1 #include "mod_skeletal_animatevertices_sse.h"
5 #ifdef MATRIX4x4_OPENGLORIENTATION
6 #error "SSE skeletal requires D3D matrix layout"
11 void Mod_Skeletal_AnimateVertices_SSE(const 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)
13 // vertex weighted skeletal
16 matrix4x4_t *bonepose;
17 matrix4x4_t *boneposerelative;
18 const blendweights_t * RESTRICT weights;
19 int num_vertices_minus_one;
21 num_vertices_minus_one = model->surfmesh.num_vertices - 1;
23 //unsigned long long ts = rdtsc();
24 bonepose = (matrix4x4_t *) Mod_Skeletal_AnimateVertices_AllocBuffers(sizeof(matrix4x4_t) * (model->num_bones*2 + model->surfmesh.num_blends));
25 boneposerelative = bonepose + model->num_bones;
27 if (skeleton && !skeleton->relativetransforms)
30 // interpolate matrices
33 for (i = 0;i < model->num_bones;i++)
35 const float * RESTRICT n = model->data_baseboneposeinverse + i * 12;
36 matrix4x4_t * RESTRICT s = &skeleton->relativetransforms[i];
37 matrix4x4_t * RESTRICT b = &bonepose[i];
38 matrix4x4_t * RESTRICT r = &boneposerelative[i];
39 __m128 b0, b1, b2, b3, r0, r1, r2, r3, nr;
40 if (model->data_bones[i].parent >= 0)
42 const matrix4x4_t * RESTRICT p = &bonepose[model->data_bones[i].parent];
43 __m128 s0 = _mm_loadu_ps(s->m[0]), s1 = _mm_loadu_ps(s->m[1]), s2 = _mm_loadu_ps(s->m[2]);
44 #ifdef OPENGLORIENTATION
45 __m128 s3 = _mm_loadu_ps(s->m[3]);
46 #define SKELETON_MATRIX(r, c) _mm_shuffle_ps(s##c, s##c, _MM_SHUFFLE(r, r, r, r))
48 #define SKELETON_MATRIX(r, c) _mm_shuffle_ps(s##r, s##r, _MM_SHUFFLE(c, c, c, c))
50 __m128 pr = _mm_load_ps(p->m[0]);
51 b0 = _mm_mul_ps(pr, SKELETON_MATRIX(0, 0));
52 b1 = _mm_mul_ps(pr, SKELETON_MATRIX(0, 1));
53 b2 = _mm_mul_ps(pr, SKELETON_MATRIX(0, 2));
54 b3 = _mm_mul_ps(pr, SKELETON_MATRIX(0, 3));
55 pr = _mm_load_ps(p->m[1]);
56 b0 = _mm_add_ps(b0, _mm_mul_ps(pr, SKELETON_MATRIX(1, 0)));
57 b1 = _mm_add_ps(b1, _mm_mul_ps(pr, SKELETON_MATRIX(1, 1)));
58 b2 = _mm_add_ps(b2, _mm_mul_ps(pr, SKELETON_MATRIX(1, 2)));
59 b3 = _mm_add_ps(b3, _mm_mul_ps(pr, SKELETON_MATRIX(1, 3)));
60 pr = _mm_load_ps(p->m[2]);
61 b0 = _mm_add_ps(b0, _mm_mul_ps(pr, SKELETON_MATRIX(2, 0)));
62 b1 = _mm_add_ps(b1, _mm_mul_ps(pr, SKELETON_MATRIX(2, 1)));
63 b2 = _mm_add_ps(b2, _mm_mul_ps(pr, SKELETON_MATRIX(2, 2)));
64 b3 = _mm_add_ps(b3, _mm_mul_ps(pr, SKELETON_MATRIX(2, 3)));
65 b3 = _mm_add_ps(b3, _mm_load_ps(p->m[3]));
69 b0 = _mm_loadu_ps(s->m[0]);
70 b1 = _mm_loadu_ps(s->m[1]);
71 b2 = _mm_loadu_ps(s->m[2]);
72 b3 = _mm_loadu_ps(s->m[3]);
73 #ifndef OPENGLORIENTATION
74 _MM_TRANSPOSE4_PS(b0, b1, b2, b3);
77 _mm_store_ps(b->m[0], b0);
78 _mm_store_ps(b->m[1], b1);
79 _mm_store_ps(b->m[2], b2);
80 _mm_store_ps(b->m[3], b3);
82 r0 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0)));
83 r1 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1)));
84 r2 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2)));
85 r3 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3)));
86 nr = _mm_loadu_ps(n+4);
87 r0 = _mm_add_ps(r0, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0))));
88 r1 = _mm_add_ps(r1, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1))));
89 r2 = _mm_add_ps(r2, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2))));
90 r3 = _mm_add_ps(r3, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3))));
91 nr = _mm_loadu_ps(n+8);
92 r0 = _mm_add_ps(r0, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0))));
93 r1 = _mm_add_ps(r1, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1))));
94 r2 = _mm_add_ps(r2, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2))));
95 r3 = _mm_add_ps(r3, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3))));
96 r3 = _mm_add_ps(r3, b3);
97 _mm_store_ps(r->m[0], r0);
98 _mm_store_ps(r->m[1], r1);
99 _mm_store_ps(r->m[2], r2);
100 _mm_store_ps(r->m[3], r3);
105 for (i = 0;i < model->num_bones;i++)
108 const short * RESTRICT firstpose7s = model->data_poses7s + 7 * (frameblend[0].subframe * model->num_bones + i);
109 float firstlerp = frameblend[0].lerp,
110 firsttx = firstpose7s[0], firstty = firstpose7s[1], firsttz = firstpose7s[2],
111 rx = firstpose7s[3] * firstlerp,
112 ry = firstpose7s[4] * firstlerp,
113 rz = firstpose7s[5] * firstlerp,
114 rw = firstpose7s[6] * firstlerp,
115 dx = firsttx*rw + firstty*rz - firsttz*ry,
116 dy = -firsttx*rz + firstty*rw + firsttz*rx,
117 dz = firsttx*ry - firstty*rx + firsttz*rw,
118 dw = -firsttx*rx - firstty*ry - firsttz*rz,
119 scale, sx, sy, sz, sw;
120 for (blends = 1;blends < MAX_FRAMEBLENDS && frameblend[blends].lerp > 0;blends++)
122 const short * RESTRICT blendpose7s = model->data_poses7s + 7 * (frameblend[blends].subframe * model->num_bones + i);
123 float blendlerp = frameblend[blends].lerp,
124 blendtx = blendpose7s[0], blendty = blendpose7s[1], blendtz = blendpose7s[2],
125 qx = blendpose7s[3], qy = blendpose7s[4], qz = blendpose7s[5], qw = blendpose7s[6];
126 if(rx*qx + ry*qy + rz*qz + rw*qw < 0) blendlerp = -blendlerp;
135 dx += blendtx*qw + blendty*qz - blendtz*qy;
136 dy += -blendtx*qz + blendty*qw + blendtz*qx;
137 dz += blendtx*qy - blendty*qx + blendtz*qw;
138 dw += -blendtx*qx - blendty*qy - blendtz*qz;
140 scale = 1.0f / (rx*rx + ry*ry + rz*rz + rw*rw);
145 m[0] = sw*rw + sx*rx - sy*ry - sz*rz;
146 m[1] = 2*(sx*ry - sw*rz);
147 m[2] = 2*(sx*rz + sw*ry);
148 m[3] = model->num_posescale*(dx*sw - dy*sz + dz*sy - dw*sx);
149 m[4] = 2*(sx*ry + sw*rz);
150 m[5] = sw*rw + sy*ry - sx*rx - sz*rz;
151 m[6] = 2*(sy*rz - sw*rx);
152 m[7] = model->num_posescale*(dx*sz + dy*sw - dz*sx - dw*sy);
153 m[8] = 2*(sx*rz - sw*ry);
154 m[9] = 2*(sy*rz + sw*rx);
155 m[10] = sw*rw + sz*rz - sx*rx - sy*ry;
156 m[11] = model->num_posescale*(dy*sx + dz*sw - dx*sy - dw*sz);
157 if (i == r_skeletal_debugbone.integer)
158 m[r_skeletal_debugbonecomponent.integer % 12] += r_skeletal_debugbonevalue.value;
159 m[3] *= r_skeletal_debugtranslatex.value;
160 m[7] *= r_skeletal_debugtranslatey.value;
161 m[11] *= r_skeletal_debugtranslatez.value;
163 const float * RESTRICT n = model->data_baseboneposeinverse + i * 12;
164 matrix4x4_t * RESTRICT b = &bonepose[i];
165 matrix4x4_t * RESTRICT r = &boneposerelative[i];
166 __m128 b0, b1, b2, b3, r0, r1, r2, r3, nr;
167 if (model->data_bones[i].parent >= 0)
169 const matrix4x4_t * RESTRICT p = &bonepose[model->data_bones[i].parent];
170 __m128 pr = _mm_load_ps(p->m[0]);
171 b0 = _mm_mul_ps(pr, _mm_set1_ps(m[0]));
172 b1 = _mm_mul_ps(pr, _mm_set1_ps(m[1]));
173 b2 = _mm_mul_ps(pr, _mm_set1_ps(m[2]));
174 b3 = _mm_mul_ps(pr, _mm_set1_ps(m[3]));
175 pr = _mm_load_ps(p->m[1]);
176 b0 = _mm_add_ps(b0, _mm_mul_ps(pr, _mm_set1_ps(m[4])));
177 b1 = _mm_add_ps(b1, _mm_mul_ps(pr, _mm_set1_ps(m[5])));
178 b2 = _mm_add_ps(b2, _mm_mul_ps(pr, _mm_set1_ps(m[6])));
179 b3 = _mm_add_ps(b3, _mm_mul_ps(pr, _mm_set1_ps(m[7])));
180 pr = _mm_load_ps(p->m[2]);
181 b0 = _mm_add_ps(b0, _mm_mul_ps(pr, _mm_set1_ps(m[8])));
182 b1 = _mm_add_ps(b1, _mm_mul_ps(pr, _mm_set1_ps(m[9])));
183 b2 = _mm_add_ps(b2, _mm_mul_ps(pr, _mm_set1_ps(m[10])));
184 b3 = _mm_add_ps(b3, _mm_mul_ps(pr, _mm_set1_ps(m[11])));
185 b3 = _mm_add_ps(b3, _mm_load_ps(p->m[3]));
189 b0 = _mm_setr_ps(m[0], m[4], m[8], 0.0f);
190 b1 = _mm_setr_ps(m[1], m[5], m[9], 0.0f);
191 b2 = _mm_setr_ps(m[2], m[6], m[10], 0.0f);
192 b3 = _mm_setr_ps(m[3], m[7], m[11], 1.0f);
194 _mm_store_ps(b->m[0], b0);
195 _mm_store_ps(b->m[1], b1);
196 _mm_store_ps(b->m[2], b2);
197 _mm_store_ps(b->m[3], b3);
198 nr = _mm_loadu_ps(n);
199 r0 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0)));
200 r1 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1)));
201 r2 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2)));
202 r3 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3)));
203 nr = _mm_loadu_ps(n+4);
204 r0 = _mm_add_ps(r0, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0))));
205 r1 = _mm_add_ps(r1, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1))));
206 r2 = _mm_add_ps(r2, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2))));
207 r3 = _mm_add_ps(r3, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3))));
208 nr = _mm_loadu_ps(n+8);
209 r0 = _mm_add_ps(r0, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0))));
210 r1 = _mm_add_ps(r1, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1))));
211 r2 = _mm_add_ps(r2, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2))));
212 r3 = _mm_add_ps(r3, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3))));
213 r3 = _mm_add_ps(r3, b3);
214 _mm_store_ps(r->m[0], r0);
215 _mm_store_ps(r->m[1], r1);
216 _mm_store_ps(r->m[2], r2);
217 _mm_store_ps(r->m[3], r3);
222 // generate matrices for all blend combinations
223 weights = model->surfmesh.data_blendweights;
224 for (i = 0;i < model->surfmesh.num_blends;i++, weights++)
226 float * RESTRICT b = &boneposerelative[model->num_bones + i].m[0][0];
227 const float * RESTRICT m = &boneposerelative[weights->index[0]].m[0][0];
228 float f = weights->influence[0] * (1.0f / 255.0f);
229 __m128 fv = _mm_set_ps1(f);
230 __m128 b0 = _mm_load_ps(m);
231 __m128 b1 = _mm_load_ps(m+4);
232 __m128 b2 = _mm_load_ps(m+8);
233 __m128 b3 = _mm_load_ps(m+12);
234 __m128 m0, m1, m2, m3;
235 b0 = _mm_mul_ps(b0, fv);
236 b1 = _mm_mul_ps(b1, fv);
237 b2 = _mm_mul_ps(b2, fv);
238 b3 = _mm_mul_ps(b3, fv);
239 for (k = 1;k < 4 && weights->influence[k];k++)
241 m = &boneposerelative[weights->index[k]].m[0][0];
242 f = weights->influence[k] * (1.0f / 255.0f);
245 m1 = _mm_load_ps(m+4);
246 m2 = _mm_load_ps(m+8);
247 m3 = _mm_load_ps(m+12);
248 m0 = _mm_mul_ps(m0, fv);
249 m1 = _mm_mul_ps(m1, fv);
250 m2 = _mm_mul_ps(m2, fv);
251 m3 = _mm_mul_ps(m3, fv);
252 b0 = _mm_add_ps(m0, b0);
253 b1 = _mm_add_ps(m1, b1);
254 b2 = _mm_add_ps(m2, b2);
255 b3 = _mm_add_ps(m3, b3);
258 _mm_store_ps(b+4, b1);
259 _mm_store_ps(b+8, b2);
260 _mm_store_ps(b+12, b3);
263 #define LOAD_MATRIX_SCALAR() const float * RESTRICT m = &boneposerelative[*b].m[0][0]
265 #define LOAD_MATRIX3() \
266 const float * RESTRICT m = &boneposerelative[*b].m[0][0]; \
267 /* bonepose array is 16 byte aligned */ \
268 __m128 m1 = _mm_load_ps((m)); \
269 __m128 m2 = _mm_load_ps((m)+4); \
270 __m128 m3 = _mm_load_ps((m)+8);
271 #define LOAD_MATRIX4() \
272 const float * RESTRICT m = &boneposerelative[*b].m[0][0]; \
273 /* bonepose array is 16 byte aligned */ \
274 __m128 m1 = _mm_load_ps((m)); \
275 __m128 m2 = _mm_load_ps((m)+4); \
276 __m128 m3 = _mm_load_ps((m)+8); \
277 __m128 m4 = _mm_load_ps((m)+12)
279 /* Note that matrix is 4x4 and transposed compared to non-USE_SSE codepath */
280 #define TRANSFORM_POSITION_SCALAR(in, out) \
281 (out)[0] = ((in)[0] * m[0] + (in)[1] * m[4] + (in)[2] * m[ 8] + m[12]); \
282 (out)[1] = ((in)[0] * m[1] + (in)[1] * m[5] + (in)[2] * m[ 9] + m[13]); \
283 (out)[2] = ((in)[0] * m[2] + (in)[1] * m[6] + (in)[2] * m[10] + m[14]);
284 #define TRANSFORM_VECTOR_SCALAR(in, out) \
285 (out)[0] = ((in)[0] * m[0] + (in)[1] * m[4] + (in)[2] * m[ 8]); \
286 (out)[1] = ((in)[0] * m[1] + (in)[1] * m[5] + (in)[2] * m[ 9]); \
287 (out)[2] = ((in)[0] * m[2] + (in)[1] * m[6] + (in)[2] * m[10]);
289 #define TRANSFORM_POSITION(in, out) { \
290 __m128 pin = _mm_loadu_ps(in); /* we ignore the value in the last element (x from the next vertex) */ \
291 __m128 x = _mm_shuffle_ps(pin, pin, 0x0); \
292 __m128 t1 = _mm_mul_ps(x, m1); \
295 __m128 y = _mm_shuffle_ps(pin, pin, 0x55); \
296 __m128 t2 = _mm_mul_ps(y, m2); \
297 __m128 t3 = _mm_add_ps(t1, t2); \
300 __m128 z = _mm_shuffle_ps(pin, pin, 0xaa); \
301 __m128 t4 = _mm_mul_ps(z, m3); \
302 __m128 t5 = _mm_add_ps(t3, t4); \
305 __m128 pout = _mm_add_ps(t5, m4); \
306 _mm_storeu_ps((out), pout); \
309 #define TRANSFORM_VECTOR(in, out) { \
310 __m128 vin = _mm_loadu_ps(in); \
313 __m128 x = _mm_shuffle_ps(vin, vin, 0x0); \
314 __m128 t1 = _mm_mul_ps(x, m1); \
317 __m128 y = _mm_shuffle_ps(vin, vin, 0x55); \
318 __m128 t2 = _mm_mul_ps(y, m2); \
319 __m128 t3 = _mm_add_ps(t1, t2); \
321 /* nz, + (ny + nx) */ \
322 __m128 z = _mm_shuffle_ps(vin, vin, 0xaa); \
323 __m128 t4 = _mm_mul_ps(z, m3); \
324 __m128 vout = _mm_add_ps(t3, t4); \
325 _mm_storeu_ps((out), vout); \
328 // transform vertex attributes by blended matrices
331 const float * RESTRICT v = model->surfmesh.data_vertex3f;
332 const unsigned short * RESTRICT b = model->surfmesh.blends;
333 // special case common combinations of attributes to avoid repeated loading of matrices
336 const float * RESTRICT n = model->surfmesh.data_normal3f;
337 if (svector3f && tvector3f)
339 const float * RESTRICT svec = model->surfmesh.data_svector3f;
340 const float * RESTRICT tvec = model->surfmesh.data_tvector3f;
342 // Note that for SSE each iteration stores one element past end, so we break one vertex short
343 // and handle that with scalars in that case
344 for (i = 0; i < num_vertices_minus_one; i++, v += 3, n += 3, svec += 3, tvec += 3, b++,
345 vertex3f += 3, normal3f += 3, svector3f += 3, tvector3f += 3)
348 TRANSFORM_POSITION(v, vertex3f);
349 TRANSFORM_VECTOR(n, normal3f);
350 TRANSFORM_VECTOR(svec, svector3f);
351 TRANSFORM_VECTOR(tvec, tvector3f);
354 // Last vertex needs to be done with scalars to avoid reading/writing 1 word past end of arrays
356 LOAD_MATRIX_SCALAR();
357 TRANSFORM_POSITION_SCALAR(v, vertex3f);
358 TRANSFORM_VECTOR_SCALAR(n, normal3f);
359 TRANSFORM_VECTOR_SCALAR(svec, svector3f);
360 TRANSFORM_VECTOR_SCALAR(tvec, tvector3f);
362 //printf("elapsed ticks: %llu\n", rdtsc() - ts); // XXX
366 for (i = 0;i < num_vertices_minus_one; i++, v += 3, n += 3, b++, vertex3f += 3, normal3f += 3)
369 TRANSFORM_POSITION(v, vertex3f);
370 TRANSFORM_VECTOR(n, normal3f);
373 LOAD_MATRIX_SCALAR();
374 TRANSFORM_POSITION_SCALAR(v, vertex3f);
375 TRANSFORM_VECTOR_SCALAR(n, normal3f);
380 for (i = 0;i < num_vertices_minus_one; i++, v += 3, b++, vertex3f += 3)
383 TRANSFORM_POSITION(v, vertex3f);
386 LOAD_MATRIX_SCALAR();
387 TRANSFORM_POSITION_SCALAR(v, vertex3f);
394 const float * RESTRICT n = model->surfmesh.data_normal3f;
395 const unsigned short * RESTRICT b = model->surfmesh.blends;
396 for (i = 0; i < num_vertices_minus_one; i++, n += 3, b++, normal3f += 3)
399 TRANSFORM_VECTOR(n, normal3f);
402 LOAD_MATRIX_SCALAR();
403 TRANSFORM_VECTOR_SCALAR(n, normal3f);
409 const float * RESTRICT svec = model->surfmesh.data_svector3f;
410 const unsigned short * RESTRICT b = model->surfmesh.blends;
411 for (i = 0; i < num_vertices_minus_one; i++, svec += 3, b++, svector3f += 3)
414 TRANSFORM_VECTOR(svec, svector3f);
417 LOAD_MATRIX_SCALAR();
418 TRANSFORM_VECTOR_SCALAR(svec, svector3f);
424 const float * RESTRICT tvec = model->surfmesh.data_tvector3f;
425 const unsigned short * RESTRICT b = model->surfmesh.blends;
426 for (i = 0; i < num_vertices_minus_one; i++, tvec += 3, b++, tvector3f += 3)
429 TRANSFORM_VECTOR(tvec, tvector3f);
432 LOAD_MATRIX_SCALAR();
433 TRANSFORM_VECTOR_SCALAR(tvec, tvector3f);
439 #undef TRANSFORM_POSITION
440 #undef TRANSFORM_VECTOR
441 #undef LOAD_MATRIX_SCALAR
442 #undef TRANSFORM_POSITION_SCALAR
443 #undef TRANSFORM_VECTOR_SCALAR