Skip to content

Commit

Permalink
Merge pull request #116 from beomki-yeo/optimize-cmath
Browse files Browse the repository at this point in the history 
Vectorize the cmath matrix operations
  • Loading branch information
@beomki-yeo
beomki-yeo authored Apr 20, 2024
2 parents 9e684de + 2360abe commit ee9cf86
Show file tree
Hide file tree
Showing 3 changed files with 49 additions and 32 deletions.
20 changes: 10 additions & 10 deletions math/cmath/include/algebra/math/impl/cmath_matrix.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -69,8 +69,8 @@ struct actor {
ALGEBRA_HOST_DEVICE void set_block(input_matrix_type &m,
const matrix_type<ROWS, COLS> &b,
size_type row, size_type col) const {
for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
element_getter()(m, i + row, j + col) = element_getter()(b, i, j);
}
}
Expand All @@ -92,8 +92,8 @@ struct actor {
ALGEBRA_HOST_DEVICE inline matrix_type<ROWS, COLS> zero() const {
matrix_type<ROWS, COLS> ret;

for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
element_getter()(ret, i, j) = 0;
}
}
Expand All @@ -106,8 +106,8 @@ struct actor {
ALGEBRA_HOST_DEVICE inline matrix_type<ROWS, COLS> identity() const {
matrix_type<ROWS, COLS> ret;

for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
if (i == j) {
element_getter()(ret, i, j) = 1;
} else {
Expand All @@ -123,8 +123,8 @@ struct actor {
template <size_type ROWS, size_type COLS>
ALGEBRA_HOST_DEVICE inline void set_zero(matrix_type<ROWS, COLS> &m) const {

for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
element_getter()(m, i, j) = 0;
}
}
Expand All @@ -135,8 +135,8 @@ struct actor {
ALGEBRA_HOST_DEVICE inline void set_identity(
matrix_type<ROWS, COLS> &m) const {

for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
if (i == j) {
element_getter()(m, i, j) = 1;
} else {
Expand Down
45 changes: 23 additions & 22 deletions math/cmath/include/algebra/math/impl/cmath_operators.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -137,8 +137,8 @@ ALGEBRA_HOST_DEVICE inline array_t<array_t<scalar_t, ROWS>, COLS> operator*(

array_t<array_t<scalar_t, ROWS>, COLS> ret;

for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
ret[j][i] = a[j][i] * static_cast<scalar_t>(s);
}
}
Expand All @@ -154,8 +154,8 @@ ALGEBRA_HOST_DEVICE inline array_t<array_t<scalar_t, ROWS>, COLS> operator*(

array_t<array_t<scalar_t, ROWS>, COLS> ret;

for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
ret[j][i] = a[j][i] * static_cast<scalar_t>(s);
}
}
Expand All @@ -171,8 +171,8 @@ ALGEBRA_HOST_DEVICE inline array_t<array_t<scalar_t, ROWS>, COLS> operator*(

array_t<array_t<scalar_t, ROWS>, COLS> ret;

for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
ret[j][i] = a[j][i] * static_cast<scalar_t>(s);
}
}
Expand All @@ -188,8 +188,8 @@ ALGEBRA_HOST_DEVICE inline array_t<array_t<scalar_t, ROWS>, COLS> operator*(

array_t<array_t<scalar_t, ROWS>, COLS> ret;

for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
ret[j][i] = a[j][i] * static_cast<scalar_t>(s);
}
}
Expand All @@ -206,16 +206,17 @@ ALGEBRA_HOST_DEVICE inline array_t<array_t<scalar_t, M>, O> operator*(

array_t<array_t<scalar_t, M>, O> C;

for (size_type i = 0; i < M; ++i) {
for (size_type j = 0; j < O; ++j) {

scalar_t val = 0;
for (size_type j = 0; j < O; ++j) {
for (size_type i = 0; i < M; ++i) {
C[j][i] = 0.f;
}
}

for (size_type k = 0; k < N; ++k) {
val += A[k][i] * B[j][k];
for (size_type i = 0; i < N; ++i) {
for (size_type j = 0; j < O; ++j) {
for (size_type k = 0; k < M; ++k) {
C[j][k] += A[i][k] * B[j][i];
}

C[j][i] = val;
}
}

Expand All @@ -231,8 +232,8 @@ ALGEBRA_HOST_DEVICE inline array_t<array_t<scalar_t, ROWS>, COLS> operator+(

array_t<array_t<scalar_t, ROWS>, COLS> C;

for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
C[j][i] = A[j][i] + B[j][i];
}
}
Expand All @@ -249,8 +250,8 @@ ALGEBRA_HOST_DEVICE inline array_t<array_t<scalar_t, ROWS>, COLS> operator-(

array_t<array_t<scalar_t, ROWS>, COLS> C;

for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
C[j][i] = A[j][i] - B[j][i];
}
}
Expand All @@ -272,8 +273,8 @@ ALGEBRA_HOST_DEVICE inline array_t<scalar_t, ROWS> operator*(

array_t<scalar_t, ROWS> ret{0};

for (size_type i = 0; i < ROWS; ++i) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type j = 0; j < COLS; ++j) {
for (size_type i = 0; i < ROWS; ++i) {
ret[i] += a[j][i] * b[j];
}
}
Expand Down
16 changes: 16 additions & 0 deletions tests/common/test_host_basics.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -226,6 +226,22 @@ TYPED_TEST_P(test_host_basics, matrix64) {
ASSERT_NEAR(algebra::getter::element(m, 0, 2), 10., this->m_epsilon);
ASSERT_NEAR(algebra::getter::element(m, 1, 2), 20., this->m_epsilon);
ASSERT_NEAR(algebra::getter::element(m, 2, 2), 30., this->m_epsilon);

typename TypeParam::template matrix<3, 3> m33;
algebra::getter::element(m33, 0, 0) = 1;
algebra::getter::element(m33, 1, 0) = 2;
algebra::getter::element(m33, 2, 0) = 3;
algebra::getter::element(m33, 0, 1) = 5;
algebra::getter::element(m33, 1, 1) = 6;
algebra::getter::element(m33, 2, 1) = 7;
algebra::getter::element(m33, 0, 2) = 9;
algebra::getter::element(m33, 1, 2) = 10;
algebra::getter::element(m33, 2, 2) = 11;

const typename TypeParam::vector3 v2 = m33 * v;
ASSERT_NEAR(v2[0], 380., this->m_epsilon);
ASSERT_NEAR(v2[1], 440., this->m_epsilon);
ASSERT_NEAR(v2[2], 500., this->m_epsilon);
}

// Test matrix operations with 3x3 matrix
Expand Down

0 comments on commit ee9cf86

Please sign in to comment.