mod_skeletal_animatevertices_generic.c

   1 #include "mod_skeletal_animatevertices_generic.h"
   2
   3 typedef struct
   4 {
   5         float f[12];
   6 }
   7 float12_t;
   8
   9 void Mod_Skeletal_AnimateVertices_Generic(const dp_model_t * RESTRICT model, const frameblend_t * RESTRICT frameblend, const skeleton_t *skeleton, float * RESTRICT vertex3f, float * RESTRICT normal3f, float * RESTRICT svector3f, float * RESTRICT tvector3f)
  10 {
  11         // vertex weighted skeletal
  12         int i, k;
  13         int blends;
  14         float12_t *bonepose;
  15         float12_t *boneposerelative;
  16         float m[12];
  17         const blendweights_t * RESTRICT weights;
  18
  19         //unsigned long long ts = rdtsc();
  20         bonepose = (float12_t *) Mod_Skeletal_AnimateVertices_AllocBuffers(sizeof(float12_t) * (model->num_bones*2 + model->surfmesh.num_blends));
  21         boneposerelative = bonepose + model->num_bones;
  22
  23         if (skeleton && !skeleton->relativetransforms)
  24                 skeleton = NULL;
  25
  26         // interpolate matrices
  27         if (skeleton)
  28         {
  29                 for (i = 0;i < model->num_bones;i++)
  30                 {
  31                         Matrix4x4_ToArray12FloatD3D(&skeleton->relativetransforms[i], m);
  32                         if (model->data_bones[i].parent >= 0)
  33                                 R_ConcatTransforms(bonepose[model->data_bones[i].parent].f, m, bonepose[i].f);
  34                         else
  35                                 memcpy(bonepose[i].f, m, sizeof(m));
  36
  37                         // create a relative deformation matrix to describe displacement
  38                         // from the base mesh, which is used by the actual weighting
  39                         R_ConcatTransforms(bonepose[i].f, model->data_baseboneposeinverse + i * 12, boneposerelative[i].f);
  40                 }
  41         }
  42         else
  43         {
  44                 float originscale = model->num_posescale;
  45                 float x,y,z,w,lerp;
  46                 const short * RESTRICT pose6s;
  47
  48                 for (i = 0;i < model->num_bones;i++)
  49                 {
  50                         memset(m, 0, sizeof(m));
  51                         for (blends = 0;blends < MAX_FRAMEBLENDS && frameblend[blends].lerp > 0;blends++)
  52                         {
  53                                 pose6s = model->data_poses6s + 6 * (frameblend[blends].subframe * model->num_bones + i);
  54                                 lerp = frameblend[blends].lerp;
  55                                 x = pose6s[3] * (1.0f / 32767.0f);
  56                                 y = pose6s[4] * (1.0f / 32767.0f);
  57                                 z = pose6s[5] * (1.0f / 32767.0f);
  58                                 w = 1.0f - (x*x+y*y+z*z);
  59                                 w = w > 0.0f ? -sqrt(w) : 0.0f;
  60                                 m[ 0] += (1-2*(y*y+z*z)) * lerp;
  61                                 m[ 1] += (  2*(x*y-z*w)) * lerp;
  62                                 m[ 2] += (  2*(x*z+y*w)) * lerp;
  63                                 m[ 3] += (pose6s[0] * originscale) * lerp;
  64                                 m[ 4] += (  2*(x*y+z*w)) * lerp;
  65                                 m[ 5] += (1-2*(x*x+z*z)) * lerp;
  66                                 m[ 6] += (  2*(y*z-x*w)) * lerp;
  67                                 m[ 7] += (pose6s[1] * originscale) * lerp;
  68                                 m[ 8] += (  2*(x*z-y*w)) * lerp;
  69                                 m[ 9] += (  2*(y*z+x*w)) * lerp;
  70                                 m[10] += (1-2*(x*x+y*y)) * lerp;
  71                                 m[11] += (pose6s[2] * originscale) * lerp;
  72                         }
  73                         VectorNormalize(m       );
  74                         VectorNormalize(m + 4);
  75                         VectorNormalize(m + 8);
  76                         if (i == r_skeletal_debugbone.integer)
  77                                 m[r_skeletal_debugbonecomponent.integer % 12] += r_skeletal_debugbonevalue.value;
  78                         m[3] *= r_skeletal_debugtranslatex.value;
  79                         m[7] *= r_skeletal_debugtranslatey.value;
  80                         m[11] *= r_skeletal_debugtranslatez.value;
  81                         if (model->data_bones[i].parent >= 0)
  82                                 R_ConcatTransforms(bonepose[model->data_bones[i].parent].f, m, bonepose[i].f);
  83                         else
  84                                 memcpy(bonepose[i].f, m, sizeof(m));
  85                         // create a relative deformation matrix to describe displacement
  86                         // from the base mesh, which is used by the actual weighting
  87                         R_ConcatTransforms(bonepose[i].f, model->data_baseboneposeinverse + i * 12, boneposerelative[i].f);
  88                 }
  89         }
  90
  91         // generate matrices for all blend combinations
  92         weights = model->surfmesh.data_blendweights;
  93         for (i = 0;i < model->surfmesh.num_blends;i++, weights++)
  94         {
  95                 float * RESTRICT b = boneposerelative[model->num_bones + i].f;
  96                 const float * RESTRICT m = boneposerelative[weights->index[0]].f;
  97                 float f = weights->influence[0] * (1.0f / 255.0f);
  98                 b[ 0] = f*m[ 0]; b[ 1] = f*m[ 1]; b[ 2] = f*m[ 2]; b[ 3] = f*m[ 3];
  99                 b[ 4] = f*m[ 4]; b[ 5] = f*m[ 5]; b[ 6] = f*m[ 6]; b[ 7] = f*m[ 7];
 100                 b[ 8] = f*m[ 8]; b[ 9] = f*m[ 9]; b[10] = f*m[10]; b[11] = f*m[11];
 101                 for (k = 1;k < 4 && weights->influence[k];k++)
 102                 {
 103                         m = boneposerelative[weights->index[k]].f;
 104                         f = weights->influence[k] * (1.0f / 255.0f);
 105                         b[ 0] += f*m[ 0]; b[ 1] += f*m[ 1]; b[ 2] += f*m[ 2]; b[ 3] += f*m[ 3];
 106                         b[ 4] += f*m[ 4]; b[ 5] += f*m[ 5]; b[ 6] += f*m[ 6]; b[ 7] += f*m[ 7];
 107                         b[ 8] += f*m[ 8]; b[ 9] += f*m[ 9]; b[10] += f*m[10]; b[11] += f*m[11];
 108                 }
 109         }
 110
 111 #define LOAD_MATRIX_SCALAR() const float * RESTRICT m = boneposerelative[*b].f
 112
 113 #define LOAD_MATRIX3() \
 114         LOAD_MATRIX_SCALAR()
 115 #define LOAD_MATRIX4() \
 116         LOAD_MATRIX_SCALAR()
 117
 118 #define TRANSFORM_POSITION_SCALAR(in, out) \
 119         (out)[0] = ((in)[0] * m[0] + (in)[1] * m[1] + (in)[2] * m[ 2] + m[3]); \
 120         (out)[1] = ((in)[0] * m[4] + (in)[1] * m[5] + (in)[2] * m[ 6] + m[7]); \
 121         (out)[2] = ((in)[0] * m[8] + (in)[1] * m[9] + (in)[2] * m[10] + m[11]);
 122 #define TRANSFORM_VECTOR_SCALAR(in, out) \
 123         (out)[0] = ((in)[0] * m[0] + (in)[1] * m[1] + (in)[2] * m[ 2]); \
 124         (out)[1] = ((in)[0] * m[4] + (in)[1] * m[5] + (in)[2] * m[ 6]); \
 125         (out)[2] = ((in)[0] * m[8] + (in)[1] * m[9] + (in)[2] * m[10]);
 126
 127 #define TRANSFORM_POSITION(in, out) \
 128         TRANSFORM_POSITION_SCALAR(in, out)
 129 #define TRANSFORM_VECTOR(in, out) \
 130         TRANSFORM_VECTOR_SCALAR(in, out)
 131
 132         // transform vertex attributes by blended matrices
 133         if (vertex3f)
 134         {
 135                 const float * RESTRICT v = model->surfmesh.data_vertex3f;
 136                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 137                 // special case common combinations of attributes to avoid repeated loading of matrices
 138                 if (normal3f)
 139                 {
 140                         const float * RESTRICT n = model->surfmesh.data_normal3f;
 141                         if (svector3f && tvector3f)
 142                         {
 143                                 const float * RESTRICT sv = model->surfmesh.data_svector3f;
 144                                 const float * RESTRICT tv = model->surfmesh.data_tvector3f;
 145
 146                                 // Note that for SSE each iteration stores one element past end, so we break one vertex short
 147                                 // and handle that with scalars in that case
 148                                 for (i = 0; i < model->surfmesh.num_vertices; i++, v += 3, n += 3, sv += 3, tv += 3, b++,
 149                                                 vertex3f += 3, normal3f += 3, svector3f += 3, tvector3f += 3)
 150                                 {
 151                                         LOAD_MATRIX4();
 152                                         TRANSFORM_POSITION(v, vertex3f);
 153                                         TRANSFORM_VECTOR(n, normal3f);
 154                                         TRANSFORM_VECTOR(sv, svector3f);
 155                                         TRANSFORM_VECTOR(tv, tvector3f);
 156                                 }
 157
 158                                 return;
 159                         }
 160
 161                         for (i = 0;i < model->surfmesh.num_vertices; i++, v += 3, n += 3, b++, vertex3f += 3, normal3f += 3)
 162                         {
 163                                 LOAD_MATRIX4();
 164                                 TRANSFORM_POSITION(v, vertex3f);
 165                                 TRANSFORM_VECTOR(n, normal3f);
 166                         }
 167                 }
 168                 else
 169                 {
 170                         for (i = 0;i < model->surfmesh.num_vertices; i++, v += 3, b++, vertex3f += 3)
 171                         {
 172                                 LOAD_MATRIX4();
 173                                 TRANSFORM_POSITION(v, vertex3f);
 174                         }
 175                 }
 176         }
 177
 178         else if (normal3f)
 179         {
 180                 const float * RESTRICT n = model->surfmesh.data_normal3f;
 181                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 182                 for (i = 0; i < model->surfmesh.num_vertices; i++, n += 3, b++, normal3f += 3)
 183                 {
 184                         LOAD_MATRIX3();
 185                         TRANSFORM_VECTOR(n, normal3f);
 186                 }
 187         }
 188
 189         if (svector3f)
 190         {
 191                 const float * RESTRICT sv = model->surfmesh.data_svector3f;
 192                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 193                 for (i = 0; i < model->surfmesh.num_vertices; i++, sv += 3, b++, svector3f += 3)
 194                 {
 195                         LOAD_MATRIX3();
 196                         TRANSFORM_VECTOR(sv, svector3f);
 197                 }
 198         }
 199
 200         if (tvector3f)
 201         {
 202                 const float * RESTRICT tv = model->surfmesh.data_tvector3f;
 203                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 204                 for (i = 0; i < model->surfmesh.num_vertices; i++, tv += 3, b++, tvector3f += 3)
 205                 {
 206                         LOAD_MATRIX3();
 207                         TRANSFORM_VECTOR(tv, tvector3f);
 208                 }
 209         }
 210 }