mirror of
https://github.com/recp/cglm.git
synced 2025-12-24 12:32:40 +00:00
sse: optimize glm_mat2_mul_sse2 with sse
This commit is contained in:
@@ -15,20 +15,23 @@
|
||||
CGLM_INLINE
|
||||
void
|
||||
glm_mat2_mul_sse2(mat2 m1, mat2 m2, mat2 dest) {
|
||||
__m128 x0, x1, x2;
|
||||
__m128 x0, x1, x2, x3, x4;
|
||||
|
||||
x1 = glmm_load(m1[0]); /* d c b a */
|
||||
x2 = glmm_load(m2[0]); /* h g f e */
|
||||
|
||||
x3 = glmm_shuff1(x2, 2, 2, 0, 0);
|
||||
x4 = glmm_shuff1(x2, 3, 3, 1, 1);
|
||||
x0 = _mm_movelh_ps(x1, x1);
|
||||
x2 = _mm_movehl_ps(x1, x1);
|
||||
|
||||
/*
|
||||
dest[0][0] = a * e + c * f;
|
||||
dest[0][1] = b * e + d * f;
|
||||
dest[1][0] = a * g + c * h;
|
||||
dest[1][1] = b * g + d * h;
|
||||
*/
|
||||
x0 = glmm_fmadd(_mm_movelh_ps(x1, x1), glmm_shuff1(x2, 2, 2, 0, 0),
|
||||
_mm_mul_ps(_mm_movehl_ps(x1, x1),
|
||||
glmm_shuff1(x2, 3, 3, 1, 1)));
|
||||
x0 = glmm_fmadd(x0, x3, _mm_mul_ps(x2, x4));
|
||||
|
||||
glmm_store(dest[0], x0);
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user