Added document
This commit is contained in:
122
src/spider/runtime/math/Matrix_Multiply.cpp
Normal file
122
src/spider/runtime/math/Matrix_Multiply.cpp
Normal file
@@ -0,0 +1,122 @@
|
||||
#include <iostream>
|
||||
|
||||
|
||||
template<typename T, int Rows, int Cols>
|
||||
struct Matrix {
|
||||
T data[Rows][Cols];
|
||||
|
||||
void setZero() {
|
||||
for (int i = 0; i < Rows; i++) {
|
||||
for (int j = 0; j < Cols; j++) {
|
||||
data[i][j] = T();
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
template<typename T, int M, int N, int P>
|
||||
Matrix<T, M, P> mat_multiply(Matrix<T, M, N> A, Matrix<T, N, P> B) {
|
||||
//Determination of resulting matrix size
|
||||
int RSize;
|
||||
if(M > P){
|
||||
RSize = M ;
|
||||
}
|
||||
else{
|
||||
RSize = P;
|
||||
}
|
||||
|
||||
//Condition for resulting matrix size rounding calculation
|
||||
int CSize;
|
||||
if (RSize <= 4){
|
||||
CSize = 4;
|
||||
}
|
||||
else if (RSize % 4 == 0){
|
||||
CSize = RSize;
|
||||
}
|
||||
else{
|
||||
CSize = 4 * (int(RSize/4) + 1);
|
||||
}
|
||||
|
||||
// SIMD width selection (based on your Python logic)
|
||||
int simd_width;
|
||||
if (RSize > 8) {
|
||||
simd_width = 8;
|
||||
}
|
||||
else if (RSize > 4 && RSize < 8){
|
||||
simd_width = 4;
|
||||
}
|
||||
else {
|
||||
simd_width = RSize;
|
||||
}
|
||||
|
||||
|
||||
// Create result square matrix with size CSize(optimization)
|
||||
Matrix<T, CSize, CSize> tempResult;
|
||||
tempResult.setZero();
|
||||
|
||||
// Matrix multiplication with SIMD-style unrolling
|
||||
for (int i = 0; i < M; i++) { // For each row in A
|
||||
for (int k = 0; k < N; k++) { // For each inner dimension
|
||||
T a_val = A.data[i][k];
|
||||
|
||||
// Process columns in chunks of simd_width
|
||||
int j = 0;
|
||||
while (j < P) {
|
||||
int remaining = P - j;
|
||||
|
||||
if (remaining >= simd_width) {
|
||||
// Full SIMD operation - unrolled loops
|
||||
if (simd_width == 8) {
|
||||
tempResult.data[i][j] += a_val * B.data[k][j];
|
||||
tempResult.data[i][j+1] += a_val * B.data[k][j+1];
|
||||
tempResult.data[i][j+2] += a_val * B.data[k][j+2];
|
||||
tempResult.data[i][j+3] += a_val * B.data[k][j+3];
|
||||
tempResult.data[i][j+4] += a_val * B.data[k][j+4];
|
||||
tempResult.data[i][j+5] += a_val * B.data[k][j+5];
|
||||
tempResult.data[i][j+6] += a_val * B.data[k][j+6];
|
||||
tempResult.data[i][j+7] += a_val * B.data[k][j+7];
|
||||
}
|
||||
else if (simd_width == 4) {
|
||||
tempResult.data[i][j] += a_val * B.data[k][j];
|
||||
tempResult.data[i][j+1] += a_val * B.data[k][j+1];
|
||||
tempResult.data[i][j+2] += a_val * B.data[k][j+2];
|
||||
tempResult.data[i][j+3] += a_val * B.data[k][j+3];
|
||||
}
|
||||
else if (simd_width == 3) {
|
||||
tempResult.data[i][j] += a_val * B.data[k][j];
|
||||
tempResult.data[i][j+1] += a_val * B.data[k][j+1];
|
||||
tempResult.data[i][j+2] += a_val * B.data[k][j+2];
|
||||
}
|
||||
else if (simd_width == 2) {
|
||||
tempResult.data[i][j] += a_val * B.data[k][j];
|
||||
tempResult.data[i][j+1] += a_val * B.data[k][j+1];
|
||||
}
|
||||
else { // simd_width == 1
|
||||
tempResult.data[i][j] += a_val * B.data[k][j];
|
||||
}
|
||||
|
||||
j += simd_width;
|
||||
}
|
||||
else {
|
||||
// Handle remaining columns that don't fit in SIMD width
|
||||
for (int s = 0; s < remaining; s++) {
|
||||
tempResult.data[i][j + s] += a_val * B.data[k][j + s];
|
||||
}
|
||||
j += remaining;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Extract the actual result (M x P) from the temporary square matrix
|
||||
Matrix<T, M, P> result;
|
||||
for (int i = 0; i < M; i++) {
|
||||
for (int j = 0; j < P; j++) {
|
||||
result.data[i][j] = tempResult.data[i][j];
|
||||
}
|
||||
}
|
||||
|
||||
return result;
|
||||
|
||||
|
||||
}
|
||||
Reference in New Issue
Block a user