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         num_vertices_minus_one = model->surfmesh.num_vertices - 1;
  24
  25         //unsigned long long ts = rdtsc();
  26         bonepose = (matrix4x4_t *) Mod_Skeletal_AnimateVertices_AllocBuffers(sizeof(matrix4x4_t) * (model->num_bones*2 + model->surfmesh.num_blends));
  27         boneposerelative = bonepose + model->num_bones;
  28
  29         if (skeleton && !skeleton->relativetransforms)
  30                 skeleton = NULL;
  31
  32         // interpolate matrices
  33         if (skeleton)
  34         {
  35                 for (i = 0;i < model->num_bones;i++)
  36                 {
  37                         // relativetransforms is in GL column-major order, which is what we need for SSE
  38                         // transposed style processing
  39                         if (model->data_bones[i].parent >= 0)
  40                                 Matrix4x4_Concat(&bonepose[i], &bonepose[model->data_bones[i].parent], &skeleton->relativetransforms[i]);
  41                         else
  42                                 memcpy(&bonepose[i], &skeleton->relativetransforms[i], sizeof(matrix4x4_t));
  43
  44                         // create a relative deformation matrix to describe displacement
  45                         // from the base mesh, which is used by the actual weighting
  46                         Matrix4x4_FromArray12FloatD3D(&mm, model->data_baseboneposeinverse + i * 12); // baseboneposeinverse is 4x3 row-major
  47                         Matrix4x4_Concat(&mm2, &bonepose[i], &mm);
  48                         Matrix4x4_Transpose(&boneposerelative[i], &mm2); // TODO: Eliminate this transpose
  49                 }
  50         }
  51         else
  52         {
  53                 float originscale = model->num_posescale;
  54                 float x,y,z,w,lerp;
  55                 const short * RESTRICT pose7s;
  56
  57                 for (i = 0;i < model->num_bones;i++)
  58                 {
  59                         memset(m, 0, sizeof(m));
  60                         for (blends = 0;blends < MAX_FRAMEBLENDS && frameblend[blends].lerp > 0;blends++)
  61                         {
  62                                 pose7s = model->data_poses7s + 7 * (frameblend[blends].subframe * model->num_bones + i);
  63                                 lerp = frameblend[blends].lerp;
  64                                 x = pose7s[3] * (1.0f / 32767.0f);
  65                                 y = pose7s[4] * (1.0f / 32767.0f);
  66                                 z = pose7s[5] * (1.0f / 32767.0f);
  67                                 w = pose7s[6] * (1.0f / 32767.0f);
  68                                 m[ 0] += (1-2*(y*y+z*z)) * lerp;
  69                                 m[ 1] += (  2*(x*y-z*w)) * lerp;
  70                                 m[ 2] += (  2*(x*z+y*w)) * lerp;
  71                                 m[ 3] += (pose7s[0] * originscale) * lerp;
  72                                 m[ 4] += (  2*(x*y+z*w)) * lerp;
  73                                 m[ 5] += (1-2*(x*x+z*z)) * lerp;
  74                                 m[ 6] += (  2*(y*z-x*w)) * lerp;
  75                                 m[ 7] += (pose7s[1] * originscale) * lerp;
  76                                 m[ 8] += (  2*(x*z-y*w)) * lerp;
  77                                 m[ 9] += (  2*(y*z+x*w)) * lerp;
  78                                 m[10] += (1-2*(x*x+y*y)) * lerp;
  79                                 m[11] += (pose7s[2] * originscale) * lerp;
  80                         }
  81                         VectorNormalize(m       );
  82                         VectorNormalize(m + 4);
  83                         VectorNormalize(m + 8);
  84                         if (i == r_skeletal_debugbone.integer)
  85                                 m[r_skeletal_debugbonecomponent.integer % 12] += r_skeletal_debugbonevalue.value;
  86                         m[3] *= r_skeletal_debugtranslatex.value;
  87                         m[7] *= r_skeletal_debugtranslatey.value;
  88                         m[11] *= r_skeletal_debugtranslatez.value;
  89                         Matrix4x4_FromArray12FloatD3D(&mm, m);
  90                         if (model->data_bones[i].parent >= 0)
  91                                 Matrix4x4_Concat(&bonepose[i], &bonepose[model->data_bones[i].parent], &mm);
  92                         else
  93                                 memcpy(&bonepose[i], &mm, sizeof(mm));
  94                         // create a relative deformation matrix to describe displacement
  95                         // from the base mesh, which is used by the actual weighting
  96                         Matrix4x4_FromArray12FloatD3D(&mm, model->data_baseboneposeinverse + i * 12); // baseboneposeinverse is 4x3 row-major
  97                         Matrix4x4_Concat(&mm2, &bonepose[i], &mm);
  98                         Matrix4x4_Transpose(&boneposerelative[i], &mm2); // TODO: Eliminate this transpose
  99                 }
 100         }
 101
 102         // generate matrices for all blend combinations
 103         weights = model->surfmesh.data_blendweights;
 104         for (i = 0;i < model->surfmesh.num_blends;i++, weights++)
 105         {
 106                 float * RESTRICT b = &boneposerelative[model->num_bones + i].m[0][0];
 107                 const float * RESTRICT m = &boneposerelative[weights->index[0]].m[0][0];
 108                 float f = weights->influence[0] * (1.0f / 255.0f);
 109                 __m128 fv = _mm_set_ps1(f);
 110                 __m128 b0 = _mm_load_ps(m);
 111                 __m128 b1 = _mm_load_ps(m+4);
 112                 __m128 b2 = _mm_load_ps(m+8);
 113                 __m128 b3 = _mm_load_ps(m+12);
 114                 __m128 m0, m1, m2, m3;
 115                 b0 = _mm_mul_ps(b0, fv);
 116                 b1 = _mm_mul_ps(b1, fv);
 117                 b2 = _mm_mul_ps(b2, fv);
 118                 b3 = _mm_mul_ps(b3, fv);
 119                 for (k = 1;k < 4 && weights->influence[k];k++)
 120                 {
 121                         m = &boneposerelative[weights->index[k]].m[0][0];
 122                         f = weights->influence[k] * (1.0f / 255.0f);
 123                         fv = _mm_set_ps1(f);
 124                         m0 = _mm_load_ps(m);
 125                         m1 = _mm_load_ps(m+4);
 126                         m2 = _mm_load_ps(m+8);
 127                         m3 = _mm_load_ps(m+12);
 128                         m0 = _mm_mul_ps(m0, fv);
 129                         m1 = _mm_mul_ps(m1, fv);
 130                         m2 = _mm_mul_ps(m2, fv);
 131                         m3 = _mm_mul_ps(m3, fv);
 132                         b0 = _mm_add_ps(m0, b0);
 133                         b1 = _mm_add_ps(m1, b1);
 134                         b2 = _mm_add_ps(m2, b2);
 135                         b3 = _mm_add_ps(m3, b3);
 136                 }
 137                 _mm_store_ps(b, b0);
 138                 _mm_store_ps(b+4, b1);
 139                 _mm_store_ps(b+8, b2);
 140                 _mm_store_ps(b+12, b3);
 141         }
 142
 143 #define LOAD_MATRIX_SCALAR() const float * RESTRICT m = &boneposerelative[*b].m[0][0]
 144
 145 #define LOAD_MATRIX3() \
 146         const float * RESTRICT m = &boneposerelative[*b].m[0][0]; \
 147         /* bonepose array is 16 byte aligned */ \
 148         __m128 m1 = _mm_load_ps((m)); \
 149         __m128 m2 = _mm_load_ps((m)+4); \
 150         __m128 m3 = _mm_load_ps((m)+8);
 151 #define LOAD_MATRIX4() \
 152         const float * RESTRICT m = &boneposerelative[*b].m[0][0]; \
 153         /* bonepose array is 16 byte aligned */ \
 154         __m128 m1 = _mm_load_ps((m)); \
 155         __m128 m2 = _mm_load_ps((m)+4); \
 156         __m128 m3 = _mm_load_ps((m)+8); \
 157         __m128 m4 = _mm_load_ps((m)+12)
 158
 159         /* Note that matrix is 4x4 and transposed compared to non-USE_SSE codepath */
 160 #define TRANSFORM_POSITION_SCALAR(in, out) \
 161         (out)[0] = ((in)[0] * m[0] + (in)[1] * m[4] + (in)[2] * m[ 8] + m[12]); \
 162         (out)[1] = ((in)[0] * m[1] + (in)[1] * m[5] + (in)[2] * m[ 9] + m[13]); \
 163         (out)[2] = ((in)[0] * m[2] + (in)[1] * m[6] + (in)[2] * m[10] + m[14]);
 164 #define TRANSFORM_VECTOR_SCALAR(in, out) \
 165         (out)[0] = ((in)[0] * m[0] + (in)[1] * m[4] + (in)[2] * m[ 8]); \
 166         (out)[1] = ((in)[0] * m[1] + (in)[1] * m[5] + (in)[2] * m[ 9]); \
 167         (out)[2] = ((in)[0] * m[2] + (in)[1] * m[6] + (in)[2] * m[10]);
 168
 169 #define TRANSFORM_POSITION(in, out) { \
 170                 __m128 pin = _mm_loadu_ps(in); /* we ignore the value in the last element (x from the next vertex) */ \
 171                 __m128 x = _mm_shuffle_ps(pin, pin, 0x0); \
 172                 __m128 t1 = _mm_mul_ps(x, m1); \
 173                 \
 174                 /* y, + x */ \
 175                 __m128 y = _mm_shuffle_ps(pin, pin, 0x55); \
 176                 __m128 t2 = _mm_mul_ps(y, m2); \
 177                 __m128 t3 = _mm_add_ps(t1, t2); \
 178                 \
 179                 /* z, + (y+x) */ \
 180                 __m128 z = _mm_shuffle_ps(pin, pin, 0xaa); \
 181                 __m128 t4 = _mm_mul_ps(z, m3); \
 182                 __m128 t5 = _mm_add_ps(t3, t4); \
 183                 \
 184                 /* + m3 */ \
 185                 __m128 pout = _mm_add_ps(t5, m4); \
 186                 _mm_storeu_ps((out), pout); \
 187         }
 188
 189 #define TRANSFORM_VECTOR(in, out) { \
 190                 __m128 vin = _mm_loadu_ps(in); \
 191                 \
 192                 /* x */ \
 193                 __m128 x = _mm_shuffle_ps(vin, vin, 0x0); \
 194                 __m128 t1 = _mm_mul_ps(x, m1); \
 195                 \
 196                 /* y, + x */ \
 197                 __m128 y = _mm_shuffle_ps(vin, vin, 0x55); \
 198                 __m128 t2 = _mm_mul_ps(y, m2); \
 199                 __m128 t3 = _mm_add_ps(t1, t2); \
 200                 \
 201                 /* nz, + (ny + nx) */ \
 202                 __m128 z = _mm_shuffle_ps(vin, vin, 0xaa); \
 203                 __m128 t4 = _mm_mul_ps(z, m3); \
 204                 __m128 vout = _mm_add_ps(t3, t4); \
 205                 _mm_storeu_ps((out), vout); \
 206         }
 207
 208         // transform vertex attributes by blended matrices
 209         if (vertex3f)
 210         {
 211                 const float * RESTRICT v = model->surfmesh.data_vertex3f;
 212                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 213                 // special case common combinations of attributes to avoid repeated loading of matrices
 214                 if (normal3f)
 215                 {
 216                         const float * RESTRICT n = model->surfmesh.data_normal3f;
 217                         if (svector3f && tvector3f)
 218                         {
 219                                 const float * RESTRICT sv = model->surfmesh.data_svector3f;
 220                                 const float * RESTRICT tv = model->surfmesh.data_tvector3f;
 221
 222                                 // Note that for SSE each iteration stores one element past end, so we break one vertex short
 223                                 // and handle that with scalars in that case
 224                                 for (i = 0; i < num_vertices_minus_one; i++, v += 3, n += 3, sv += 3, tv += 3, b++,
 225                                                 vertex3f += 3, normal3f += 3, svector3f += 3, tvector3f += 3)
 226                                 {
 227                                         LOAD_MATRIX4();
 228                                         TRANSFORM_POSITION(v, vertex3f);
 229                                         TRANSFORM_VECTOR(n, normal3f);
 230                                         TRANSFORM_VECTOR(sv, svector3f);
 231                                         TRANSFORM_VECTOR(tv, tvector3f);
 232                                 }
 233
 234                                 // Last vertex needs to be done with scalars to avoid reading/writing 1 word past end of arrays
 235                                 {
 236                                         LOAD_MATRIX_SCALAR();
 237                                         TRANSFORM_POSITION_SCALAR(v, vertex3f);
 238                                         TRANSFORM_VECTOR_SCALAR(n, normal3f);
 239                                         TRANSFORM_VECTOR_SCALAR(sv, svector3f);
 240                                         TRANSFORM_VECTOR_SCALAR(tv, tvector3f);
 241                                 }
 242                                 //printf("elapsed ticks: %llu\n", rdtsc() - ts); // XXX
 243                                 return;
 244                         }
 245
 246                         for (i = 0;i < num_vertices_minus_one; i++, v += 3, n += 3, b++, vertex3f += 3, normal3f += 3)
 247                         {
 248                                 LOAD_MATRIX4();
 249                                 TRANSFORM_POSITION(v, vertex3f);
 250                                 TRANSFORM_VECTOR(n, normal3f);
 251                         }
 252                         {
 253                                 LOAD_MATRIX_SCALAR();
 254                                 TRANSFORM_POSITION_SCALAR(v, vertex3f);
 255                                 TRANSFORM_VECTOR_SCALAR(n, normal3f);
 256                         }
 257                 }
 258                 else
 259                 {
 260                         for (i = 0;i < num_vertices_minus_one; i++, v += 3, b++, vertex3f += 3)
 261                         {
 262                                 LOAD_MATRIX4();
 263                                 TRANSFORM_POSITION(v, vertex3f);
 264                         }
 265                         {
 266                                 LOAD_MATRIX_SCALAR();
 267                                 TRANSFORM_POSITION_SCALAR(v, vertex3f);
 268                         }
 269                 }
 270         }
 271
 272         else if (normal3f)
 273         {
 274                 const float * RESTRICT n = model->surfmesh.data_normal3f;
 275                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 276                 for (i = 0; i < num_vertices_minus_one; i++, n += 3, b++, normal3f += 3)
 277                 {
 278                         LOAD_MATRIX3();
 279                         TRANSFORM_VECTOR(n, normal3f);
 280                 }
 281                 {
 282                         LOAD_MATRIX_SCALAR();
 283                         TRANSFORM_VECTOR_SCALAR(n, normal3f);
 284                 }
 285         }
 286
 287         if (svector3f)
 288         {
 289                 const float * RESTRICT sv = model->surfmesh.data_svector3f;
 290                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 291                 for (i = 0; i < num_vertices_minus_one; i++, sv += 3, b++, svector3f += 3)
 292                 {
 293                         LOAD_MATRIX3();
 294                         TRANSFORM_VECTOR(sv, svector3f);
 295                 }
 296                 {
 297                         LOAD_MATRIX_SCALAR();
 298                         TRANSFORM_VECTOR_SCALAR(sv, svector3f);
 299                 }
 300         }
 301
 302         if (tvector3f)
 303         {
 304                 const float * RESTRICT tv = model->surfmesh.data_tvector3f;
 305                 const unsigned short * RESTRICT b = model->surfmesh.blends;
 306                 for (i = 0; i < num_vertices_minus_one; i++, tv += 3, b++, tvector3f += 3)
 307                 {
 308                         LOAD_MATRIX3();
 309                         TRANSFORM_VECTOR(tv, tvector3f);
 310                 }
 311                 {
 312                         LOAD_MATRIX_SCALAR();
 313                         TRANSFORM_VECTOR_SCALAR(tv, tvector3f);
 314                 }
 315         }
 316
 317 #undef LOAD_MATRIX3
 318 #undef LOAD_MATRIX4
 319 #undef TRANSFORM_POSITION
 320 #undef TRANSFORM_VECTOR
 321 #undef LOAD_MATRIX_SCALAR
 322 #undef TRANSFORM_POSITION_SCALAR
 323 #undef TRANSFORM_VECTOR_SCALAR
 324 }
 325
 326 #endif