mod_skeletal_animatevertices_sse.c

   1 #include "mod_skeletal_animatevertices_sse.h"
   2
   3 #ifdef SSE_POSSIBLE
   4
   5 #ifdef MATRIX4x4_OPENGLORIENTATION
   6 #error "SSE skeletal requires D3D matrix layout"
   7 #endif
   8
   9 #include <xmmintrin.h>
  10
  11 void Mod_Skeletal_AnimateVertices_SSE(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)
  12 {
  13         // vertex weighted skeletal
  14         int i, k;
  15         int blends;
  16         matrix4x4_t *bonepose;
  17         matrix4x4_t *boneposerelative;
  18         float m[12];
  19         matrix4x4_t mm, mm2;
  20         const blendweights_t * RESTRICT weights;
  21         int num_vertices_minus_one;
  22
  23         if (!model->surfmesh.num_vertices)
  24                 return;
  25
  26         if (!model->num_bones)
  27         {
  28                 if (vertex3f) memcpy(vertex3f, model->surfmesh.data_vertex3f, model->surfmesh.num_vertices*sizeof(float[3]));
  29                 if (normal3f) memcpy(normal3f, model->surfmesh.data_normal3f, model->surfmesh.num_vertices*sizeof(float[3]));
  30                 if (svector3f) memcpy(svector3f, model->surfmesh.data_svector3f, model->surfmesh.num_vertices*sizeof(float[3]));
  31                 if (tvector3f) memcpy(tvector3f, model->surfmesh.data_tvector3f, model->surfmesh.num_vertices*sizeof(float[3]));
  32                 return;
  33         }
  34
  35         num_vertices_minus_one = model->surfmesh.num_vertices - 1;
  36
  37         //unsigned long long ts = rdtsc();
  38         bonepose = (matrix4x4_t *) Mod_Skeletal_AnimateVertices_AllocBuffers(sizeof(matrix4x4_t) * (model->num_bones*2 + model->surfmesh.num_blends));
  39         boneposerelative = bonepose + model->num_bones;
  40
  41         if (skeleton && !skeleton->relativetransforms)
  42                 skeleton = NULL;
  43
  44         // interpolate matrices
  45         if (skeleton)
  46         {
  47                 for (i = 0;i < model->num_bones;i++)
  48                 {
  49                         // relativetransforms is in GL column-major order, which is what we need for SSE
  50                         // transposed style processing
  51                         if (model->data_bones[i].parent >= 0)
  52                                 Matrix4x4_Concat(&bonepose[i], &bonepose[model->data_bones[i].parent], &skeleton->relativetransforms[i]);
  53                         else
  54                                 memcpy(&bonepose[i], &skeleton->relativetransforms[i], sizeof(matrix4x4_t));
  55
  56                         // create a relative deformation matrix to describe displacement
  57                         // from the base mesh, which is used by the actual weighting
  58                         Matrix4x4_FromArray12FloatD3D(&mm, model->data_baseboneposeinverse + i * 12); // baseboneposeinverse is 4x3 row-major
  59                         Matrix4x4_Concat(&mm2, &bonepose[i], &mm);
  60                         Matrix4x4_Transpose(&boneposerelative[i], &mm2); // TODO: Eliminate this transpose
  61                 }
  62         }
  63         else
  64         {
  65                 float originscale = model->num_posescale;
  66                 float x,y,z,w,lerp;
  67                 const short * RESTRICT pose6s;
  68
  69                 for (i = 0;i < model->num_bones;i++)
  70                 {
  71                         memset(m, 0, sizeof(m));
  72                         for (blends = 0;blends < MAX_FRAMEBLENDS && frameblend[blends].lerp > 0;blends++)
  73                         {
  74                                 pose6s = model->data_poses6s + 6 * (frameblend[blends].subframe * model->num_bones + i);
  75                                 lerp = frameblend[blends].lerp;
  76                                 x = pose6s[3] * (1.0f / 32767.0f);
  77                                 y = pose6s[4] * (1.0f / 32767.0f);
  78                                 z = pose6s[5] * (1.0f / 32767.0f);
  79                                 w = 1.0f - (x*x+y*y+z*z);
  80                                 w = w > 0.0f ? -sqrt(w) : 0.0f;
  81                                 m[ 0] += (1-2*(y*y+z*z)) * lerp;
  82                                 m[ 1] += (  2*(x*y-z*w)) * lerp;
  83                                 m[ 2] += (  2*(x*z+y*w)) * lerp;
  84                                 m[ 3] += (pose6s[0] * originscale) * lerp;
  85                                 m[ 4] += (  2*(x*y+z*w)) * lerp;
  86                                 m[ 5] += (1-2*(x*x+z*z)) * lerp;
  87                                 m[ 6] += (  2*(y*z-x*w)) * lerp;
  88                                 m[ 7] += (pose6s[1] * originscale) * lerp;
  89                                 m[ 8] += (  2*(x*z-y*w)) * lerp;
  90                                 m[ 9] += (  2*(y*z+x*w)) * lerp;
  91                                 m[10] += (1-2*(x*x+y*y)) * lerp;
  92                                 m[11] += (pose6s[2] * originscale) * lerp;
  93                         }
  94                         VectorNormalize(m       );
  95                         VectorNormalize(m + 4);
  96                         VectorNormalize(m + 8);
  97                         if (i == r_skeletal_debugbone.integer)
  98                                 m[r_skeletal_debugbonecomponent.integer % 12] += r_skeletal_debugbonevalue.value;
  99                         m[3] *= r_skeletal_debugtranslatex.value;
 100                         m[7] *= r_skeletal_debugtranslatey.value;
 101                         m[11] *= r_skeletal_debugtranslatez.value;
 102                         Matrix4x4_FromArray12FloatD3D(&mm, m);
 103                         if (model->data_bones[i].parent >= 0)
 104                                 Matrix4x4_Concat(&bonepose[i], &bonepose[model->data_bones[i].parent], &mm);
 105                         else
 106                                 memcpy(&bonepose[i], &mm, sizeof(mm));
 107                         // create a relative deformation matrix to describe displacement
 108                         // from the base mesh, which is used by the actual weighting
 109                         Matrix4x4_FromArray12FloatD3D(&mm, model->data_baseboneposeinverse + i * 12); // baseboneposeinverse is 4x3 row-major
 110                         Matrix4x4_Concat(&mm2, &bonepose[i], &mm);
 111                         Matrix4x4_Transpose(&boneposerelative[i], &mm2); // TODO: Eliminate this transpose
 112                 }
 113         }
 114
 115         // generate matrices for all blend combinations
 116         weights = model->surfmesh.data_blendweights;
 117         for (i = 0;i < model->surfmesh.num_blends;i++, weights++)
 118         {
 119                 float * RESTRICT b = &boneposerelative[model->num_bones + i].m[0][0];
 120                 const float * RESTRICT m = &boneposerelative[weights->index[0]].m[0][0];
 121                 float f = weights->influence[0] * (1.0f / 255.0f);
 122                 __m128 fv = _mm_set_ps1(f);
 123                 __m128 b0 = _mm_load_ps(m);
 124                 __m128 b1 = _mm_load_ps(m+4);
 125                 __m128 b2 = _mm_load_ps(m+8);
 126                 __m128 b3 = _mm_load_ps(m+12);
 127                 __m128 m0, m1, m2, m3;
 128                 b0 = _mm_mul_ps(b0, fv);
 129                 b1 = _mm_mul_ps(b1, fv);
 130                 b2 = _mm_mul_ps(b2, fv);
 131                 b3 = _mm_mul_ps(b3, fv);
 132                 for (k = 1;k < 4 && weights->influence[k];k++)
 133                 {
 134                         m = &boneposerelative[weights->index[k]].m[0][0];
 135                         f = weights->influence[k] * (1.0f / 255.0f);
 136                         fv = _mm_set_ps1(f);
 137                         m0 = _mm_load_ps(m);
 138                         m1 = _mm_load_ps(m+4);
 139                         m2 = _mm_load_ps(m+8);
 140                         m3 = _mm_load_ps(m+12);
 141                         m0 = _mm_mul_ps(m0, fv);
 142                         m1 = _mm_mul_ps(m1, fv);
 143                         m2 = _mm_mul_ps(m2, fv);
 144                         m3 = _mm_mul_ps(m3, fv);
 145                         b0 = _mm_add_ps(m0, b0);
 146                         b1 = _mm_add_ps(m1, b1);
 147                         b2 = _mm_add_ps(m2, b2);
 148                         b3 = _mm_add_ps(m3, b3);
 149                 }
 150                 _mm_store_ps(b, b0);
 151                 _mm_store_ps(b+4, b1);
 152                 _mm_store_ps(b+8, b2);
 153                 _mm_store_ps(b+12, b3);
 154         }
 155
 156 #define LOAD_MATRIX_SCALAR() const float * RESTRICT m = &boneposerelative[*b].m[0][0]
 157
 158 #define LOAD_MATRIX3() \
 159         const float * RESTRICT m = &boneposerelative[*b].m[0][0]; \
 160         /* bonepose array is 16 byte aligned */ \
 161         __m128 m1 = _mm_load_ps((m)); \
 162         __m128 m2 = _mm_load_ps((m)+4); \
 163         __m128 m3 = _mm_load_ps((m)+8);
 164 #define LOAD_MATRIX4() \
 165         const float * RESTRICT m = &boneposerelative[*b].m[0][0]; \
 166         /* bonepose array is 16 byte aligned */ \
 167         __m128 m1 = _mm_load_ps((m)); \
 168         __m128 m2 = _mm_load_ps((m)+4); \
 169         __m128 m3 = _mm_load_ps((m)+8); \
 170         __m128 m4 = _mm_load_ps((m)+12)
 171
 172         /* Note that matrix is 4x4 and transposed compared to non-USE_SSE codepath */
 173 #define TRANSFORM_POSITION_SCALAR(in, out) \
 174         (out)[0] = ((in)[0] * m[0] + (in)[1] * m[4] + (in)[2] * m[ 8] + m[12]); \
 175         (out)[1] = ((in)[0] * m[1] + (in)[1] * m[5] + (in)[2] * m[ 9] + m[13]); \
 176         (out)[2] = ((in)[0] * m[2] + (in)[1] * m[6] + (in)[2] * m[10] + m[14]);
 177 #define TRANSFORM_VECTOR_SCALAR(in, out) \
 178         (out)[0] = ((in)[0] * m[0] + (in)[1] * m[4] + (in)[2] * m[ 8]); \
 179         (out)[1] = ((in)[0] * m[1] + (in)[1] * m[5] + (in)[2] * m[ 9]); \
 180         (out)[2] = ((in)[0] * m[2] + (in)[1] * m[6] + (in)[2] * m[10]);
 181
 182 #define TRANSFORM_POSITION(in, out) { \
 183                 __m128 pin = _mm_loadu_ps(in); /* we ignore the value in the last element (x from the next vertex) */ \
 184                 __m128 x = _mm_shuffle_ps(pin, pin, 0x0); \
 185                 __m128 t1 = _mm_mul_ps(x, m1); \
 186                 \
 187                 /* y, + x */ \
 188                 __m128 y = _mm_shuffle_ps(pin, pin, 0x55); \
 189                 __m128 t2 = _mm_mul_ps(y, m2); \
 190                 __m128 t3 = _mm_add_ps(t1, t2); \
 191                 \
 192                 /* z, + (y+x) */ \
 193                 __m128 z = _mm_shuffle_ps(pin, pin, 0xaa); \
 194                 __m128 t4 = _mm_mul_ps(z, m3); \
 195                 __m128 t5 = _mm_add_ps(t3, t4); \
 196                 \
 197                 /* + m3 */ \
 198                 __m128 pout = _mm_add_ps(t5, m4); \
 199                 _mm_storeu_ps((out), pout); \
 200         }
 201
 202 #define TRANSFORM_VECTOR(in, out) { \
 203                 __m128 vin = _mm_loadu_ps(in); \
 204                 \
 205                 /* x */ \
 206                 __m128 x = _mm_shuffle_ps(vin, vin, 0x0); \
 207                 __m128 t1 = _mm_mul_ps(x, m1); \
 208                 \
 209                 /* y, + x */ \
 210                 __m128 y = _mm_shuffle_ps(vin, vin, 0x55); \
 211                 __m128 t2 = _mm_mul_ps(y, m2); \
 212                 __m128 t3 = _mm_add_ps(t1, t2); \
 213                 \
 214                 /* nz, + (ny + nx) */ \
 215                 __m128 z = _mm_shuffle_ps(vin, vin, 0xaa); \
 216                 __m128 t4 = _mm_mul_ps(z, m3); \
 217                 __m128 vout = _mm_add_ps(t3, t4); \
 218                 _mm_storeu_ps((out), vout); \
 219         }
 220
 221         // transform vertex attributes by blended matrices
 222         if (vertex3f)
 223         {
 224                 const float * RESTRICT v = model->surfmesh.data_vertex3f;
 225                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 226                 // special case common combinations of attributes to avoid repeated loading of matrices
 227                 if (normal3f)
 228                 {
 229                         const float * RESTRICT n = model->surfmesh.data_normal3f;
 230                         if (svector3f && tvector3f)
 231                         {
 232                                 const float * RESTRICT sv = model->surfmesh.data_svector3f;
 233                                 const float * RESTRICT tv = model->surfmesh.data_tvector3f;
 234
 235                                 // Note that for SSE each iteration stores one element past end, so we break one vertex short
 236                                 // and handle that with scalars in that case
 237                                 for (i = 0; i < num_vertices_minus_one; i++, v += 3, n += 3, sv += 3, tv += 3, b++,
 238                                                 vertex3f += 3, normal3f += 3, svector3f += 3, tvector3f += 3)
 239                                 {
 240                                         LOAD_MATRIX4();
 241                                         TRANSFORM_POSITION(v, vertex3f);
 242                                         TRANSFORM_VECTOR(n, normal3f);
 243                                         TRANSFORM_VECTOR(sv, svector3f);
 244                                         TRANSFORM_VECTOR(tv, tvector3f);
 245                                 }
 246
 247                                 // Last vertex needs to be done with scalars to avoid reading/writing 1 word past end of arrays
 248                                 {
 249                                         LOAD_MATRIX_SCALAR();
 250                                         TRANSFORM_POSITION_SCALAR(v, vertex3f);
 251                                         TRANSFORM_VECTOR_SCALAR(n, normal3f);
 252                                         TRANSFORM_VECTOR_SCALAR(sv, svector3f);
 253                                         TRANSFORM_VECTOR_SCALAR(tv, tvector3f);
 254                                 }
 255                                 //printf("elapsed ticks: %llu\n", rdtsc() - ts); // XXX
 256                                 return;
 257                         }
 258
 259                         for (i = 0;i < num_vertices_minus_one; i++, v += 3, n += 3, b++, vertex3f += 3, normal3f += 3)
 260                         {
 261                                 LOAD_MATRIX4();
 262                                 TRANSFORM_POSITION(v, vertex3f);
 263                                 TRANSFORM_VECTOR(n, normal3f);
 264                         }
 265                         {
 266                                 LOAD_MATRIX_SCALAR();
 267                                 TRANSFORM_POSITION_SCALAR(v, vertex3f);
 268                                 TRANSFORM_VECTOR_SCALAR(n, normal3f);
 269                         }
 270                 }
 271                 else
 272                 {
 273                         for (i = 0;i < num_vertices_minus_one; i++, v += 3, b++, vertex3f += 3)
 274                         {
 275                                 LOAD_MATRIX4();
 276                                 TRANSFORM_POSITION(v, vertex3f);
 277                         }
 278                         {
 279                                 LOAD_MATRIX_SCALAR();
 280                                 TRANSFORM_POSITION_SCALAR(v, vertex3f);
 281                         }
 282                 }
 283         }
 284
 285         else if (normal3f)
 286         {
 287                 const float * RESTRICT n = model->surfmesh.data_normal3f;
 288                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 289                 for (i = 0; i < num_vertices_minus_one; i++, n += 3, b++, normal3f += 3)
 290                 {
 291                         LOAD_MATRIX3();
 292                         TRANSFORM_VECTOR(n, normal3f);
 293                 }
 294                 {
 295                         LOAD_MATRIX_SCALAR();
 296                         TRANSFORM_VECTOR_SCALAR(n, normal3f);
 297                 }
 298         }
 299
 300         if (svector3f)
 301         {
 302                 const float * RESTRICT sv = model->surfmesh.data_svector3f;
 303                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 304                 for (i = 0; i < num_vertices_minus_one; i++, sv += 3, b++, svector3f += 3)
 305                 {
 306                         LOAD_MATRIX3();
 307                         TRANSFORM_VECTOR(sv, svector3f);
 308                 }
 309                 {
 310                         LOAD_MATRIX_SCALAR();
 311                         TRANSFORM_VECTOR_SCALAR(sv, svector3f);
 312                 }
 313         }
 314
 315         if (tvector3f)
 316         {
 317                 const float * RESTRICT tv = model->surfmesh.data_tvector3f;
 318                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 319                 for (i = 0; i < num_vertices_minus_one; i++, tv += 3, b++, tvector3f += 3)
 320                 {
 321                         LOAD_MATRIX3();
 322                         TRANSFORM_VECTOR(tv, tvector3f);
 323                 }
 324                 {
 325                         LOAD_MATRIX_SCALAR();
 326                         TRANSFORM_VECTOR_SCALAR(tv, tvector3f);
 327                 }
 328         }
 329
 330 #undef LOAD_MATRIX3
 331 #undef LOAD_MATRIX4
 332 #undef TRANSFORM_POSITION
 333 #undef TRANSFORM_VECTOR
 334 #undef LOAD_MATRIX_SCALAR
 335 #undef TRANSFORM_POSITION_SCALAR
 336 #undef TRANSFORM_VECTOR_SCALAR
 337 }
 338
 339 #endif