mirror/glm
1
0
Fork 0
mirror of https://github.com/g-truc/glm.git synced 2024-11-14 06:04:34 +00:00
Code Issues Packages Projects Releases Wiki Activity

Remove the SSE3 implementation in operator*(fquatSIMD, fquatSIMD).

Browse Source 
The SSE2 version is now running faster than the SSE3 version.
This commit is contained in:
Dave Reid 2013-04-24 08:51:17 +10:00
parent 197b6c96d8
commit 13837e1079
1 changed files with 14 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
glm/gtx/simd_quat.inl
View File

@ -115,11 +115,6 @@ GLM_FUNC_QUALIFIER fquatSIMD operator* (fquatSIMD const & q1, fquatSIMD const &
// 8 mul // 8 mul
// 8 add // 8 add
// SSE3 STATS:
// 3 shuffle
// 8 mul
// 8 add
// SSE4 STATS: // SSE4 STATS:
// 3 shuffle // 3 shuffle
// 4 mul // 4 mul
@ -137,31 +132,21 @@ GLM_FUNC_QUALIFIER fquatSIMD operator* (fquatSIMD const & q1, fquatSIMD const &
__m128 add3 = _mm_dp_ps(mul3, _mm_set_ps(1.0f, 1.0f, -1.0f, 1.0f), 0xff); __m128 add3 = _mm_dp_ps(mul3, _mm_set_ps(1.0f, 1.0f, -1.0f, 1.0f), 0xff);
# else # else
mul0 = _mm_mul_ps(mul0, _mm_set_ps(1.0f, -1.0f, -1.0f, -1.0f)); mul0 = _mm_mul_ps(mul0, _mm_set_ps(1.0f, -1.0f, -1.0f, -1.0f));
mul1 = _mm_mul_ps(mul1, _mm_set_ps(1.0f, -1.0f, 1.0f, 1.0f));
mul2 = _mm_mul_ps(mul2, _mm_set_ps(1.0f, 1.0f, 1.0f, -1.0f));
mul3 = _mm_mul_ps(mul3, _mm_set_ps(1.0f, 1.0f, -1.0f, 1.0f));
# if((GLM_ARCH & GLM_ARCH_SSE3))
__m128 add0 = _mm_hadd_ps(mul0, mul0);
add0 = _mm_hadd_ps(add0, add0);
__m128 add1 = _mm_hadd_ps(mul1, mul1);
add1 = _mm_hadd_ps(add1, add1);
__m128 add2 = _mm_hadd_ps(mul2, mul2);
add2 = _mm_hadd_ps(add2, add2);
__m128 add3 = _mm_hadd_ps(mul3, mul3);
add3 = _mm_hadd_ps(add3, add3);
# else
__m128 add0 = _mm_add_ps(mul0, _mm_movehl_ps(mul0, mul0)); __m128 add0 = _mm_add_ps(mul0, _mm_movehl_ps(mul0, mul0));
add0 = _mm_add_ss(add0, _mm_shuffle_ps(add0, add0, 1)); add0 = _mm_add_ss(add0, _mm_shuffle_ps(add0, add0, 1));
mul1 = _mm_mul_ps(mul1, _mm_set_ps(1.0f, -1.0f, 1.0f, 1.0f));
__m128 add1 = _mm_add_ps(mul1, _mm_movehl_ps(mul1, mul1)); __m128 add1 = _mm_add_ps(mul1, _mm_movehl_ps(mul1, mul1));
add1 = _mm_add_ss(add1, _mm_shuffle_ps(add1, add1, 1)); add1 = _mm_add_ss(add1, _mm_shuffle_ps(add1, add1, 1));
mul2 = _mm_mul_ps(mul2, _mm_set_ps(1.0f, 1.0f, 1.0f, -1.0f));
__m128 add2 = _mm_add_ps(mul2, _mm_movehl_ps(mul2, mul2)); __m128 add2 = _mm_add_ps(mul2, _mm_movehl_ps(mul2, mul2));
add2 = _mm_add_ss(add2, _mm_shuffle_ps(add2, add2, 1)); add2 = _mm_add_ss(add2, _mm_shuffle_ps(add2, add2, 1));
mul3 = _mm_mul_ps(mul3, _mm_set_ps(1.0f, 1.0f, -1.0f, 1.0f));
__m128 add3 = _mm_add_ps(mul3, _mm_movehl_ps(mul3, mul3)); __m128 add3 = _mm_add_ps(mul3, _mm_movehl_ps(mul3, mul3));
add3 = _mm_add_ss(add3, _mm_shuffle_ps(add3, add3, 1)); add3 = _mm_add_ss(add3, _mm_shuffle_ps(add3, add3, 1));
#endif #endif
#endif
// I had tried something clever here using shuffles to produce the final result, but it turns out that using // I had tried something clever here using shuffles to produce the final result, but it turns out that using