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 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)
13 // vertex weighted skeletal
16 matrix4x4_t *bonepose;
17 matrix4x4_t *boneposerelative;
19 const blendweights_t * RESTRICT weights;
20 int num_vertices_minus_one;
22 num_vertices_minus_one = model->surfmesh.num_vertices - 1;
24 //unsigned long long ts = rdtsc();
25 bonepose = (matrix4x4_t *) Mod_Skeletal_AnimateVertices_AllocBuffers(sizeof(matrix4x4_t) * (model->num_bones*2 + model->surfmesh.num_blends));
26 boneposerelative = bonepose + model->num_bones;
28 if (skeleton && !skeleton->relativetransforms)
31 // interpolate matrices
34 for (i = 0;i < model->num_bones;i++)
36 const float * RESTRICT n = model->data_baseboneposeinverse + i * 12;
37 matrix4x4_t * RESTRICT s = &skeleton->relativetransforms[i];
38 matrix4x4_t * RESTRICT b = &bonepose[i];
39 matrix4x4_t * RESTRICT r = &boneposerelative[i];
40 __m128 b0, b1, b2, b3, r0, r1, r2, r3, nr;
41 if (model->data_bones[i].parent >= 0)
43 const matrix4x4_t * RESTRICT p = &bonepose[model->data_bones[i].parent];
44 __m128 s0 = _mm_loadu_ps(s->m[0]), s1 = _mm_loadu_ps(s->m[1]), s2 = _mm_loadu_ps(s->m[2]);
45 #ifdef OPENGLORIENTATION
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 _mm_store_ps(r->m[0], r0);
97 _mm_store_ps(r->m[1], r1);
98 _mm_store_ps(r->m[2], r2);
99 _mm_store_ps(r->m[3], r3);
104 for (i = 0;i < model->num_bones;i++)
106 const short * RESTRICT pose7s = model->data_poses7s + 7 * (frameblend[0].subframe * model->num_bones + i);
107 float lerp = frameblend[0].lerp,
108 tx = pose7s[0], ty = pose7s[1], tz = pose7s[2],
109 rx = pose7s[3] * lerp,
110 ry = pose7s[4] * lerp,
111 rz = pose7s[5] * lerp,
112 rw = pose7s[6] * lerp,
113 dx = tx*rw + ty*rz - tz*ry,
114 dy = -tx*rz + ty*rw + tz*rx,
115 dz = tx*ry - ty*rx + tz*rw,
116 dw = -tx*rx - ty*ry - tz*rz,
118 for (blends = 1;blends < MAX_FRAMEBLENDS && frameblend[blends].lerp > 0;blends++)
120 const short * RESTRICT pose7s = model->data_poses7s + 7 * (frameblend[blends].subframe * model->num_bones + i);
121 float lerp = frameblend[blends].lerp,
122 tx = pose7s[0], ty = pose7s[1], tz = pose7s[2],
123 qx = pose7s[3], qy = pose7s[4], qz = pose7s[5], qw = pose7s[6];
124 if(rx*qx + ry*qy + rz*qz + rw*qw < 0) lerp = -lerp;
133 dx += tx*qw + ty*qz - tz*qy;
134 dy += -tx*qz + ty*qw + tz*qx;
135 dz += tx*qy - ty*qx + tz*qw;
136 dw += -tx*qx - ty*qy - tz*qz;
138 scale = 1.0f / (rx*rx + ry*ry + rz*rz + rw*rw);
139 originscale = -model->num_posescale * scale;
140 m[0] = scale*(rw*rw + rx*rx - ry*ry - rz*rz);
141 m[1] = 2*scale*(rx*ry - rw*rz);
142 m[2] = 2*scale*(rx*rz + rw*ry);
143 m[3] = originscale*(dw*rx - dx*rw + dy*rz - dz*ry);
144 m[4] = 2*scale*(rx*ry + rw*rz);
145 m[5] = scale*(rw*rw + ry*ry - rx*rx - rz*rz);
146 m[6] = 2*scale*(ry*rz - rw*rx);
147 m[7] = originscale*(dw*ry - dx*rz - dy*rw + dz*rx);
148 m[8] = 2*scale*(rx*rz - rw*ry);
149 m[9] = 2*scale*(ry*rz + rw*rx);
150 m[10] = scale*(rw*rw + rz*rz - rx*rx - ry*ry);
151 m[11] = originscale*(dw*rz + dx*ry - dy*rx - dz*rw);
152 if (i == r_skeletal_debugbone.integer)
153 m[r_skeletal_debugbonecomponent.integer % 12] += r_skeletal_debugbonevalue.value;
154 m[3] *= r_skeletal_debugtranslatex.value;
155 m[7] *= r_skeletal_debugtranslatey.value;
156 m[11] *= r_skeletal_debugtranslatez.value;
158 const float * RESTRICT n = model->data_baseboneposeinverse + i * 12;
159 matrix4x4_t * RESTRICT b = &bonepose[i];
160 matrix4x4_t * RESTRICT r = &boneposerelative[i];
161 __m128 b0, b1, b2, b3, r0, r1, r2, r3, nr;
162 if (model->data_bones[i].parent >= 0)
164 const matrix4x4_t * RESTRICT p = &bonepose[model->data_bones[i].parent];
165 __m128 pr = _mm_load_ps(p->m[0]);
166 b0 = _mm_mul_ps(pr, _mm_set1_ps(m[0]));
167 b1 = _mm_mul_ps(pr, _mm_set1_ps(m[1]));
168 b2 = _mm_mul_ps(pr, _mm_set1_ps(m[2]));
169 b3 = _mm_mul_ps(pr, _mm_set1_ps(m[3]));
170 pr = _mm_load_ps(p->m[1]);
171 b0 = _mm_add_ps(b0, _mm_mul_ps(pr, _mm_set1_ps(m[4])));
172 b1 = _mm_add_ps(b1, _mm_mul_ps(pr, _mm_set1_ps(m[5])));
173 b2 = _mm_add_ps(b2, _mm_mul_ps(pr, _mm_set1_ps(m[6])));
174 b3 = _mm_add_ps(b3, _mm_mul_ps(pr, _mm_set1_ps(m[7])));
175 pr = _mm_load_ps(p->m[2]);
176 b0 = _mm_add_ps(b0, _mm_mul_ps(pr, _mm_set1_ps(m[8])));
177 b1 = _mm_add_ps(b1, _mm_mul_ps(pr, _mm_set1_ps(m[9])));
178 b2 = _mm_add_ps(b2, _mm_mul_ps(pr, _mm_set1_ps(m[10])));
179 b3 = _mm_add_ps(b3, _mm_mul_ps(pr, _mm_set1_ps(m[11])));
180 b3 = _mm_add_ps(b3, _mm_load_ps(p->m[3]));
184 b0 = _mm_setr_ps(m[0], m[4], m[8], 0.0f);
185 b1 = _mm_setr_ps(m[1], m[5], m[9], 0.0f);
186 b2 = _mm_setr_ps(m[2], m[6], m[10], 0.0f);
187 b3 = _mm_setr_ps(m[3], m[7], m[11], 1.0f);
189 _mm_store_ps(b->m[0], b0);
190 _mm_store_ps(b->m[1], b1);
191 _mm_store_ps(b->m[2], b2);
192 _mm_store_ps(b->m[3], b3);
193 nr = _mm_loadu_ps(n);
194 r0 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0)));
195 r1 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1)));
196 r2 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2)));
197 r3 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3)));
198 nr = _mm_loadu_ps(n+4);
199 r0 = _mm_add_ps(r0, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0))));
200 r1 = _mm_add_ps(r1, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1))));
201 r2 = _mm_add_ps(r2, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2))));
202 r3 = _mm_add_ps(r3, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3))));
203 nr = _mm_loadu_ps(n+8);
204 r0 = _mm_add_ps(r0, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0))));
205 r1 = _mm_add_ps(r1, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1))));
206 r2 = _mm_add_ps(r2, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2))));
207 r3 = _mm_add_ps(r3, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3))));
208 r3 = _mm_add_ps(r3, b3);
209 _mm_store_ps(r->m[0], r0);
210 _mm_store_ps(r->m[1], r1);
211 _mm_store_ps(r->m[2], r2);
212 _mm_store_ps(r->m[3], r3);
217 // generate matrices for all blend combinations
218 weights = model->surfmesh.data_blendweights;
219 for (i = 0;i < model->surfmesh.num_blends;i++, weights++)
221 float * RESTRICT b = &boneposerelative[model->num_bones + i].m[0][0];
222 const float * RESTRICT m = &boneposerelative[weights->index[0]].m[0][0];
223 float f = weights->influence[0] * (1.0f / 255.0f);
224 __m128 fv = _mm_set_ps1(f);
225 __m128 b0 = _mm_load_ps(m);
226 __m128 b1 = _mm_load_ps(m+4);
227 __m128 b2 = _mm_load_ps(m+8);
228 __m128 b3 = _mm_load_ps(m+12);
229 __m128 m0, m1, m2, m3;
230 b0 = _mm_mul_ps(b0, fv);
231 b1 = _mm_mul_ps(b1, fv);
232 b2 = _mm_mul_ps(b2, fv);
233 b3 = _mm_mul_ps(b3, fv);
234 for (k = 1;k < 4 && weights->influence[k];k++)
236 m = &boneposerelative[weights->index[k]].m[0][0];
237 f = weights->influence[k] * (1.0f / 255.0f);
240 m1 = _mm_load_ps(m+4);
241 m2 = _mm_load_ps(m+8);
242 m3 = _mm_load_ps(m+12);
243 m0 = _mm_mul_ps(m0, fv);
244 m1 = _mm_mul_ps(m1, fv);
245 m2 = _mm_mul_ps(m2, fv);
246 m3 = _mm_mul_ps(m3, fv);
247 b0 = _mm_add_ps(m0, b0);
248 b1 = _mm_add_ps(m1, b1);
249 b2 = _mm_add_ps(m2, b2);
250 b3 = _mm_add_ps(m3, b3);
253 _mm_store_ps(b+4, b1);
254 _mm_store_ps(b+8, b2);
255 _mm_store_ps(b+12, b3);
258 #define LOAD_MATRIX_SCALAR() const float * RESTRICT m = &boneposerelative[*b].m[0][0]
260 #define LOAD_MATRIX3() \
261 const float * RESTRICT m = &boneposerelative[*b].m[0][0]; \
262 /* bonepose array is 16 byte aligned */ \
263 __m128 m1 = _mm_load_ps((m)); \
264 __m128 m2 = _mm_load_ps((m)+4); \
265 __m128 m3 = _mm_load_ps((m)+8);
266 #define LOAD_MATRIX4() \
267 const float * RESTRICT m = &boneposerelative[*b].m[0][0]; \
268 /* bonepose array is 16 byte aligned */ \
269 __m128 m1 = _mm_load_ps((m)); \
270 __m128 m2 = _mm_load_ps((m)+4); \
271 __m128 m3 = _mm_load_ps((m)+8); \
272 __m128 m4 = _mm_load_ps((m)+12)
274 /* Note that matrix is 4x4 and transposed compared to non-USE_SSE codepath */
275 #define TRANSFORM_POSITION_SCALAR(in, out) \
276 (out)[0] = ((in)[0] * m[0] + (in)[1] * m[4] + (in)[2] * m[ 8] + m[12]); \
277 (out)[1] = ((in)[0] * m[1] + (in)[1] * m[5] + (in)[2] * m[ 9] + m[13]); \
278 (out)[2] = ((in)[0] * m[2] + (in)[1] * m[6] + (in)[2] * m[10] + m[14]);
279 #define TRANSFORM_VECTOR_SCALAR(in, out) \
280 (out)[0] = ((in)[0] * m[0] + (in)[1] * m[4] + (in)[2] * m[ 8]); \
281 (out)[1] = ((in)[0] * m[1] + (in)[1] * m[5] + (in)[2] * m[ 9]); \
282 (out)[2] = ((in)[0] * m[2] + (in)[1] * m[6] + (in)[2] * m[10]);
284 #define TRANSFORM_POSITION(in, out) { \
285 __m128 pin = _mm_loadu_ps(in); /* we ignore the value in the last element (x from the next vertex) */ \
286 __m128 x = _mm_shuffle_ps(pin, pin, 0x0); \
287 __m128 t1 = _mm_mul_ps(x, m1); \
290 __m128 y = _mm_shuffle_ps(pin, pin, 0x55); \
291 __m128 t2 = _mm_mul_ps(y, m2); \
292 __m128 t3 = _mm_add_ps(t1, t2); \
295 __m128 z = _mm_shuffle_ps(pin, pin, 0xaa); \
296 __m128 t4 = _mm_mul_ps(z, m3); \
297 __m128 t5 = _mm_add_ps(t3, t4); \
300 __m128 pout = _mm_add_ps(t5, m4); \
301 _mm_storeu_ps((out), pout); \
304 #define TRANSFORM_VECTOR(in, out) { \
305 __m128 vin = _mm_loadu_ps(in); \
308 __m128 x = _mm_shuffle_ps(vin, vin, 0x0); \
309 __m128 t1 = _mm_mul_ps(x, m1); \
312 __m128 y = _mm_shuffle_ps(vin, vin, 0x55); \
313 __m128 t2 = _mm_mul_ps(y, m2); \
314 __m128 t3 = _mm_add_ps(t1, t2); \
316 /* nz, + (ny + nx) */ \
317 __m128 z = _mm_shuffle_ps(vin, vin, 0xaa); \
318 __m128 t4 = _mm_mul_ps(z, m3); \
319 __m128 vout = _mm_add_ps(t3, t4); \
320 _mm_storeu_ps((out), vout); \
323 // transform vertex attributes by blended matrices
326 const float * RESTRICT v = model->surfmesh.data_vertex3f;
327 const unsigned short * RESTRICT b = model->surfmesh.blends;
328 // special case common combinations of attributes to avoid repeated loading of matrices
331 const float * RESTRICT n = model->surfmesh.data_normal3f;
332 if (svector3f && tvector3f)
334 const float * RESTRICT sv = model->surfmesh.data_svector3f;
335 const float * RESTRICT tv = model->surfmesh.data_tvector3f;
337 // Note that for SSE each iteration stores one element past end, so we break one vertex short
338 // and handle that with scalars in that case
339 for (i = 0; i < num_vertices_minus_one; i++, v += 3, n += 3, sv += 3, tv += 3, b++,
340 vertex3f += 3, normal3f += 3, svector3f += 3, tvector3f += 3)
343 TRANSFORM_POSITION(v, vertex3f);
344 TRANSFORM_VECTOR(n, normal3f);
345 TRANSFORM_VECTOR(sv, svector3f);
346 TRANSFORM_VECTOR(tv, tvector3f);
349 // Last vertex needs to be done with scalars to avoid reading/writing 1 word past end of arrays
351 LOAD_MATRIX_SCALAR();
352 TRANSFORM_POSITION_SCALAR(v, vertex3f);
353 TRANSFORM_VECTOR_SCALAR(n, normal3f);
354 TRANSFORM_VECTOR_SCALAR(sv, svector3f);
355 TRANSFORM_VECTOR_SCALAR(tv, tvector3f);
357 //printf("elapsed ticks: %llu\n", rdtsc() - ts); // XXX
361 for (i = 0;i < num_vertices_minus_one; i++, v += 3, n += 3, b++, vertex3f += 3, normal3f += 3)
364 TRANSFORM_POSITION(v, vertex3f);
365 TRANSFORM_VECTOR(n, normal3f);
368 LOAD_MATRIX_SCALAR();
369 TRANSFORM_POSITION_SCALAR(v, vertex3f);
370 TRANSFORM_VECTOR_SCALAR(n, normal3f);
375 for (i = 0;i < num_vertices_minus_one; i++, v += 3, b++, vertex3f += 3)
378 TRANSFORM_POSITION(v, vertex3f);
381 LOAD_MATRIX_SCALAR();
382 TRANSFORM_POSITION_SCALAR(v, vertex3f);
389 const float * RESTRICT n = model->surfmesh.data_normal3f;
390 const unsigned short * RESTRICT b = model->surfmesh.blends;
391 for (i = 0; i < num_vertices_minus_one; i++, n += 3, b++, normal3f += 3)
394 TRANSFORM_VECTOR(n, normal3f);
397 LOAD_MATRIX_SCALAR();
398 TRANSFORM_VECTOR_SCALAR(n, normal3f);
404 const float * RESTRICT sv = model->surfmesh.data_svector3f;
405 const unsigned short * RESTRICT b = model->surfmesh.blends;
406 for (i = 0; i < num_vertices_minus_one; i++, sv += 3, b++, svector3f += 3)
409 TRANSFORM_VECTOR(sv, svector3f);
412 LOAD_MATRIX_SCALAR();
413 TRANSFORM_VECTOR_SCALAR(sv, svector3f);
419 const float * RESTRICT tv = model->surfmesh.data_tvector3f;
420 const unsigned short * RESTRICT b = model->surfmesh.blends;
421 for (i = 0; i < num_vertices_minus_one; i++, tv += 3, b++, tvector3f += 3)
424 TRANSFORM_VECTOR(tv, tvector3f);
427 LOAD_MATRIX_SCALAR();
428 TRANSFORM_VECTOR_SCALAR(tv, tvector3f);
434 #undef TRANSFORM_POSITION
435 #undef TRANSFORM_VECTOR
436 #undef LOAD_MATRIX_SCALAR
437 #undef TRANSFORM_POSITION_SCALAR
438 #undef TRANSFORM_VECTOR_SCALAR