size, a_shape, a_shape_size);
113   c_shape[c_shape_size - 1] = b_shape[b_shape_size - 1];
114   if (del_start) {
115     int erase_ret = ShapeErase(c_shape, &c_shape_size, 0);
116     if (erase_ret != NNACL_OK) {
117       NNACL_LOG_ERROR("ShapeErase failed,erase_ret != NNACL_OK");
118       return NNACL_ERR;
119     }
120   }
121   if (del_end) {
122     c_shape_size--;
123   }
124 
125   for (size_t i = 0; i < (a_shape_size - 2) && i < (b_shape_size - 2); ++i) {
126     c_shape[i] = MSMAX(a_shape[i], b_shape[i]);
127   }
128 
129   SetShapeArray(output, c_shape, c_shape_size);
130   return NNACL_OK;
131 }
132 
MatmulInferShape(const TensorC * const * inputs,size_t inputs_size,TensorC ** outputs,size_t outputs_size,OpParameter * parameter)133 int MatmulInferShape(const TensorC *const *inputs, size_t inputs_size, TensorC **outputs, size_t outputs_size,
134                      OpParameter *parameter) {
135   int check_ret = CheckAugmentNullSizeInputTwo(inputs, inputs_size, outputs, outputs_size, parameter, 2, 3, 1);
136   if (check_ret != NNACL_OK) {
137     return check_ret;
138   }
139 
140   TensorC *input0 = (TensorC *)inputs[0];
141   TensorC *input1 = (TensorC *)inputs[1];
142   TensorC *output = outputs[0];
143 
144   TensorC *input = input1->data_ == NULL ? input1 : input0;  // transfer the input which comes from the other node.
145   SetDataTypeFormat(output, input);
146   if (input->data_type_ == kNumberTypeInt8 && parameter->quant_type_ == Quant_QuantDynamic) {
147     output->data_type_ = kNumberTypeFloat32;
148   }
149   if (!InferFlag(inputs, inputs_size)) {
150     return NNACL_INFER_INVALID;
151   }
152   return SetShape(inputs, inputs_size, outputs, outputs_size, parameter);
153 }
154 
155 REG_INFER(MatMul, PrimType_MatMulFusion, MatmulInferShape)
156