sunwen/Aurora
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update convertfp16tofloat

Browse Source
This commit is contained in:
kradchen
2023-05-09 17:56:15 +08:00
parent e1a917609d
commit 61baf0dc4a
4 changed files with 23 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
src/Function1D.cpp
View File

@@ -753,26 +753,26 @@ void Aurora::padding(Matrix &aMatrix, int aIndex, double aValue)
aMatrix = Matrix::New(newData,size,1,1,aMatrix.getValueType());
}
Matrix Aurora::convertfp16tofloat(const Matrix& aMatrix)
Matrix Aurora::convertfp16tofloat(short* aData, int aRows, int aColumns)
{
auto input = aMatrix.getData();
auto input = aData;
size_t size = aRows*aColumns;
size_t quaterSize = size/4;
//uint16变换为float(32位)输出大小翻倍
auto output = malloc(aMatrix.getDataSize()*4);
size_t rows = aMatrix.getDataSize()*sizeof(double)/sizeof(short);
size_t total_count = aMatrix.getDataSize();
auto output = malloc(size);
// #pragma omp parallel for
for (size_t i = 0; i < total_count; i+=8) {
#pragma omp parallel for
for (size_t i = 0; i < quaterSize; i+=8) {
//循环展开以避免过度的线程调用
if (i < total_count)::convertValue2((short*)(input+i), output + (i) * 4);
if (i+1 < total_count)::convertValue2((short*)(input+i+1), output + (i+1) * 4);
if (i+2 < total_count)::convertValue2((short*)(input+i+2), output + (i+2) * 4);
if (i+3 < total_count)::convertValue2((short*)(input+i+3), output + (i+3) * 4);
if (i+4 < total_count)::convertValue2((short*)(input+i+4), output + (i+4) * 4);
if (i+5 < total_count)::convertValue2((short*)(input+i+5), output + (i+5) * 4);
if (i+6 < total_count)::convertValue2((short*)(input+i+6), output + (i+6) * 4);
if (i+7 < total_count)::convertValue2((short*)(input+i+7), output + (i+7) * 4);
if (i < quaterSize)::convertValue2((short*)(input+i*4), output + (i) * 4);
if (i+1 < quaterSize)::convertValue2((short*)(input+(i+1)*4), output + (i+1) * 4);
if (i+2 < quaterSize)::convertValue2((short*)(input+(i+2)*4), output + (i+2) * 4);
if (i+3 < quaterSize)::convertValue2((short*)(input+(i+3)*4), output + (i+3) * 4);
if (i+4 < quaterSize)::convertValue2((short*)(input+(i+4)*4), output + (i+4) * 4);
if (i+5 < quaterSize)::convertValue2((short*)(input+(i+5)*4), output + (i+5) * 4);
if (i+6 < quaterSize)::convertValue2((short*)(input+(i+6)*4), output + (i+6) * 4);
if (i+7 < quaterSize)::convertValue2((short*)(input+(i+7)*4), output + (i+7) * 4);
}
return Matrix::New(output,aMatrix.getDimSize(0),aMatrix.getDimSize(1),aMatrix.getDimSize(2));
return Matrix::New(output,aRows,aColumns,1);
}

2
src/Function1D.h
View File

@@ -113,7 +113,7 @@ namespace Aurora {
*/
void padding(Matrix& aMatrix, int aIndex, double aValue);
Matrix convertfp16tofloat(const Matrix& aMatrix);
Matrix convertfp16tofloat(short* aData, int aRows, int aColumns);
};