scottgs::FloatMatrix scottgs::MatrixMultiply::multiply(const scottgs::FloatMatrix& lhs, const scottgs::FloatMatrix& rhs) const
{
// Verify acceptable dimensions
if (lhs.size2() != rhs.size1())
throw std::logic_error("matrix incompatible lhs.size2() != rhs.size1()");
return boost::numeric::ublas::prod(lhs,rhs);
}
scottgs::FloatMatrix scottgs::MatrixMultiply::operator()(const scottgs::FloatMatrix& lhs, const scottgs::FloatMatrix& rhs) const
{
// Verify acceptable dimensions
if (lhs.size2() != rhs.size1())
throw std::logic_error("matrix incompatible lhs.size2() != rhs.size1()");
scottgs::FloatMatrix result(lhs.size1(),rhs.size2());
// YOUR ALGORIHM WITH COMMENTS GOES HERE:
int i, j, k;
scottgs::FloatMatrix tMatrix = scottgs::MatrixMultiply::transpose(rhs);
std::vector<float> vlhs = makeVector(lhs);
std::vector<float> vrhs = makeVector(tMatrix);
for (i = 0; i < lhs.size1(); ++i)
{
for (j = 0; j < rhs.size2(); ++j)
{
for (k = 0; k < lhs.size2(); ++k)
{
//result(i, j) += lhs(i,k) * rhs(k,j);
//result(i, j) += vlhs[k + i * lhs.size2()] * vrhs[j + k * rhs.size2() - 1];
result(i, j) += vlhs[k + i * lhs.size2()] * vrhs[k + j * rhs.size1()];
}
}
}
return result;
}
scottgs::FloatMatrix scottgs::MatrixMultiply::transpose(const scottgs::FloatMatrix& matrix) const
{
scottgs::FloatMatrix tMatrix(matrix.size2(), matrix.size1());
int i, j;
for (i = 0; i < matrix.size1(); ++i)
{
for (j = 0; j < matrix.size2(); ++j)
{
tMatrix(j, i) = matrix(i, j);
}
}
return tMatrix;
}
std::vector<float> scottgs::MatrixMultiply::makeVectorTransposed(const scottgs::FloatMatrix& matrix) const
{
std::vector<float> vmatrix;
int i;
int j;
for (i = 0; i < matrix.size1(); ++i)
{
for (j = 0; j < matrix.size2(); ++j)
{
vmatrix.push_back(matrix(j, i));
}
}
return vmatrix;
}
scottgs::FloatMatrix scottgs::MatrixMultiply::operator()(const scottgs::FloatMatrix& lhs, const scottgs::FloatMatrix& rhs) const
{
// Verify acceptable dimensions
if (lhs.size2() != rhs.size1())
throw std::logic_error("matrix incompatible lhs.size2() != rhs.size1()");
//create the matrix
scottgs::FloatMatrix result(lhs.size1(),rhs.size2());
//create the unsigned ints used for the three for loops
unsigned int i,j,k,ii,jj,kk; //matrices indexes
//get the sizes I need and put them into a constant
const unsigned int m1_num_row = lhs.size1(); //# of row of matrix 1
const unsigned int m1_num_col = lhs.size2(); //# of col of matrix 1
const unsigned int m2_num_row = rhs.size1(); //# of row of matrix 2
const unsigned int m2_num_col = rhs.size2(); //# of column of matrix 2
//block size calculation
//2 * (blockSize)^2 * 4 = 32768 (L1 cache)
//SQRT(32768/8) = blockSize = 64
const int block_size = 64;
//get a reference of the matrix's first element ( this will be a pointer to the first element )
const float *m1 = &lhs(0,0);
const float *m2 = &rhs(0,0);
//get a copy of the first element.
float *r = &result(0,0);
for (i = 0; i < m1_num_row; ++i)
//loop through each column of matrix 2
for (j = 0; j < m2_num_col; ++j)
//loop through each column of matrix 1
for (k = 0; k < m2_num_row; ++k)
r[i*m2_num_col + j] = r[i*m2_num_col + j] + m1[i*m1_num_col + k] *m2[k*m2_num_col + j];
return result;
}
scottgs::FloatMatrix scottgs::MatrixMultiply::operator()(const scottgs::FloatMatrix& lhs, const scottgs::FloatMatrix& rhs) const
{
// Verify acceptable dimensions
if (lhs.size2() != rhs.size1())
throw std::logic_error("matrix incompatible lhs.size2() != rhs.size1()");
scottgs::FloatMatrix result(lhs.size1(),rhs.size2());
//create the unsigned ints used for the three for loops
unsigned int i,j,k,ii,jj,kk; //matrices indexes
//get the sizes I need and put them into a constant
const unsigned int m1_num_row = lhs.size1(); //# of row of matrix 1
const unsigned int m1_num_col = lhs.size2(); //# of col of matrix 1
const unsigned int m2_num_row = rhs.size1(); //# of row of matrix 2
const unsigned int m2_num_col = rhs.size2(); //# of column of matrix 2
//block size calculation
//2 * (blockSize)^2 * 4 = 32768 (L1 cache)
//SQRT(32768/8) = blockSize = 64
const int block_size = 64;
//get a reference of the matrix's first element ( this will be a pointer to the first element )
const float *m1 = &lhs(0,0);
//malloc memory for the soon to be transposed matrix ( matrix #2)
//float *transposed = (float*)malloc(sizeof(float)*m2_num_col*m2_num_row);
std::vector<float> transposed;
transposed.resize(m2_num_col*m2_num_row);
//get a copy of the first element.
float *r = &result(0,0);
//transpose the second matrix
for (j = 0; j < m2_num_col; ++j)
for (i = 0; i < m2_num_row; ++i)
transposed[j*m2_num_row + i] = m1[i*m2_num_col + j];
if(m1_num_row > 100 || m1_num_col > 100 || m2_num_row > 100 || m2_num_col > 100){
for (i = 0; i < m1_num_row; ++i)
//loop through each column of matrix 2
for (j = 0; j < m2_num_col; ++j)
//loop through each column of matrix 1
for (k = 0; k < m1_num_col; ++k)
r[i*m2_num_col + j] = r[i*m2_num_col + j] + m1[i*m1_num_col + k] * transposed[j*m1_num_col + k];
}else{
//loop through each row of matrix 1
for (ii = 0; ii < m1_num_row; ii+=block_size)
//loop through each column of matrix 2
for (jj = 0; jj < m2_num_col; jj+=block_size)
//loop through each column of matrix 1
for (kk = 0; kk < m1_num_col; kk+=block_size)
for (i = ii; i < std::min(m1_num_row, ii+block_size); ++i)
//loop through each column of matrix 2
for (j = jj; j < std::min(m2_num_col, jj+block_size); ++j)
//loop through each column of matrix 1
for (k = kk; k < std::min(m1_num_col, kk+block_size); ++k)
r[i*m2_num_col + j] = r[i*m2_num_col + j] + m1[i*m1_num_col + k] *transposed[j*m1_num_col + k];
}
return result;
}
