External/cglm
1
0
Fork 0
mirror of https://github.com/recp/cglm.git synced 2025-10-04 09:08:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

simd, sse: optimize mat4 inv with sse

Browse Source 
* reduce a few shufflees
* re-oder instructions for ILP
This commit is contained in:
Recep Aslantas
2021-04-30 21:12:17 +03:00
parent 1b3b91fe0b
commit c065d71a2f
1 changed files with 104 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

140
include/cglm/simd/sse2/mat4.h
View File

@@ -153,87 +153,111 @@ glm_mat4_inv_fast_sse2(mat4 mat, mat4 dest) {
r2 = glmm_load(mat[2]); /* l k j i */
r3 = glmm_load(mat[3]); /* p o n m */
x0 = _mm_shuffle_ps(r2, r3, _MM_SHUFFLE(3, 2, 3, 2)); /* p o l k */
x0 = _mm_movehl_ps(r3, r2); /* p o l k */
x3 = _mm_movelh_ps(r2, r3); /* n m j i */
x1 = glmm_shuff1(x0, 1, 3, 3 ,3); /* l p p p */
x2 = glmm_shuff1(x0, 0, 2, 2, 2); /* k o o o */
x0 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(3, 3, 3, 3)); /* h h l l */
x4 = glmm_shuff1(x3, 1, 3, 3, 3); /* j n n n */
x7 = glmm_shuff1(x3, 0, 2, 2, 2); /* i m m m */
x6 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(0, 0, 0, 0)); /* e e i i */
x5 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(1, 1, 1, 1)); /* f f j j */
x3 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(2, 2, 2, 2)); /* g g k k */
x0 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(3, 3, 3, 3)); /* h h l l */
t0 = _mm_mul_ps(x3, x1);
t1 = _mm_mul_ps(x5, x1);
t2 = _mm_mul_ps(x5, x2);
t3 = _mm_mul_ps(x6, x1);
t4 = _mm_mul_ps(x6, x2);
t5 = _mm_mul_ps(x6, x4);
/* t1[0] = k * p - o * l;
t1[0] = k * p - o * l;
t2[0] = g * p - o * h;
t3[0] = g * l - k * h; */
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 */
x5 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(1, 1, 1, 1)); /* f f j j */
t0 = glmm_fnmadd(x2, x0, t0);
/* t1[1] = j * p - n * l;
t1[1] = j * p - n * l;
t2[1] = f * p - n * h;
t3[1] = f * l - j * h; */
t1 = glmm_fnmadd(x4, x0, _mm_mul_ps(x5, x1));
t1 = glmm_fnmadd(x4, x0, t1);
/* 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 = 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 */
t2 = glmm_fnmadd(x4, x3, t2);
/* t1[3] = i * p - m * l;
t1[3] = i * p - m * l;
t2[3] = e * p - m * h;
t3[3] = e * l - i * h; */
t3 = glmm_fnmadd(x7, x0, _mm_mul_ps(x6, x1));
t3 = glmm_fnmadd(x7, x0, t3);
/* 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 = glmm_fnmadd(x7, x3, _mm_mul_ps(x6, x2));
t4 = glmm_fnmadd(x7, x3, t4);
/* 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 = glmm_fnmadd(x7, x5, _mm_mul_ps(x6, x4));
t5 = glmm_fnmadd(x7, x5, t5);
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 */
x2 = glmm_shuff2(r1, r0, 2, 2, 2, 2, 2, 2, 2, 0); /* c c c g */
x3 = glmm_shuff2(r1, r0, 3, 3, 3, 3, 2, 2, 2, 0); /* d d d h */
x4 = _mm_movelh_ps(r0, r1); /* f e b a */
x5 = _mm_movehl_ps(r1, r0); /* h g d c */
x0 = glmm_shuff1(x4, 0, 0, 0, 2); /* a a a e */
x1 = glmm_shuff1(x4, 1, 1, 1, 3); /* b b b f */
x2 = glmm_shuff1(x5, 0, 0, 0, 2); /* c c c g */
x3 = glmm_shuff1(x5, 1, 1, 1, 3); /* d d d h */
v2 = _mm_mul_ps(x0, t1);
v1 = _mm_mul_ps(x0, t0);
v3 = _mm_mul_ps(x0, t2);
v0 = _mm_mul_ps(x1, t0);
v2 = glmm_fnmadd(x1, t3, v2);
v3 = glmm_fnmadd(x1, t4, v3);
v0 = glmm_fnmadd(x2, t1, v0);
v1 = glmm_fnmadd(x2, t3, v1);
v3 = glmm_fmadd(x2, t5, v3);
v0 = glmm_fmadd(x3, t2, v0);
v2 = glmm_fmadd(x3, t5, v2);
v1 = glmm_fmadd(x3, t4, v1);
/*
dest[0][0] = f * t1[0] - g * t1[1] + h * t1[2];
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_xor_ps(glmm_fmadd(x3, t2, glmm_fnmadd(x2, t1, _mm_mul_ps(x1, t0))), x8);
v0 = _mm_xor_ps(v0, 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_xor_ps(glmm_fmadd(x3, t5, glmm_fnmadd(x1, t3, _mm_mul_ps(x0, t1))), x8);
v2 = _mm_xor_ps(v2, 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);
v1 = _mm_xor_ps(v1, 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_xor_ps(glmm_fmadd(x2, t5, glmm_fnmadd(x1, t4, _mm_mul_ps(x0, t2))), x9);
v3 = _mm_xor_ps(v3, x9);
/* determinant */
x0 = _mm_shuffle_ps(v0, v1, _MM_SHUFFLE(0, 0, 0, 0));
@@ -265,87 +289,111 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
r2 = glmm_load(mat[2]); /* l k j i */
r3 = glmm_load(mat[3]); /* p o n m */
x0 = _mm_shuffle_ps(r2, r3, _MM_SHUFFLE(3, 2, 3, 2)); /* p o l k */
x0 = _mm_movehl_ps(r3, r2); /* p o l k */
x3 = _mm_movelh_ps(r2, r3); /* n m j i */
x1 = glmm_shuff1(x0, 1, 3, 3 ,3); /* l p p p */
x2 = glmm_shuff1(x0, 0, 2, 2, 2); /* k o o o */
x0 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(3, 3, 3, 3)); /* h h l l */
x4 = glmm_shuff1(x3, 1, 3, 3, 3); /* j n n n */
x7 = glmm_shuff1(x3, 0, 2, 2, 2); /* i m m m */
x6 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(0, 0, 0, 0)); /* e e i i */
x5 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(1, 1, 1, 1)); /* f f j j */
x3 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(2, 2, 2, 2)); /* g g k k */
x0 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(3, 3, 3, 3)); /* h h l l */
t0 = _mm_mul_ps(x3, x1);
t1 = _mm_mul_ps(x5, x1);
t2 = _mm_mul_ps(x5, x2);
t3 = _mm_mul_ps(x6, x1);
t4 = _mm_mul_ps(x6, x2);
t5 = _mm_mul_ps(x6, x4);
/* t1[0] = k * p - o * l;
t1[0] = k * p - o * l;
t2[0] = g * p - o * h;
t3[0] = g * l - k * h; */
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 */
x5 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(1, 1, 1, 1)); /* f f j j */
t0 = glmm_fnmadd(x2, x0, t0);
/* t1[1] = j * p - n * l;
t1[1] = j * p - n * l;
t2[1] = f * p - n * h;
t3[1] = f * l - j * h; */
t1 = glmm_fnmadd(x4, x0, _mm_mul_ps(x5, x1));
t1 = glmm_fnmadd(x4, x0, t1);
/* 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 = 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 */
t2 = glmm_fnmadd(x4, x3, t2);
/* t1[3] = i * p - m * l;
t1[3] = i * p - m * l;
t2[3] = e * p - m * h;
t3[3] = e * l - i * h; */
t3 = glmm_fnmadd(x7, x0, _mm_mul_ps(x6, x1));
t3 = glmm_fnmadd(x7, x0, t3);
/* 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 = glmm_fnmadd(x7, x3, _mm_mul_ps(x6, x2));
t4 = glmm_fnmadd(x7, x3, t4);
/* 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 = glmm_fnmadd(x7, x5, _mm_mul_ps(x6, x4));
t5 = glmm_fnmadd(x7, x5, t5);
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 */
x2 = glmm_shuff2(r1, r0, 2, 2, 2, 2, 2, 2, 2, 0); /* c c c g */
x3 = glmm_shuff2(r1, r0, 3, 3, 3, 3, 2, 2, 2, 0); /* d d d h */
x4 = _mm_movelh_ps(r0, r1); /* f e b a */
x5 = _mm_movehl_ps(r1, r0); /* h g d c */
x0 = glmm_shuff1(x4, 0, 0, 0, 2); /* a a a e */
x1 = glmm_shuff1(x4, 1, 1, 1, 3); /* b b b f */
x2 = glmm_shuff1(x5, 0, 0, 0, 2); /* c c c g */
x3 = glmm_shuff1(x5, 1, 1, 1, 3); /* d d d h */
v2 = _mm_mul_ps(x0, t1);
v1 = _mm_mul_ps(x0, t0);
v3 = _mm_mul_ps(x0, t2);
v0 = _mm_mul_ps(x1, t0);
v2 = glmm_fnmadd(x1, t3, v2);
v3 = glmm_fnmadd(x1, t4, v3);
v0 = glmm_fnmadd(x2, t1, v0);
v1 = glmm_fnmadd(x2, t3, v1);
v3 = glmm_fmadd(x2, t5, v3);
v0 = glmm_fmadd(x3, t2, v0);
v2 = glmm_fmadd(x3, t5, v2);
v1 = glmm_fmadd(x3, t4, v1);
/*
dest[0][0] = f * t1[0] - g * t1[1] + h * t1[2];
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_xor_ps(glmm_fmadd(x3, t2, glmm_fnmadd(x2, t1, _mm_mul_ps(x1, t0))), x8);
v0 = _mm_xor_ps(v0, 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_xor_ps(glmm_fmadd(x3, t5, glmm_fnmadd(x1, t3, _mm_mul_ps(x0, t1))), x8);
v2 = _mm_xor_ps(v2, 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);
v1 = _mm_xor_ps(v1, 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_xor_ps(glmm_fmadd(x2, t5, glmm_fnmadd(x1, t4, _mm_mul_ps(x0, t2))), x9);
v3 = _mm_xor_ps(v3, x9);
/* determinant */
x0 = _mm_shuffle_ps(v0, v1, _MM_SHUFFLE(0, 0, 0, 0));