diff options
Diffstat (limited to 'source/blender/render/intern/raytrace/bvh.h')
| -rw-r--r-- | source/blender/render/intern/raytrace/bvh.h | 12 |
1 files changed, 6 insertions, 6 deletions
diff --git a/source/blender/render/intern/raytrace/bvh.h b/source/blender/render/intern/raytrace/bvh.h index b17a072f3a6..2d75cd800a7 100644 --- a/source/blender/render/intern/raytrace/bvh.h +++ b/source/blender/render/intern/raytrace/bvh.h @@ -196,18 +196,18 @@ static int bvh_node_stack_raycast_simd(Node *root, Isect *isec) const float *bb2 = stack[stack_pos+2]->bb; const float *bb3 = stack[stack_pos+3]->bb; - const __m128 x0y0x1y1 = _mm_shuffle_ps( _mm_loadu_ps(bb0), _mm_loadu_ps(bb1), _MM_SHUFFLE(1,0,1,0) ); - const __m128 x2y2x3y3 = _mm_shuffle_ps( _mm_loadu_ps(bb2), _mm_loadu_ps(bb3), _MM_SHUFFLE(1,0,1,0) ); + const __m128 x0y0x1y1 = _mm_shuffle_ps( _mm_load_ps(bb0), _mm_load_ps(bb1), _MM_SHUFFLE(1,0,1,0) ); + const __m128 x2y2x3y3 = _mm_shuffle_ps( _mm_load_ps(bb2), _mm_load_ps(bb3), _MM_SHUFFLE(1,0,1,0) ); t_bb[0] = _mm_shuffle_ps( x0y0x1y1, x2y2x3y3, _MM_SHUFFLE(2,0,2,0) ); t_bb[1] = _mm_shuffle_ps( x0y0x1y1, x2y2x3y3, _MM_SHUFFLE(3,1,3,1) ); - const __m128 z0X0z1X1 = _mm_shuffle_ps( _mm_loadu_ps(bb0+2), _mm_loadu_ps(bb1+2), _MM_SHUFFLE(1,0,1,0) ); - const __m128 z2X2z3X3 = _mm_shuffle_ps( _mm_loadu_ps(bb2+2), _mm_loadu_ps(bb3+2), _MM_SHUFFLE(1,0,1,0) ); + const __m128 z0X0z1X1 = _mm_shuffle_ps( _mm_load_ps(bb0), _mm_load_ps(bb1), _MM_SHUFFLE(3,1,3,1) ); + const __m128 z2X2z3X3 = _mm_shuffle_ps( _mm_load_ps(bb2), _mm_load_ps(bb3), _MM_SHUFFLE(3,1,3,1) ); t_bb[2] = _mm_shuffle_ps( z0X0z1X1, z2X2z3X3, _MM_SHUFFLE(2,0,2,0) ); t_bb[3] = _mm_shuffle_ps( z0X0z1X1, z2X2z3X3, _MM_SHUFFLE(3,1,3,1) ); - const __m128 Y0Z0Y1Z1 = _mm_shuffle_ps( _mm_loadu_ps(bb0+4), _mm_loadu_ps(bb1+4), _MM_SHUFFLE(1,0,1,0) ); - const __m128 Y2Z2Y3Z3 = _mm_shuffle_ps( _mm_loadu_ps(bb2+4), _mm_loadu_ps(bb3+4), _MM_SHUFFLE(1,0,1,0) ); + const __m128 Y0Z0Y1Z1 = _mm_shuffle_ps( _mm_load_ps(bb0+4), _mm_load_ps(bb1+4), _MM_SHUFFLE(1,0,1,0) ); + const __m128 Y2Z2Y3Z3 = _mm_shuffle_ps( _mm_load_ps(bb2+4), _mm_load_ps(bb3+4), _MM_SHUFFLE(1,0,1,0) ); t_bb[4] = _mm_shuffle_ps( Y0Z0Y1Z1, Y2Z2Y3Z3, _MM_SHUFFLE(2,0,2,0) ); t_bb[5] = _mm_shuffle_ps( Y0Z0Y1Z1, Y2Z2Y3Z3, _MM_SHUFFLE(3,1,3,1) ); /* |