Set NvOptimusEnablement.

[xonotic/darkplaces.git] / mod_skeletal_animatevertices_sse.c

diff --git a/mod_skeletal_animatevertices_sse.c b/mod_skeletal_animatevertices_sse.c
index a25a9224143d4919697305b6bafce3faba23a671..648ab31a15460b7f333e6a8686d8df28bbbc554c 100644 (file)
--- a/mod_skeletal_animatevertices_sse.c
+++ b/mod_skeletal_animatevertices_sse.c
@@ -43,6 +43,7 @@ void Mod_Skeletal_AnimateVertices_SSE(const dp_model_t * RESTRICT model, const f
                                const matrix4x4_t * RESTRICT p = &bonepose[model->data_bones[i].parent];
                                __m128 s0 = _mm_loadu_ps(s->m[0]), s1 = _mm_loadu_ps(s->m[1]), s2 = _mm_loadu_ps(s->m[2]);
 #ifdef OPENGLORIENTATION
+                               __m128 s3 = _mm_loadu_ps(s->m[3]);
 #define SKELETON_MATRIX(r, c) _mm_shuffle_ps(s##c, s##c, _MM_SHUFFLE(r, r, r, r))
 #else
 #define SKELETON_MATRIX(r, c) _mm_shuffle_ps(s##r, s##r, _MM_SHUFFLE(c, c, c, c))
@@ -93,6 +94,7 @@ void Mod_Skeletal_AnimateVertices_SSE(const dp_model_t * RESTRICT model, const f
                        r1 = _mm_add_ps(r1, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1))));
                        r2 = _mm_add_ps(r2, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2))));
                        r3 = _mm_add_ps(r3, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3))));
+                       r3 = _mm_add_ps(r3, b3);
                        _mm_store_ps(r->m[0], r0);
                        _mm_store_ps(r->m[1], r1);
                        _mm_store_ps(r->m[2], r2);
@@ -114,7 +116,7 @@ void Mod_Skeletal_AnimateVertices_SSE(const dp_model_t * RESTRICT model, const f
                                dy = -tx*rz + ty*rw + tz*rx,
                                dz = tx*ry - ty*rx + tz*rw,
                                dw = -tx*rx - ty*ry - tz*rz,
-                               scale, originscale;
+                               scale, sx, sy, sz, sw;
                        for (blends = 1;blends < MAX_FRAMEBLENDS && frameblend[blends].lerp > 0;blends++)
                        {
                                const short * RESTRICT pose7s = model->data_poses7s + 7 * (frameblend[blends].subframe * model->num_bones + i);
@@ -136,19 +138,22 @@ void Mod_Skeletal_AnimateVertices_SSE(const dp_model_t * RESTRICT model, const f
                                dw += -tx*qx - ty*qy - tz*qz;
                        }
                        scale = 1.0f / (rx*rx + ry*ry + rz*rz + rw*rw);
-                       originscale = -model->num_posescale * scale;
-                       m[0] = scale*(rw*rw + rx*rx - ry*ry - rz*rz);
-                       m[1] = 2*scale*(rx*ry - rw*rz);
-                       m[2] = 2*scale*(rx*rz + rw*ry);
-                       m[3] = originscale*(dw*rx - dx*rw + dy*rz - dz*ry);
-                       m[4] = 2*scale*(rx*ry + rw*rz);
-                       m[5] = scale*(rw*rw + ry*ry - rx*rx - rz*rz);
-                       m[6] = 2*scale*(ry*rz - rw*rx);
-                       m[7] = originscale*(dw*ry - dx*rz - dy*rw + dz*rx);
-                       m[8] = 2*scale*(rx*rz - rw*ry);
-                       m[9] = 2*scale*(ry*rz + rw*rx);
-                       m[10] = scale*(rw*rw + rz*rz - rx*rx - ry*ry);
-                       m[11] = originscale*(dw*rz + dx*ry - dy*rx - dz*rw);
+                       sx = rx * scale;
+                       sy = ry * scale;
+                       sz = rz * scale;
+                       sw = rw * scale;
+                       m[0] = sw*rw + sx*rx - sy*ry - sz*rz;
+                       m[1] = 2*(sx*ry - sw*rz);
+                       m[2] = 2*(sx*rz + sw*ry);
+                       m[3] = model->num_posescale*(dx*sw - dy*sz + dz*sy - dw*sx);
+                       m[4] = 2*(sx*ry + sw*rz);
+                       m[5] = sw*rw + sy*ry - sx*rx - sz*rz;
+                       m[6] = 2*(sy*rz - sw*rx);
+                       m[7] = model->num_posescale*(dx*sz + dy*sw - dz*sx - dw*sy);
+                       m[8] = 2*(sx*rz - sw*ry);
+                       m[9] = 2*(sy*rz + sw*rx);
+                       m[10] = sw*rw + sz*rz - sx*rx - sy*ry;
+                       m[11] = model->num_posescale*(dy*sx + dz*sw - dx*sy - dw*sz);
                        if (i == r_skeletal_debugbone.integer)
                                m[r_skeletal_debugbonecomponent.integer % 12] += r_skeletal_debugbonevalue.value;
                        m[3] *= r_skeletal_debugtranslatex.value;