Files
TencentOS-tiny/components/ai/onnx/operator_int/matmul.c
2021-09-06 21:50:44 +08:00

64 lines
1.6 KiB
C

#include "onnx.h"
void matmul(const int *input, // pointer to vector
const int *weight, // pointer to matrix
const uint16_t dim_vec, // length of the vector
const uint16_t num_of_rows, // numCol of A
int *output)
{
for (int i = 0; i < num_of_rows; i++)
{
int ip_out = 0;
for (int j = 0; j < dim_vec; j++)
{
ip_out += input[j] * weight[i * dim_vec + j];
}
output[i] = ip_out;
}
}
int* matmul_layer(Onnx__GraphProto* graph, const int *input, int64_t* shapeInput, int64_t* shapeOutput, const char* layer_name)
{
//assert(graph != NULL && input != NULL && layer_name != "" );
Onnx__NodeProto* node = onnx_graph_get_node_by_name(graph, layer_name);
const char* weight = node->input[1];
int64_t* shapeW = onnx_graph_get_dims_by_name(graph, weight);
if(shapeW == NULL)
{
return NULL;
}
int64_t dimW = onnx_graph_get_dim_by_name(graph, weight);
if(dimW < 0)
{
return NULL;
}
//assert(shapeW[0] == shapeInput[1]);
int64_t permW_t[] = {1, 0};
int* W = onnx_graph_get_weights_by_name(graph, weight);
if(W == NULL)
{
return NULL;
}
int* W_t = transpose(W, shapeW, dimW, permW_t);
int* output = (int*) malloc(sizeof(int)*shapeW[1]);
if(output == NULL)
{
// No memory
return NULL;
}
memset(output, 0, sizeof(sizeof(int)*shapeW[1]));
matmul(input, W_t, shapeW[0], shapeW[1], output);
shapeOutput[0] = shapeInput[0];
shapeOutput[1] = shapeW[1];
free(W_t);
return output;
}