我目前正在try 优化我为一个24x24矩阵三角化程序编写的一些MIPS汇编程序.我目前的目标是利用延迟分支和手动循环展开来try 减少循环.Note: I am using 32-bit single precision for all the matrix arithmetic.
算法的一部分涉及我试图展开的以下循环(N始终为24)
...
float inv = 1/A[k][k]
for (j = k + 1; j < N; j++) {
/* divide by pivot element */
A[k][j] = A[k][j] * inv;
}
...
我想要
...
float inv = 1/A[k][k]
for (j = k + 1; j < N; j +=2) {
/* divide by pivot element */
A[k][j] = A[k][j] * inv;
A[k][j + 1] = A[k][j + 1] * inv;
}
...
但它产生了错误的结果，我不知道为什么.有趣的是，使用循环展开的版本正确地生成了矩阵的第一行，但其余的不正确.没有循环展开的版本可以正确地对矩阵进行三角化.
以下是我的try .
...
# No loop unrolling
loop_2:
move $a3, $t2 # column number b = j (getelem A[k][j])
jal getelem # Addr of A[k][j] in $v0 and val in $f0
addiu $t2, $t2, 1 ## j += 2
mul.s $f0, $f0, $f2 # Perform A[k][j] * inv
bltu $t2, 24, loop_2 # if j < N, jump to loop_2
swc1 $f0, 0($v0) ## Perform A[k][j] := A[k][j] * inv
# The matrix triangulates without problem with this original code.
...
...
# One loop unrolling
loop_2:
move $a3, $t2 # column number b = j (getelem A[k][j])
jal getelem # Addr of A[k][j] in $v0 and val in $f0
addiu $t2, $t2, 2 ## j += 2
lwc1 $f1, 4($v0) # $f1 <- A[k][j + 1]
mul.s $f0, $f0, $f2 # Perform A[k][j] * inv
mul.s $f1, $f1, $f2 # Perform A[k][j+1] * inv
swc1 $f0, 0($v0) # Perform A[k][j] := A[k][j] * inv
bltu $t2, 24, loop_2 # if j < N, jump to loop_2
swc1 $f1, 4($v0) ## Perform A[k][j + 1] := A[k][j + 1] * inv
# The first row in the resulting matrix is correct, but the remaining ones not when using this once unrolled loop code.
...