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         if (!model->surfmesh.num_vertices)
  20                 return;
  21
  22         //unsigned long long ts = rdtsc();
  23         bonepose = (float12_t *) Mod_Skeletal_AnimateVertices_AllocBuffers(sizeof(float12_t) * (model->num_bones*2 + model->surfmesh.num_blends));
  24         boneposerelative = bonepose + model->num_bones;
  25
  26         if (skeleton && !skeleton->relativetransforms)
  27                 skeleton = NULL;
  28
  29         // interpolate matrices
  30         if (skeleton)
  31         {
  32                 for (i = 0;i < model->num_bones;i++)
  33                 {
  34                         Matrix4x4_ToArray12FloatD3D(&skeleton->relativetransforms[i], m);
  35                         if (model->data_bones[i].parent >= 0)
  36                                 R_ConcatTransforms(bonepose[model->data_bones[i].parent].f, m, bonepose[i].f);
  37                         else
  38                                 memcpy(bonepose[i].f, m, sizeof(m));
  39
  40                         // create a relative deformation matrix to describe displacement
  41                         // from the base mesh, which is used by the actual weighting
  42                         R_ConcatTransforms(bonepose[i].f, model->data_baseboneposeinverse + i * 12, boneposerelative[i].f);
  43                 }
  44         }
  45         else
  46         {
  47                 float originscale = model->num_posescale;
  48                 float x,y,z,w,lerp;
  49                 const short * RESTRICT pose6s;
  50
  51                 for (i = 0;i < model->num_bones;i++)
  52                 {
  53                         memset(m, 0, sizeof(m));
  54                         for (blends = 0;blends < MAX_FRAMEBLENDS && frameblend[blends].lerp > 0;blends++)
  55                         {
  56                                 pose6s = model->data_poses6s + 6 * (frameblend[blends].subframe * model->num_bones + i);
  57                                 lerp = frameblend[blends].lerp;
  58                                 x = pose6s[3] * (1.0f / 32767.0f);
  59                                 y = pose6s[4] * (1.0f / 32767.0f);
  60                                 z = pose6s[5] * (1.0f / 32767.0f);
  61                                 w = 1.0f - (x*x+y*y+z*z);
  62                                 w = w > 0.0f ? -sqrt(w) : 0.0f;
  63                                 m[ 0] += (1-2*(y*y+z*z)) * lerp;
  64                                 m[ 1] += (  2*(x*y-z*w)) * lerp;
  65                                 m[ 2] += (  2*(x*z+y*w)) * lerp;
  66                                 m[ 3] += (pose6s[0] * originscale) * lerp;
  67                                 m[ 4] += (  2*(x*y+z*w)) * lerp;
  68                                 m[ 5] += (1-2*(x*x+z*z)) * lerp;
  69                                 m[ 6] += (  2*(y*z-x*w)) * lerp;
  70                                 m[ 7] += (pose6s[1] * originscale) * lerp;
  71                                 m[ 8] += (  2*(x*z-y*w)) * lerp;
  72                                 m[ 9] += (  2*(y*z+x*w)) * lerp;
  73                                 m[10] += (1-2*(x*x+y*y)) * lerp;
  74                                 m[11] += (pose6s[2] * originscale) * lerp;
  75                         }
  76                         VectorNormalize(m       );
  77                         VectorNormalize(m + 4);
  78                         VectorNormalize(m + 8);
  79                         if (i == r_skeletal_debugbone.integer)
  80                                 m[r_skeletal_debugbonecomponent.integer % 12] += r_skeletal_debugbonevalue.value;
  81                         m[3] *= r_skeletal_debugtranslatex.value;
  82                         m[7] *= r_skeletal_debugtranslatey.value;
  83                         m[11] *= r_skeletal_debugtranslatez.value;
  84                         if (model->data_bones[i].parent >= 0)
  85                                 R_ConcatTransforms(bonepose[model->data_bones[i].parent].f, m, bonepose[i].f);
  86                         else
  87                                 memcpy(bonepose[i].f, m, sizeof(m));
  88                         // create a relative deformation matrix to describe displacement
  89                         // from the base mesh, which is used by the actual weighting
  90                         R_ConcatTransforms(bonepose[i].f, model->data_baseboneposeinverse + i * 12, boneposerelative[i].f);
  91                 }
  92         }
  93
  94         // generate matrices for all blend combinations
  95         weights = model->surfmesh.data_blendweights;
  96         for (i = 0;i < model->surfmesh.num_blends;i++, weights++)
  97         {
  98                 float * RESTRICT b = boneposerelative[model->num_bones + i].f;
  99                 const float * RESTRICT m = boneposerelative[weights->index[0]].f;
 100                 float f = weights->influence[0] * (1.0f / 255.0f);
 101                 b[ 0] = f*m[ 0]; b[ 1] = f*m[ 1]; b[ 2] = f*m[ 2]; b[ 3] = f*m[ 3];
 102                 b[ 4] = f*m[ 4]; b[ 5] = f*m[ 5]; b[ 6] = f*m[ 6]; b[ 7] = f*m[ 7];
 103                 b[ 8] = f*m[ 8]; b[ 9] = f*m[ 9]; b[10] = f*m[10]; b[11] = f*m[11];
 104                 for (k = 1;k < 4 && weights->influence[k];k++)
 105                 {
 106                         m = boneposerelative[weights->index[k]].f;
 107                         f = weights->influence[k] * (1.0f / 255.0f);
 108                         b[ 0] += f*m[ 0]; b[ 1] += f*m[ 1]; b[ 2] += f*m[ 2]; b[ 3] += f*m[ 3];
 109                         b[ 4] += f*m[ 4]; b[ 5] += f*m[ 5]; b[ 6] += f*m[ 6]; b[ 7] += f*m[ 7];
 110                         b[ 8] += f*m[ 8]; b[ 9] += f*m[ 9]; b[10] += f*m[10]; b[11] += f*m[11];
 111                 }
 112         }
 113
 114 #define LOAD_MATRIX_SCALAR() const float * RESTRICT m = boneposerelative[*b].f
 115
 116 #define LOAD_MATRIX3() \
 117         LOAD_MATRIX_SCALAR()
 118 #define LOAD_MATRIX4() \
 119         LOAD_MATRIX_SCALAR()
 120
 121 #define TRANSFORM_POSITION_SCALAR(in, out) \
 122         (out)[0] = ((in)[0] * m[0] + (in)[1] * m[1] + (in)[2] * m[ 2] + m[3]); \
 123         (out)[1] = ((in)[0] * m[4] + (in)[1] * m[5] + (in)[2] * m[ 6] + m[7]); \
 124         (out)[2] = ((in)[0] * m[8] + (in)[1] * m[9] + (in)[2] * m[10] + m[11]);
 125 #define TRANSFORM_VECTOR_SCALAR(in, out) \
 126         (out)[0] = ((in)[0] * m[0] + (in)[1] * m[1] + (in)[2] * m[ 2]); \
 127         (out)[1] = ((in)[0] * m[4] + (in)[1] * m[5] + (in)[2] * m[ 6]); \
 128         (out)[2] = ((in)[0] * m[8] + (in)[1] * m[9] + (in)[2] * m[10]);
 129
 130 #define TRANSFORM_POSITION(in, out) \
 131         TRANSFORM_POSITION_SCALAR(in, out)
 132 #define TRANSFORM_VECTOR(in, out) \
 133         TRANSFORM_VECTOR_SCALAR(in, out)
 134
 135         // transform vertex attributes by blended matrices
 136         if (vertex3f)
 137         {
 138                 const float * RESTRICT v = model->surfmesh.data_vertex3f;
 139                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 140                 // special case common combinations of attributes to avoid repeated loading of matrices
 141                 if (normal3f)
 142                 {
 143                         const float * RESTRICT n = model->surfmesh.data_normal3f;
 144                         if (svector3f && tvector3f)
 145                         {
 146                                 const float * RESTRICT sv = model->surfmesh.data_svector3f;
 147                                 const float * RESTRICT tv = model->surfmesh.data_tvector3f;
 148
 149                                 // Note that for SSE each iteration stores one element past end, so we break one vertex short
 150                                 // and handle that with scalars in that case
 151                                 for (i = 0; i < model->surfmesh.num_vertices; i++, v += 3, n += 3, sv += 3, tv += 3, b++,
 152                                                 vertex3f += 3, normal3f += 3, svector3f += 3, tvector3f += 3)
 153                                 {
 154                                         LOAD_MATRIX4();
 155                                         TRANSFORM_POSITION(v, vertex3f);
 156                                         TRANSFORM_VECTOR(n, normal3f);
 157                                         TRANSFORM_VECTOR(sv, svector3f);
 158                                         TRANSFORM_VECTOR(tv, tvector3f);
 159                                 }
 160
 161                                 return;
 162                         }
 163
 164                         for (i = 0;i < model->surfmesh.num_vertices; i++, v += 3, n += 3, b++, vertex3f += 3, normal3f += 3)
 165                         {
 166                                 LOAD_MATRIX4();
 167                                 TRANSFORM_POSITION(v, vertex3f);
 168                                 TRANSFORM_VECTOR(n, normal3f);
 169                         }
 170                 }
 171                 else
 172                 {
 173                         for (i = 0;i < model->surfmesh.num_vertices; i++, v += 3, b++, vertex3f += 3)
 174                         {
 175                                 LOAD_MATRIX4();
 176                                 TRANSFORM_POSITION(v, vertex3f);
 177                         }
 178                 }
 179         }
 180
 181         else if (normal3f)
 182         {
 183                 const float * RESTRICT n = model->surfmesh.data_normal3f;
 184                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 185                 for (i = 0; i < model->surfmesh.num_vertices; i++, n += 3, b++, normal3f += 3)
 186                 {
 187                         LOAD_MATRIX3();
 188                         TRANSFORM_VECTOR(n, normal3f);
 189                 }
 190         }
 191
 192         if (svector3f)
 193         {
 194                 const float * RESTRICT sv = model->surfmesh.data_svector3f;
 195                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 196                 for (i = 0; i < model->surfmesh.num_vertices; i++, sv += 3, b++, svector3f += 3)
 197                 {
 198                         LOAD_MATRIX3();
 199                         TRANSFORM_VECTOR(sv, svector3f);
 200                 }
 201         }
 202
 203         if (tvector3f)
 204         {
 205                 const float * RESTRICT tv = model->surfmesh.data_tvector3f;
 206                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 207                 for (i = 0; i < model->surfmesh.num_vertices; i++, tv += 3, b++, tvector3f += 3)
 208                 {
 209                         LOAD_MATRIX3();
 210                         TRANSFORM_VECTOR(tv, tvector3f);
 211                 }
 212         }
 213 }