+#include "mod_skeletal_animatevertices_sse.h"
+
+#ifdef SSE_POSSIBLE
+
+#ifdef MATRIX4x4_OPENGLORIENTATION
+#error "SSE skeletal requires D3D matrix layout"
+#endif
+
+#include <xmmintrin.h>
+
+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)
+{
+ // vertex weighted skeletal
+ int i, k;
+ int blends;
+ matrix4x4_t *bonepose;
+ matrix4x4_t *boneposerelative;
+ float m[12];
+ matrix4x4_t mm, mm2;
+ const blendweights_t * RESTRICT weights;
+ int num_vertices_minus_one;
+
+ if (!model->surfmesh.num_vertices)
+ return;
+
+ num_vertices_minus_one = model->surfmesh.num_vertices - 1;
+
+ //unsigned long long ts = rdtsc();
+ bonepose = (matrix4x4_t *) Mod_Skeletal_AnimateVertices_AllocBuffers(sizeof(matrix4x4_t) * (model->num_bones*2 + model->surfmesh.num_blends));
+ boneposerelative = bonepose + model->num_bones;
+
+ if (skeleton && !skeleton->relativetransforms)
+ skeleton = NULL;
+
+ // interpolate matrices
+ if (skeleton)
+ {
+ for (i = 0;i < model->num_bones;i++)
+ {
+ // relativetransforms is in GL column-major order, which is what we need for SSE
+ // transposed style processing
+ if (model->data_bones[i].parent >= 0)
+ Matrix4x4_Concat(&bonepose[i], &bonepose[model->data_bones[i].parent], &skeleton->relativetransforms[i]);
+ else
+ memcpy(&bonepose[i], &skeleton->relativetransforms[i], sizeof(matrix4x4_t));
+
+ // create a relative deformation matrix to describe displacement
+ // from the base mesh, which is used by the actual weighting
+ Matrix4x4_FromArray12FloatD3D(&mm, model->data_baseboneposeinverse + i * 12); // baseboneposeinverse is 4x3 row-major
+ Matrix4x4_Concat(&boneposerelative[i], &bonepose[i], &mm);
+ }
+ }
+ else
+ {
+ float originscale = model->num_posescale;
+ float x,y,z,w,lerp;
+ const short * RESTRICT pose6s;
+
+ for (i = 0;i < model->num_bones;i++)
+ {
+ memset(m, 0, sizeof(m));
+ for (blends = 0;blends < MAX_FRAMEBLENDS && frameblend[blends].lerp > 0;blends++)
+ {
+ pose6s = model->data_poses6s + 6 * (frameblend[blends].subframe * model->num_bones + i);
+ lerp = frameblend[blends].lerp;
+ x = pose6s[3] * (1.0f / 32767.0f);
+ y = pose6s[4] * (1.0f / 32767.0f);
+ z = pose6s[5] * (1.0f / 32767.0f);
+ w = 1.0f - (x*x+y*y+z*z);
+ w = w > 0.0f ? -sqrt(w) : 0.0f;
+ m[ 0] += (1-2*(y*y+z*z)) * lerp;
+ m[ 1] += ( 2*(x*y-z*w)) * lerp;
+ m[ 2] += ( 2*(x*z+y*w)) * lerp;
+ m[ 3] += (pose6s[0] * originscale) * lerp;
+ m[ 4] += ( 2*(x*y+z*w)) * lerp;
+ m[ 5] += (1-2*(x*x+z*z)) * lerp;
+ m[ 6] += ( 2*(y*z-x*w)) * lerp;
+ m[ 7] += (pose6s[1] * originscale) * lerp;
+ m[ 8] += ( 2*(x*z-y*w)) * lerp;
+ m[ 9] += ( 2*(y*z+x*w)) * lerp;
+ m[10] += (1-2*(x*x+y*y)) * lerp;
+ m[11] += (pose6s[2] * originscale) * lerp;
+ }
+ VectorNormalize(m );
+ VectorNormalize(m + 4);
+ VectorNormalize(m + 8);
+ if (i == r_skeletal_debugbone.integer)
+ m[r_skeletal_debugbonecomponent.integer % 12] += r_skeletal_debugbonevalue.value;
+ m[3] *= r_skeletal_debugtranslatex.value;
+ m[7] *= r_skeletal_debugtranslatey.value;
+ m[11] *= r_skeletal_debugtranslatez.value;
+ Matrix4x4_FromArray12FloatD3D(&mm, m);
+ if (model->data_bones[i].parent >= 0)
+ Matrix4x4_Concat(&bonepose[i], &bonepose[model->data_bones[i].parent], &mm);
+ else
+ memcpy(&bonepose[i], &mm, sizeof(mm));
+ // create a relative deformation matrix to describe displacement
+ // from the base mesh, which is used by the actual weighting
+ Matrix4x4_FromArray12FloatD3D(&mm, model->data_baseboneposeinverse + i * 12); // baseboneposeinverse is 4x3 row-major
+ Matrix4x4_Concat(&mm2, &bonepose[i], &mm);
+ Matrix4x4_Transpose(&boneposerelative[i], &mm2); // TODO: Eliminate this transpose
+ }
+ }
+
+ // generate matrices for all blend combinations
+ weights = model->surfmesh.data_blendweights;
+ for (i = 0;i < model->surfmesh.num_blends;i++, weights++)
+ {
+ float * RESTRICT b = &boneposerelative[model->num_bones + i].m[0][0];
+ const float * RESTRICT m = &boneposerelative[weights->index[0]].m[0][0];
+ float f = weights->influence[0] * (1.0f / 255.0f);
+ __m128 fv = _mm_set_ps1(f);
+ __m128 b0 = _mm_load_ps(m);
+ __m128 b1 = _mm_load_ps(m+4);
+ __m128 b2 = _mm_load_ps(m+8);
+ __m128 b3 = _mm_load_ps(m+12);
+ __m128 m0, m1, m2, m3;
+ b0 = _mm_mul_ps(b0, fv);
+ b1 = _mm_mul_ps(b1, fv);
+ b2 = _mm_mul_ps(b2, fv);
+ b3 = _mm_mul_ps(b3, fv);
+ for (k = 1;k < 4 && weights->influence[k];k++)
+ {
+ m = &boneposerelative[weights->index[k]].m[0][0];
+ f = weights->influence[k] * (1.0f / 255.0f);
+ fv = _mm_set_ps1(f);
+ m0 = _mm_load_ps(m);
+ m1 = _mm_load_ps(m+4);
+ m2 = _mm_load_ps(m+8);
+ m3 = _mm_load_ps(m+12);
+ m0 = _mm_mul_ps(m0, fv);
+ m1 = _mm_mul_ps(m1, fv);
+ m2 = _mm_mul_ps(m2, fv);
+ m3 = _mm_mul_ps(m3, fv);
+ b0 = _mm_add_ps(m0, b0);
+ b1 = _mm_add_ps(m1, b1);
+ b2 = _mm_add_ps(m2, b2);
+ b3 = _mm_add_ps(m3, b3);
+ }
+ _mm_store_ps(b, b0);
+ _mm_store_ps(b+4, b1);
+ _mm_store_ps(b+8, b2);
+ _mm_store_ps(b+12, b3);
+ }
+
+#define LOAD_MATRIX_SCALAR() const float * RESTRICT m = &boneposerelative[*b].m[0][0]
+
+#define LOAD_MATRIX3() \
+ const float * RESTRICT m = &boneposerelative[*b].m[0][0]; \
+ /* bonepose array is 16 byte aligned */ \
+ __m128 m1 = _mm_load_ps((m)); \
+ __m128 m2 = _mm_load_ps((m)+4); \
+ __m128 m3 = _mm_load_ps((m)+8);
+#define LOAD_MATRIX4() \
+ const float * RESTRICT m = &boneposerelative[*b].m[0][0]; \
+ /* bonepose array is 16 byte aligned */ \
+ __m128 m1 = _mm_load_ps((m)); \
+ __m128 m2 = _mm_load_ps((m)+4); \
+ __m128 m3 = _mm_load_ps((m)+8); \
+ __m128 m4 = _mm_load_ps((m)+12)
+
+ /* Note that matrix is 4x4 and transposed compared to non-USE_SSE codepath */
+#define TRANSFORM_POSITION_SCALAR(in, out) \
+ (out)[0] = ((in)[0] * m[0] + (in)[1] * m[4] + (in)[2] * m[ 8] + m[12]); \
+ (out)[1] = ((in)[0] * m[1] + (in)[1] * m[5] + (in)[2] * m[ 9] + m[13]); \
+ (out)[2] = ((in)[0] * m[2] + (in)[1] * m[6] + (in)[2] * m[10] + m[14]);
+#define TRANSFORM_VECTOR_SCALAR(in, out) \
+ (out)[0] = ((in)[0] * m[0] + (in)[1] * m[4] + (in)[2] * m[ 8]); \
+ (out)[1] = ((in)[0] * m[1] + (in)[1] * m[5] + (in)[2] * m[ 9]); \
+ (out)[2] = ((in)[0] * m[2] + (in)[1] * m[6] + (in)[2] * m[10]);
+
+#define TRANSFORM_POSITION(in, out) { \
+ __m128 pin = _mm_loadu_ps(in); /* we ignore the value in the last element (x from the next vertex) */ \
+ __m128 x = _mm_shuffle_ps(pin, pin, 0x0); \
+ __m128 t1 = _mm_mul_ps(x, m1); \
+ \
+ /* y, + x */ \
+ __m128 y = _mm_shuffle_ps(pin, pin, 0x55); \
+ __m128 t2 = _mm_mul_ps(y, m2); \
+ __m128 t3 = _mm_add_ps(t1, t2); \
+ \
+ /* z, + (y+x) */ \
+ __m128 z = _mm_shuffle_ps(pin, pin, 0xaa); \
+ __m128 t4 = _mm_mul_ps(z, m3); \
+ __m128 t5 = _mm_add_ps(t3, t4); \
+ \
+ /* + m3 */ \
+ __m128 pout = _mm_add_ps(t5, m4); \
+ _mm_storeu_ps((out), pout); \
+ }
+
+#define TRANSFORM_VECTOR(in, out) { \
+ __m128 vin = _mm_loadu_ps(in); \
+ \
+ /* x */ \
+ __m128 x = _mm_shuffle_ps(vin, vin, 0x0); \
+ __m128 t1 = _mm_mul_ps(x, m1); \
+ \
+ /* y, + x */ \
+ __m128 y = _mm_shuffle_ps(vin, vin, 0x55); \
+ __m128 t2 = _mm_mul_ps(y, m2); \
+ __m128 t3 = _mm_add_ps(t1, t2); \
+ \
+ /* nz, + (ny + nx) */ \
+ __m128 z = _mm_shuffle_ps(vin, vin, 0xaa); \
+ __m128 t4 = _mm_mul_ps(z, m3); \
+ __m128 vout = _mm_add_ps(t3, t4); \
+ _mm_storeu_ps((out), vout); \
+ }
+
+ // transform vertex attributes by blended matrices
+ if (vertex3f)
+ {
+ const float * RESTRICT v = model->surfmesh.data_vertex3f;
+ const unsigned short * RESTRICT b = model->surfmesh.blends;
+ // special case common combinations of attributes to avoid repeated loading of matrices
+ if (normal3f)
+ {
+ const float * RESTRICT n = model->surfmesh.data_normal3f;
+ if (svector3f && tvector3f)
+ {
+ const float * RESTRICT sv = model->surfmesh.data_svector3f;
+ const float * RESTRICT tv = model->surfmesh.data_tvector3f;
+
+ // Note that for SSE each iteration stores one element past end, so we break one vertex short
+ // and handle that with scalars in that case
+ for (i = 0; i < num_vertices_minus_one; i++, v += 3, n += 3, sv += 3, tv += 3, b++,
+ vertex3f += 3, normal3f += 3, svector3f += 3, tvector3f += 3)
+ {
+ LOAD_MATRIX4();
+ TRANSFORM_POSITION(v, vertex3f);
+ TRANSFORM_VECTOR(n, normal3f);
+ TRANSFORM_VECTOR(sv, svector3f);
+ TRANSFORM_VECTOR(tv, tvector3f);
+ }
+
+ // Last vertex needs to be done with scalars to avoid reading/writing 1 word past end of arrays
+ {
+ LOAD_MATRIX_SCALAR();
+ TRANSFORM_POSITION_SCALAR(v, vertex3f);
+ TRANSFORM_VECTOR_SCALAR(n, normal3f);
+ TRANSFORM_VECTOR_SCALAR(sv, svector3f);
+ TRANSFORM_VECTOR_SCALAR(tv, tvector3f);
+ }
+ //printf("elapsed ticks: %llu\n", rdtsc() - ts); // XXX
+ return;
+ }
+
+ for (i = 0;i < num_vertices_minus_one; i++, v += 3, n += 3, b++, vertex3f += 3, normal3f += 3)
+ {
+ LOAD_MATRIX4();
+ TRANSFORM_POSITION(v, vertex3f);
+ TRANSFORM_VECTOR(n, normal3f);
+ }
+ {
+ LOAD_MATRIX_SCALAR();
+ TRANSFORM_POSITION_SCALAR(v, vertex3f);
+ TRANSFORM_VECTOR_SCALAR(n, normal3f);
+ }
+ }
+ else
+ {
+ for (i = 0;i < num_vertices_minus_one; i++, v += 3, b++, vertex3f += 3)
+ {
+ LOAD_MATRIX4();
+ TRANSFORM_POSITION(v, vertex3f);
+ }
+ {
+ LOAD_MATRIX_SCALAR();
+ TRANSFORM_POSITION_SCALAR(v, vertex3f);
+ }
+ }
+ }
+
+ else if (normal3f)
+ {
+ const float * RESTRICT n = model->surfmesh.data_normal3f;
+ const unsigned short * RESTRICT b = model->surfmesh.blends;
+ for (i = 0; i < num_vertices_minus_one; i++, n += 3, b++, normal3f += 3)
+ {
+ LOAD_MATRIX3();
+ TRANSFORM_VECTOR(n, normal3f);
+ }
+ {
+ LOAD_MATRIX_SCALAR();
+ TRANSFORM_VECTOR_SCALAR(n, normal3f);
+ }
+ }
+
+ if (svector3f)
+ {
+ const float * RESTRICT sv = model->surfmesh.data_svector3f;
+ const unsigned short * RESTRICT b = model->surfmesh.blends;
+ for (i = 0; i < num_vertices_minus_one; i++, sv += 3, b++, svector3f += 3)
+ {
+ LOAD_MATRIX3();
+ TRANSFORM_VECTOR(sv, svector3f);
+ }
+ {
+ LOAD_MATRIX_SCALAR();
+ TRANSFORM_VECTOR_SCALAR(sv, svector3f);
+ }
+ }
+
+ if (tvector3f)
+ {
+ const float * RESTRICT tv = model->surfmesh.data_tvector3f;
+ const unsigned short * RESTRICT b = model->surfmesh.blends;
+ for (i = 0; i < num_vertices_minus_one; i++, tv += 3, b++, tvector3f += 3)
+ {
+ LOAD_MATRIX3();
+ TRANSFORM_VECTOR(tv, tvector3f);
+ }
+ {
+ LOAD_MATRIX_SCALAR();
+ TRANSFORM_VECTOR_SCALAR(tv, tvector3f);
+ }
+ }
+
+#undef LOAD_MATRIX3
+#undef LOAD_MATRIX4
+#undef TRANSFORM_POSITION
+#undef TRANSFORM_VECTOR
+#undef LOAD_MATRIX_SCALAR
+#undef TRANSFORM_POSITION_SCALAR
+#undef TRANSFORM_VECTOR_SCALAR
+}
+
+#endif