External/cglm
1
0
Fork 0
mirror of https://github.com/recp/cglm.git synced 2025-10-04 01:00:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

opitimize mat4 operations with fma

Browse Source
This commit is contained in:
Recep Aslantas
2021-04-18 13:50:51 +03:00
parent c5655bbd2e
commit abe29a788a
1 changed files with 95 additions and 135 deletions
Download Patch File Download Diff File Expand all files Collapse all files

194
include/cglm/simd/sse2/mat4.h
View File

@@ -56,46 +56,37 @@ glm_mat4_mul_sse2(mat4 m1, mat4 m2, mat4 dest) {
l2 = glmm_load(m1[2]);
l3 = glmm_load(m1[3]);
r = glmm_load(m2[0]);
glmm_store(dest[0],
_mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
_mm_mul_ps(glmm_shuff1x(r, 1), l1)),
_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 2), l2),
_mm_mul_ps(glmm_shuff1x(r, 3), l3))));
r = glmm_load(m2[1]);
glmm_store(dest[1],
_mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
_mm_mul_ps(glmm_shuff1x(r, 1), l1)),
_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 2), l2),
_mm_mul_ps(glmm_shuff1x(r, 3), l3))));
r = glmm_load(m2[2]);
glmm_store(dest[2],
_mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
_mm_mul_ps(glmm_shuff1x(r, 1), l1)),
_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 2), l2),
_mm_mul_ps(glmm_shuff1x(r, 3), l3))));
#define XX(C) \
\
r = glmm_load(m2[C]); \
glmm_store(dest[C], \
glmm_fmadd(glmm_shuff1x(r, 0), l0, \
glmm_fmadd(glmm_shuff1x(r, 1), l1, \
glmm_fmadd(glmm_shuff1x(r, 2), l2, \
_mm_mul_ps(glmm_shuff1x(r, 3), \
l3)))));
r = glmm_load(m2[3]);
glmm_store(dest[3],
_mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
_mm_mul_ps(glmm_shuff1x(r, 1), l1)),
_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 2), l2),
_mm_mul_ps(glmm_shuff1x(r, 3), l3))));
XX(0);
XX(1);
XX(2);
XX(3);
#undef XX
}
CGLM_INLINE
void
glm_mat4_mulv_sse2(mat4 m, vec4 v, vec4 dest) {
__m128 x0, x1, x2;
__m128 x0, x1;
x0 = glmm_load(v);
x1 = _mm_add_ps(_mm_mul_ps(glmm_load(m[0]), glmm_shuff1x(x0, 0)),
_mm_mul_ps(glmm_load(m[1]), glmm_shuff1x(x0, 1)));
x1 = glmm_fmadd(glmm_load(m[0]), glmm_shuff1x(x0, 0),
glmm_fmadd(glmm_load(m[1]), glmm_shuff1x(x0, 1),
glmm_fmadd(glmm_load(m[2]), glmm_shuff1x(x0, 2),
_mm_mul_ps(glmm_load(m[3]),
glmm_shuff1x(x0, 3)))));
x2 = _mm_add_ps(_mm_mul_ps(glmm_load(m[2]), glmm_shuff1x(x0, 2)),
_mm_mul_ps(glmm_load(m[3]), glmm_shuff1x(x0, 3)));
glmm_store(dest, _mm_add_ps(x1, x2));
glmm_store(dest, x1);
}
CGLM_INLINE
@@ -115,20 +106,18 @@ glm_mat4_det_sse2(mat4 mat) {
t[3] = i * p - m * l;
t[4] = i * o - m * k;
*/
x0 = _mm_sub_ps(_mm_mul_ps(glmm_shuff1(r2, 0, 0, 1, 1),
glmm_shuff1(r3, 2, 3, 2, 3)),
_mm_mul_ps(glmm_shuff1(r3, 0, 0, 1, 1),
glmm_shuff1(r2, 2, 3, 2, 3)));
x0 = glmm_fnmadd(glmm_shuff1(r3, 0, 0, 1, 1), glmm_shuff1(r2, 2, 3, 2, 3),
_mm_mul_ps(glmm_shuff1(r2, 0, 0, 1, 1),
glmm_shuff1(r3, 2, 3, 2, 3)));
/*
t[0] = k * p - o * l;
t[0] = k * p - o * l;
t[5] = i * n - m * j;
t[5] = i * n - m * j;
*/
x1 = _mm_sub_ps(_mm_mul_ps(glmm_shuff1(r2, 0, 0, 2, 2),
glmm_shuff1(r3, 1, 1, 3, 3)),
_mm_mul_ps(glmm_shuff1(r3, 0, 0, 2, 2),
glmm_shuff1(r2, 1, 1, 3, 3)));
x1 = glmm_fnmadd(glmm_shuff1(r3, 0, 0, 2, 2), glmm_shuff1(r2, 1, 1, 3, 3),
_mm_mul_ps(glmm_shuff1(r2, 0, 0, 2, 2),
glmm_shuff1(r3, 1, 1, 3, 3)));
/*
a * (f * t[0] - g * t[1] + h * t[2])
@@ -136,21 +125,16 @@ glm_mat4_det_sse2(mat4 mat) {
+ c * (e * t[1] - f * t[3] + h * t[5])
- d * (e * t[2] - f * t[4] + g * t[5])
*/
x2 = _mm_sub_ps(_mm_mul_ps(glmm_shuff1(r1, 0, 0, 0, 1),
_mm_shuffle_ps(x1, x0, _MM_SHUFFLE(1, 0, 0, 0))),
_mm_mul_ps(glmm_shuff1(r1, 1, 1, 2, 2),
glmm_shuff1(x0, 3, 2, 2, 0)));
x2 = glmm_fnmadd(glmm_shuff1(r1, 1, 1, 2, 2), glmm_shuff1(x0, 3, 2, 2, 0),
_mm_mul_ps(glmm_shuff1(r1, 0, 0, 0, 1),
_mm_shuffle_ps(x1, x0, _MM_SHUFFLE(1, 0, 0, 0))));
x2 = glmm_fmadd(glmm_shuff1(r1, 2, 3, 3, 3),
_mm_shuffle_ps(x0, x1, _MM_SHUFFLE(2, 2, 3, 1)),
x2);
x2 = _mm_add_ps(x2,
_mm_mul_ps(glmm_shuff1(r1, 2, 3, 3, 3),
_mm_shuffle_ps(x0, x1, _MM_SHUFFLE(2, 2, 3, 1))));
x2 = _mm_xor_ps(x2, _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
x0 = _mm_mul_ps(r0, x2);
x0 = _mm_add_ps(x0, glmm_shuff1(x0, 0, 1, 2, 3));
x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 3, 3, 1));
return _mm_cvtss_f32(x0);
return glmm_hadd(_mm_mul_ps(x2, r0));
}
CGLM_INLINE
@@ -159,7 +143,10 @@ glm_mat4_inv_fast_sse2(mat4 mat, mat4 dest) {
__m128 r0, r1, r2, r3,
v0, v1, v2, v3,
t0, t1, t2, t3, t4, t5,
x0, x1, x2, x3, x4, x5, x6, x7;
x0, x1, x2, x3, x4, x5, x6, x7, x8, x9;
x8 = _mm_set_ps(-0.f, 0.f, -0.f, 0.f);
x9 = glmm_shuff1(x8, 2, 1, 2, 1);
/* 127 <- 0 */
r0 = glmm_load(mat[0]); /* d c b a */
@@ -177,7 +164,7 @@ glm_mat4_inv_fast_sse2(mat4 mat, mat4 dest) {
t1[0] = k * p - o * l;
t2[0] = g * p - o * h;
t3[0] = g * l - k * h; */
t0 = _mm_sub_ps(_mm_mul_ps(x3, x1), _mm_mul_ps(x2, x0));
t0 = glmm_fnmadd(x2, x0, _mm_mul_ps(x3, x1));
x4 = _mm_shuffle_ps(r2, r3, _MM_SHUFFLE(2, 1, 2, 1)); /* o n k j */
x4 = glmm_shuff1(x4, 0, 2, 2, 2); /* j n n n */
@@ -187,13 +174,13 @@ glm_mat4_inv_fast_sse2(mat4 mat, mat4 dest) {
t1[1] = j * p - n * l;
t2[1] = f * p - n * h;
t3[1] = f * l - j * h; */
t1 = _mm_sub_ps(_mm_mul_ps(x5, x1), _mm_mul_ps(x4, x0));
t1 = glmm_fnmadd(x4, x0, _mm_mul_ps(x5, x1));
/* t1[2] = j * o - n * k
t1[2] = j * o - n * k;
t2[2] = f * o - n * g;
t3[2] = f * k - j * g; */
t2 = _mm_sub_ps(_mm_mul_ps(x5, x2), _mm_mul_ps(x4, x3));
t2 = glmm_fnmadd(x4, x3, _mm_mul_ps(x5, x2));
x6 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(0, 0, 0, 0)); /* e e i i */
x7 = glmm_shuff2(r3, r2, 0, 0, 0, 0, 2, 0, 0, 0); /* i m m m */
@@ -202,19 +189,19 @@ glm_mat4_inv_fast_sse2(mat4 mat, mat4 dest) {
t1[3] = i * p - m * l;
t2[3] = e * p - m * h;
t3[3] = e * l - i * h; */
t3 = _mm_sub_ps(_mm_mul_ps(x6, x1), _mm_mul_ps(x7, x0));
t3 = glmm_fnmadd(x7, x0, _mm_mul_ps(x6, x1));
/* t1[4] = i * o - m * k;
t1[4] = i * o - m * k;
t2[4] = e * o - m * g;
t3[4] = e * k - i * g; */
t4 = _mm_sub_ps(_mm_mul_ps(x6, x2), _mm_mul_ps(x7, x3));
t4 = glmm_fnmadd(x7, x3, _mm_mul_ps(x6, x2));
/* t1[5] = i * n - m * j;
t1[5] = i * n - m * j;
t2[5] = e * n - m * f;
t3[5] = e * j - i * f; */
t5 = _mm_sub_ps(_mm_mul_ps(x6, x4), _mm_mul_ps(x7, x5));
t5 = glmm_fnmadd(x7, x5, _mm_mul_ps(x6, x4));
x0 = glmm_shuff2(r1, r0, 0, 0, 0, 0, 2, 2, 2, 0); /* a a a e */
x1 = glmm_shuff2(r1, r0, 1, 1, 1, 1, 2, 2, 2, 0); /* b b b f */
@@ -226,50 +213,35 @@ glm_mat4_inv_fast_sse2(mat4 mat, mat4 dest) {
dest[0][1] =-(b * t1[0] - c * t1[1] + d * t1[2]);
dest[0][2] = b * t2[0] - c * t2[1] + d * t2[2];
dest[0][3] =-(b * t3[0] - c * t3[1] + d * t3[2]); */
v0 = _mm_add_ps(_mm_mul_ps(x3, t2),
_mm_sub_ps(_mm_mul_ps(x1, t0),
_mm_mul_ps(x2, t1)));
v0 = _mm_xor_ps(v0, _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
/*
dest[1][0] =-(e * t1[0] - g * t1[3] + h * t1[4]);
dest[1][1] = a * t1[0] - c * t1[3] + d * t1[4];
dest[1][2] =-(a * t2[0] - c * t2[3] + d * t2[4]);
dest[1][3] = a * t3[0] - c * t3[3] + d * t3[4]; */
v1 = _mm_add_ps(_mm_mul_ps(x3, t4),
_mm_sub_ps(_mm_mul_ps(x0, t0),
_mm_mul_ps(x2, t3)));
v1 = _mm_xor_ps(v1, _mm_set_ps(0.f, -0.f, 0.f, -0.f));
v0 = _mm_xor_ps(glmm_fmadd(x3, t2, glmm_fnmadd(x2, t1, _mm_mul_ps(x1, t0))), x8);
/*
dest[2][0] = e * t1[1] - f * t1[3] + h * t1[5];
dest[2][1] =-(a * t1[1] - b * t1[3] + d * t1[5]);
dest[2][2] = a * t2[1] - b * t2[3] + d * t2[5];
dest[2][3] =-(a * t3[1] - b * t3[3] + d * t3[5]);*/
v2 = _mm_add_ps(_mm_mul_ps(x3, t5),
_mm_sub_ps(_mm_mul_ps(x0, t1),
_mm_mul_ps(x1, t3)));
v2 = _mm_xor_ps(v2, _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
v2 = _mm_xor_ps(glmm_fmadd(x3, t5, glmm_fnmadd(x1, t3, _mm_mul_ps(x0, t1))), x8);
/*
dest[1][0] =-(e * t1[0] - g * t1[3] + h * t1[4]);
dest[1][1] = a * t1[0] - c * t1[3] + d * t1[4];
dest[1][2] =-(a * t2[0] - c * t2[3] + d * t2[4]);
dest[1][3] = a * t3[0] - c * t3[3] + d * t3[4]; */
v1 = _mm_xor_ps(glmm_fmadd(x3, t4, glmm_fnmadd(x2, t3, _mm_mul_ps(x0, t0))), x9);
/*
dest[3][0] =-(e * t1[2] - f * t1[4] + g * t1[5]);
dest[3][1] = a * t1[2] - b * t1[4] + c * t1[5];
dest[3][2] =-(a * t2[2] - b * t2[4] + c * t2[5]);
dest[3][3] = a * t3[2] - b * t3[4] + c * t3[5]; */
v3 = _mm_add_ps(_mm_mul_ps(x2, t5),
_mm_sub_ps(_mm_mul_ps(x0, t2),
_mm_mul_ps(x1, t4)));
v3 = _mm_xor_ps(v3, _mm_set_ps(0.f, -0.f, 0.f, -0.f));
v3 = _mm_xor_ps(glmm_fmadd(x2, t5, glmm_fnmadd(x1, t4, _mm_mul_ps(x0, t2))), x9);
/* determinant */
x0 = _mm_shuffle_ps(v0, v1, _MM_SHUFFLE(0, 0, 0, 0));
x1 = _mm_shuffle_ps(v2, v3, _MM_SHUFFLE(0, 0, 0, 0));
x0 = _mm_shuffle_ps(x0, x1, _MM_SHUFFLE(2, 0, 2, 0));
x0 = _mm_mul_ps(x0, r0);
x0 = _mm_add_ps(x0, glmm_shuff1(x0, 0, 1, 2, 3));
x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 0, 0, 1));
x0 = _mm_rcp_ps(x0);
x0 = _mm_rcp_ps(glmm_vhadd(_mm_mul_ps(x0, r0)));
glmm_store(dest[0], _mm_mul_ps(v0, x0));
glmm_store(dest[1], _mm_mul_ps(v1, x0));
@@ -283,7 +255,10 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
__m128 r0, r1, r2, r3,
v0, v1, v2, v3,
t0, t1, t2, t3, t4, t5,
x0, x1, x2, x3, x4, x5, x6, x7;
x0, x1, x2, x3, x4, x5, x6, x7, x8, x9;
x8 = _mm_set_ps(-0.f, 0.f, -0.f, 0.f);
x9 = glmm_shuff1(x8, 2, 1, 2, 1);
/* 127 <- 0 */
r0 = glmm_load(mat[0]); /* d c b a */
@@ -301,7 +276,7 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
t1[0] = k * p - o * l;
t2[0] = g * p - o * h;
t3[0] = g * l - k * h; */
t0 = _mm_sub_ps(_mm_mul_ps(x3, x1), _mm_mul_ps(x2, x0));
t0 = glmm_fnmadd(x2, x0, _mm_mul_ps(x3, x1));
x4 = _mm_shuffle_ps(r2, r3, _MM_SHUFFLE(2, 1, 2, 1)); /* o n k j */
x4 = glmm_shuff1(x4, 0, 2, 2, 2); /* j n n n */
@@ -311,13 +286,13 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
t1[1] = j * p - n * l;
t2[1] = f * p - n * h;
t3[1] = f * l - j * h; */
t1 = _mm_sub_ps(_mm_mul_ps(x5, x1), _mm_mul_ps(x4, x0));
t1 = glmm_fnmadd(x4, x0, _mm_mul_ps(x5, x1));
/* t1[2] = j * o - n * k
t1[2] = j * o - n * k;
t2[2] = f * o - n * g;
t3[2] = f * k - j * g; */
t2 = _mm_sub_ps(_mm_mul_ps(x5, x2), _mm_mul_ps(x4, x3));
t2 = glmm_fnmadd(x4, x3, _mm_mul_ps(x5, x2));
x6 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(0, 0, 0, 0)); /* e e i i */
x7 = glmm_shuff2(r3, r2, 0, 0, 0, 0, 2, 0, 0, 0); /* i m m m */
@@ -326,19 +301,19 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
t1[3] = i * p - m * l;
t2[3] = e * p - m * h;
t3[3] = e * l - i * h; */
t3 = _mm_sub_ps(_mm_mul_ps(x6, x1), _mm_mul_ps(x7, x0));
t3 = glmm_fnmadd(x7, x0, _mm_mul_ps(x6, x1));
/* t1[4] = i * o - m * k;
t1[4] = i * o - m * k;
t2[4] = e * o - m * g;
t3[4] = e * k - i * g; */
t4 = _mm_sub_ps(_mm_mul_ps(x6, x2), _mm_mul_ps(x7, x3));
t4 = glmm_fnmadd(x7, x3, _mm_mul_ps(x6, x2));
/* t1[5] = i * n - m * j;
t1[5] = i * n - m * j;
t2[5] = e * n - m * f;
t3[5] = e * j - i * f; */
t5 = _mm_sub_ps(_mm_mul_ps(x6, x4), _mm_mul_ps(x7, x5));
t5 = glmm_fnmadd(x7, x5, _mm_mul_ps(x6, x4));
x0 = glmm_shuff2(r1, r0, 0, 0, 0, 0, 2, 2, 2, 0); /* a a a e */
x1 = glmm_shuff2(r1, r0, 1, 1, 1, 1, 2, 2, 2, 0); /* b b b f */
@@ -350,50 +325,35 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
dest[0][1] =-(b * t1[0] - c * t1[1] + d * t1[2]);
dest[0][2] = b * t2[0] - c * t2[1] + d * t2[2];
dest[0][3] =-(b * t3[0] - c * t3[1] + d * t3[2]); */
v0 = _mm_add_ps(_mm_mul_ps(x3, t2),
_mm_sub_ps(_mm_mul_ps(x1, t0),
_mm_mul_ps(x2, t1)));
v0 = _mm_xor_ps(v0, _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
/*
dest[1][0] =-(e * t1[0] - g * t1[3] + h * t1[4]);
dest[1][1] = a * t1[0] - c * t1[3] + d * t1[4];
dest[1][2] =-(a * t2[0] - c * t2[3] + d * t2[4]);
dest[1][3] = a * t3[0] - c * t3[3] + d * t3[4]; */
v1 = _mm_add_ps(_mm_mul_ps(x3, t4),
_mm_sub_ps(_mm_mul_ps(x0, t0),
_mm_mul_ps(x2, t3)));
v1 = _mm_xor_ps(v1, _mm_set_ps(0.f, -0.f, 0.f, -0.f));
v0 = _mm_xor_ps(glmm_fmadd(x3, t2, glmm_fnmadd(x2, t1, _mm_mul_ps(x1, t0))), x8);
/*
dest[2][0] = e * t1[1] - f * t1[3] + h * t1[5];
dest[2][1] =-(a * t1[1] - b * t1[3] + d * t1[5]);
dest[2][2] = a * t2[1] - b * t2[3] + d * t2[5];
dest[2][3] =-(a * t3[1] - b * t3[3] + d * t3[5]);*/
v2 = _mm_add_ps(_mm_mul_ps(x3, t5),
_mm_sub_ps(_mm_mul_ps(x0, t1),
_mm_mul_ps(x1, t3)));
v2 = _mm_xor_ps(v2, _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
v2 = _mm_xor_ps(glmm_fmadd(x3, t5, glmm_fnmadd(x1, t3, _mm_mul_ps(x0, t1))), x8);
/*
dest[1][0] =-(e * t1[0] - g * t1[3] + h * t1[4]);
dest[1][1] = a * t1[0] - c * t1[3] + d * t1[4];
dest[1][2] =-(a * t2[0] - c * t2[3] + d * t2[4]);
dest[1][3] = a * t3[0] - c * t3[3] + d * t3[4]; */
v1 = _mm_xor_ps(glmm_fmadd(x3, t4, glmm_fnmadd(x2, t3, _mm_mul_ps(x0, t0))), x9);
/*
dest[3][0] =-(e * t1[2] - f * t1[4] + g * t1[5]);
dest[3][1] = a * t1[2] - b * t1[4] + c * t1[5];
dest[3][2] =-(a * t2[2] - b * t2[4] + c * t2[5]);
dest[3][3] = a * t3[2] - b * t3[4] + c * t3[5]; */
v3 = _mm_add_ps(_mm_mul_ps(x2, t5),
_mm_sub_ps(_mm_mul_ps(x0, t2),
_mm_mul_ps(x1, t4)));
v3 = _mm_xor_ps(v3, _mm_set_ps(0.f, -0.f, 0.f, -0.f));
v3 = _mm_xor_ps(glmm_fmadd(x2, t5, glmm_fnmadd(x1, t4, _mm_mul_ps(x0, t2))), x9);
/* determinant */
x0 = _mm_shuffle_ps(v0, v1, _MM_SHUFFLE(0, 0, 0, 0));
x1 = _mm_shuffle_ps(v2, v3, _MM_SHUFFLE(0, 0, 0, 0));
x0 = _mm_shuffle_ps(x0, x1, _MM_SHUFFLE(2, 0, 2, 0));
x0 = _mm_mul_ps(x0, r0);
x0 = _mm_add_ps(x0, glmm_shuff1(x0, 0, 1, 2, 3));
x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 0, 0, 1));
x0 = _mm_div_ps(_mm_set1_ps(1.0f), x0);
x0 = _mm_div_ps(_mm_set1_ps(1.0f), glmm_vhadd(_mm_mul_ps(x0, r0)));
glmm_store(dest[0], _mm_mul_ps(v0, x0));
glmm_store(dest[1], _mm_mul_ps(v1, x0));