X-Git-Url: http://de.git.xonotic.org/?p=xonotic%2Fdarkplaces.git;a=blobdiff_plain;f=model_alias.c;h=cdf0edeee0928722f85bf5373004f98fd1353929;hp=448b0c4328dcf09648c71272f9b954e548324c24;hb=35b7d263c31d6576b488a3df4c1b06ad5acb3ecd;hpb=deb439e4d0f954489242e932c0fefee5cc2b3f77

diff --git a/model_alias.c b/model_alias.c
index 448b0c43..cdf0edee 100644
--- a/model_alias.c
+++ b/model_alias.c
@@ -21,7 +21,15 @@ Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
 #include "quakedef.h"
 #include "image.h"
 #include "r_shadow.h"
+#include "mod_skeletal_animatevertices_generic.h"
+#ifdef SSE_POSSIBLE
+#include "mod_skeletal_animatevertices_sse.h"
+#endif
 
+#ifdef SSE_POSSIBLE
+static qboolean r_skeletal_use_sse_defined = false;
+cvar_t r_skeletal_use_sse = {0, "r_skeletal_use_sse", "1", "use SSE for skeletal model animation"};
+#endif
 cvar_t r_skeletal_debugbone = {0, "r_skeletal_debugbone", "-1", "development cvar for testing skeletal model code"};
 cvar_t r_skeletal_debugbonecomponent = {0, "r_skeletal_debugbonecomponent", "3", "development cvar for testing skeletal model code"};
 cvar_t r_skeletal_debugbonevalue = {0, "r_skeletal_debugbonevalue", "100", "development cvar for testing skeletal model code"};
@@ -29,33 +37,39 @@ cvar_t r_skeletal_debugtranslatex = {0, "r_skeletal_debugtranslatex", "1", "deve
 cvar_t r_skeletal_debugtranslatey = {0, "r_skeletal_debugtranslatey", "1", "development cvar for testing skeletal model code"};
 cvar_t r_skeletal_debugtranslatez = {0, "r_skeletal_debugtranslatez", "1", "development cvar for testing skeletal model code"};
 cvar_t mod_alias_supporttagscale = {0, "mod_alias_supporttagscale", "1", "support scaling factors in bone/tag attachment matrices as supported by MD3"};
+cvar_t mod_alias_force_animated = {0, "mod_alias_force_animated", "", "if set to an non-empty string, overrides the is-animated flag of any alias models (for benchmarking)"};
 
 float mod_md3_sin[320];
 
-void Mod_AliasInit (void)
+static size_t Mod_Skeletal_AnimateVertices_maxbonepose = 0;
+static void *Mod_Skeletal_AnimateVertices_bonepose = NULL;
+void Mod_Skeletal_FreeBuffers(void)
 {
-	int i;
-	Cvar_RegisterVariable(&r_skeletal_debugbone);
-	Cvar_RegisterVariable(&r_skeletal_debugbonecomponent);
-	Cvar_RegisterVariable(&r_skeletal_debugbonevalue);
-	Cvar_RegisterVariable(&r_skeletal_debugtranslatex);
-	Cvar_RegisterVariable(&r_skeletal_debugtranslatey);
-	Cvar_RegisterVariable(&r_skeletal_debugtranslatez);
-	Cvar_RegisterVariable(&mod_alias_supporttagscale);
-	for (i = 0;i < 320;i++)
-		mod_md3_sin[i] = sin(i * M_PI * 2.0f / 256.0);
+	if(Mod_Skeletal_AnimateVertices_bonepose)
+		Mem_Free(Mod_Skeletal_AnimateVertices_bonepose);
+	Mod_Skeletal_AnimateVertices_maxbonepose = 0;
+	Mod_Skeletal_AnimateVertices_bonepose = NULL;
+}
+void *Mod_Skeletal_AnimateVertices_AllocBuffers(size_t nbytes)
+{
+	if(Mod_Skeletal_AnimateVertices_maxbonepose < nbytes)
+	{
+		if(Mod_Skeletal_AnimateVertices_bonepose)
+			Mem_Free(Mod_Skeletal_AnimateVertices_bonepose);
+		Mod_Skeletal_AnimateVertices_bonepose = Z_Malloc(nbytes);
+		Mod_Skeletal_AnimateVertices_maxbonepose = nbytes;
+	}
+	return Mod_Skeletal_AnimateVertices_bonepose;
 }
 
-void Mod_Skeletal_AnimateVertices(const dp_model_t *model, const frameblend_t *frameblend, const skeleton_t *skeleton, float *vertex3f, float *normal3f, float *svector3f, float *tvector3f)
+void Mod_Skeletal_BuildTransforms(const dp_model_t * RESTRICT model, const frameblend_t * RESTRICT frameblend, const skeleton_t *skeleton, float * RESTRICT bonepose, float * RESTRICT boneposerelative)
 {
-#define MAX_BONES 256
-	// vertex weighted skeletal
-	int i, k;
-	int blends;
-	float desiredscale[3];
-	float boneposerelative[MAX_BONES][12];
-	float *matrix, m[12], bonepose[MAX_BONES][12];
+	int i, blends;
+	float m[12];
 
+	if (!bonepose)
+		bonepose = (float * RESTRICT) Mod_Skeletal_AnimateVertices_AllocBuffers(sizeof(float[12]) * model->num_bones);
+		
 	if (skeleton && !skeleton->relativetransforms)
 		skeleton = NULL;
 
@@ -66,192 +80,200 @@ void Mod_Skeletal_AnimateVertices(const dp_model_t *model, const frameblend_t *f
 		{
 			Matrix4x4_ToArray12FloatD3D(&skeleton->relativetransforms[i], m);
 			if (model->data_bones[i].parent >= 0)
-				R_ConcatTransforms(bonepose[model->data_bones[i].parent], m, bonepose[i]);
+				R_ConcatTransforms(bonepose + model->data_bones[i].parent * 12, m, bonepose + i * 12);
 			else
-				for (k = 0;k < 12;k++)
-					bonepose[i][k] = m[k];
-	
+				memcpy(bonepose + i * 12, m, sizeof(m));
+
 			// create a relative deformation matrix to describe displacement
 			// from the base mesh, which is used by the actual weighting
-			R_ConcatTransforms(bonepose[i], model->data_baseboneposeinverse + i * 12, boneposerelative[i]);
+			R_ConcatTransforms(bonepose + i * 12, model->data_baseboneposeinverse + i * 12, boneposerelative + i * 12);
 		}
 	}
 	else
 	{
 		for (i = 0;i < model->num_bones;i++)
 		{
-			for (k = 0;k < 12;k++)
-				m[k] = 0;
-			VectorClear(desiredscale);
-			for (blends = 0;blends < MAX_FRAMEBLENDS && frameblend[blends].lerp > 0;blends++)
+			// blend by transform each quaternion/translation into a dual-quaternion first, then blending
+			const short * RESTRICT firstpose7s = model->data_poses7s + 7 * (frameblend[0].subframe * model->num_bones + i);
+			float firstlerp = frameblend[0].lerp,
+				firsttx = firstpose7s[0], firstty = firstpose7s[1], firsttz = firstpose7s[2],
+				rx = firstpose7s[3] * firstlerp,
+				ry = firstpose7s[4] * firstlerp,
+				rz = firstpose7s[5] * firstlerp,
+				rw = firstpose7s[6] * firstlerp,
+				dx = firsttx*rw + firstty*rz - firsttz*ry,
+				dy = -firsttx*rz + firstty*rw + firsttz*rx,
+				dz = firsttx*ry - firstty*rx + firsttz*rw,
+				dw = -firsttx*rx - firstty*ry - firsttz*rz,
+				scale, sx, sy, sz, sw;
+			for (blends = 1;blends < MAX_FRAMEBLENDS && frameblend[blends].lerp > 0;blends++)
 			{
-				matrix = model->data_poses + (frameblend[blends].subframe * model->num_bones + i) * 12;
-				for (k = 0;k < 12;k++)
-					m[k] += matrix[k] * frameblend[blends].lerp;
-				desiredscale[0] += frameblend[blends].lerp * VectorLength(matrix    );
-				desiredscale[1] += frameblend[blends].lerp * VectorLength(matrix + 4);
-				desiredscale[2] += frameblend[blends].lerp * VectorLength(matrix + 8);
+				const short * RESTRICT blendpose7s = model->data_poses7s + 7 * (frameblend[blends].subframe * model->num_bones + i);
+				float blendlerp = frameblend[blends].lerp,
+					blendtx = blendpose7s[0], blendty = blendpose7s[1], blendtz = blendpose7s[2],
+					qx = blendpose7s[3], qy = blendpose7s[4], qz = blendpose7s[5], qw = blendpose7s[6];
+				if(rx*qx + ry*qy + rz*qz + rw*qw < 0) blendlerp = -blendlerp;
+				qx *= blendlerp;
+				qy *= blendlerp;
+				qz *= blendlerp;
+				qw *= blendlerp;
+				rx += qx;
+				ry += qy;
+				rz += qz;
+				rw += qw;
+				dx += blendtx*qw + blendty*qz - blendtz*qy;
+				dy += -blendtx*qz + blendty*qw + blendtz*qx;
+				dz += blendtx*qy - blendty*qx + blendtz*qw;
+				dw += -blendtx*qx - blendty*qy - blendtz*qz;
 			}
-			VectorNormalize(m    );
-			VectorNormalize(m + 4);
-			VectorNormalize(m + 8);
-			VectorScale(m    , desiredscale[0], m    );
-			VectorScale(m + 4, desiredscale[1], m + 4);
-			VectorScale(m + 8, desiredscale[2], m + 8);
+			// generate a matrix from the dual-quaternion, implicitly normalizing it in the process
+			scale = 1.0f / (rx*rx + ry*ry + rz*rz + rw*rw);
+			sx = rx * scale;
+			sy = ry * scale;
+			sz = rz * scale;
+			sw = rw * scale;
+			m[0] = sw*rw + sx*rx - sy*ry - sz*rz;
+			m[1] = 2*(sx*ry - sw*rz);
+			m[2] = 2*(sx*rz + sw*ry);
+			m[3] = model->num_posescale*(dx*sw - dy*sz + dz*sy - dw*sx);
+			m[4] = 2*(sx*ry + sw*rz);
+			m[5] = sw*rw + sy*ry - sx*rx - sz*rz;
+			m[6] = 2*(sy*rz - sw*rx);
+			m[7] = model->num_posescale*(dx*sz + dy*sw - dz*sx - dw*sy);
+			m[8] = 2*(sx*rz - sw*ry);
+			m[9] = 2*(sy*rz + sw*rx);
+			m[10] = sw*rw + sz*rz - sx*rx - sy*ry;
+			m[11] = model->num_posescale*(dy*sx + dz*sw - dx*sy - dw*sz);
 			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;
 			if (model->data_bones[i].parent >= 0)
-				R_ConcatTransforms(bonepose[model->data_bones[i].parent], m, bonepose[i]);
+				R_ConcatTransforms(bonepose + model->data_bones[i].parent * 12, m, bonepose + i * 12);
 			else
-				for (k = 0;k < 12;k++)
-					bonepose[i][k] = m[k];
+				memcpy(bonepose + i * 12, m, sizeof(m));
 			// create a relative deformation matrix to describe displacement
 			// from the base mesh, which is used by the actual weighting
-			R_ConcatTransforms(bonepose[i], model->data_baseboneposeinverse + i * 12, boneposerelative[i]);
+			R_ConcatTransforms(bonepose + i * 12, model->data_baseboneposeinverse + i * 12, boneposerelative + i * 12);
 		}
 	}
+}
+
+static void Mod_Skeletal_AnimateVertices(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)
+{
 
-	// blend the vertex bone weights
-	// special case for the extremely common wf[0] == 1 because it saves 3 multiplies per array when compared to the other case (w[0] is always 1 if only one bone controls this vertex, artists only use multiple bones for certain special cases)
-	// special case for the first bone because it avoids the need to memset the arrays before filling
-	if (vertex3f)
+	if (!model->surfmesh.num_vertices)
+		return;
+
+	if (!model->num_bones)
 	{
-		const float *v = model->surfmesh.data_vertex3f;
-		const int *wi = model->surfmesh.data_vertexweightindex4i;
-		const float *wf = model->surfmesh.data_vertexweightinfluence4f;
-		memset(vertex3f, 0, sizeof(float[3]) * model->surfmesh.num_vertices);
-		for (i = 0;i < model->surfmesh.num_vertices;i++, v += 3, wi += 4, wf += 4, vertex3f += 3)
-		{
-			if (wf[0] == 1)
-			{
-				const float *m = boneposerelative[wi[0]];
-				vertex3f[0] = (v[0] * m[0] + v[1] * m[1] + v[2] * m[ 2] + m[ 3]);
-				vertex3f[1] = (v[0] * m[4] + v[1] * m[5] + v[2] * m[ 6] + m[ 7]);
-				vertex3f[2] = (v[0] * m[8] + v[1] * m[9] + v[2] * m[10] + m[11]);
-			}
-			else
-			{
-				const float *m = boneposerelative[wi[0]];
-				float f = wf[0];
-				vertex3f[0] = f * (v[0] * m[0] + v[1] * m[1] + v[2] * m[ 2] + m[ 3]);
-				vertex3f[1] = f * (v[0] * m[4] + v[1] * m[5] + v[2] * m[ 6] + m[ 7]);
-				vertex3f[2] = f * (v[0] * m[8] + v[1] * m[9] + v[2] * m[10] + m[11]);
-				for (k = 1;k < 4 && wf[k];k++)
-				{
-					const float *m = boneposerelative[wi[k]];
-					float f = wf[k];
-					vertex3f[0] += f * (v[0] * m[0] + v[1] * m[1] + v[2] * m[ 2] + m[ 3]);
-					vertex3f[1] += f * (v[0] * m[4] + v[1] * m[5] + v[2] * m[ 6] + m[ 7]);
-					vertex3f[2] += f * (v[0] * m[8] + v[1] * m[9] + v[2] * m[10] + m[11]);
-				}
-			}
-		}
+		if (vertex3f) memcpy(vertex3f, model->surfmesh.data_vertex3f, model->surfmesh.num_vertices*sizeof(float[3]));
+		if (normal3f) memcpy(normal3f, model->surfmesh.data_normal3f, model->surfmesh.num_vertices*sizeof(float[3]));
+		if (svector3f) memcpy(svector3f, model->surfmesh.data_svector3f, model->surfmesh.num_vertices*sizeof(float[3]));
+		if (tvector3f) memcpy(tvector3f, model->surfmesh.data_tvector3f, model->surfmesh.num_vertices*sizeof(float[3]));
+		return;
 	}
-	if (normal3f)
-	{
-		const float *n = model->surfmesh.data_normal3f;
-		const int *wi = model->surfmesh.data_vertexweightindex4i;
-		const float *wf = model->surfmesh.data_vertexweightinfluence4f;
-		memset(normal3f, 0, sizeof(float[3]) * model->surfmesh.num_vertices);
-		for (i = 0;i < model->surfmesh.num_vertices;i++, n += 3, wi += 4, wf += 4, normal3f += 3)
+
+#ifdef SSE_POSSIBLE
+	if(r_skeletal_use_sse_defined)
+		if(r_skeletal_use_sse.integer)
 		{
-			if (wf[0] == 1)
-			{
-				const float *m = boneposerelative[wi[0]];
-				normal3f[0] = (n[0] * m[0] + n[1] * m[1] + n[2] * m[ 2]);
-				normal3f[1] = (n[0] * m[4] + n[1] * m[5] + n[2] * m[ 6]);
-				normal3f[2] = (n[0] * m[8] + n[1] * m[9] + n[2] * m[10]);
-			}
-			else
-			{
-				const float *m = boneposerelative[wi[0]];
-				float f = wf[0];
-				normal3f[0] = f * (n[0] * m[0] + n[1] * m[1] + n[2] * m[ 2]);
-				normal3f[1] = f * (n[0] * m[4] + n[1] * m[5] + n[2] * m[ 6]);
-				normal3f[2] = f * (n[0] * m[8] + n[1] * m[9] + n[2] * m[10]);
-				for (k = 1;k < 4 && wf[k];k++)
-				{
-					const float *m = boneposerelative[wi[k]];
-					float f = wf[k];
-					normal3f[0] += f * (n[0] * m[0] + n[1] * m[1] + n[2] * m[ 2]);
-					normal3f[1] += f * (n[0] * m[4] + n[1] * m[5] + n[2] * m[ 6]);
-					normal3f[2] += f * (n[0] * m[8] + n[1] * m[9] + n[2] * m[10]);
-				}
-			}
+			Mod_Skeletal_AnimateVertices_SSE(model, frameblend, skeleton, vertex3f, normal3f, svector3f, tvector3f);
+			return;
 		}
-	}
-	if (svector3f)
+#endif
+	Mod_Skeletal_AnimateVertices_Generic(model, frameblend, skeleton, vertex3f, normal3f, svector3f, tvector3f);
+}
+
+void Mod_AliasInit (void)
+{
+	int i;
+	Cvar_RegisterVariable(&r_skeletal_debugbone);
+	Cvar_RegisterVariable(&r_skeletal_debugbonecomponent);
+	Cvar_RegisterVariable(&r_skeletal_debugbonevalue);
+	Cvar_RegisterVariable(&r_skeletal_debugtranslatex);
+	Cvar_RegisterVariable(&r_skeletal_debugtranslatey);
+	Cvar_RegisterVariable(&r_skeletal_debugtranslatez);
+	Cvar_RegisterVariable(&mod_alias_supporttagscale);
+	Cvar_RegisterVariable(&mod_alias_force_animated);
+	for (i = 0;i < 320;i++)
+		mod_md3_sin[i] = sin(i * M_PI * 2.0f / 256.0);
+#ifdef SSE_POSSIBLE
+	if(Sys_HaveSSE())
 	{
-		const float *sv = model->surfmesh.data_svector3f;
-		const int *wi = model->surfmesh.data_vertexweightindex4i;
-		const float *wf = model->surfmesh.data_vertexweightinfluence4f;
-		memset(svector3f, 0, sizeof(float[3]) * model->surfmesh.num_vertices);
-		for (i = 0;i < model->surfmesh.num_vertices;i++, sv += 3, wi += 4, wf += 4, svector3f += 3)
+		Con_Printf("Skeletal animation uses SSE code path\n");
+		r_skeletal_use_sse_defined = true;
+		Cvar_RegisterVariable(&r_skeletal_use_sse);
+	}
+	else
+		Con_Printf("Skeletal animation uses generic code path (SSE disabled or not detected)\n");
+#else
+	Con_Printf("Skeletal animation uses generic code path (SSE not compiled in)\n");
+#endif
+}
+
+static int Mod_Skeletal_AddBlend(dp_model_t *model, const blendweights_t *newweights)
+{
+	int i;
+	blendweights_t *weights;
+	if(!newweights->influence[1])
+		return newweights->index[0];
+	weights = model->surfmesh.data_blendweights;
+	for (i = 0;i < model->surfmesh.num_blends;i++, weights++)
+	{
+		if (!memcmp(weights, newweights, sizeof(blendweights_t)))
+			return model->num_bones + i;
+	}
+	model->surfmesh.num_blends++;
+	memcpy(weights, newweights, sizeof(blendweights_t));
+	return model->num_bones + i;
+}
+
+static int Mod_Skeletal_CompressBlend(dp_model_t *model, const int *newindex, const float *newinfluence)
+{
+	int i, total;
+	float scale;
+	blendweights_t newweights;
+	if(!newinfluence[1])
+		return newindex[0];
+	scale = 0;
+	for (i = 0;i < 4;i++)
+		scale += newinfluence[i];
+	scale = 255.0f / scale;
+	total = 0;
+	for (i = 0;i < 4;i++)
+	{
+		newweights.index[i] = newindex[i];
+		newweights.influence[i] = (unsigned char)(newinfluence[i] * scale);
+		total += newweights.influence[i];
+	}	
+	while (total > 255)
+	{
+		for (i = 0;i < 4;i++)
 		{
-			if (wf[0] == 1)
-			{
-				const float *m = boneposerelative[wi[0]];
-				svector3f[0] = (sv[0] * m[0] + sv[1] * m[1] + sv[2] * m[ 2]);
-				svector3f[1] = (sv[0] * m[4] + sv[1] * m[5] + sv[2] * m[ 6]);
-				svector3f[2] = (sv[0] * m[8] + sv[1] * m[9] + sv[2] * m[10]);
-			}
-			else
-			{
-				const float *m = boneposerelative[wi[0]];
-				float f = wf[0];
-				svector3f[0] = f * (sv[0] * m[0] + sv[1] * m[1] + sv[2] * m[ 2]);
-				svector3f[1] = f * (sv[0] * m[4] + sv[1] * m[5] + sv[2] * m[ 6]);
-				svector3f[2] = f * (sv[0] * m[8] + sv[1] * m[9] + sv[2] * m[10]);
-				for (k = 1;k < 4 && wf[k];k++)
-				{
-					const float *m = boneposerelative[wi[k]];
-					float f = wf[k];
-					svector3f[0] += f * (sv[0] * m[0] + sv[1] * m[1] + sv[2] * m[ 2]);
-					svector3f[1] += f * (sv[0] * m[4] + sv[1] * m[5] + sv[2] * m[ 6]);
-					svector3f[2] += f * (sv[0] * m[8] + sv[1] * m[9] + sv[2] * m[10]);
-				}
+			if(newweights.influence[i] > 0 && total > 255) 
+			{ 
+				newweights.influence[i]--;
+				total--; 
 			}
 		}
 	}
-	if (tvector3f)
+	while (total < 255)
 	{
-		const float *tv = model->surfmesh.data_tvector3f;
-		const int *wi = model->surfmesh.data_vertexweightindex4i;
-		const float *wf = model->surfmesh.data_vertexweightinfluence4f;
-		memset(tvector3f, 0, sizeof(float[3]) * model->surfmesh.num_vertices);
-		for (i = 0;i < model->surfmesh.num_vertices;i++, tv += 3, wi += 4, wf += 4, tvector3f += 3)
+		for (i = 0; i < 4;i++)
 		{
-			if (wf[0] == 1)
-			{
-				const float *m = boneposerelative[wi[0]];
-				tvector3f[0] = (tv[0] * m[0] + tv[1] * m[1] + tv[2] * m[ 2]);
-				tvector3f[1] = (tv[0] * m[4] + tv[1] * m[5] + tv[2] * m[ 6]);
-				tvector3f[2] = (tv[0] * m[8] + tv[1] * m[9] + tv[2] * m[10]);
-			}
-			else
-			{
-				const float *m = boneposerelative[wi[0]];
-				float f = wf[0];
-				tvector3f[0] = f * (tv[0] * m[0] + tv[1] * m[1] + tv[2] * m[ 2]);
-				tvector3f[1] = f * (tv[0] * m[4] + tv[1] * m[5] + tv[2] * m[ 6]);
-				tvector3f[2] = f * (tv[0] * m[8] + tv[1] * m[9] + tv[2] * m[10]);
-				for (k = 1;k < 4 && wf[k];k++)
-				{
-					const float *m = boneposerelative[wi[k]];
-					float f = wf[k];
-					tvector3f[0] += f * (tv[0] * m[0] + tv[1] * m[1] + tv[2] * m[ 2]);
-					tvector3f[1] += f * (tv[0] * m[4] + tv[1] * m[5] + tv[2] * m[ 6]);
-					tvector3f[2] += f * (tv[0] * m[8] + tv[1] * m[9] + tv[2] * m[10]);
-				}
+			if(newweights.influence[i] < 255 && total < 255) 
+			{ 
+				newweights.influence[i]++; 
+				total++; 
 			}
 		}
 	}
+	return Mod_Skeletal_AddBlend(model, &newweights);
 }
 
-void Mod_MD3_AnimateVertices(const dp_model_t *model, const frameblend_t *frameblend, const skeleton_t *skeleton, float *vertex3f, float *normal3f, float *svector3f, float *tvector3f)
+static void Mod_MD3_AnimateVertices(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 morph
 	int i, numblends, blendnum;
@@ -291,7 +313,7 @@ void Mod_MD3_AnimateVertices(const dp_model_t *model, const frameblend_t *frameb
 		}
 		// the yaw and pitch stored in md3 models are 8bit quantized angles
 		// (0-255), and as such a lookup table is very well suited to
-		// decoding them, and since cosine is equivilant to sine with an
+		// decoding them, and since cosine is equivalent to sine with an
 		// extra 45 degree rotation, this uses one lookup table for both
 		// sine and cosine with a +64 bias to get cosine.
 		if (normal3f)
@@ -339,8 +361,7 @@ void Mod_MD3_AnimateVertices(const dp_model_t *model, const frameblend_t *frameb
 		}
 	}
 }
-
-void Mod_MDL_AnimateVertices(const dp_model_t *model, const frameblend_t *frameblend, const skeleton_t *skeleton, float *vertex3f, float *normal3f, float *svector3f, float *tvector3f)
+static void Mod_MDL_AnimateVertices(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 morph
 	int i, numblends, blendnum;
@@ -440,13 +461,16 @@ void Mod_MDL_AnimateVertices(const dp_model_t *model, const frameblend_t *frameb
 int Mod_Alias_GetTagMatrix(const dp_model_t *model, const frameblend_t *frameblend, const skeleton_t *skeleton, int tagindex, matrix4x4_t *outmatrix)
 {
 	matrix4x4_t temp;
-	const float *boneframe;
-	const float *input;
+	matrix4x4_t parentbonematrix;
+	matrix4x4_t tempbonematrix;
+	matrix4x4_t bonematrix;
+	matrix4x4_t blendmatrix;
 	int blendindex;
 	int parenttagindex;
 	int k;
 	float lerp;
-	float tempbonematrix[12], bonematrix[12], blendmatrix[12];
+	const float *input;
+	float blendtag[12];
 	*outmatrix = identitymatrix;
 	if (skeleton && skeleton->relativetransforms)
 	{
@@ -463,42 +487,36 @@ int Mod_Alias_GetTagMatrix(const dp_model_t *model, const frameblend_t *frameble
 	{
 		if (tagindex < 0 || tagindex >= model->num_bones)
 			return 4;
-		for (k = 0;k < 12;k++)
-			blendmatrix[k] = 0;
+		Matrix4x4_Clear(&blendmatrix);
 		for (blendindex = 0;blendindex < MAX_FRAMEBLENDS && frameblend[blendindex].lerp > 0;blendindex++)
 		{
 			lerp = frameblend[blendindex].lerp;
-			boneframe = model->data_poses + frameblend[blendindex].subframe * model->num_bones * 12;
-			input = boneframe + tagindex * 12;
-			for (k = 0;k < 12;k++)
-				bonematrix[k] = input[k];
+			Matrix4x4_FromBonePose7s(&bonematrix, model->num_posescale, model->data_poses7s + 7 * (frameblend[blendindex].subframe * model->num_bones + tagindex));
 			parenttagindex = tagindex;
 			while ((parenttagindex = model->data_bones[parenttagindex].parent) >= 0)
 			{
-				for (k = 0;k < 12;k++)
-					tempbonematrix[k] = bonematrix[k];
-				input = boneframe + parenttagindex * 12;
-				R_ConcatTransforms(input, tempbonematrix, bonematrix);
+				Matrix4x4_FromBonePose7s(&parentbonematrix, model->num_posescale, model->data_poses7s + 7 * (frameblend[blendindex].subframe * model->num_bones + parenttagindex));
+				tempbonematrix = bonematrix;
+				Matrix4x4_Concat(&bonematrix, &parentbonematrix, &tempbonematrix);
 			}
-			for (k = 0;k < 12;k++)
-				blendmatrix[k] += bonematrix[k] * lerp;
+			Matrix4x4_Accumulate(&blendmatrix, &bonematrix, lerp);
 		}
-		Matrix4x4_FromArray12FloatD3D(outmatrix, blendmatrix);
+		*outmatrix = blendmatrix;
 	}
 	else if (model->num_tags)
 	{
 		if (tagindex < 0 || tagindex >= model->num_tags)
 			return 4;
 		for (k = 0;k < 12;k++)
-			blendmatrix[k] = 0;
+			blendtag[k] = 0;
 		for (blendindex = 0;blendindex < MAX_FRAMEBLENDS && frameblend[blendindex].lerp > 0;blendindex++)
 		{
 			lerp = frameblend[blendindex].lerp;
 			input = model->data_tags[frameblend[blendindex].subframe * model->num_tags + tagindex].matrixgl;
 			for (k = 0;k < 12;k++)
-				blendmatrix[k] += input[k] * lerp;
+				blendtag[k] += input[k] * lerp;
 		}
-		Matrix4x4_FromArray12FloatGL(outmatrix, blendmatrix);
+		Matrix4x4_FromArray12FloatGL(outmatrix, blendtag);
 	}
 
 	if(!mod_alias_supporttagscale.integer)
@@ -509,12 +527,13 @@ int Mod_Alias_GetTagMatrix(const dp_model_t *model, const frameblend_t *frameble
 
 int Mod_Alias_GetExtendedTagInfoForIndex(const dp_model_t *model, unsigned int skin, const frameblend_t *frameblend, const skeleton_t *skeleton, int tagindex, int *parentindex, const char **tagname, matrix4x4_t *tag_localmatrix)
 {
-	const float *boneframe;
-	const float *input;
 	int blendindex;
 	int k;
 	float lerp;
-	float blendmatrix[12];
+	matrix4x4_t bonematrix;
+	matrix4x4_t blendmatrix;
+	const float *input;
+	float blendtag[12];
 
 	if (skeleton && skeleton->relativetransforms)
 	{
@@ -527,35 +546,36 @@ int Mod_Alias_GetExtendedTagInfoForIndex(const dp_model_t *model, unsigned int s
 	}
 	else if (model->num_bones)
 	{
-		if(tagindex >= model->num_bones || tagindex < 0)
+		if (tagindex < 0 || tagindex >= model->num_bones)
 			return 1;
 		*parentindex = model->data_bones[tagindex].parent;
 		*tagname = model->data_bones[tagindex].name;
+		Matrix4x4_Clear(&blendmatrix);
 		for (blendindex = 0;blendindex < MAX_FRAMEBLENDS && frameblend[blendindex].lerp > 0;blendindex++)
 		{
 			lerp = frameblend[blendindex].lerp;
-			boneframe = model->data_poses + frameblend[blendindex].subframe * model->num_bones * 12;
-			input = boneframe + tagindex * 12;
-			for (k = 0;k < 12;k++)
-				blendmatrix[k] += input[k] * lerp;
+			Matrix4x4_FromBonePose7s(&bonematrix, model->num_posescale, model->data_poses7s + 7 * (frameblend[blendindex].subframe * model->num_bones + tagindex));
+			Matrix4x4_Accumulate(&blendmatrix, &bonematrix, lerp);
 		}
-		Matrix4x4_FromArray12FloatD3D(tag_localmatrix, blendmatrix);
+		*tag_localmatrix = blendmatrix;
 		return 0;
 	}
 	else if (model->num_tags)
 	{
-		if(tagindex >= model->num_tags || tagindex < 0)
+		if (tagindex < 0 || tagindex >= model->num_tags)
 			return 1;
 		*parentindex = -1;
 		*tagname = model->data_tags[tagindex].name;
+		for (k = 0;k < 12;k++)
+			blendtag[k] = 0;
 		for (blendindex = 0;blendindex < MAX_FRAMEBLENDS && frameblend[blendindex].lerp > 0;blendindex++)
 		{
 			lerp = frameblend[blendindex].lerp;
 			input = model->data_tags[frameblend[blendindex].subframe * model->num_tags + tagindex].matrixgl;
 			for (k = 0;k < 12;k++)
-				blendmatrix[k] += input[k] * lerp;
+				blendtag[k] += input[k] * lerp;
 		}
-		Matrix4x4_FromArray12FloatGL(tag_localmatrix, blendmatrix);
+		Matrix4x4_FromArray12FloatGL(tag_localmatrix, blendtag);
 		return 0;
 	}
 
@@ -580,70 +600,93 @@ int Mod_Alias_GetTagIndexForName(const dp_model_t *model, unsigned int skin, con
 
 static void Mod_BuildBaseBonePoses(void)
 {
-	int i, k;
-	double scale;
-	float *basebonepose;
-	float *in12f = loadmodel->data_poses;
-	float *out12f;
-	float *outinv12f = loadmodel->data_baseboneposeinverse;
+	int boneindex;
+	matrix4x4_t *basebonepose;
+	float *outinvmatrix = loadmodel->data_baseboneposeinverse;
+	matrix4x4_t bonematrix;
+	matrix4x4_t tempbonematrix;
 	if (!loadmodel->num_bones)
 		return;
-	out12f = basebonepose = (float *) Mem_Alloc(tempmempool, loadmodel->num_bones * sizeof(float[12]));
-	for (i = 0;i < loadmodel->num_bones;i++, in12f += 12, out12f += 12, outinv12f += 12)
+	basebonepose = (matrix4x4_t *)Mem_Alloc(tempmempool, loadmodel->num_bones * sizeof(matrix4x4_t));
+	for (boneindex = 0;boneindex < loadmodel->num_bones;boneindex++)
 	{
-		if (loadmodel->data_bones[i].parent >= 0)
-			R_ConcatTransforms(basebonepose + 12 * loadmodel->data_bones[i].parent, in12f, out12f);
-		else
-			for (k = 0;k < 12;k++)
-				out12f[k] = in12f[k];
-
-		// invert The Matrix
-
-		// we only support uniform scaling, so assume the first row is enough
-		// (note the lack of sqrt here, because we're trying to undo the scaling,
-		// this means multiplying by the inverse scale twice - squaring it, which
-		// makes the sqrt a waste of time)
-		scale = 1.0 / (out12f[ 0] * out12f[ 0] + out12f[ 1] * out12f[ 1] + out12f[ 2] * out12f[ 2]);
-
-		// invert the rotation by transposing and multiplying by the squared
-		// recipricol of the input matrix scale as described above
-		outinv12f[ 0] = (float)(out12f[ 0] * scale);
-		outinv12f[ 1] = (float)(out12f[ 4] * scale);
-		outinv12f[ 2] = (float)(out12f[ 8] * scale);
-		outinv12f[ 4] = (float)(out12f[ 1] * scale);
-		outinv12f[ 5] = (float)(out12f[ 5] * scale);
-		outinv12f[ 6] = (float)(out12f[ 9] * scale);
-		outinv12f[ 8] = (float)(out12f[ 2] * scale);
-		outinv12f[ 9] = (float)(out12f[ 6] * scale);
-		outinv12f[10] = (float)(out12f[10] * scale);
-
-		// invert the translate
-		outinv12f[ 3] = -(out12f[ 3] * outinv12f[ 0] + out12f[ 7] * outinv12f[ 1] + out12f[11] * outinv12f[ 2]);
-		outinv12f[ 7] = -(out12f[ 3] * outinv12f[ 4] + out12f[ 7] * outinv12f[ 5] + out12f[11] * outinv12f[ 6]);
-		outinv12f[11] = -(out12f[ 3] * outinv12f[ 8] + out12f[ 7] * outinv12f[ 9] + out12f[11] * outinv12f[10]);
+		Matrix4x4_FromBonePose7s(&bonematrix, loadmodel->num_posescale, loadmodel->data_poses7s + 7 * boneindex);
+		if (loadmodel->data_bones[boneindex].parent >= 0)
+		{
+			tempbonematrix = bonematrix;
+			Matrix4x4_Concat(&bonematrix, basebonepose + loadmodel->data_bones[boneindex].parent, &tempbonematrix);
+		}
+		basebonepose[boneindex] = bonematrix;
+		Matrix4x4_Invert_Simple(&tempbonematrix, basebonepose + boneindex);
+		Matrix4x4_ToArray12FloatD3D(&tempbonematrix, outinvmatrix + 12*boneindex);
 	}
 	Mem_Free(basebonepose);
 }
 
-static void Mod_Alias_CalculateBoundingBox(void)
+static qboolean Mod_Alias_CalculateBoundingBox(void)
 {
 	int vnum;
 	qboolean firstvertex = true;
 	float dist, yawradius, radius;
 	float *v;
-	float *vertex3f;
-	frameblend_t frameblend[MAX_FRAMEBLENDS];
-	memset(frameblend, 0, sizeof(frameblend));
-	frameblend[0].lerp = 1;
-	vertex3f = (float *) Mem_Alloc(loadmodel->mempool, loadmodel->surfmesh.num_vertices * sizeof(float[3]));
+	qboolean isanimated = false;
 	VectorClear(loadmodel->normalmins);
 	VectorClear(loadmodel->normalmaxs);
 	yawradius = 0;
 	radius = 0;
-	for (frameblend[0].subframe = 0;frameblend[0].subframe < loadmodel->num_poses;frameblend[0].subframe++)
+	if (loadmodel->AnimateVertices)
+	{
+		float *vertex3f, *refvertex3f;
+		frameblend_t frameblend[MAX_FRAMEBLENDS];
+		memset(frameblend, 0, sizeof(frameblend));
+		frameblend[0].lerp = 1;
+		vertex3f = (float *) Mem_Alloc(loadmodel->mempool, loadmodel->surfmesh.num_vertices * sizeof(float[3]) * 2);
+		refvertex3f = NULL;
+		for (frameblend[0].subframe = 0;frameblend[0].subframe < loadmodel->num_poses;frameblend[0].subframe++)
+		{
+			loadmodel->AnimateVertices(loadmodel, frameblend, NULL, vertex3f, NULL, NULL, NULL);
+			if (!refvertex3f)
+			{
+				// make a copy of the first frame for comparing all others
+				refvertex3f = vertex3f + loadmodel->surfmesh.num_vertices * 3;
+				memcpy(refvertex3f, vertex3f, loadmodel->surfmesh.num_vertices * sizeof(float[3]));
+			}
+			else
+			{
+				if (!isanimated && memcmp(refvertex3f, vertex3f, loadmodel->surfmesh.num_vertices * sizeof(float[3])))
+					isanimated = true;
+			}
+			for (vnum = 0, v = vertex3f;vnum < loadmodel->surfmesh.num_vertices;vnum++, v += 3)
+			{
+				if (firstvertex)
+				{
+					firstvertex = false;
+					VectorCopy(v, loadmodel->normalmins);
+					VectorCopy(v, loadmodel->normalmaxs);
+				}
+				else
+				{
+					if (loadmodel->normalmins[0] > v[0]) loadmodel->normalmins[0] = v[0];
+					if (loadmodel->normalmins[1] > v[1]) loadmodel->normalmins[1] = v[1];
+					if (loadmodel->normalmins[2] > v[2]) loadmodel->normalmins[2] = v[2];
+					if (loadmodel->normalmaxs[0] < v[0]) loadmodel->normalmaxs[0] = v[0];
+					if (loadmodel->normalmaxs[1] < v[1]) loadmodel->normalmaxs[1] = v[1];
+					if (loadmodel->normalmaxs[2] < v[2]) loadmodel->normalmaxs[2] = v[2];
+				}
+				dist = v[0] * v[0] + v[1] * v[1];
+				if (yawradius < dist)
+					yawradius = dist;
+				dist += v[2] * v[2];
+				if (radius < dist)
+					radius = dist;
+			}
+		}
+		if (vertex3f)
+			Mem_Free(vertex3f);
+	}
+	else
 	{
-		loadmodel->AnimateVertices(loadmodel, frameblend, NULL, vertex3f, NULL, NULL, NULL);
-		for (vnum = 0, v = vertex3f;vnum < loadmodel->surfmesh.num_vertices;vnum++, v += 3)
+		for (vnum = 0, v = loadmodel->surfmesh.data_vertex3f;vnum < loadmodel->surfmesh.num_vertices;vnum++, v += 3)
 		{
 			if (firstvertex)
 			{
@@ -668,8 +711,6 @@ static void Mod_Alias_CalculateBoundingBox(void)
 				radius = dist;
 		}
 	}
-	if (vertex3f)
-		Mem_Free(vertex3f);
 	radius = sqrt(radius);
 	yawradius = sqrt(yawradius);
 	loadmodel->yawmins[0] = loadmodel->yawmins[1] = -yawradius;
@@ -680,6 +721,7 @@ static void Mod_Alias_CalculateBoundingBox(void)
 	loadmodel->rotatedmaxs[0] = loadmodel->rotatedmaxs[1] = loadmodel->rotatedmaxs[2] = radius;
 	loadmodel->radius = radius;
 	loadmodel->radius2 = radius * radius;
+	return isanimated;
 }
 
 static void Mod_Alias_MorphMesh_CompileFrames(void)
@@ -710,44 +752,53 @@ static void Mod_Alias_MorphMesh_CompileFrames(void)
 	}
 }
 
-static void Mod_MDLMD2MD3_TraceLine(dp_model_t *model, const frameblend_t *frameblend, const skeleton_t *skeleton, trace_t *trace, const vec3_t start, const vec3_t end, int hitsupercontentsmask)
+static void Mod_MDLMD2MD3_TraceLine(dp_model_t *model, const frameblend_t *frameblend, const skeleton_t *skeleton, trace_t *trace, const vec3_t start, const vec3_t end, int hitsupercontentsmask, int skipsupercontentsmask, int skipmaterialflagsmask)
 {
 	int i;
 	float segmentmins[3], segmentmaxs[3];
 	msurface_t *surface;
-	static int maxvertices = 0;
-	static float *vertex3f = NULL;
+	float vertex3fbuf[1024 * 3];
+	float *vertex3f = vertex3fbuf;
+	float *freevertex3f = NULL;
+	// for static cases we can just call CollisionBIH which is much faster
+	if ((frameblend == NULL || (frameblend[0].subframe == 0 && frameblend[1].lerp == 0)) && (skeleton == NULL || skeleton->relativetransforms == NULL))
+	{
+		Mod_CollisionBIH_TraceLine(model, frameblend, skeleton, trace, start, end, hitsupercontentsmask, skipsupercontentsmask, skipmaterialflagsmask);
+		return;
+	}
 	memset(trace, 0, sizeof(*trace));
 	trace->fraction = 1;
-	trace->realfraction = 1;
 	trace->hitsupercontentsmask = hitsupercontentsmask;
-	if (maxvertices < model->surfmesh.num_vertices)
-	{
-		if (vertex3f)
-			Z_Free(vertex3f);
-		maxvertices = (model->surfmesh.num_vertices + 255) & ~255;
-		vertex3f = (float *)Z_Malloc(maxvertices * sizeof(float[3]));
-	}
+	trace->skipsupercontentsmask = skipsupercontentsmask;
+	trace->skipmaterialflagsmask = skipmaterialflagsmask;
 	segmentmins[0] = min(start[0], end[0]) - 1;
 	segmentmins[1] = min(start[1], end[1]) - 1;
 	segmentmins[2] = min(start[2], end[2]) - 1;
 	segmentmaxs[0] = max(start[0], end[0]) + 1;
 	segmentmaxs[1] = max(start[1], end[1]) + 1;
 	segmentmaxs[2] = max(start[2], end[2]) + 1;
-	model->AnimateVertices(model, frameblend, skeleton, vertex3f, NULL, NULL, NULL);
+	if (frameblend == NULL || frameblend[0].subframe != 0 || frameblend[0].lerp != 0 || skeleton != NULL)
+	{
+		if (model->surfmesh.num_vertices > 1024)
+			vertex3f = freevertex3f = (float *)Mem_Alloc(tempmempool, model->surfmesh.num_vertices * sizeof(float[3]));
+		model->AnimateVertices(model, frameblend, skeleton, vertex3f, NULL, NULL, NULL);
+	}
+	else
+		vertex3f = model->surfmesh.data_vertex3f;
 	for (i = 0, surface = model->data_surfaces;i < model->num_surfaces;i++, surface++)
 		Collision_TraceLineTriangleMeshFloat(trace, start, end, model->surfmesh.num_triangles, model->surfmesh.data_element3i, vertex3f, 0, NULL, SUPERCONTENTS_SOLID | (surface->texture->basematerialflags & MATERIALFLAGMASK_TRANSLUCENT ? 0 : SUPERCONTENTS_OPAQUE), 0, surface->texture, segmentmins, segmentmaxs);
+	if (freevertex3f)
+		Mem_Free(freevertex3f);
 }
 
-static int maxvertices = 0;
-static float *vertex3f = NULL;
-
-static void Mod_MDLMD2MD3_TraceBox(dp_model_t *model, const frameblend_t *frameblend, const skeleton_t *skeleton, trace_t *trace, const vec3_t start, const vec3_t boxmins, const vec3_t boxmaxs, const vec3_t end, int hitsupercontentsmask)
+static void Mod_MDLMD2MD3_TraceBox(dp_model_t *model, const frameblend_t *frameblend, const skeleton_t *skeleton, trace_t *trace, const vec3_t start, const vec3_t boxmins, const vec3_t boxmaxs, const vec3_t end, int hitsupercontentsmask, int skipsupercontentsmask, int skipmaterialflagsmask)
 {
 	int i;
 	vec3_t shiftstart, shiftend;
 	float segmentmins[3], segmentmaxs[3];
 	msurface_t *surface;
+	float vertex3fbuf[1024*3];
+	float *vertex3f = vertex3fbuf;
 	colboxbrushf_t thisbrush_start, thisbrush_end;
 	vec3_t boxstartmins, boxstartmaxs, boxendmins, boxendmaxs;
 
@@ -755,23 +806,26 @@ static void Mod_MDLMD2MD3_TraceBox(dp_model_t *model, const frameblend_t *frameb
 	{
 		VectorAdd(start, boxmins, shiftstart);
 		VectorAdd(end, boxmins, shiftend);
-		Mod_MDLMD2MD3_TraceLine(model, frameblend, skeleton, trace, start, end, hitsupercontentsmask);
+		Mod_MDLMD2MD3_TraceLine(model, frameblend, skeleton, trace, shiftstart, shiftend, hitsupercontentsmask, skipsupercontentsmask, skipmaterialflagsmask);
 		VectorSubtract(trace->endpos, boxmins, trace->endpos);
 		return;
 	}
 
+	// for static cases we can just call CollisionBIH which is much faster
+	if ((frameblend == NULL || (frameblend[0].subframe == 0 && frameblend[1].lerp == 0)) && (skeleton == NULL || skeleton->relativetransforms == NULL))
+	{
+		Mod_CollisionBIH_TraceBox(model, frameblend, skeleton, trace, start, boxmins, boxmaxs, end, hitsupercontentsmask, skipsupercontentsmask, skipmaterialflagsmask);
+		return;
+	}
+
 	// box trace, performed as brush trace
 	memset(trace, 0, sizeof(*trace));
 	trace->fraction = 1;
-	trace->realfraction = 1;
 	trace->hitsupercontentsmask = hitsupercontentsmask;
-	if (maxvertices < model->surfmesh.num_vertices)
-	{
-		if (vertex3f)
-			Z_Free(vertex3f);
-		maxvertices = (model->surfmesh.num_vertices + 255) & ~255;
-		vertex3f = (float *)Z_Malloc(maxvertices * sizeof(float[3]));
-	}
+	trace->skipsupercontentsmask = skipsupercontentsmask;
+	trace->skipmaterialflagsmask = skipmaterialflagsmask;
+	if (model->surfmesh.num_vertices > 1024)
+		vertex3f = (float *)Mem_Alloc(tempmempool, model->surfmesh.num_vertices * sizeof(float[3]));
 	segmentmins[0] = min(start[0], end[0]) + boxmins[0] - 1;
 	segmentmins[1] = min(start[1], end[1]) + boxmins[1] - 1;
 	segmentmins[2] = min(start[2], end[2]) + boxmins[2] - 1;
@@ -784,16 +838,11 @@ static void Mod_MDLMD2MD3_TraceBox(dp_model_t *model, const frameblend_t *frameb
 	VectorAdd(end, boxmaxs, boxendmaxs);
 	Collision_BrushForBox(&thisbrush_start, boxstartmins, boxstartmaxs, 0, 0, NULL);
 	Collision_BrushForBox(&thisbrush_end, boxendmins, boxendmaxs, 0, 0, NULL);
-	if (maxvertices < model->surfmesh.num_vertices)
-	{
-		if (vertex3f)
-			Z_Free(vertex3f);
-		maxvertices = (model->surfmesh.num_vertices + 255) & ~255;
-		vertex3f = (float *)Z_Malloc(maxvertices * sizeof(float[3]));
-	}
 	model->AnimateVertices(model, frameblend, skeleton, vertex3f, NULL, NULL, NULL);
 	for (i = 0, surface = model->data_surfaces;i < model->num_surfaces;i++, surface++)
 		Collision_TraceBrushTriangleMeshFloat(trace, &thisbrush_start.brush, &thisbrush_end.brush, model->surfmesh.num_triangles, model->surfmesh.data_element3i, vertex3f, 0, NULL, SUPERCONTENTS_SOLID | (surface->texture->basematerialflags & MATERIALFLAGMASK_TRANSLUCENT ? 0 : SUPERCONTENTS_OPAQUE), 0, surface->texture, segmentmins, segmentmaxs);
+	if (vertex3f != vertex3fbuf)
+		Mem_Free(vertex3f);
 }
 
 static void Mod_ConvertAliasVerts (int inverts, trivertx_t *v, trivertx_t *out, int *vertremap)
@@ -865,7 +914,6 @@ static void Mod_MDL_LoadFrames (unsigned char* datapointer, int inverts, int *ve
 		// read frames
 		for (i = 0;i < groupframes;i++)
 		{
-			pinframe = (daliasframe_t *)datapointer;
 			datapointer += sizeof(daliasframe_t);
 			Mod_ConvertAliasVerts(inverts, (trivertx_t *)datapointer, loadmodel->surfmesh.data_morphmdlvertex + pose * loadmodel->surfmesh.num_vertices, vertremap);
 			datapointer += sizeof(trivertx_t) * inverts;
@@ -874,40 +922,13 @@ static void Mod_MDL_LoadFrames (unsigned char* datapointer, int inverts, int *ve
 	}
 }
 
-static void Mod_BuildAliasSkinFromSkinFrame(texture_t *texture, skinframe_t *skinframe)
-{
-	if (cls.state == ca_dedicated)
-		return;
-	// hack
-	if (!skinframe)
-		skinframe = R_SkinFrame_LoadMissing();
-	memset(texture, 0, sizeof(*texture));
-	texture->currentframe = texture;
-	//texture->animated = false;
-	texture->numskinframes = 1;
-	texture->skinframerate = 1;
-	texture->skinframes[0] = skinframe;
-	texture->currentskinframe = skinframe;
-	//texture->backgroundnumskinframes = 0;
-	//texture->customblendfunc[0] = 0;
-	//texture->customblendfunc[1] = 0;
-	//texture->surfaceflags = 0;
-	//texture->supercontents = 0;
-	//texture->surfaceparms = 0;
-	//texture->textureflags = 0;
-
-	texture->basematerialflags = MATERIALFLAG_WALL;
-	if (texture->currentskinframe->hasalpha)
-		texture->basematerialflags |= MATERIALFLAG_ALPHA | MATERIALFLAG_BLENDED | MATERIALFLAG_NOSHADOW;
-	texture->currentmaterialflags = texture->basematerialflags;
-	texture->specularscalemod = 1;
-	texture->specularpowermod = 1;
-}
-
-static void Mod_BuildAliasSkinsFromSkinFiles(texture_t *skin, skinfile_t *skinfile, const char *meshname, const char *shadername)
+void Mod_BuildAliasSkinsFromSkinFiles(texture_t *skin, skinfile_t *skinfile, const char *meshname, const char *shadername)
 {
 	int i;
+	char stripbuf[MAX_QPATH];
 	skinfileitem_t *skinfileitem;
+	if(developer_extra.integer)
+		Con_DPrintf("Looking up texture for %s (default: %s)\n", meshname, shadername);
 	if (skinfile)
 	{
 		// the skin += loadmodel->num_surfaces part of this is because data_textures on alias models is arranged as [numskins][numsurfaces]
@@ -920,20 +941,30 @@ static void Mod_BuildAliasSkinsFromSkinFiles(texture_t *skin, skinfile_t *skinfi
 				// leave the skin unitialized (nodraw) if the replacement is "common/nodraw" or "textures/common/nodraw"
 				if (!strcmp(skinfileitem->name, meshname))
 				{
-					Mod_LoadTextureFromQ3Shader(skin, skinfileitem->replacement, true, true, (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_ALPHA | TEXF_PICMIP | TEXF_COMPRESS);
+					Image_StripImageExtension(skinfileitem->replacement, stripbuf, sizeof(stripbuf));
+					if(developer_extra.integer)
+						Con_DPrintf("--> got %s from skin file\n", stripbuf);
+					Mod_LoadTextureFromQ3Shader(loadmodel->mempool, loadmodel->name, skin, stripbuf, true, true, (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_ALPHA | TEXF_PICMIP | TEXF_COMPRESS, MATERIALFLAG_WALL);
 					break;
 				}
 			}
 			if (!skinfileitem)
 			{
 				// don't render unmentioned meshes
-				Mod_BuildAliasSkinFromSkinFrame(skin, NULL);
+				Mod_LoadCustomMaterial(loadmodel->mempool, skin, meshname, SUPERCONTENTS_SOLID, MATERIALFLAG_WALL, R_SkinFrame_LoadMissing());
+				if(developer_extra.integer)
+					Con_DPrintf("--> skipping\n");
 				skin->basematerialflags = skin->currentmaterialflags = MATERIALFLAG_NOSHADOW | MATERIALFLAG_NODRAW;
 			}
 		}
 	}
 	else
-		Mod_LoadTextureFromQ3Shader(skin, shadername, true, true, (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_ALPHA | TEXF_PICMIP | TEXF_COMPRESS);
+	{
+		if(developer_extra.integer)
+			Con_DPrintf("--> using default\n");
+		Image_StripImageExtension(shadername, stripbuf, sizeof(stripbuf));
+		Mod_LoadTextureFromQ3Shader(loadmodel->mempool, loadmodel->name, skin, stripbuf, true, true, (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_ALPHA | TEXF_PICMIP | TEXF_COMPRESS, MATERIALFLAG_WALL);
+	}
 }
 
 #define BOUNDI(VALUE,MIN,MAX) if (VALUE < MIN || VALUE >= MAX) Host_Error("model %s has an invalid ##VALUE (%d exceeds %d - %d)", loadmodel->name, VALUE, MIN, MAX);
@@ -973,7 +1004,6 @@ void Mod_IDP0_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->modeldatatypestring = "MDL";
 
 	loadmodel->type = mod_alias;
-	loadmodel->AnimateVertices = Mod_MDL_AnimateVertices;
 	loadmodel->DrawSky = NULL;
 	loadmodel->DrawAddWaterPlanes = NULL;
 	loadmodel->Draw = R_Q1BSP_Draw;
@@ -987,7 +1017,9 @@ void Mod_IDP0_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->DrawLight = R_Q1BSP_DrawLight;
 	loadmodel->TraceBox = Mod_MDLMD2MD3_TraceBox;
 	loadmodel->TraceLine = Mod_MDLMD2MD3_TraceLine;
+	// FIXME add TraceBrush!
 	loadmodel->PointSuperContents = NULL;
+	loadmodel->AnimateVertices = Mod_MDL_AnimateVertices;
 
 	loadmodel->num_surfaces = 1;
 	loadmodel->nummodelsurfaces = loadmodel->num_surfaces;
@@ -1009,7 +1041,7 @@ void Mod_IDP0_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->numframes = LittleLong(pinmodel->numframes);
 	BOUNDI(loadmodel->numframes,0,65536);
 	loadmodel->synctype = (synctype_t)LittleLong (pinmodel->synctype);
-	BOUNDI(loadmodel->synctype,0,2);
+	BOUNDI((int)loadmodel->synctype,0,2);
 	// convert model flags to EF flags (MF_ROCKET becomes EF_ROCKET, etc)
 	i = LittleLong (pinmodel->flags);
 	loadmodel->effects = ((i & 255) << 24) | (i & 0x00FFFF00);
@@ -1085,8 +1117,8 @@ void Mod_IDP0_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	for (i = 0;i < numverts;i++)
 	{
 		vertonseam[i] = LittleLong(pinstverts[i].onseam);
-		vertst[i*2+0] = (LittleLong(pinstverts[i].s) + 0.5) * scales;
-		vertst[i*2+1] = (LittleLong(pinstverts[i].t) + 0.5) * scalet;
+		vertst[i*2+0] = LittleLong(pinstverts[i].s) * scales;
+		vertst[i*2+1] = LittleLong(pinstverts[i].t) * scalet;
 		vertst[(i+numverts)*2+0] = vertst[i*2+0] + 0.5;
 		vertst[(i+numverts)*2+1] = vertst[i*2+1];
 	}
@@ -1148,10 +1180,14 @@ void Mod_IDP0_Load(dp_model_t *mod, void *buffer, void *bufferend)
 // load the frames
 	loadmodel->animscenes = (animscene_t *)Mem_Alloc(loadmodel->mempool, sizeof(animscene_t) * loadmodel->numframes);
 	loadmodel->surfmesh.data_morphmdlvertex = (trivertx_t *)Mem_Alloc(loadmodel->mempool, sizeof(trivertx_t) * loadmodel->surfmesh.num_morphframes * loadmodel->surfmesh.num_vertices);
-	loadmodel->surfmesh.data_neighbor3i = (int *)Mem_Alloc(loadmodel->mempool, loadmodel->surfmesh.num_triangles * sizeof(int[3]));
+	if (r_enableshadowvolumes.integer)
+	{
+		loadmodel->surfmesh.data_neighbor3i = (int *)Mem_Alloc(loadmodel->mempool, loadmodel->surfmesh.num_triangles * sizeof(int[3]));
+	}
 	Mod_MDL_LoadFrames (startframes, numverts, vertremap);
-	Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
-	Mod_Alias_CalculateBoundingBox();
+	if (loadmodel->surfmesh.data_neighbor3i)
+		Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
+	loadmodel->surfmesh.isanimated = Mod_Alias_CalculateBoundingBox();
 	Mod_Alias_MorphMesh_CompileFrames();
 
 	Mem_Free(vertst);
@@ -1223,8 +1259,8 @@ void Mod_IDP0_Load(dp_model_t *mod, void *buffer, void *bufferend)
 					dpsnprintf (name, sizeof(name), "%s_%i_%i", loadmodel->name, i, j);
 				else
 					dpsnprintf (name, sizeof(name), "%s_%i", loadmodel->name, i);
-				if (!Mod_LoadTextureFromQ3Shader(loadmodel->data_textures + totalskins * loadmodel->num_surfaces, name, false, true, (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_ALPHA | TEXF_PICMIP | TEXF_COMPRESS))
-					Mod_BuildAliasSkinFromSkinFrame(loadmodel->data_textures + totalskins * loadmodel->num_surfaces, R_SkinFrame_LoadInternalQuake(name, (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_PICMIP, true, r_fullbrights.integer, (unsigned char *)datapointer, skinwidth, skinheight));
+				if (!Mod_LoadTextureFromQ3Shader(loadmodel->mempool, loadmodel->name, loadmodel->data_textures + totalskins * loadmodel->num_surfaces, name, false, false, (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_ALPHA | TEXF_PICMIP | TEXF_COMPRESS, MATERIALFLAG_WALL))
+					Mod_LoadCustomMaterial(loadmodel->mempool, loadmodel->data_textures + totalskins * loadmodel->num_surfaces, name, SUPERCONTENTS_SOLID, MATERIALFLAG_WALL, R_SkinFrame_LoadInternalQuake(name, (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_PICMIP, true, r_fullbrights.integer, (unsigned char *)datapointer, skinwidth, skinheight));
 				datapointer += skinwidth * skinheight;
 				totalskins++;
 			}
@@ -1232,8 +1268,12 @@ void Mod_IDP0_Load(dp_model_t *mod, void *buffer, void *bufferend)
 		// check for skins that don't exist in the model, but do exist as external images
 		// (this was added because yummyluv kept pestering me about support for it)
 		// TODO: support shaders here?
-		while ((tempskinframe = R_SkinFrame_LoadExternal(va("%s_%i", loadmodel->name, loadmodel->numskins), (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_ALPHA | TEXF_PICMIP | TEXF_COMPRESS, false)))
+		for (;;)
 		{
+			dpsnprintf(name, sizeof(name), "%s_%i", loadmodel->name, loadmodel->numskins);
+			tempskinframe = R_SkinFrame_LoadExternal(name, (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_ALPHA | TEXF_PICMIP | TEXF_COMPRESS, false, false);
+			if (!tempskinframe)
+				break;
 			// expand the arrays to make room
 			tempskinscenes = loadmodel->skinscenes;
 			loadmodel->skinscenes = (animscene_t *)Mem_Alloc(loadmodel->mempool, (loadmodel->numskins + 1) * sizeof(animscene_t));
@@ -1246,7 +1286,7 @@ void Mod_IDP0_Load(dp_model_t *mod, void *buffer, void *bufferend)
 			Mem_Free(tempaliasskins);
 
 			// store the info about the new skin
-			Mod_BuildAliasSkinFromSkinFrame(loadmodel->data_textures + totalskins * loadmodel->num_surfaces, tempskinframe);
+			Mod_LoadCustomMaterial(loadmodel->mempool, loadmodel->data_textures + totalskins * loadmodel->num_surfaces, name, SUPERCONTENTS_SOLID, MATERIALFLAG_WALL, tempskinframe);
 			strlcpy(loadmodel->skinscenes[loadmodel->numskins].name, name, sizeof(loadmodel->skinscenes[loadmodel->numskins].name));
 			loadmodel->skinscenes[loadmodel->numskins].firstframe = totalskins;
 			loadmodel->skinscenes[loadmodel->numskins].framecount = 1;
@@ -1254,6 +1294,7 @@ void Mod_IDP0_Load(dp_model_t *mod, void *buffer, void *bufferend)
 			loadmodel->skinscenes[loadmodel->numskins].loop = true;
 
 			//increase skin counts
+			loadmodel->num_textures++;
 			loadmodel->numskins++;
 			totalskins++;
 
@@ -1271,7 +1312,28 @@ void Mod_IDP0_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	surface->num_firstvertex = 0;
 	surface->num_vertices = loadmodel->surfmesh.num_vertices;
 
-	loadmodel->surfmesh.isanimated = loadmodel->numframes > 1 || loadmodel->animscenes[0].framecount > 1;
+	if(mod_alias_force_animated.string[0])
+		loadmodel->surfmesh.isanimated = mod_alias_force_animated.integer != 0;
+
+	// Always make a BIH for the first frame, we can use it where possible.
+	Mod_MakeCollisionBIH(loadmodel, true, &loadmodel->collision_bih);
+	if (!loadmodel->surfmesh.isanimated)
+	{
+		loadmodel->TraceBox = Mod_CollisionBIH_TraceBox;
+		loadmodel->TraceBrush = Mod_CollisionBIH_TraceBrush;
+		loadmodel->TraceLine = Mod_CollisionBIH_TraceLine;
+		loadmodel->TracePoint = Mod_CollisionBIH_TracePoint_Mesh;
+		loadmodel->PointSuperContents = Mod_CollisionBIH_PointSuperContents_Mesh;
+	}
+
+	// because shaders can do somewhat unexpected things, check for unusual features now
+	for (i = 0;i < loadmodel->num_textures;i++)
+	{
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_SKY))
+			mod->DrawSky = R_Q1BSP_DrawSky;
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_WATERSHADER | MATERIALFLAG_REFRACTION | MATERIALFLAG_REFLECTION | MATERIALFLAG_CAMERA))
+			mod->DrawAddWaterPlanes = R_Q1BSP_DrawAddWaterPlanes;
+	}
 }
 
 void Mod_IDP2_Load(dp_model_t *mod, void *buffer, void *bufferend)
@@ -1306,7 +1368,6 @@ void Mod_IDP2_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->modeldatatypestring = "MD2";
 
 	loadmodel->type = mod_alias;
-	loadmodel->AnimateVertices = Mod_MDL_AnimateVertices;
 	loadmodel->DrawSky = NULL;
 	loadmodel->DrawAddWaterPlanes = NULL;
 	loadmodel->Draw = R_Q1BSP_Draw;
@@ -1321,6 +1382,7 @@ void Mod_IDP2_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->TraceBox = Mod_MDLMD2MD3_TraceBox;
 	loadmodel->TraceLine = Mod_MDLMD2MD3_TraceLine;
 	loadmodel->PointSuperContents = NULL;
+	loadmodel->AnimateVertices = Mod_MDL_AnimateVertices;
 
 	if (LittleLong(pinmodel->num_tris) < 1 || LittleLong(pinmodel->num_tris) > 65536)
 		Host_Error ("%s has invalid number of triangles: %i", loadmodel->name, LittleLong(pinmodel->num_tris));
@@ -1357,14 +1419,17 @@ void Mod_IDP2_Load(dp_model_t *mod, void *buffer, void *bufferend)
 
 	loadmodel->num_surfaces = 1;
 	loadmodel->nummodelsurfaces = loadmodel->num_surfaces;
-	data = (unsigned char *)Mem_Alloc(loadmodel->mempool, loadmodel->num_surfaces * sizeof(msurface_t) + loadmodel->num_surfaces * sizeof(int) + loadmodel->numframes * sizeof(animscene_t) + loadmodel->numframes * sizeof(float[6]) + loadmodel->surfmesh.num_triangles * sizeof(int[3]) + loadmodel->surfmesh.num_triangles * sizeof(int[3]));
+	data = (unsigned char *)Mem_Alloc(loadmodel->mempool, loadmodel->num_surfaces * sizeof(msurface_t) + loadmodel->num_surfaces * sizeof(int) + loadmodel->numframes * sizeof(animscene_t) + loadmodel->numframes * sizeof(float[6]) + loadmodel->surfmesh.num_triangles * sizeof(int[3]) + (r_enableshadowvolumes.integer ? loadmodel->surfmesh.num_triangles * sizeof(int[3]) : 0));
 	loadmodel->data_surfaces = (msurface_t *)data;data += loadmodel->num_surfaces * sizeof(msurface_t);
 	loadmodel->sortedmodelsurfaces = (int *)data;data += loadmodel->num_surfaces * sizeof(int);
 	loadmodel->sortedmodelsurfaces[0] = 0;
 	loadmodel->animscenes = (animscene_t *)data;data += loadmodel->numframes * sizeof(animscene_t);
 	loadmodel->surfmesh.data_morphmd2framesize6f = (float *)data;data += loadmodel->numframes * sizeof(float[6]);
 	loadmodel->surfmesh.data_element3i = (int *)data;data += loadmodel->surfmesh.num_triangles * sizeof(int[3]);
-	loadmodel->surfmesh.data_neighbor3i = (int *)data;data += loadmodel->surfmesh.num_triangles * sizeof(int[3]);
+	if (r_enableshadowvolumes.integer)
+	{
+		loadmodel->surfmesh.data_neighbor3i = (int *)data;data += loadmodel->surfmesh.num_triangles * sizeof(int[3]);
+	}
 
 	loadmodel->synctype = ST_RAND;
 
@@ -1386,7 +1451,7 @@ void Mod_IDP2_Load(dp_model_t *mod, void *buffer, void *bufferend)
 		loadmodel->num_texturesperskin = loadmodel->num_surfaces;
 		loadmodel->data_textures = (texture_t *)Mem_Alloc(loadmodel->mempool, loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t));
 		for (i = 0;i < loadmodel->numskins;i++, inskin += MD2_SKINNAME)
-			Mod_LoadTextureFromQ3Shader(loadmodel->data_textures + i * loadmodel->num_surfaces, inskin, true, true, (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_ALPHA | TEXF_PICMIP | TEXF_COMPRESS);
+			Mod_LoadTextureFromQ3Shader(loadmodel->mempool, loadmodel->name, loadmodel->data_textures + i * loadmodel->num_surfaces, inskin, true, true, (r_mipskins.integer ? TEXF_MIPMAP : 0) | TEXF_ALPHA | TEXF_PICMIP | TEXF_COMPRESS, MATERIALFLAG_WALL);
 	}
 	else
 	{
@@ -1395,7 +1460,7 @@ void Mod_IDP2_Load(dp_model_t *mod, void *buffer, void *bufferend)
 		loadmodel->num_textures = loadmodel->num_surfaces * loadmodel->numskins;
 		loadmodel->num_texturesperskin = loadmodel->num_surfaces;
 		loadmodel->data_textures = (texture_t *)Mem_Alloc(loadmodel->mempool, loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t));
-		Mod_BuildAliasSkinFromSkinFrame(loadmodel->data_textures, NULL);
+		Mod_LoadCustomMaterial(loadmodel->mempool, loadmodel->data_textures, loadmodel->name, SUPERCONTENTS_SOLID, MATERIALFLAG_WALL, R_SkinFrame_LoadMissing());
 	}
 
 	loadmodel->skinscenes = (animscene_t *)Mem_Alloc(loadmodel->mempool, sizeof(animscene_t) * loadmodel->numskins);
@@ -1508,9 +1573,12 @@ void Mod_IDP2_Load(dp_model_t *mod, void *buffer, void *bufferend)
 
 	Mem_Free(vertremap);
 
-	Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
-	Mod_Alias_CalculateBoundingBox();
+	if (loadmodel->surfmesh.data_neighbor3i)
+		Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
+	loadmodel->surfmesh.isanimated = Mod_Alias_CalculateBoundingBox();
 	Mod_Alias_MorphMesh_CompileFrames();
+	if(mod_alias_force_animated.string[0])
+		loadmodel->surfmesh.isanimated = mod_alias_force_animated.integer != 0;
 
 	surface = loadmodel->data_surfaces;
 	surface->texture = loadmodel->data_textures;
@@ -1519,7 +1587,25 @@ void Mod_IDP2_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	surface->num_firstvertex = 0;
 	surface->num_vertices = loadmodel->surfmesh.num_vertices;
 
-	loadmodel->surfmesh.isanimated = loadmodel->numframes > 1 || loadmodel->animscenes[0].framecount > 1;
+	// Always make a BIH for the first frame, we can use it where possible.
+	Mod_MakeCollisionBIH(loadmodel, true, &loadmodel->collision_bih);
+	if (!loadmodel->surfmesh.isanimated)
+	{
+		loadmodel->TraceBox = Mod_CollisionBIH_TraceBox;
+		loadmodel->TraceBrush = Mod_CollisionBIH_TraceBrush;
+		loadmodel->TraceLine = Mod_CollisionBIH_TraceLine;
+		loadmodel->TracePoint = Mod_CollisionBIH_TracePoint_Mesh;
+		loadmodel->PointSuperContents = Mod_CollisionBIH_PointSuperContents_Mesh;
+	}
+
+	// because shaders can do somewhat unexpected things, check for unusual features now
+	for (i = 0;i < loadmodel->num_textures;i++)
+	{
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_SKY))
+			mod->DrawSky = R_Q1BSP_DrawSky;
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_WATERSHADER | MATERIALFLAG_REFRACTION | MATERIALFLAG_REFLECTION | MATERIALFLAG_CAMERA))
+			mod->DrawAddWaterPlanes = R_Q1BSP_DrawAddWaterPlanes;
+	}
 }
 
 void Mod_IDP3_Load(dp_model_t *mod, void *buffer, void *bufferend)
@@ -1549,7 +1635,6 @@ void Mod_IDP3_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->modeldatatypestring = "MD3";
 
 	loadmodel->type = mod_alias;
-	loadmodel->AnimateVertices = Mod_MD3_AnimateVertices;
 	loadmodel->DrawSky = NULL;
 	loadmodel->DrawAddWaterPlanes = NULL;
 	loadmodel->Draw = R_Q1BSP_Draw;
@@ -1564,6 +1649,7 @@ void Mod_IDP3_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->TraceBox = Mod_MDLMD2MD3_TraceBox;
 	loadmodel->TraceLine = Mod_MDLMD2MD3_TraceLine;
 	loadmodel->PointSuperContents = NULL;
+	loadmodel->AnimateVertices = Mod_MD3_AnimateVertices;
 	loadmodel->synctype = ST_RAND;
 	// convert model flags to EF flags (MF_ROCKET becomes EF_ROCKET, etc)
 	i = LittleLong (pinmodel->flags);
@@ -1624,7 +1710,7 @@ void Mod_IDP3_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->nummodelsurfaces = loadmodel->num_surfaces;
 	loadmodel->num_textures = loadmodel->num_surfaces * loadmodel->numskins;
 	loadmodel->num_texturesperskin = loadmodel->num_surfaces;
-	data = (unsigned char *)Mem_Alloc(loadmodel->mempool, loadmodel->num_surfaces * sizeof(msurface_t) + loadmodel->num_surfaces * sizeof(int) + loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t) + meshtriangles * sizeof(int[3]) + meshtriangles * sizeof(int[3]) + (meshvertices <= 65536 ? meshtriangles * sizeof(unsigned short[3]) : 0) + meshvertices * sizeof(float[2]) + meshvertices * loadmodel->numframes * sizeof(md3vertex_t));
+	data = (unsigned char *)Mem_Alloc(loadmodel->mempool, loadmodel->num_surfaces * sizeof(msurface_t) + loadmodel->num_surfaces * sizeof(int) + loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t) + meshtriangles * sizeof(int[3]) + (r_enableshadowvolumes.integer ? meshtriangles * sizeof(int[3]) : 0) + (meshvertices <= 65536 ? meshtriangles * sizeof(unsigned short[3]) : 0) + meshvertices * sizeof(float[2]) + meshvertices * loadmodel->numframes * sizeof(md3vertex_t));
 	loadmodel->data_surfaces = (msurface_t *)data;data += loadmodel->num_surfaces * sizeof(msurface_t);
 	loadmodel->sortedmodelsurfaces = (int *)data;data += loadmodel->num_surfaces * sizeof(int);
 	loadmodel->data_textures = (texture_t *)data;data += loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t);
@@ -1633,11 +1719,16 @@ void Mod_IDP3_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->surfmesh.num_morphframes = loadmodel->numframes; // TODO: remove?
 	loadmodel->num_poses = loadmodel->surfmesh.num_morphframes;
 	loadmodel->surfmesh.data_element3i = (int *)data;data += meshtriangles * sizeof(int[3]);
-	loadmodel->surfmesh.data_neighbor3i = (int *)data;data += meshtriangles * sizeof(int[3]);
+	if (r_enableshadowvolumes.integer)
+	{
+		loadmodel->surfmesh.data_neighbor3i = (int *)data;data += meshtriangles * sizeof(int[3]);
+	}
 	loadmodel->surfmesh.data_texcoordtexture2f = (float *)data;data += meshvertices * sizeof(float[2]);
 	loadmodel->surfmesh.data_morphmd3vertex = (md3vertex_t *)data;data += meshvertices * loadmodel->numframes * sizeof(md3vertex_t);
 	if (meshvertices <= 65536)
+	{
 		loadmodel->surfmesh.data_element3s = (unsigned short *)data;data += meshtriangles * sizeof(unsigned short[3]);
+	}
 
 	meshvertices = 0;
 	meshtriangles = 0;
@@ -1683,14 +1774,34 @@ void Mod_IDP3_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	if (loadmodel->surfmesh.data_element3s)
 		for (i = 0;i < loadmodel->surfmesh.num_triangles*3;i++)
 			loadmodel->surfmesh.data_element3s[i] = loadmodel->surfmesh.data_element3i[i];
-	Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
+	if (loadmodel->surfmesh.data_neighbor3i)
+		Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
 	Mod_Alias_MorphMesh_CompileFrames();
-	Mod_Alias_CalculateBoundingBox();
+	loadmodel->surfmesh.isanimated = Mod_Alias_CalculateBoundingBox();
 	Mod_FreeSkinFiles(skinfiles);
 	Mod_MakeSortedSurfaces(loadmodel);
+	if(mod_alias_force_animated.string[0])
+		loadmodel->surfmesh.isanimated = mod_alias_force_animated.integer != 0;
+
+	// Always make a BIH for the first frame, we can use it where possible.
+	Mod_MakeCollisionBIH(loadmodel, true, &loadmodel->collision_bih);
+	if (!loadmodel->surfmesh.isanimated)
+	{
+		loadmodel->TraceBox = Mod_CollisionBIH_TraceBox;
+		loadmodel->TraceBrush = Mod_CollisionBIH_TraceBrush;
+		loadmodel->TraceLine = Mod_CollisionBIH_TraceLine;
+		loadmodel->TracePoint = Mod_CollisionBIH_TracePoint_Mesh;
+		loadmodel->PointSuperContents = Mod_CollisionBIH_PointSuperContents_Mesh;
+	}
 
-	loadmodel->surfmesh.isanimated = loadmodel->numframes > 1
-	     || (loadmodel->animscenes && loadmodel->animscenes[0].framecount > 1);
+	// because shaders can do somewhat unexpected things, check for unusual features now
+	for (i = 0;i < loadmodel->num_textures;i++)
+	{
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_SKY))
+			mod->DrawSky = R_Q1BSP_DrawSky;
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_WATERSHADER | MATERIALFLAG_REFRACTION | MATERIALFLAG_REFLECTION | MATERIALFLAG_CAMERA))
+			mod->DrawAddWaterPlanes = R_Q1BSP_DrawAddWaterPlanes;
+	}
 }
 
 void Mod_ZYMOTICMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
@@ -1698,7 +1809,7 @@ void Mod_ZYMOTICMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	zymtype1header_t *pinmodel, *pheader;
 	unsigned char *pbase;
 	int i, j, k, numposes, meshvertices, meshtriangles, *bonecount, *vertbonecounts, count, *renderlist, *renderlistend, *outelements;
-	float modelradius, corner[2], *poses, *intexcoord2f, *outtexcoord2f, *bonepose;
+	float modelradius, corner[2], *poses, *intexcoord2f, *outtexcoord2f, *bonepose, f, biggestorigin, tempvec[3], modelscale;
 	zymvertex_t *verts, *vertdata;
 	zymscene_t *scene;
 	zymbone_t *bone;
@@ -1765,7 +1876,6 @@ void Mod_ZYMOTICMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 		return;
 	}
 
-	loadmodel->AnimateVertices = Mod_Skeletal_AnimateVertices;
 	loadmodel->DrawSky = NULL;
 	loadmodel->DrawAddWaterPlanes = NULL;
 	loadmodel->Draw = R_Q1BSP_Draw;
@@ -1780,6 +1890,7 @@ void Mod_ZYMOTICMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->TraceBox = Mod_MDLMD2MD3_TraceBox;
 	loadmodel->TraceLine = Mod_MDLMD2MD3_TraceLine;
 	loadmodel->PointSuperContents = NULL;
+	loadmodel->AnimateVertices = Mod_Skeletal_AnimateVertices;
 
 	loadmodel->numframes = pheader->numscenes;
 	loadmodel->num_surfaces = pheader->numshaders;
@@ -1799,6 +1910,7 @@ void Mod_ZYMOTICMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	}
 
 	// model bbox
+	// LordHavoc: actually we blow this away later with Mod_Alias_CalculateBoundingBox()
 	modelradius = pheader->radius;
 	for (i = 0;i < 3;i++)
 	{
@@ -1871,30 +1983,75 @@ void Mod_ZYMOTICMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->nummodelsurfaces = loadmodel->num_surfaces;
 	loadmodel->num_textures = loadmodel->num_surfaces * loadmodel->numskins;
 	loadmodel->num_texturesperskin = loadmodel->num_surfaces;
-	data = (unsigned char *)Mem_Alloc(loadmodel->mempool, loadmodel->num_surfaces * sizeof(msurface_t) + loadmodel->num_surfaces * sizeof(int) + loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t) + meshtriangles * sizeof(int[3]) + meshtriangles * sizeof(int[3]) + (meshvertices <= 65536 ? meshtriangles * sizeof(unsigned short[3]) : 0) + meshvertices * sizeof(float[14]) + meshvertices * sizeof(int[4]) + meshvertices * sizeof(float[4]) + loadmodel->num_poses * loadmodel->num_bones * sizeof(float[12]) + loadmodel->num_bones * sizeof(float[12]));
+	data = (unsigned char *)Mem_Alloc(loadmodel->mempool, loadmodel->num_surfaces * sizeof(msurface_t) + loadmodel->num_surfaces * sizeof(int) + loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t) + meshtriangles * sizeof(int[3]) + (r_enableshadowvolumes.integer ? meshtriangles * sizeof(int[3]) : 0) + (meshvertices <= 65536 ? meshtriangles * sizeof(unsigned short[3]) : 0) + meshvertices * (sizeof(float[14]) + sizeof(unsigned short) + sizeof(unsigned char[2][4])) + loadmodel->num_poses * loadmodel->num_bones * sizeof(short[7]) + loadmodel->num_bones * sizeof(float[12]));
 	loadmodel->data_surfaces = (msurface_t *)data;data += loadmodel->num_surfaces * sizeof(msurface_t);
 	loadmodel->sortedmodelsurfaces = (int *)data;data += loadmodel->num_surfaces * sizeof(int);
 	loadmodel->data_textures = (texture_t *)data;data += loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t);
 	loadmodel->surfmesh.num_vertices = meshvertices;
 	loadmodel->surfmesh.num_triangles = meshtriangles;
 	loadmodel->surfmesh.data_element3i = (int *)data;data += meshtriangles * sizeof(int[3]);
-	loadmodel->surfmesh.data_neighbor3i = (int *)data;data += meshtriangles * sizeof(int[3]);
+	if (r_enableshadowvolumes.integer)
+	{
+		loadmodel->surfmesh.data_neighbor3i = (int *)data;data += meshtriangles * sizeof(int[3]);
+	}
 	loadmodel->surfmesh.data_vertex3f = (float *)data;data += meshvertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_svector3f = (float *)data;data += meshvertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_tvector3f = (float *)data;data += meshvertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_normal3f = (float *)data;data += meshvertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_texcoordtexture2f = (float *)data;data += meshvertices * sizeof(float[2]);
-	loadmodel->surfmesh.data_vertexweightindex4i = (int *)data;data += meshvertices * sizeof(int[4]);
-	loadmodel->surfmesh.data_vertexweightinfluence4f = (float *)data;data += meshvertices * sizeof(float[4]);
-	loadmodel->data_poses = (float *)data;data += loadmodel->num_poses * loadmodel->num_bones * sizeof(float[12]);
+	loadmodel->surfmesh.data_skeletalindex4ub = (unsigned char *)data;data += meshvertices * sizeof(unsigned char[4]);
+	loadmodel->surfmesh.data_skeletalweight4ub = (unsigned char *)data;data += meshvertices * sizeof(unsigned char[4]);
 	loadmodel->data_baseboneposeinverse = (float *)data;data += loadmodel->num_bones * sizeof(float[12]);
+	loadmodel->surfmesh.num_blends = 0;
+	loadmodel->surfmesh.blends = (unsigned short *)data;data += meshvertices * sizeof(unsigned short);
 	if (loadmodel->surfmesh.num_vertices <= 65536)
+	{
 		loadmodel->surfmesh.data_element3s = (unsigned short *)data;data += loadmodel->surfmesh.num_triangles * sizeof(unsigned short[3]);
+	}
+	loadmodel->data_poses7s = (short *)data;data += loadmodel->num_poses * loadmodel->num_bones * sizeof(short[7]);
+	loadmodel->surfmesh.data_blendweights = NULL;
 
 	//zymlump_t lump_poses; // float pose[numposes][numbones][3][4]; // animation data
 	poses = (float *) (pheader->lump_poses.start + pbase);
-	for (i = 0;i < pheader->lump_poses.length / 4;i++)
-		loadmodel->data_poses[i] = BigFloat(poses[i]);
+	// figure out scale of model from root bone, for compatibility with old zmodel versions
+	tempvec[0] = BigFloat(poses[0]);
+	tempvec[1] = BigFloat(poses[1]);
+	tempvec[2] = BigFloat(poses[2]);
+	modelscale = VectorLength(tempvec);
+	biggestorigin = 0;
+	for (i = 0;i < loadmodel->num_bones * numposes * 12;i++)
+	{
+		f = fabs(BigFloat(poses[i]));
+		biggestorigin = max(biggestorigin, f);
+	}
+	loadmodel->num_posescale = biggestorigin / 32767.0f;
+	loadmodel->num_poseinvscale = 1.0f / loadmodel->num_posescale;
+	for (i = 0;i < numposes;i++)
+	{
+		const float *frameposes = (float *) (pheader->lump_poses.start + pbase) + 12*i*loadmodel->num_bones;
+		for (j = 0;j < loadmodel->num_bones;j++)
+		{
+			float pose[12];
+			matrix4x4_t posematrix;
+			for (k = 0;k < 12;k++)
+				pose[k] = BigFloat(frameposes[j*12+k]);
+			//if (j < loadmodel->num_bones)
+			//	Con_Printf("%s: bone %i = %f %f %f %f : %f %f %f %f : %f %f %f %f : scale = %f\n", loadmodel->name, j, pose[0], pose[1], pose[2], pose[3], pose[4], pose[5], pose[6], pose[7], pose[8], pose[9], pose[10], pose[11], VectorLength(pose));
+			// scale child bones to match the root scale
+			if (loadmodel->data_bones[j].parent >= 0)
+			{
+				pose[3] *= modelscale;
+				pose[7] *= modelscale;
+				pose[11] *= modelscale;
+			}
+			// normalize rotation matrix
+			VectorNormalize(pose + 0);
+			VectorNormalize(pose + 4);
+			VectorNormalize(pose + 8);
+			Matrix4x4_FromArray12FloatD3D(&posematrix, pose);
+			Matrix4x4_ToBonePose7s(&posematrix, loadmodel->num_poseinvscale, loadmodel->data_poses7s + 7*(i*loadmodel->num_bones+j));
+		}
+	}
 
 	//zymlump_t lump_verts; // zymvertex_t vert[numvertices]; // see vertex struct
 	verts = (zymvertex_t *)Mem_Alloc(loadmodel->mempool, pheader->lump_verts.length);
@@ -1905,7 +2062,9 @@ void Mod_ZYMOTICMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	bonepose = (float *)Z_Malloc(loadmodel->num_bones * sizeof(float[12]));
 	for (i = 0;i < loadmodel->num_bones;i++)
 	{
-		const float *m = loadmodel->data_poses + i * 12;
+		float m[12];
+		for (k = 0;k < 12;k++)
+			m[k] = BigFloat(poses[i*12+k]);
 		if (loadmodel->data_bones[i].parent >= 0)
 			R_ConcatTransforms(bonepose + 12 * loadmodel->data_bones[i].parent, m, bonepose + 12 * i);
 		else
@@ -1928,8 +2087,15 @@ void Mod_ZYMOTICMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 		loadmodel->surfmesh.data_vertex3f[j*3+1] = relativeorigin[0] * m[4] + relativeorigin[1] * m[5] + relativeorigin[2] * m[ 6] + m[ 7];
 		loadmodel->surfmesh.data_vertex3f[j*3+2] = relativeorigin[0] * m[8] + relativeorigin[1] * m[9] + relativeorigin[2] * m[10] + m[11];
 		// store the weight as the primary weight on this vertex
-		loadmodel->surfmesh.data_vertexweightindex4i[j*4+0] = boneindex;
-		loadmodel->surfmesh.data_vertexweightinfluence4f[j*4+0] = 1;
+		loadmodel->surfmesh.blends[j] = boneindex;
+		loadmodel->surfmesh.data_skeletalindex4ub[j*4  ] = boneindex;
+		loadmodel->surfmesh.data_skeletalindex4ub[j*4+1] = 0;
+		loadmodel->surfmesh.data_skeletalindex4ub[j*4+2] = 0;
+		loadmodel->surfmesh.data_skeletalindex4ub[j*4+3] = 0;
+		loadmodel->surfmesh.data_skeletalweight4ub[j*4  ] = 255;
+		loadmodel->surfmesh.data_skeletalweight4ub[j*4+1] = 0;
+		loadmodel->surfmesh.data_skeletalweight4ub[j*4+2] = 0;
+		loadmodel->surfmesh.data_skeletalweight4ub[j*4+3] = 0;
 	}
 	Z_Free(bonepose);
 	// normals and tangents are calculated after elements are loaded
@@ -2010,17 +2176,39 @@ void Mod_ZYMOTICMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 			loadmodel->surfmesh.data_element3s[i] = loadmodel->surfmesh.data_element3i[i];
 	Mod_ValidateElements(loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles, 0, loadmodel->surfmesh.num_vertices, __FILE__, __LINE__);
 	Mod_BuildBaseBonePoses();
-	Mod_BuildNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_normal3f, true);
-	Mod_BuildTextureVectorsFromNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_texcoordtexture2f, loadmodel->surfmesh.data_normal3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_svector3f, loadmodel->surfmesh.data_tvector3f, true);
-	Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
+	Mod_BuildNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_normal3f, r_smoothnormals_areaweighting.integer != 0);
+	Mod_BuildTextureVectorsFromNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_texcoordtexture2f, loadmodel->surfmesh.data_normal3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_svector3f, loadmodel->surfmesh.data_tvector3f, r_smoothnormals_areaweighting.integer != 0);
+	if (loadmodel->surfmesh.data_neighbor3i)
+		Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
+	loadmodel->surfmesh.isanimated = Mod_Alias_CalculateBoundingBox();
+	if(mod_alias_force_animated.string[0])
+		loadmodel->surfmesh.isanimated = mod_alias_force_animated.integer != 0;
+
+	// Always make a BIH for the first frame, we can use it where possible.
+	Mod_MakeCollisionBIH(loadmodel, true, &loadmodel->collision_bih);
+	if (!loadmodel->surfmesh.isanimated)
+	{
+		loadmodel->TraceBox = Mod_CollisionBIH_TraceBox;
+		loadmodel->TraceBrush = Mod_CollisionBIH_TraceBrush;
+		loadmodel->TraceLine = Mod_CollisionBIH_TraceLine;
+		loadmodel->TracePoint = Mod_CollisionBIH_TracePoint_Mesh;
+		loadmodel->PointSuperContents = Mod_CollisionBIH_PointSuperContents_Mesh;
+	}
 
-	loadmodel->surfmesh.isanimated = loadmodel->numframes > 1 || loadmodel->animscenes[0].framecount > 1;
+	// because shaders can do somewhat unexpected things, check for unusual features now
+	for (i = 0;i < loadmodel->num_textures;i++)
+	{
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_SKY))
+			mod->DrawSky = R_Q1BSP_DrawSky;
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_WATERSHADER | MATERIALFLAG_REFRACTION | MATERIALFLAG_REFLECTION | MATERIALFLAG_CAMERA))
+			mod->DrawAddWaterPlanes = R_Q1BSP_DrawAddWaterPlanes;
+	}
 }
 
 void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 {
 	dpmheader_t *pheader;
-	dpmframe_t *frame;
+	dpmframe_t *frames;
 	dpmbone_t *bone;
 	dpmmesh_t *dpmmesh;
 	unsigned char *pbase;
@@ -2028,6 +2216,9 @@ void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	skinfile_t *skinfiles;
 	unsigned char *data;
 	float *bonepose;
+	float biggestorigin, tempvec[3], modelscale;
+	float f;
+	float *poses;
 
 	pheader = (dpmheader_t *)buffer;
 	pbase = (unsigned char *)buffer;
@@ -2070,7 +2261,6 @@ void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 		return;
 	}
 
-	loadmodel->AnimateVertices = Mod_Skeletal_AnimateVertices;
 	loadmodel->DrawSky = NULL;
 	loadmodel->DrawAddWaterPlanes = NULL;
 	loadmodel->Draw = R_Q1BSP_Draw;
@@ -2085,8 +2275,10 @@ void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->TraceBox = Mod_MDLMD2MD3_TraceBox;
 	loadmodel->TraceLine = Mod_MDLMD2MD3_TraceLine;
 	loadmodel->PointSuperContents = NULL;
+	loadmodel->AnimateVertices = Mod_Skeletal_AnimateVertices;
 
 	// model bbox
+	// LordHavoc: actually we blow this away later with Mod_Alias_CalculateBoundingBox()
 	for (i = 0;i < 3;i++)
 	{
 		loadmodel->normalmins[i] = pheader->mins[i];
@@ -2124,28 +2316,36 @@ void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->num_textures = loadmodel->num_surfaces * loadmodel->numskins;
 	loadmodel->num_texturesperskin = loadmodel->num_surfaces;
 	// do most allocations as one merged chunk
-	data = (unsigned char *)Mem_Alloc(loadmodel->mempool, loadmodel->num_surfaces * sizeof(msurface_t) + loadmodel->num_surfaces * sizeof(int) + loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t) + meshtriangles * sizeof(int[3]) + (meshvertices <= 65536 ? meshtriangles * sizeof(unsigned short[3]) : 0) + meshtriangles * sizeof(int[3]) + meshvertices * (sizeof(float[14]) + sizeof(int[4]) + sizeof(float[4])) + loadmodel->num_poses * loadmodel->num_bones * sizeof(float[12]) + loadmodel->num_bones * sizeof(float[12]) + loadmodel->numskins * sizeof(animscene_t) + loadmodel->num_bones * sizeof(aliasbone_t) + loadmodel->numframes * sizeof(animscene_t));
+	data = (unsigned char *)Mem_Alloc(loadmodel->mempool, loadmodel->num_surfaces * sizeof(msurface_t) + loadmodel->num_surfaces * sizeof(int) + loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t) + meshtriangles * sizeof(int[3]) + (meshvertices <= 65536 ? meshtriangles * sizeof(unsigned short[3]) : 0) + (r_enableshadowvolumes.integer ? meshtriangles * sizeof(int[3]) : 0) + meshvertices * (sizeof(float[14]) + sizeof(unsigned short) + sizeof(unsigned char[2][4])) + loadmodel->num_poses * loadmodel->num_bones * sizeof(short[7]) + loadmodel->num_bones * sizeof(float[12]) + loadmodel->numskins * sizeof(animscene_t) + loadmodel->num_bones * sizeof(aliasbone_t) + loadmodel->numframes * sizeof(animscene_t));
 	loadmodel->data_surfaces = (msurface_t *)data;data += loadmodel->num_surfaces * sizeof(msurface_t);
 	loadmodel->sortedmodelsurfaces = (int *)data;data += loadmodel->num_surfaces * sizeof(int);
 	loadmodel->data_textures = (texture_t *)data;data += loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t);
 	loadmodel->surfmesh.num_vertices = meshvertices;
 	loadmodel->surfmesh.num_triangles = meshtriangles;
 	loadmodel->surfmesh.data_element3i = (int *)data;data += meshtriangles * sizeof(int[3]);
-	loadmodel->surfmesh.data_neighbor3i = (int *)data;data += meshtriangles * sizeof(int[3]);
+	if (r_enableshadowvolumes.integer)
+	{
+		loadmodel->surfmesh.data_neighbor3i = (int *)data;data += meshtriangles * sizeof(int[3]);
+	}
 	loadmodel->surfmesh.data_vertex3f = (float *)data;data += meshvertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_svector3f = (float *)data;data += meshvertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_tvector3f = (float *)data;data += meshvertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_normal3f = (float *)data;data += meshvertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_texcoordtexture2f = (float *)data;data += meshvertices * sizeof(float[2]);
-	loadmodel->surfmesh.data_vertexweightindex4i = (int *)data;data += meshvertices * sizeof(int[4]);
-	loadmodel->surfmesh.data_vertexweightinfluence4f = (float *)data;data += meshvertices * sizeof(float[4]);
-	loadmodel->data_poses = (float *)data;data += loadmodel->num_poses * loadmodel->num_bones * sizeof(float[12]);
+	loadmodel->surfmesh.data_skeletalindex4ub = (unsigned char *)data;data += meshvertices * sizeof(unsigned char[4]);
+	loadmodel->surfmesh.data_skeletalweight4ub = (unsigned char *)data;data += meshvertices * sizeof(unsigned char[4]);
 	loadmodel->data_baseboneposeinverse = (float *)data;data += loadmodel->num_bones * sizeof(float[12]);
 	loadmodel->skinscenes = (animscene_t *)data;data += loadmodel->numskins * sizeof(animscene_t);
 	loadmodel->data_bones = (aliasbone_t *)data;data += loadmodel->num_bones * sizeof(aliasbone_t);
 	loadmodel->animscenes = (animscene_t *)data;data += loadmodel->numframes * sizeof(animscene_t);
+	loadmodel->surfmesh.num_blends = 0;
+	loadmodel->surfmesh.blends = (unsigned short *)data;data += meshvertices * sizeof(unsigned short);
 	if (meshvertices <= 65536)
+	{
 		loadmodel->surfmesh.data_element3s = (unsigned short *)data;data += meshtriangles * sizeof(unsigned short[3]);
+	}
+	loadmodel->data_poses7s = (short *)data;data += loadmodel->num_poses * loadmodel->num_bones * sizeof(short[7]);
+	loadmodel->surfmesh.data_blendweights = (blendweights_t *)Mem_Alloc(loadmodel->mempool, meshvertices * sizeof(blendweights_t));
 
 	for (i = 0;i < loadmodel->numskins;i++)
 	{
@@ -2167,21 +2367,55 @@ void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	}
 
 	// load the frames
-	frame = (dpmframe_t *) (pbase + pheader->ofs_frames);
+	frames = (dpmframe_t *) (pbase + pheader->ofs_frames);
+	// figure out scale of model from root bone, for compatibility with old dpmodel versions
+	poses = (float *) (pbase + BigLong(frames[0].ofs_bonepositions));
+	tempvec[0] = BigFloat(poses[0]);
+	tempvec[1] = BigFloat(poses[1]);
+	tempvec[2] = BigFloat(poses[2]);
+	modelscale = VectorLength(tempvec);
+	biggestorigin = 0;
 	for (i = 0;i < loadmodel->numframes;i++)
 	{
-		const float *poses;
-		memcpy(loadmodel->animscenes[i].name, frame->name, sizeof(frame->name));
+		memcpy(loadmodel->animscenes[i].name, frames[i].name, sizeof(frames[i].name));
 		loadmodel->animscenes[i].firstframe = i;
 		loadmodel->animscenes[i].framecount = 1;
 		loadmodel->animscenes[i].loop = true;
 		loadmodel->animscenes[i].framerate = 10;
 		// load the bone poses for this frame
-		poses = (float *) (pbase + BigLong(frame->ofs_bonepositions));
+		poses = (float *) (pbase + BigLong(frames[i].ofs_bonepositions));
 		for (j = 0;j < loadmodel->num_bones*12;j++)
-			loadmodel->data_poses[i * loadmodel->num_bones*12 + j] = BigFloat(poses[j]);
+		{
+			f = fabs(BigFloat(poses[j]));
+			biggestorigin = max(biggestorigin, f);
+		}
 		// stuff not processed here: mins, maxs, yawradius, allradius
-		frame++;
+	}
+	loadmodel->num_posescale = biggestorigin / 32767.0f;
+	loadmodel->num_poseinvscale = 1.0f / loadmodel->num_posescale;
+	for (i = 0;i < loadmodel->numframes;i++)
+	{
+		const float *frameposes = (float *) (pbase + BigLong(frames[i].ofs_bonepositions));
+		for (j = 0;j < loadmodel->num_bones;j++)
+		{
+			float pose[12];
+			matrix4x4_t posematrix;
+			for (k = 0;k < 12;k++)
+				pose[k] = BigFloat(frameposes[j*12+k]);
+			// scale child bones to match the root scale
+			if (loadmodel->data_bones[j].parent >= 0)
+			{
+				pose[3] *= modelscale;
+				pose[7] *= modelscale;
+				pose[11] *= modelscale;
+			}
+			// normalize rotation matrix
+			VectorNormalize(pose + 0);
+			VectorNormalize(pose + 4);
+			VectorNormalize(pose + 8);
+			Matrix4x4_FromArray12FloatD3D(&posematrix, pose);
+			Matrix4x4_ToBonePose7s(&posematrix, loadmodel->num_poseinvscale, loadmodel->data_poses7s + 7*(i*loadmodel->num_bones+j));
+		}
 	}
 
 	// load the meshes now
@@ -2191,10 +2425,13 @@ void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	// reconstruct frame 0 matrices to allow reconstruction of the base mesh
 	// (converting from weight-blending skeletal animation to
 	//  deformation-based skeletal animation)
+	poses = (float *) (pbase + BigLong(frames[0].ofs_bonepositions));
 	bonepose = (float *)Z_Malloc(loadmodel->num_bones * sizeof(float[12]));
 	for (i = 0;i < loadmodel->num_bones;i++)
 	{
-		const float *m = loadmodel->data_poses + i * 12;
+		float m[12];
+		for (k = 0;k < 12;k++)
+			m[k] = BigFloat(poses[i*12+k]);
 		if (loadmodel->data_bones[i].parent >= 0)
 			R_ConcatTransforms(bonepose + 12 * loadmodel->data_bones[i].parent, m, bonepose + 12 * i);
 		else
@@ -2237,7 +2474,8 @@ void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 		data = (unsigned char *) (pbase + BigLong(dpmmesh->ofs_verts));
 		for (j = surface->num_firstvertex;j < surface->num_firstvertex + surface->num_vertices;j++)
 		{
-			float sum;
+			int weightindex[4] = { 0, 0, 0, 0 };
+			float weightinfluence[4] = { 0, 0, 0, 0 };
 			int l;
 			int numweights = BigLong(((dpmvertex_t *)data)->numbones);
 			data += sizeof(dpmvertex_t);
@@ -2264,8 +2502,8 @@ void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 				if (!k)
 				{
 					// store the first (and often only) weight
-					loadmodel->surfmesh.data_vertexweightinfluence4f[j*4+0] = influence;
-					loadmodel->surfmesh.data_vertexweightindex4i[j*4+0] = boneindex;
+					weightinfluence[0] = influence;
+					weightindex[0] = boneindex;
 				}
 				else
 				{
@@ -2273,33 +2511,33 @@ void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 					// (which only accepts up to 4 bones per vertex)
 					for (l = 0;l < 4;l++)
 					{
-						if (loadmodel->surfmesh.data_vertexweightinfluence4f[j*4+l] < influence)
+						if (weightinfluence[l] < influence)
 						{
 							// move weaker influence weights out of the way first
 							int l2;
 							for (l2 = 3;l2 > l;l2--)
 							{
-								loadmodel->surfmesh.data_vertexweightinfluence4f[j*4+l2] = loadmodel->surfmesh.data_vertexweightinfluence4f[j*4+l2-1];
-								loadmodel->surfmesh.data_vertexweightindex4i[j*4+l2] = loadmodel->surfmesh.data_vertexweightindex4i[j*4+l2-1];
+								weightinfluence[l2] = weightinfluence[l2-1];
+								weightindex[l2] = weightindex[l2-1];
 							}
 							// store the new weight
-							loadmodel->surfmesh.data_vertexweightinfluence4f[j*4+l] = influence;
-							loadmodel->surfmesh.data_vertexweightindex4i[j*4+l] = boneindex;
+							weightinfluence[l] = influence;
+							weightindex[l] = boneindex;
 							break;
 						}
 					}
 				}
 				data += sizeof(dpmbonevert_t);
 			}
-			sum = 0;
-			for (l = 0;l < 4;l++)
-				sum += loadmodel->surfmesh.data_vertexweightinfluence4f[j*4+l];
-			if (sum && fabs(sum - 1) > (1.0f / 256.0f))
-			{
-				float f = 1.0f / sum;
-				for (l = 0;l < 4;l++)
-					loadmodel->surfmesh.data_vertexweightinfluence4f[j*4+l] *= f;
-			}
+			loadmodel->surfmesh.blends[j] = Mod_Skeletal_CompressBlend(loadmodel, weightindex, weightinfluence);
+			loadmodel->surfmesh.data_skeletalindex4ub[j*4  ] = weightindex[0];
+			loadmodel->surfmesh.data_skeletalindex4ub[j*4+1] = weightindex[1];
+			loadmodel->surfmesh.data_skeletalindex4ub[j*4+2] = weightindex[2];
+			loadmodel->surfmesh.data_skeletalindex4ub[j*4+3] = weightindex[3];
+			loadmodel->surfmesh.data_skeletalweight4ub[j*4  ] = (unsigned char)(weightinfluence[0]*255.0f);
+			loadmodel->surfmesh.data_skeletalweight4ub[j*4+1] = (unsigned char)(weightinfluence[1]*255.0f);
+			loadmodel->surfmesh.data_skeletalweight4ub[j*4+2] = (unsigned char)(weightinfluence[2]*255.0f);
+			loadmodel->surfmesh.data_skeletalweight4ub[j*4+3] = (unsigned char)(weightinfluence[3]*255.0f);
 		}
 
 		// since dpm models do not have named sections, reuse their shader name as the section name
@@ -2307,6 +2545,8 @@ void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 
 		Mod_ValidateElements(loadmodel->surfmesh.data_element3i + surface->num_firsttriangle * 3, surface->num_triangles, surface->num_firstvertex, surface->num_vertices, __FILE__, __LINE__);
 	}
+	if (loadmodel->surfmesh.num_blends < meshvertices)
+		loadmodel->surfmesh.data_blendweights = (blendweights_t *)Mem_Realloc(loadmodel->mempool, loadmodel->surfmesh.data_blendweights, loadmodel->surfmesh.num_blends * sizeof(blendweights_t));
 	Z_Free(bonepose);
 	Mod_FreeSkinFiles(skinfiles);
 	Mod_MakeSortedSurfaces(loadmodel);
@@ -2316,10 +2556,32 @@ void Mod_DARKPLACESMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 		for (i = 0;i < loadmodel->surfmesh.num_triangles*3;i++)
 			loadmodel->surfmesh.data_element3s[i] = loadmodel->surfmesh.data_element3i[i];
 	Mod_BuildBaseBonePoses();
-	Mod_BuildTextureVectorsFromNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_texcoordtexture2f, loadmodel->surfmesh.data_normal3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_svector3f, loadmodel->surfmesh.data_tvector3f, true);
-	Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
+	Mod_BuildTextureVectorsFromNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_texcoordtexture2f, loadmodel->surfmesh.data_normal3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_svector3f, loadmodel->surfmesh.data_tvector3f, r_smoothnormals_areaweighting.integer != 0);
+	if (loadmodel->surfmesh.data_neighbor3i)
+		Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
+	loadmodel->surfmesh.isanimated = Mod_Alias_CalculateBoundingBox();
+	if(mod_alias_force_animated.string[0])
+		loadmodel->surfmesh.isanimated = mod_alias_force_animated.integer != 0;
 
-	loadmodel->surfmesh.isanimated = loadmodel->numframes > 1 || loadmodel->animscenes[0].framecount > 1;
+	// Always make a BIH for the first frame, we can use it where possible.
+	Mod_MakeCollisionBIH(loadmodel, true, &loadmodel->collision_bih);
+	if (!loadmodel->surfmesh.isanimated)
+	{
+		loadmodel->TraceBox = Mod_CollisionBIH_TraceBox;
+		loadmodel->TraceBrush = Mod_CollisionBIH_TraceBrush;
+		loadmodel->TraceLine = Mod_CollisionBIH_TraceLine;
+		loadmodel->TracePoint = Mod_CollisionBIH_TracePoint_Mesh;
+		loadmodel->PointSuperContents = Mod_CollisionBIH_PointSuperContents_Mesh;
+	}
+
+	// because shaders can do somewhat unexpected things, check for unusual features now
+	for (i = 0;i < loadmodel->num_textures;i++)
+	{
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_SKY))
+			mod->DrawSky = R_Q1BSP_DrawSky;
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_WATERSHADER | MATERIALFLAG_REFRACTION | MATERIALFLAG_REFLECTION | MATERIALFLAG_CAMERA))
+			mod->DrawAddWaterPlanes = R_Q1BSP_DrawAddWaterPlanes;
+	}
 }
 
 // no idea why PSK/PSA files contain weird quaternions but they do...
@@ -2335,7 +2597,7 @@ void Mod_PSKMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	pskmatt_t *matts;
 	pskboneinfo_t *bones;
 	pskrawweights_t *rawweights;
-	pskboneinfo_t *animbones;
+	//pskboneinfo_t *animbones;
 	pskaniminfo_t *anims;
 	pskanimkeys_t *animkeys;
 	void *animfilebuffer, *animbuffer, *animbufferend;
@@ -2344,6 +2606,7 @@ void Mod_PSKMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	skinfile_t *skinfiles;
 	char animname[MAX_QPATH];
 	size_t size;
+	float biggestorigin;
 
 	pchunk = (pskchunk_t *)buffer;
 	if (strcmp(pchunk->id, "ACTRHEAD"))
@@ -2352,7 +2615,6 @@ void Mod_PSKMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->modeldatatypestring = "PSK";
 
 	loadmodel->type = mod_alias;
-	loadmodel->AnimateVertices = Mod_Skeletal_AnimateVertices;
 	loadmodel->DrawSky = NULL;
 	loadmodel->DrawAddWaterPlanes = NULL;
 	loadmodel->Draw = R_Q1BSP_Draw;
@@ -2367,14 +2629,15 @@ void Mod_PSKMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->TraceBox = Mod_MDLMD2MD3_TraceBox;
 	loadmodel->TraceLine = Mod_MDLMD2MD3_TraceLine;
 	loadmodel->PointSuperContents = NULL;
+	loadmodel->AnimateVertices = Mod_Skeletal_AnimateVertices;
 	loadmodel->synctype = ST_RAND;
 
 	FS_StripExtension(loadmodel->name, animname, sizeof(animname));
 	strlcat(animname, ".psa", sizeof(animname));
 	animbuffer = animfilebuffer = FS_LoadFile(animname, loadmodel->mempool, false, &filesize);
 	animbufferend = (void *)((unsigned char*)animbuffer + (int)filesize);
-	if (animbuffer == NULL)
-		Host_Error("%s: can't find .psa file (%s)", loadmodel->name, animname);
+	if (!animbuffer)
+		animbufferend = animbuffer;
 
 	numpnts = 0;
 	pnts = NULL;
@@ -2400,8 +2663,8 @@ void Mod_PSKMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 		version = LittleLong(pchunk->version);
 		recordsize = LittleLong(pchunk->recordsize);
 		numrecords = LittleLong(pchunk->numrecords);
-		if (developer.integer >= 100)
-			Con_Printf("%s: %s %x: %i * %i = %i\n", loadmodel->name, pchunk->id, version, recordsize, numrecords, recordsize * numrecords);
+		if (developer_extra.integer)
+			Con_DPrintf("%s: %s %x: %i * %i = %i\n", loadmodel->name, pchunk->id, version, recordsize, numrecords, recordsize * numrecords);
 		if (version != 0x1e83b9 && version != 0x1e9179 && version != 0x2e && version != 0x12f2bc && version != 0x12f2f0)
 			Con_Printf ("%s: chunk %s has unknown version %x (0x1e83b9, 0x1e9179, 0x2e, 0x12f2bc, 0x12f2f0 are currently supported), trying to load anyway!\n", loadmodel->name, pchunk->id, version);
 		if (!strcmp(pchunk->id, "ACTRHEAD"))
@@ -2571,8 +2834,8 @@ void Mod_PSKMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 		version = LittleLong(pchunk->version);
 		recordsize = LittleLong(pchunk->recordsize);
 		numrecords = LittleLong(pchunk->numrecords);
-		if (developer.integer >= 100)
-			Con_Printf("%s: %s %x: %i * %i = %i\n", animname, pchunk->id, version, recordsize, numrecords, recordsize * numrecords);
+		if (developer_extra.integer)
+			Con_DPrintf("%s: %s %x: %i * %i = %i\n", animname, pchunk->id, version, recordsize, numrecords, recordsize * numrecords);
 		if (version != 0x1e83b9 && version != 0x1e9179 && version != 0x2e && version != 0x12f2bc && version != 0x12f2f0)
 			Con_Printf ("%s: chunk %s has unknown version %x (0x1e83b9, 0x1e9179, 0x2e, 0x12f2bc, 0x12f2f0 are currently supported), trying to load anyway!\n", animname, pchunk->id, version);
 		if (!strcmp(pchunk->id, "ANIMHEAD"))
@@ -2586,7 +2849,7 @@ void Mod_PSKMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 				Host_Error("%s: %s has unsupported recordsize", animname, pchunk->id);
 			// byteswap in place and keep the pointer
 			numanimbones = numrecords;
-			animbones = (pskboneinfo_t *)animbuffer;
+			//animbones = (pskboneinfo_t *)animbuffer;
 			// NOTE: supposedly psa does not need to match the psk model, the
 			// bones missing from the psa would simply use their base
 			// positions from the psk, but this is hard for me to implement
@@ -2692,15 +2955,19 @@ void Mod_PSKMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 			Con_Printf("%s: unknown chunk ID \"%s\"\n", animname, pchunk->id);
 	}
 
-	if (!numpnts || !pnts || !numvtxw || !vtxw || !numfaces || !faces || !nummatts || !matts || !numbones || !bones || !numrawweights || !rawweights || !numanims || !anims || !numanimkeys || !animkeys)
+	if (!numpnts || !pnts || !numvtxw || !vtxw || !numfaces || !faces || !nummatts || !matts || !numbones || !bones || !numrawweights || !rawweights)
 		Host_Error("%s: missing required chunks", loadmodel->name);
 
-	loadmodel->numframes = 0;
-	for (index = 0;index < numanims;index++)
-		loadmodel->numframes += anims[index].numframes;
-
-	if (numanimkeys != numbones * loadmodel->numframes)
-		Host_Error("%s: %s has incorrect number of animation keys", animname, pchunk->id);
+	if (numanims)
+	{
+		loadmodel->numframes = 0;
+		for (index = 0;index < numanims;index++)
+			loadmodel->numframes += anims[index].numframes;
+		if (numanimkeys != numbones * loadmodel->numframes)
+			Host_Error("%s: %s has incorrect number of animation keys", animname, pchunk->id);
+	}
+	else
+		loadmodel->numframes = loadmodel->num_poses = 1;
 
 	meshvertices = numvtxw;
 	meshtriangles = numfaces;
@@ -2717,27 +2984,35 @@ void Mod_PSKMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 	loadmodel->surfmesh.num_vertices = meshvertices;
 	loadmodel->surfmesh.num_triangles = meshtriangles;
 	// do most allocations as one merged chunk
-	size = loadmodel->num_surfaces * sizeof(msurface_t) + loadmodel->num_surfaces * sizeof(int) + loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t) + loadmodel->surfmesh.num_triangles * sizeof(int[3]) + loadmodel->surfmesh.num_triangles * sizeof(int[3]) + loadmodel->surfmesh.num_vertices * sizeof(float[3]) + loadmodel->surfmesh.num_vertices * sizeof(float[3]) + loadmodel->surfmesh.num_vertices * sizeof(float[3]) + loadmodel->surfmesh.num_vertices * sizeof(float[3]) + loadmodel->surfmesh.num_vertices * sizeof(float[2]) + loadmodel->surfmesh.num_vertices * sizeof(int[4]) + loadmodel->surfmesh.num_vertices * sizeof(float[4]) + loadmodel->num_poses * loadmodel->num_bones * sizeof(float[12]) + loadmodel->num_bones * sizeof(float[12]) + loadmodel->numskins * sizeof(animscene_t) + loadmodel->num_bones * sizeof(aliasbone_t) + loadmodel->numframes * sizeof(animscene_t) + ((loadmodel->surfmesh.num_vertices <= 65536) ? (loadmodel->surfmesh.num_triangles * sizeof(unsigned short[3])) : 0);
+	size = loadmodel->num_surfaces * sizeof(msurface_t) + loadmodel->num_surfaces * sizeof(int) + loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t) + loadmodel->surfmesh.num_triangles * sizeof(int[3]) + (r_enableshadowvolumes.integer ? loadmodel->surfmesh.num_triangles * sizeof(int[3]) : 0)  + loadmodel->surfmesh.num_vertices * sizeof(float[3]) + loadmodel->surfmesh.num_vertices * sizeof(float[3]) + loadmodel->surfmesh.num_vertices * sizeof(float[3]) + loadmodel->surfmesh.num_vertices * sizeof(float[3]) + loadmodel->surfmesh.num_vertices * sizeof(float[2]) + loadmodel->surfmesh.num_vertices * sizeof(unsigned char[4]) + loadmodel->surfmesh.num_vertices * sizeof(unsigned char[4]) + loadmodel->surfmesh.num_vertices * sizeof(unsigned short) + loadmodel->num_poses * loadmodel->num_bones * sizeof(short[7]) + loadmodel->num_bones * sizeof(float[12]) + loadmodel->numskins * sizeof(animscene_t) + loadmodel->num_bones * sizeof(aliasbone_t) + loadmodel->numframes * sizeof(animscene_t) + ((loadmodel->surfmesh.num_vertices <= 65536) ? (loadmodel->surfmesh.num_triangles * sizeof(unsigned short[3])) : 0);
 	data = (unsigned char *)Mem_Alloc(loadmodel->mempool, size);
 	loadmodel->data_surfaces = (msurface_t *)data;data += loadmodel->num_surfaces * sizeof(msurface_t);
 	loadmodel->sortedmodelsurfaces = (int *)data;data += loadmodel->num_surfaces * sizeof(int);
 	loadmodel->data_textures = (texture_t *)data;data += loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t);
 	loadmodel->surfmesh.data_element3i = (int *)data;data += loadmodel->surfmesh.num_triangles * sizeof(int[3]);
-	loadmodel->surfmesh.data_neighbor3i = (int *)data;data += loadmodel->surfmesh.num_triangles * sizeof(int[3]);
+	if (r_enableshadowvolumes.integer)
+	{
+		loadmodel->surfmesh.data_neighbor3i = (int *)data;data += loadmodel->surfmesh.num_triangles * sizeof(int[3]);
+	}
 	loadmodel->surfmesh.data_vertex3f = (float *)data;data += loadmodel->surfmesh.num_vertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_svector3f = (float *)data;data += loadmodel->surfmesh.num_vertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_tvector3f = (float *)data;data += loadmodel->surfmesh.num_vertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_normal3f = (float *)data;data += loadmodel->surfmesh.num_vertices * sizeof(float[3]);
 	loadmodel->surfmesh.data_texcoordtexture2f = (float *)data;data += loadmodel->surfmesh.num_vertices * sizeof(float[2]);
-	loadmodel->surfmesh.data_vertexweightindex4i = (int *)data;data += loadmodel->surfmesh.num_vertices * sizeof(int[4]);
-	loadmodel->surfmesh.data_vertexweightinfluence4f = (float *)data;data += loadmodel->surfmesh.num_vertices * sizeof(float[4]);
-	loadmodel->data_poses = (float *)data;data += loadmodel->num_poses * loadmodel->num_bones * sizeof(float[12]);
+	loadmodel->surfmesh.data_skeletalindex4ub = (unsigned char *)data;data += loadmodel->surfmesh.num_vertices * sizeof(unsigned char[4]);
+	loadmodel->surfmesh.data_skeletalweight4ub = (unsigned char *)data;data += loadmodel->surfmesh.num_vertices * sizeof(unsigned char[4]);
 	loadmodel->data_baseboneposeinverse = (float *)data;data += loadmodel->num_bones * sizeof(float[12]);
 	loadmodel->skinscenes = (animscene_t *)data;data += loadmodel->numskins * sizeof(animscene_t);
 	loadmodel->data_bones = (aliasbone_t *)data;data += loadmodel->num_bones * sizeof(aliasbone_t);
 	loadmodel->animscenes = (animscene_t *)data;data += loadmodel->numframes * sizeof(animscene_t);
+	loadmodel->surfmesh.num_blends = 0;
+	loadmodel->surfmesh.blends = (unsigned short *)data;data += meshvertices * sizeof(unsigned short);
 	if (loadmodel->surfmesh.num_vertices <= 65536)
+	{
 		loadmodel->surfmesh.data_element3s = (unsigned short *)data;data += loadmodel->surfmesh.num_triangles * sizeof(unsigned short[3]);
+	}
+	loadmodel->data_poses7s = (short *)data;data += loadmodel->num_poses * loadmodel->num_bones * sizeof(short[7]);
+	loadmodel->surfmesh.data_blendweights = (blendweights_t *)Mem_Alloc(loadmodel->mempool, loadmodel->surfmesh.num_vertices * sizeof(blendweights_t));
 
 	for (i = 0;i < loadmodel->numskins;i++)
 	{
@@ -2794,12 +3069,37 @@ void Mod_PSKMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 			Host_Error("%s bone[%i].parent >= %i", loadmodel->name, index, index);
 	}
 
+	// convert the basepose data
+	if (loadmodel->num_bones)
+	{
+		int boneindex;
+		matrix4x4_t *basebonepose;
+		float *outinvmatrix = loadmodel->data_baseboneposeinverse;
+		matrix4x4_t bonematrix;
+		matrix4x4_t tempbonematrix;
+		basebonepose = (matrix4x4_t *)Mem_Alloc(tempmempool, loadmodel->num_bones * sizeof(matrix4x4_t));
+		for (boneindex = 0;boneindex < loadmodel->num_bones;boneindex++)
+		{
+			Matrix4x4_FromOriginQuat(&bonematrix, bones[boneindex].basepose.origin[0], bones[boneindex].basepose.origin[1], bones[boneindex].basepose.origin[2], bones[boneindex].basepose.quat[0], bones[boneindex].basepose.quat[1], bones[boneindex].basepose.quat[2], bones[boneindex].basepose.quat[3]);
+			if (loadmodel->data_bones[boneindex].parent >= 0)
+			{
+				tempbonematrix = bonematrix;
+				Matrix4x4_Concat(&bonematrix, basebonepose + loadmodel->data_bones[boneindex].parent, &tempbonematrix);
+			}
+			basebonepose[boneindex] = bonematrix;
+			Matrix4x4_Invert_Simple(&tempbonematrix, basebonepose + boneindex);
+			Matrix4x4_ToArray12FloatD3D(&tempbonematrix, outinvmatrix + 12*boneindex);
+		}
+		Mem_Free(basebonepose);
+	}
+
 	// sort the psk point weights into the vertex weight tables
 	// (which only accept up to 4 bones per vertex)
 	for (index = 0;index < numvtxw;index++)
 	{
+		int weightindex[4] = { 0, 0, 0, 0 };
+		float weightinfluence[4] = { 0, 0, 0, 0 };
 		int l;
-		float sum;
 		for (j = 0;j < numrawweights;j++)
 		{
 			if (rawweights[j].pntsindex == vtxw[index].pntsindex)
@@ -2808,70 +3108,929 @@ void Mod_PSKMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
 				float influence = rawweights[j].weight;
 				for (l = 0;l < 4;l++)
 				{
-					if (loadmodel->surfmesh.data_vertexweightinfluence4f[index*4+l] < influence)
+					if (weightinfluence[l] < influence)
 					{
 						// move lower influence weights out of the way first
 						int l2;
 						for (l2 = 3;l2 > l;l2--)
 						{
-							loadmodel->surfmesh.data_vertexweightinfluence4f[index*4+l2] = loadmodel->surfmesh.data_vertexweightinfluence4f[index*4+l2-1];
-							loadmodel->surfmesh.data_vertexweightindex4i[index*4+l2] = loadmodel->surfmesh.data_vertexweightindex4i[index*4+l2-1];
+							weightinfluence[l2] = weightinfluence[l2-1];
+							weightindex[l2] = weightindex[l2-1];
 						}
 						// store the new weight
-						loadmodel->surfmesh.data_vertexweightinfluence4f[index*4+l] = influence;
-						loadmodel->surfmesh.data_vertexweightindex4i[index*4+l] = boneindex;
+						weightinfluence[l] = influence;
+						weightindex[l] = boneindex;
 						break;
 					}
 				}
 			}
 		}
-		sum = 0;
-		for (l = 0;l < 4;l++)
-			sum += loadmodel->surfmesh.data_vertexweightinfluence4f[index*4+l];
-		if (sum && fabs(sum - 1) > (1.0f / 256.0f))
+		loadmodel->surfmesh.blends[index] = Mod_Skeletal_CompressBlend(loadmodel, weightindex, weightinfluence);
+		loadmodel->surfmesh.data_skeletalindex4ub[index*4  ] = weightindex[0];
+		loadmodel->surfmesh.data_skeletalindex4ub[index*4+1] = weightindex[1];
+		loadmodel->surfmesh.data_skeletalindex4ub[index*4+2] = weightindex[2];
+		loadmodel->surfmesh.data_skeletalindex4ub[index*4+3] = weightindex[3];
+		loadmodel->surfmesh.data_skeletalweight4ub[index*4  ] = (unsigned char)(weightinfluence[0]*255.0f);
+		loadmodel->surfmesh.data_skeletalweight4ub[index*4+1] = (unsigned char)(weightinfluence[1]*255.0f);
+		loadmodel->surfmesh.data_skeletalweight4ub[index*4+2] = (unsigned char)(weightinfluence[2]*255.0f);
+		loadmodel->surfmesh.data_skeletalweight4ub[index*4+3] = (unsigned char)(weightinfluence[3]*255.0f);
+	}
+	if (loadmodel->surfmesh.num_blends < loadmodel->surfmesh.num_vertices)
+		loadmodel->surfmesh.data_blendweights = (blendweights_t *)Mem_Realloc(loadmodel->mempool, loadmodel->surfmesh.data_blendweights, loadmodel->surfmesh.num_blends * sizeof(blendweights_t));
+
+	// set up the animscenes based on the anims
+	if (numanims)
+	{
+		for (index = 0, i = 0;index < numanims;index++)
+		{
+			for (j = 0;j < anims[index].numframes;j++, i++)
+			{
+				dpsnprintf(loadmodel->animscenes[i].name, sizeof(loadmodel->animscenes[i].name), "%s_%d", anims[index].name, j);
+				loadmodel->animscenes[i].firstframe = i;
+				loadmodel->animscenes[i].framecount = 1;
+				loadmodel->animscenes[i].loop = true;
+				loadmodel->animscenes[i].framerate = anims[index].fps;
+			}
+		}
+		// calculate the scaling value for bone origins so they can be compressed to short
+		biggestorigin = 0;
+		for (index = 0;index < numanimkeys;index++)
+		{
+			pskanimkeys_t *k = animkeys + index;
+			biggestorigin = max(biggestorigin, fabs(k->origin[0]));
+			biggestorigin = max(biggestorigin, fabs(k->origin[1]));
+			biggestorigin = max(biggestorigin, fabs(k->origin[2]));
+		}
+		loadmodel->num_posescale = biggestorigin / 32767.0f;
+		loadmodel->num_poseinvscale = 1.0f / loadmodel->num_posescale;
+	
+		// load the poses from the animkeys
+		for (index = 0;index < numanimkeys;index++)
+		{
+			pskanimkeys_t *k = animkeys + index;
+			float quat[4];
+			Vector4Copy(k->quat, quat);
+			if (quat[3] > 0)
+				Vector4Negate(quat, quat);
+			Vector4Normalize2(quat, quat);
+			// compress poses to the short[7] format for longterm storage
+			loadmodel->data_poses7s[index*7+0] = k->origin[0] * loadmodel->num_poseinvscale;
+			loadmodel->data_poses7s[index*7+1] = k->origin[1] * loadmodel->num_poseinvscale;
+			loadmodel->data_poses7s[index*7+2] = k->origin[2] * loadmodel->num_poseinvscale;
+			loadmodel->data_poses7s[index*7+3] = quat[0] * 32767.0f;
+			loadmodel->data_poses7s[index*7+4] = quat[1] * 32767.0f;
+			loadmodel->data_poses7s[index*7+5] = quat[2] * 32767.0f;
+			loadmodel->data_poses7s[index*7+6] = quat[3] * 32767.0f;
+		}
+	}
+	else
+	{
+		strlcpy(loadmodel->animscenes[0].name, "base", sizeof(loadmodel->animscenes[0].name));
+		loadmodel->animscenes[0].firstframe = 0;
+		loadmodel->animscenes[0].framecount = 1;
+		loadmodel->animscenes[0].loop = true;
+		loadmodel->animscenes[0].framerate = 10;
+
+		// calculate the scaling value for bone origins so they can be compressed to short
+		biggestorigin = 0;
+		for (index = 0;index < numbones;index++)
+		{
+			pskboneinfo_t *p = bones + index;
+			biggestorigin = max(biggestorigin, fabs(p->basepose.origin[0]));
+			biggestorigin = max(biggestorigin, fabs(p->basepose.origin[1]));
+			biggestorigin = max(biggestorigin, fabs(p->basepose.origin[2]));
+		}
+		loadmodel->num_posescale = biggestorigin / 32767.0f;
+		loadmodel->num_poseinvscale = 1.0f / loadmodel->num_posescale;
+	
+		// load the basepose as a frame
+		for (index = 0;index < numbones;index++)
+		{
+			pskboneinfo_t *p = bones + index;
+			float quat[4];
+			Vector4Copy(p->basepose.quat, quat);
+			if (quat[3] > 0)
+				Vector4Negate(quat, quat);
+			Vector4Normalize2(quat, quat);
+			// compress poses to the short[7] format for longterm storage
+			loadmodel->data_poses7s[index*7+0] = p->basepose.origin[0] * loadmodel->num_poseinvscale;
+			loadmodel->data_poses7s[index*7+1] = p->basepose.origin[1] * loadmodel->num_poseinvscale;
+			loadmodel->data_poses7s[index*7+2] = p->basepose.origin[2] * loadmodel->num_poseinvscale;
+			loadmodel->data_poses7s[index*7+3] = quat[0] * 32767.0f;
+			loadmodel->data_poses7s[index*7+4] = quat[1] * 32767.0f;
+			loadmodel->data_poses7s[index*7+5] = quat[2] * 32767.0f;
+			loadmodel->data_poses7s[index*7+6] = quat[3] * 32767.0f;
+		}
+	}
+
+	Mod_FreeSkinFiles(skinfiles);
+	if (animfilebuffer)
+		Mem_Free(animfilebuffer);
+	Mod_MakeSortedSurfaces(loadmodel);
+
+	// compute all the mesh information that was not loaded from the file
+	// TODO: honor smoothing groups somehow?
+	if (loadmodel->surfmesh.data_element3s)
+		for (i = 0;i < loadmodel->surfmesh.num_triangles*3;i++)
+			loadmodel->surfmesh.data_element3s[i] = loadmodel->surfmesh.data_element3i[i];
+	Mod_ValidateElements(loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles, 0, loadmodel->surfmesh.num_vertices, __FILE__, __LINE__);
+	Mod_BuildNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_normal3f, r_smoothnormals_areaweighting.integer != 0);
+	Mod_BuildTextureVectorsFromNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_texcoordtexture2f, loadmodel->surfmesh.data_normal3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_svector3f, loadmodel->surfmesh.data_tvector3f, r_smoothnormals_areaweighting.integer != 0);
+	if (loadmodel->surfmesh.data_neighbor3i)
+		Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
+	loadmodel->surfmesh.isanimated = Mod_Alias_CalculateBoundingBox();
+	if(mod_alias_force_animated.string[0])
+		loadmodel->surfmesh.isanimated = mod_alias_force_animated.integer != 0;
+
+	// Always make a BIH for the first frame, we can use it where possible.
+	Mod_MakeCollisionBIH(loadmodel, true, &loadmodel->collision_bih);
+	if (!loadmodel->surfmesh.isanimated)
+	{
+		loadmodel->TraceBox = Mod_CollisionBIH_TraceBox;
+		loadmodel->TraceBrush = Mod_CollisionBIH_TraceBrush;
+		loadmodel->TraceLine = Mod_CollisionBIH_TraceLine;
+		loadmodel->TracePoint = Mod_CollisionBIH_TracePoint_Mesh;
+		loadmodel->PointSuperContents = Mod_CollisionBIH_PointSuperContents_Mesh;
+	}
+
+	// because shaders can do somewhat unexpected things, check for unusual features now
+	for (i = 0;i < loadmodel->num_textures;i++)
+	{
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_SKY))
+			mod->DrawSky = R_Q1BSP_DrawSky;
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_WATERSHADER | MATERIALFLAG_REFRACTION | MATERIALFLAG_REFLECTION | MATERIALFLAG_CAMERA))
+			mod->DrawAddWaterPlanes = R_Q1BSP_DrawAddWaterPlanes;
+	}
+}
+
+void Mod_INTERQUAKEMODEL_Load(dp_model_t *mod, void *buffer, void *bufferend)
+{
+	unsigned char *data;
+	const char *text;
+	const unsigned char *pbase, *pend;
+	iqmheader_t header;
+	skinfile_t *skinfiles;
+	int i, j, k, meshvertices, meshtriangles;
+	float biggestorigin;
+	const unsigned int *inelements;
+	int *outelements;
+	const int *inneighbors;
+	int *outneighbors;
+	float *outvertex, *outnormal, *outtexcoord, *outsvector, *outtvector, *outcolor;
+	// this pointers into the file data are read only through Little* functions so they can be unaligned memory
+	const float *vnormal = NULL;
+	const float *vposition = NULL;
+	const float *vtangent = NULL;
+	const float *vtexcoord = NULL;
+	const float *vcolor4f = NULL;
+	const unsigned char *vblendindexes = NULL;
+	const unsigned char *vblendweights = NULL;
+	const unsigned char *vcolor4ub = NULL;
+	const unsigned short *framedata = NULL;
+	// temporary memory allocations (because the data in the file may be misaligned)
+	iqmanim_t *anims = NULL;
+	iqmbounds_t *bounds = NULL;
+	iqmjoint1_t *joint1 = NULL;
+	iqmjoint_t *joint = NULL;
+	iqmmesh_t *meshes = NULL;
+	iqmpose1_t *pose1 = NULL;
+	iqmpose_t *pose = NULL;
+	iqmvertexarray_t *vas = NULL;
+
+	pbase = (unsigned char *)buffer;
+	pend = (unsigned char *)bufferend;
+
+	if (pbase + sizeof(iqmheader_t) > pend)
+		Host_Error ("Mod_INTERQUAKEMODEL_Load: %s is not an Inter-Quake Model %d", loadmodel->name, (int)(pend - pbase));
+
+	// copy struct (otherwise it may be misaligned)
+	// LordHavoc: okay it's definitely not misaligned here, but for consistency...
+	memcpy(&header, pbase, sizeof(iqmheader_t));
+
+	if (memcmp(header.id, "INTERQUAKEMODEL", 16))
+		Host_Error ("Mod_INTERQUAKEMODEL_Load: %s is not an Inter-Quake Model", loadmodel->name);
+	if (LittleLong(header.version) != 1 && LittleLong(header.version) != 2)
+		Host_Error ("Mod_INTERQUAKEMODEL_Load: only version 1 and 2 models are currently supported (name = %s)", loadmodel->name);
+
+	loadmodel->modeldatatypestring = "IQM";
+
+	loadmodel->type = mod_alias;
+	loadmodel->synctype = ST_RAND;
+
+	// byteswap header
+	header.version = LittleLong(header.version);
+	header.filesize = LittleLong(header.filesize);
+	header.flags = LittleLong(header.flags);
+	header.num_text = LittleLong(header.num_text);
+	header.ofs_text = LittleLong(header.ofs_text);
+	header.num_meshes = LittleLong(header.num_meshes);
+	header.ofs_meshes = LittleLong(header.ofs_meshes);
+	header.num_vertexarrays = LittleLong(header.num_vertexarrays);
+	header.num_vertexes = LittleLong(header.num_vertexes);
+	header.ofs_vertexarrays = LittleLong(header.ofs_vertexarrays);
+	header.num_triangles = LittleLong(header.num_triangles);
+	header.ofs_triangles = LittleLong(header.ofs_triangles);
+	header.ofs_neighbors = LittleLong(header.ofs_neighbors);
+	header.num_joints = LittleLong(header.num_joints);
+	header.ofs_joints = LittleLong(header.ofs_joints);
+	header.num_poses = LittleLong(header.num_poses);
+	header.ofs_poses = LittleLong(header.ofs_poses);
+	header.num_anims = LittleLong(header.num_anims);
+	header.ofs_anims = LittleLong(header.ofs_anims);
+	header.num_frames = LittleLong(header.num_frames);
+	header.num_framechannels = LittleLong(header.num_framechannels);
+	header.ofs_frames = LittleLong(header.ofs_frames);
+	header.ofs_bounds = LittleLong(header.ofs_bounds);
+	header.num_comment = LittleLong(header.num_comment);
+	header.ofs_comment = LittleLong(header.ofs_comment);
+	header.num_extensions = LittleLong(header.num_extensions);
+	header.ofs_extensions = LittleLong(header.ofs_extensions);
+
+	if (header.version == 1)
+	{
+		if (pbase + header.ofs_joints + header.num_joints*sizeof(iqmjoint1_t) > pend ||
+			pbase + header.ofs_poses + header.num_poses*sizeof(iqmpose1_t) > pend)
+		{
+			Con_Printf("%s has invalid size or offset information\n", loadmodel->name);
+			return;
+		}
+	}
+	else
+	{
+		if (pbase + header.ofs_joints + header.num_joints*sizeof(iqmjoint_t) > pend ||
+			pbase + header.ofs_poses + header.num_poses*sizeof(iqmpose_t) > pend)
+		{
+			Con_Printf("%s has invalid size or offset information\n", loadmodel->name);
+			return;
+		}
+	}
+	if (pbase + header.ofs_text + header.num_text > pend ||
+		pbase + header.ofs_meshes + header.num_meshes*sizeof(iqmmesh_t) > pend ||
+		pbase + header.ofs_vertexarrays + header.num_vertexarrays*sizeof(iqmvertexarray_t) > pend ||
+		pbase + header.ofs_triangles + header.num_triangles*sizeof(int[3]) > pend ||
+		(header.ofs_neighbors && pbase + header.ofs_neighbors + header.num_triangles*sizeof(int[3]) > pend) ||
+		pbase + header.ofs_anims + header.num_anims*sizeof(iqmanim_t) > pend ||
+		pbase + header.ofs_frames + header.num_frames*header.num_framechannels*sizeof(unsigned short) > pend ||
+		(header.ofs_bounds && pbase + header.ofs_bounds + header.num_frames*sizeof(iqmbounds_t) > pend) ||
+		pbase + header.ofs_comment + header.num_comment > pend)
+	{
+		Con_Printf("%s has invalid size or offset information\n", loadmodel->name);
+		return;
+	}
+
+	// copy structs to make them aligned in memory (otherwise we crash on Sparc and PowerPC and others)
+	if (header.num_vertexarrays)
+		vas = (iqmvertexarray_t *)(pbase + header.ofs_vertexarrays);
+	if (header.num_anims)
+		anims = (iqmanim_t *)(pbase + header.ofs_anims);
+	if (header.ofs_bounds)
+		bounds = (iqmbounds_t *)(pbase + header.ofs_bounds);
+	if (header.num_meshes)
+		meshes = (iqmmesh_t *)(pbase + header.ofs_meshes);
+
+	for (i = 0;i < (int)header.num_vertexarrays;i++)
+	{
+		iqmvertexarray_t va;
+		size_t vsize;
+		va.type = LittleLong(vas[i].type);
+		va.flags = LittleLong(vas[i].flags);
+		va.format = LittleLong(vas[i].format);
+		va.size = LittleLong(vas[i].size);
+		va.offset = LittleLong(vas[i].offset);
+		vsize = header.num_vertexes*va.size;
+		switch (va.format)
+		{ 
+		case IQM_FLOAT: vsize *= sizeof(float); break;
+		case IQM_UBYTE: vsize *= sizeof(unsigned char); break;
+		default: continue;
+		}
+		if (pbase + va.offset + vsize > pend)
+			continue;
+		// no need to copy the vertex data for alignment because LittleLong/LittleShort will be invoked on reading them, and the destination is aligned
+		switch (va.type)
+		{
+		case IQM_POSITION:
+			if (va.format == IQM_FLOAT && va.size == 3)
+				vposition = (const float *)(pbase + va.offset);
+			break;
+		case IQM_TEXCOORD:
+			if (va.format == IQM_FLOAT && va.size == 2)
+				vtexcoord = (const float *)(pbase + va.offset);
+			break;
+		case IQM_NORMAL:
+			if (va.format == IQM_FLOAT && va.size == 3)
+				vnormal = (const float *)(pbase + va.offset);
+			break;
+		case IQM_TANGENT:
+			if (va.format == IQM_FLOAT && va.size == 4)
+				vtangent = (const float *)(pbase + va.offset);
+			break;
+		case IQM_BLENDINDEXES:
+			if (va.format == IQM_UBYTE && va.size == 4)
+				vblendindexes = (const unsigned char *)(pbase + va.offset);
+			break;
+		case IQM_BLENDWEIGHTS:
+			if (va.format == IQM_UBYTE && va.size == 4)
+				vblendweights = (const unsigned char *)(pbase + va.offset);
+			break;
+		case IQM_COLOR:
+			if (va.format == IQM_FLOAT && va.size == 4)
+				vcolor4f = (const float *)(pbase + va.offset);
+			if (va.format == IQM_UBYTE && va.size == 4)
+				vcolor4ub = (const unsigned char *)(pbase + va.offset);
+			break;
+		}
+	}
+	if (header.num_vertexes > 0 && (!vposition || !vtexcoord || ((header.num_frames > 0 || header.num_anims > 0) && (!vblendindexes || !vblendweights))))
+	{
+		Con_Printf("%s is missing vertex array data\n", loadmodel->name);
+		return;
+	}
+
+	text = header.num_text && header.ofs_text ? (const char *)(pbase + header.ofs_text) : "";
+
+	loadmodel->DrawSky = NULL;
+	loadmodel->DrawAddWaterPlanes = NULL;
+	loadmodel->Draw = R_Q1BSP_Draw;
+	loadmodel->DrawDepth = R_Q1BSP_DrawDepth;
+	loadmodel->DrawDebug = R_Q1BSP_DrawDebug;
+	loadmodel->DrawPrepass = R_Q1BSP_DrawPrepass;
+	loadmodel->CompileShadowMap = R_Q1BSP_CompileShadowMap;
+	loadmodel->DrawShadowMap = R_Q1BSP_DrawShadowMap;
+	loadmodel->CompileShadowVolume = R_Q1BSP_CompileShadowVolume;
+	loadmodel->DrawShadowVolume = R_Q1BSP_DrawShadowVolume;
+	loadmodel->DrawLight = R_Q1BSP_DrawLight;
+	loadmodel->TraceBox = Mod_MDLMD2MD3_TraceBox;
+	loadmodel->TraceLine = Mod_MDLMD2MD3_TraceLine;
+	loadmodel->PointSuperContents = NULL;
+	loadmodel->AnimateVertices = Mod_Skeletal_AnimateVertices;
+
+	// load external .skin files if present
+	skinfiles = Mod_LoadSkinFiles();
+	if (loadmodel->numskins < 1)
+		loadmodel->numskins = 1;
+
+	loadmodel->numframes = max(header.num_anims, 1);
+	loadmodel->num_bones = header.num_joints;
+	loadmodel->num_poses = max(header.num_frames, 1);
+	loadmodel->nummodelsurfaces = loadmodel->num_surfaces = header.num_meshes;
+	loadmodel->num_textures = loadmodel->num_surfaces * loadmodel->numskins;
+	loadmodel->num_texturesperskin = loadmodel->num_surfaces;
+
+	meshvertices = header.num_vertexes;
+	meshtriangles = header.num_triangles;
+
+	// do most allocations as one merged chunk
+	data = (unsigned char *)Mem_Alloc(loadmodel->mempool, loadmodel->num_surfaces * sizeof(msurface_t) + loadmodel->num_surfaces * sizeof(int) + loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t) + meshtriangles * sizeof(int[3]) + (meshvertices <= 65536 ? meshtriangles * sizeof(unsigned short[3]) : 0) + (r_enableshadowvolumes.integer ? meshtriangles * sizeof(int[3]) : 0) + meshvertices * (sizeof(float[14]) + (vcolor4f || vcolor4ub ? sizeof(float[4]) : 0)) + (vblendindexes && vblendweights ? meshvertices * (sizeof(unsigned short) + sizeof(unsigned char[2][4])) : 0) + loadmodel->num_poses * loadmodel->num_bones * sizeof(short[7]) + loadmodel->num_bones * sizeof(float[12]) + loadmodel->numskins * sizeof(animscene_t) + loadmodel->num_bones * sizeof(aliasbone_t) + loadmodel->numframes * sizeof(animscene_t));
+	loadmodel->data_surfaces = (msurface_t *)data;data += loadmodel->num_surfaces * sizeof(msurface_t);
+	loadmodel->sortedmodelsurfaces = (int *)data;data += loadmodel->num_surfaces * sizeof(int);
+	loadmodel->data_textures = (texture_t *)data;data += loadmodel->num_surfaces * loadmodel->numskins * sizeof(texture_t);
+	loadmodel->surfmesh.num_vertices = meshvertices;
+	loadmodel->surfmesh.num_triangles = meshtriangles;
+	loadmodel->surfmesh.data_element3i = (int *)data;data += meshtriangles * sizeof(int[3]);
+	if (r_enableshadowvolumes.integer)
+	{
+		loadmodel->surfmesh.data_neighbor3i = (int *)data;data += meshtriangles * sizeof(int[3]);
+	}
+	loadmodel->surfmesh.data_vertex3f = (float *)data;data += meshvertices * sizeof(float[3]);
+	loadmodel->surfmesh.data_svector3f = (float *)data;data += meshvertices * sizeof(float[3]);
+	loadmodel->surfmesh.data_tvector3f = (float *)data;data += meshvertices * sizeof(float[3]);
+	loadmodel->surfmesh.data_normal3f = (float *)data;data += meshvertices * sizeof(float[3]);
+	loadmodel->surfmesh.data_texcoordtexture2f = (float *)data;data += meshvertices * sizeof(float[2]);
+	if (vcolor4f || vcolor4ub)
+	{
+		loadmodel->surfmesh.data_lightmapcolor4f = (float *)data;data += meshvertices * sizeof(float[4]);
+	}
+	if (vblendindexes && vblendweights)
+	{
+		loadmodel->surfmesh.data_skeletalindex4ub = (unsigned char *)data;data += meshvertices * sizeof(unsigned char[4]);
+		loadmodel->surfmesh.data_skeletalweight4ub = (unsigned char *)data;data += meshvertices * sizeof(unsigned char[4]);
+	}
+	loadmodel->data_baseboneposeinverse = (float *)data;data += loadmodel->num_bones * sizeof(float[12]);
+	loadmodel->skinscenes = (animscene_t *)data;data += loadmodel->numskins * sizeof(animscene_t);
+	loadmodel->data_bones = (aliasbone_t *)data;data += loadmodel->num_bones * sizeof(aliasbone_t);
+	loadmodel->animscenes = (animscene_t *)data;data += loadmodel->numframes * sizeof(animscene_t);
+	if (vblendindexes && vblendweights)
+	{
+		loadmodel->surfmesh.num_blends = 0;
+		loadmodel->surfmesh.blends = (unsigned short *)data;data += meshvertices * sizeof(unsigned short);
+	}
+	if (meshvertices <= 65536)
+	{
+		loadmodel->surfmesh.data_element3s = (unsigned short *)data;data += meshtriangles * sizeof(unsigned short[3]);
+	}
+	loadmodel->data_poses7s = (short *)data;data += loadmodel->num_poses * loadmodel->num_bones * sizeof(short[7]);
+	if (vblendindexes && vblendweights)
+		loadmodel->surfmesh.data_blendweights = (blendweights_t *)Mem_Alloc(loadmodel->mempool, meshvertices * sizeof(blendweights_t));
+
+	for (i = 0;i < loadmodel->numskins;i++)
+	{
+		loadmodel->skinscenes[i].firstframe = i;
+		loadmodel->skinscenes[i].framecount = 1;
+		loadmodel->skinscenes[i].loop = true;
+		loadmodel->skinscenes[i].framerate = 10;
+	}
+
+	// load the bone info
+	if (header.version == 1)
+	{
+		iqmjoint1_t *injoint1 = (iqmjoint1_t *)(pbase + header.ofs_joints);
+		if (loadmodel->num_bones)
+			joint1 = (iqmjoint1_t *)Mem_Alloc(loadmodel->mempool, loadmodel->num_bones * sizeof(iqmjoint1_t));
+		for (i = 0;i < loadmodel->num_bones;i++)
+		{
+			matrix4x4_t relbase, relinvbase, pinvbase, invbase;
+			joint1[i].name = LittleLong(injoint1[i].name);
+			joint1[i].parent = LittleLong(injoint1[i].parent);
+			for (j = 0;j < 3;j++)
+			{
+				joint1[i].origin[j] = LittleFloat(injoint1[i].origin[j]);
+				joint1[i].rotation[j] = LittleFloat(injoint1[i].rotation[j]);
+				joint1[i].scale[j] = LittleFloat(injoint1[i].scale[j]);
+			}
+			strlcpy(loadmodel->data_bones[i].name, &text[joint1[i].name], sizeof(loadmodel->data_bones[i].name));
+			loadmodel->data_bones[i].parent = joint1[i].parent;
+			if (loadmodel->data_bones[i].parent >= i)
+				Host_Error("%s bone[%i].parent >= %i", loadmodel->name, i, i);
+			Matrix4x4_FromDoom3Joint(&relbase, joint1[i].origin[0], joint1[i].origin[1], joint1[i].origin[2], joint1[i].rotation[0], joint1[i].rotation[1], joint1[i].rotation[2]);
+			Matrix4x4_Invert_Simple(&relinvbase, &relbase);
+			if (loadmodel->data_bones[i].parent >= 0)
+			{
+				Matrix4x4_FromArray12FloatD3D(&pinvbase, loadmodel->data_baseboneposeinverse + 12*loadmodel->data_bones[i].parent);
+				Matrix4x4_Concat(&invbase, &relinvbase, &pinvbase);
+				Matrix4x4_ToArray12FloatD3D(&invbase, loadmodel->data_baseboneposeinverse + 12*i);
+			}
+			else Matrix4x4_ToArray12FloatD3D(&relinvbase, loadmodel->data_baseboneposeinverse + 12*i);
+		}
+	}
+	else
+	{
+		iqmjoint_t *injoint = (iqmjoint_t *)(pbase + header.ofs_joints);
+		if (header.num_joints)
+			joint = (iqmjoint_t *)Mem_Alloc(loadmodel->mempool, loadmodel->num_bones * sizeof(iqmjoint_t));
+		for (i = 0;i < loadmodel->num_bones;i++)
 		{
-			float f = 1.0f / sum;
-			for (l = 0;l < 4;l++)
-				loadmodel->surfmesh.data_vertexweightinfluence4f[index*4+l] *= f;
+			matrix4x4_t relbase, relinvbase, pinvbase, invbase;
+			joint[i].name = LittleLong(injoint[i].name);
+			joint[i].parent = LittleLong(injoint[i].parent);
+			for (j = 0;j < 3;j++)
+			{
+				joint[i].origin[j] = LittleFloat(injoint[i].origin[j]);
+				joint[i].rotation[j] = LittleFloat(injoint[i].rotation[j]);
+				joint[i].scale[j] = LittleFloat(injoint[i].scale[j]);
+			}
+			joint[i].rotation[3] = LittleFloat(injoint[i].rotation[3]);
+			strlcpy(loadmodel->data_bones[i].name, &text[joint[i].name], sizeof(loadmodel->data_bones[i].name));
+			loadmodel->data_bones[i].parent = joint[i].parent;
+			if (loadmodel->data_bones[i].parent >= i)
+				Host_Error("%s bone[%i].parent >= %i", loadmodel->name, i, i);
+			if (joint[i].rotation[3] > 0)
+				Vector4Negate(joint[i].rotation, joint[i].rotation);
+			Vector4Normalize2(joint[i].rotation, joint[i].rotation);
+			Matrix4x4_FromDoom3Joint(&relbase, joint[i].origin[0], joint[i].origin[1], joint[i].origin[2], joint[i].rotation[0], joint[i].rotation[1], joint[i].rotation[2]);
+			Matrix4x4_Invert_Simple(&relinvbase, &relbase);
+			if (loadmodel->data_bones[i].parent >= 0)
+			{
+				Matrix4x4_FromArray12FloatD3D(&pinvbase, loadmodel->data_baseboneposeinverse + 12*loadmodel->data_bones[i].parent);
+				Matrix4x4_Concat(&invbase, &relinvbase, &pinvbase);
+				Matrix4x4_ToArray12FloatD3D(&invbase, loadmodel->data_baseboneposeinverse + 12*i);
+			}	
+			else Matrix4x4_ToArray12FloatD3D(&relinvbase, loadmodel->data_baseboneposeinverse + 12*i);
 		}
 	}
 
 	// set up the animscenes based on the anims
-	for (index = 0, i = 0;index < numanims;index++)
+	for (i = 0;i < (int)header.num_anims;i++)
+	{
+		iqmanim_t anim;
+		anim.name = LittleLong(anims[i].name);
+		anim.first_frame = LittleLong(anims[i].first_frame);
+		anim.num_frames = LittleLong(anims[i].num_frames);
+		anim.framerate = LittleFloat(anims[i].framerate);
+		anim.flags = LittleLong(anims[i].flags);
+		strlcpy(loadmodel->animscenes[i].name, &text[anim.name], sizeof(loadmodel->animscenes[i].name));
+		loadmodel->animscenes[i].firstframe = anim.first_frame;
+		loadmodel->animscenes[i].framecount = anim.num_frames;
+		loadmodel->animscenes[i].loop = ((anim.flags & IQM_LOOP) != 0);
+		loadmodel->animscenes[i].framerate = anim.framerate;
+	}
+	if (header.num_anims <= 0)
+	{
+		strlcpy(loadmodel->animscenes[0].name, "static", sizeof(loadmodel->animscenes[0].name));
+		loadmodel->animscenes[0].firstframe = 0;
+		loadmodel->animscenes[0].framecount = 1;
+		loadmodel->animscenes[0].loop = true;
+		loadmodel->animscenes[0].framerate = 10;
+	}
+
+	loadmodel->surfmesh.isanimated = loadmodel->num_bones > 1 || loadmodel->numframes > 1 || (loadmodel->animscenes && loadmodel->animscenes[0].framecount > 1);
+	if(mod_alias_force_animated.string[0])
+		loadmodel->surfmesh.isanimated = mod_alias_force_animated.integer != 0;
+
+	biggestorigin = 0;
+	if (header.version == 1)
+	{
+		iqmpose1_t *inpose1 = (iqmpose1_t *)(pbase + header.ofs_poses);
+		if (header.num_poses)
+			pose1 = (iqmpose1_t *)Mem_Alloc(loadmodel->mempool, header.num_poses * sizeof(iqmpose1_t));
+		for (i = 0;i < (int)header.num_poses;i++)
+		{
+			float f;
+			pose1[i].parent = LittleLong(inpose1[i].parent);
+			pose1[i].channelmask = LittleLong(inpose1[i].channelmask);
+			for (j = 0;j < 9;j++)
+			{
+				pose1[i].channeloffset[j] = LittleFloat(inpose1[i].channeloffset[j]);
+				pose1[i].channelscale[j] = LittleFloat(inpose1[i].channelscale[j]);
+			}
+			f = fabs(pose1[i].channeloffset[0]); biggestorigin = max(biggestorigin, f);
+			f = fabs(pose1[i].channeloffset[1]); biggestorigin = max(biggestorigin, f);
+			f = fabs(pose1[i].channeloffset[2]); biggestorigin = max(biggestorigin, f);
+			f = fabs(pose1[i].channeloffset[0] + 0xFFFF*pose1[i].channelscale[0]); biggestorigin = max(biggestorigin, f);
+			f = fabs(pose1[i].channeloffset[1] + 0xFFFF*pose1[i].channelscale[1]); biggestorigin = max(biggestorigin, f);
+			f = fabs(pose1[i].channeloffset[2] + 0xFFFF*pose1[i].channelscale[2]); biggestorigin = max(biggestorigin, f);
+		}
+		if (header.num_frames <= 0)
+		{
+			for (i = 0;i < loadmodel->num_bones;i++)
+			{
+				float f;
+				f = fabs(joint1[i].origin[0]); biggestorigin = max(biggestorigin, f);
+				f = fabs(joint1[i].origin[1]); biggestorigin = max(biggestorigin, f);
+				f = fabs(joint1[i].origin[2]); biggestorigin = max(biggestorigin, f);
+			}
+		}
+	}
+	else
 	{
-		for (j = 0;j < anims[index].numframes;j++, i++)
+		iqmpose_t *inpose = (iqmpose_t *)(pbase + header.ofs_poses);
+		if (header.num_poses)
+			pose = (iqmpose_t *)Mem_Alloc(loadmodel->mempool, header.num_poses * sizeof(iqmpose_t));
+		for (i = 0;i < (int)header.num_poses;i++)
+		{
+			float f;
+			pose[i].parent = LittleLong(inpose[i].parent);
+			pose[i].channelmask = LittleLong(inpose[i].channelmask);
+			for (j = 0;j < 10;j++)
+			{
+				pose[i].channeloffset[j] = LittleFloat(inpose[i].channeloffset[j]);
+				pose[i].channelscale[j] = LittleFloat(inpose[i].channelscale[j]);
+			}
+			f = fabs(pose[i].channeloffset[0]); biggestorigin = max(biggestorigin, f);
+			f = fabs(pose[i].channeloffset[1]); biggestorigin = max(biggestorigin, f);
+			f = fabs(pose[i].channeloffset[2]); biggestorigin = max(biggestorigin, f);
+			f = fabs(pose[i].channeloffset[0] + 0xFFFF*pose[i].channelscale[0]); biggestorigin = max(biggestorigin, f);
+			f = fabs(pose[i].channeloffset[1] + 0xFFFF*pose[i].channelscale[1]); biggestorigin = max(biggestorigin, f);
+			f = fabs(pose[i].channeloffset[2] + 0xFFFF*pose[i].channelscale[2]); biggestorigin = max(biggestorigin, f);
+		}
+		if (header.num_frames <= 0)
 		{
-			dpsnprintf(loadmodel->animscenes[i].name, sizeof(loadmodel->animscenes[i].name), "%s_%d", anims[index].name, j);
-			loadmodel->animscenes[i].firstframe = i;
-			loadmodel->animscenes[i].framecount = 1;
-			loadmodel->animscenes[i].loop = true;
-			loadmodel->animscenes[i].framerate = 10;
+			for (i = 0;i < loadmodel->num_bones;i++)
+			{
+				float f;
+				f = fabs(joint[i].origin[0]); biggestorigin = max(biggestorigin, f);
+				f = fabs(joint[i].origin[1]); biggestorigin = max(biggestorigin, f);
+				f = fabs(joint[i].origin[2]); biggestorigin = max(biggestorigin, f);
+			}
 		}
 	}
+	loadmodel->num_posescale = biggestorigin / 32767.0f;
+	loadmodel->num_poseinvscale = 1.0f / loadmodel->num_posescale;
 
-	// load the poses from the animkeys
-	for (index = 0;index < numanimkeys;index++)
+	// load the pose data
+	// this unaligned memory access is safe (LittleShort reads as bytes)
+	framedata = (const unsigned short *)(pbase + header.ofs_frames);
+	if (header.version == 1)
 	{
-		pskanimkeys_t *k = animkeys + index;
-		matrix4x4_t matrix;
-		Matrix4x4_FromOriginQuat(&matrix, k->origin[0], k->origin[1], k->origin[2], k->quat[0], k->quat[1], k->quat[2], k->quat[3]);
-		Matrix4x4_ToArray12FloatD3D(&matrix, loadmodel->data_poses + index*12);
+		for (i = 0, k = 0;i < (int)header.num_frames;i++)
+		{
+			for (j = 0;j < (int)header.num_poses;j++, k++)
+			{
+				float qx, qy, qz, qw;
+				loadmodel->data_poses7s[k*7 + 0] = loadmodel->num_poseinvscale * (pose1[j].channeloffset[0] + (pose1[j].channelmask&1 ? (unsigned short)LittleShort(*framedata++) * pose1[j].channelscale[0] : 0));
+				loadmodel->data_poses7s[k*7 + 1] = loadmodel->num_poseinvscale * (pose1[j].channeloffset[1] + (pose1[j].channelmask&2 ? (unsigned short)LittleShort(*framedata++) * pose1[j].channelscale[1] : 0));
+				loadmodel->data_poses7s[k*7 + 2] = loadmodel->num_poseinvscale * (pose1[j].channeloffset[2] + (pose1[j].channelmask&4 ? (unsigned short)LittleShort(*framedata++) * pose1[j].channelscale[2] : 0));
+				qx = pose1[j].channeloffset[3] + (pose1[j].channelmask&8 ? (unsigned short)LittleShort(*framedata++) * pose1[j].channelscale[3] : 0);
+				qy = pose1[j].channeloffset[4] + (pose1[j].channelmask&16 ? (unsigned short)LittleShort(*framedata++) * pose1[j].channelscale[4] : 0);
+				qz = pose1[j].channeloffset[5] + (pose1[j].channelmask&32 ? (unsigned short)LittleShort(*framedata++) * pose1[j].channelscale[5] : 0);
+				qw = 1.0f - (qx*qx + qy*qy + qz*qz);
+				qw = qw > 0.0f ? -sqrt(qw) : 0.0f;
+				loadmodel->data_poses7s[k*7 + 3] = 32767.0f * qx;
+				loadmodel->data_poses7s[k*7 + 4] = 32767.0f * qy;
+				loadmodel->data_poses7s[k*7 + 5] = 32767.0f * qz;
+				loadmodel->data_poses7s[k*7 + 6] = 32767.0f * qw;
+				// skip scale data for now
+				if(pose1[j].channelmask&64) framedata++;
+				if(pose1[j].channelmask&128) framedata++;
+				if(pose1[j].channelmask&256) framedata++;
+			}
+		}
+		if (header.num_frames <= 0)
+		{
+			for (i = 0;i < loadmodel->num_bones;i++)
+			{
+				float qx, qy, qz, qw;
+				loadmodel->data_poses7s[i*7 + 0] = loadmodel->num_poseinvscale * joint1[i].origin[0];
+				loadmodel->data_poses7s[i*7 + 1] = loadmodel->num_poseinvscale * joint1[i].origin[1];
+				loadmodel->data_poses7s[i*7 + 2] = loadmodel->num_poseinvscale * joint1[i].origin[2];
+				qx = joint1[i].rotation[0];
+				qy = joint1[i].rotation[1];
+				qz = joint1[i].rotation[2];
+				qw = 1.0f - (qx*qx + qy*qy + qz*qz);
+				qw = qw > 0.0f ? -sqrt(qw) : 0.0f;
+				loadmodel->data_poses7s[i*7 + 3] = 32767.0f * qx;
+				loadmodel->data_poses7s[i*7 + 4] = 32767.0f * qy;
+				loadmodel->data_poses7s[i*7 + 5] = 32767.0f * qz;
+				loadmodel->data_poses7s[i*7 + 6] = 32767.0f * qw;
+			}
+		}
 	}
+	else
+	{
+		for (i = 0, k = 0;i < (int)header.num_frames;i++)	
+		{
+			for (j = 0;j < (int)header.num_poses;j++, k++)
+			{
+				float rot[4];
+				loadmodel->data_poses7s[k*7 + 0] = loadmodel->num_poseinvscale * (pose[j].channeloffset[0] + (pose[j].channelmask&1 ? (unsigned short)LittleShort(*framedata++) * pose[j].channelscale[0] : 0));
+				loadmodel->data_poses7s[k*7 + 1] = loadmodel->num_poseinvscale * (pose[j].channeloffset[1] + (pose[j].channelmask&2 ? (unsigned short)LittleShort(*framedata++) * pose[j].channelscale[1] : 0));
+				loadmodel->data_poses7s[k*7 + 2] = loadmodel->num_poseinvscale * (pose[j].channeloffset[2] + (pose[j].channelmask&4 ? (unsigned short)LittleShort(*framedata++) * pose[j].channelscale[2] : 0));
+				rot[0] = pose[j].channeloffset[3] + (pose[j].channelmask&8 ? (unsigned short)LittleShort(*framedata++) * pose[j].channelscale[3] : 0);
+				rot[1] = pose[j].channeloffset[4] + (pose[j].channelmask&16 ? (unsigned short)LittleShort(*framedata++) * pose[j].channelscale[4] : 0);
+				rot[2] = pose[j].channeloffset[5] + (pose[j].channelmask&32 ? (unsigned short)LittleShort(*framedata++) * pose[j].channelscale[5] : 0);
+				rot[3] = pose[j].channeloffset[6] + (pose[j].channelmask&64 ? (unsigned short)LittleShort(*framedata++) * pose[j].channelscale[6] : 0);
+				if (rot[3] > 0)
+					Vector4Negate(rot, rot);
+				Vector4Normalize2(rot, rot);
+				loadmodel->data_poses7s[k*7 + 3] = 32767.0f * rot[0];
+				loadmodel->data_poses7s[k*7 + 4] = 32767.0f * rot[1];
+				loadmodel->data_poses7s[k*7 + 5] = 32767.0f * rot[2];
+				loadmodel->data_poses7s[k*7 + 6] = 32767.0f * rot[3];
+				// skip scale data for now
+				if(pose[j].channelmask&128) framedata++;
+				if(pose[j].channelmask&256) framedata++;
+				if(pose[j].channelmask&512) framedata++;
+			}
+		}
+		if (header.num_frames <= 0)
+		{
+			for (i = 0;i < loadmodel->num_bones;i++)
+			{
+				loadmodel->data_poses7s[i*7 + 0] = loadmodel->num_poseinvscale * joint[i].origin[0];
+				loadmodel->data_poses7s[i*7 + 1] = loadmodel->num_poseinvscale * joint[i].origin[1];
+				loadmodel->data_poses7s[i*7 + 2] = loadmodel->num_poseinvscale * joint[i].origin[2];
+				loadmodel->data_poses7s[i*7 + 3] = 32767.0f * joint[i].rotation[0];
+				loadmodel->data_poses7s[i*7 + 4] = 32767.0f * joint[i].rotation[1];
+				loadmodel->data_poses7s[i*7 + 5] = 32767.0f * joint[i].rotation[2];
+				loadmodel->data_poses7s[i*7 + 6] = 32767.0f * joint[i].rotation[3];
+			}
+		}
+	}
+
+	// load bounding box data
+	if (header.ofs_bounds)
+	{
+		float xyradius = 0, radius = 0;
+		VectorClear(loadmodel->normalmins);
+		VectorClear(loadmodel->normalmaxs);
+		for (i = 0; i < (int)header.num_frames;i++)
+		{
+			iqmbounds_t bound;
+			bound.mins[0] = LittleFloat(bounds[i].mins[0]);
+			bound.mins[1] = LittleFloat(bounds[i].mins[1]);
+			bound.mins[2] = LittleFloat(bounds[i].mins[2]);
+			bound.maxs[0] = LittleFloat(bounds[i].maxs[0]);			
+			bound.maxs[1] = LittleFloat(bounds[i].maxs[1]);	
+			bound.maxs[2] = LittleFloat(bounds[i].maxs[2]);	
+			bound.xyradius = LittleFloat(bounds[i].xyradius);
+			bound.radius = LittleFloat(bounds[i].radius);
+			if (!i)
+			{
+				VectorCopy(bound.mins, loadmodel->normalmins);
+				VectorCopy(bound.maxs, loadmodel->normalmaxs);
+			}
+			else
+			{
+				if (loadmodel->normalmins[0] > bound.mins[0]) loadmodel->normalmins[0] = bound.mins[0];
+				if (loadmodel->normalmins[1] > bound.mins[1]) loadmodel->normalmins[1] = bound.mins[1];
+				if (loadmodel->normalmins[2] > bound.mins[2]) loadmodel->normalmins[2] = bound.mins[2];
+				if (loadmodel->normalmaxs[0] < bound.maxs[0]) loadmodel->normalmaxs[0] = bound.maxs[0];
+				if (loadmodel->normalmaxs[1] < bound.maxs[1]) loadmodel->normalmaxs[1] = bound.maxs[1];
+				if (loadmodel->normalmaxs[2] < bound.maxs[2]) loadmodel->normalmaxs[2] = bound.maxs[2];
+			}
+			if (bound.xyradius > xyradius)
+				xyradius = bound.xyradius;
+			if (bound.radius > radius)
+				radius = bound.radius;
+		}
+		loadmodel->yawmins[0] = loadmodel->yawmins[1] = -xyradius;
+		loadmodel->yawmaxs[0] = loadmodel->yawmaxs[1] = xyradius;
+		loadmodel->yawmins[2] = loadmodel->normalmins[2];
+		loadmodel->yawmaxs[2] = loadmodel->normalmaxs[2];
+		loadmodel->rotatedmins[0] = loadmodel->rotatedmins[1] = loadmodel->rotatedmins[2] = -radius;
+		loadmodel->rotatedmaxs[0] = loadmodel->rotatedmaxs[1] = loadmodel->rotatedmaxs[2] = radius;
+		loadmodel->radius = radius;
+		loadmodel->radius2 = radius * radius;
+	}
+
+	// load triangle data
+	// this unaligned memory access is safe (LittleLong reads as bytes)
+	inelements = (const unsigned int *)(pbase + header.ofs_triangles);
+	outelements = loadmodel->surfmesh.data_element3i;
+	for (i = 0;i < (int)header.num_triangles;i++)
+	{
+		outelements[0] = LittleLong(inelements[0]);
+		outelements[1] = LittleLong(inelements[1]);
+		outelements[2] = LittleLong(inelements[2]);
+		outelements += 3;
+		inelements += 3;
+	}
+	Mod_ValidateElements(loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles, 0, header.num_vertexes, __FILE__, __LINE__);
+
+	if (header.ofs_neighbors && loadmodel->surfmesh.data_neighbor3i)
+	{
+		// this unaligned memory access is safe (LittleLong reads as bytes)
+		inneighbors = (const int *)(pbase + header.ofs_neighbors);
+		outneighbors = loadmodel->surfmesh.data_neighbor3i;
+		for (i = 0;i < (int)header.num_triangles;i++)
+		{
+			outneighbors[0] = LittleLong(inneighbors[0]);
+			outneighbors[1] = LittleLong(inneighbors[1]);
+			outneighbors[2] = LittleLong(inneighbors[2]);
+			outneighbors += 3;
+			inneighbors += 3;
+		}
+	}
+
+	// load vertex data
+	// this unaligned memory access is safe (LittleFloat reads as bytes)
+	outvertex = loadmodel->surfmesh.data_vertex3f;
+	for (i = 0;i < (int)header.num_vertexes;i++)
+	{
+		outvertex[0] = LittleFloat(vposition[0]);
+		outvertex[1] = LittleFloat(vposition[1]);
+		outvertex[2] = LittleFloat(vposition[2]);
+		vposition += 3;
+		outvertex += 3;
+	}
+
+	outtexcoord = loadmodel->surfmesh.data_texcoordtexture2f;
+	// this unaligned memory access is safe (LittleFloat reads as bytes)
+	for (i = 0;i < (int)header.num_vertexes;i++)
+	{
+		outtexcoord[0] = LittleFloat(vtexcoord[0]);
+		outtexcoord[1] = LittleFloat(vtexcoord[1]);
+		vtexcoord += 2;
+		outtexcoord += 2;
+	}
+
+	// this unaligned memory access is safe (LittleFloat reads as bytes)
+	if(vnormal)
+	{
+		outnormal = loadmodel->surfmesh.data_normal3f;
+		for (i = 0;i < (int)header.num_vertexes;i++)
+		{
+			outnormal[0] = LittleFloat(vnormal[0]);
+			outnormal[1] = LittleFloat(vnormal[1]);
+			outnormal[2] = LittleFloat(vnormal[2]);
+			vnormal += 3;
+			outnormal += 3;
+		}
+	}
+
+	// this unaligned memory access is safe (LittleFloat reads as bytes)
+	if(vnormal && vtangent)
+	{
+		outnormal = loadmodel->surfmesh.data_normal3f;
+		outsvector = loadmodel->surfmesh.data_svector3f;
+		outtvector = loadmodel->surfmesh.data_tvector3f;
+		for (i = 0;i < (int)header.num_vertexes;i++)
+		{
+			outsvector[0] = LittleFloat(vtangent[0]);
+			outsvector[1] = LittleFloat(vtangent[1]);
+			outsvector[2] = LittleFloat(vtangent[2]);
+			if(LittleFloat(vtangent[3]) < 0)
+				CrossProduct(outsvector, outnormal, outtvector);
+			else
+				CrossProduct(outnormal, outsvector, outtvector);
+			vtangent += 4;
+			outnormal += 3;
+			outsvector += 3;
+			outtvector += 3;
+		}
+	}
+
+	// this unaligned memory access is safe (all bytes)
+	if (vblendindexes && vblendweights)
+	{
+		for (i = 0; i < (int)header.num_vertexes;i++)
+		{
+			blendweights_t weights;
+			memcpy(weights.index, vblendindexes + i*4, 4);
+			memcpy(weights.influence, vblendweights + i*4, 4);
+			loadmodel->surfmesh.blends[i] = Mod_Skeletal_AddBlend(loadmodel, &weights);
+			loadmodel->surfmesh.data_skeletalindex4ub[i*4  ] = weights.index[0];
+			loadmodel->surfmesh.data_skeletalindex4ub[i*4+1] = weights.index[1];
+			loadmodel->surfmesh.data_skeletalindex4ub[i*4+2] = weights.index[2];
+			loadmodel->surfmesh.data_skeletalindex4ub[i*4+3] = weights.index[3];
+			loadmodel->surfmesh.data_skeletalweight4ub[i*4  ] = weights.influence[0];
+			loadmodel->surfmesh.data_skeletalweight4ub[i*4+1] = weights.influence[1];
+			loadmodel->surfmesh.data_skeletalweight4ub[i*4+2] = weights.influence[2];
+			loadmodel->surfmesh.data_skeletalweight4ub[i*4+3] = weights.influence[3];
+		}
+	}
+
+	if (vcolor4f)
+	{
+		outcolor = loadmodel->surfmesh.data_lightmapcolor4f;
+		// this unaligned memory access is safe (LittleFloat reads as bytes)
+		for (i = 0;i < (int)header.num_vertexes;i++)
+		{
+			outcolor[0] = LittleFloat(vcolor4f[0]);
+			outcolor[1] = LittleFloat(vcolor4f[1]);
+			outcolor[2] = LittleFloat(vcolor4f[2]);
+			outcolor[3] = LittleFloat(vcolor4f[3]);
+			vcolor4f += 4;
+			outcolor += 4;
+		}
+	}
+	else if (vcolor4ub)
+	{
+		outcolor = loadmodel->surfmesh.data_lightmapcolor4f;
+		// this unaligned memory access is safe (all bytes)
+		for (i = 0;i < (int)header.num_vertexes;i++)
+		{
+			outcolor[0] = vcolor4ub[0] * (1.0f / 255.0f);
+			outcolor[1] = vcolor4ub[1] * (1.0f / 255.0f);
+			outcolor[2] = vcolor4ub[2] * (1.0f / 255.0f);
+			outcolor[3] = vcolor4ub[3] * (1.0f / 255.0f);
+			vcolor4ub += 4;
+			outcolor += 4;
+		}
+	}
+
+	// load meshes
+	for (i = 0;i < (int)header.num_meshes;i++)
+	{
+		iqmmesh_t mesh;
+		msurface_t *surface;
+
+		mesh.name = LittleLong(meshes[i].name);
+		mesh.material = LittleLong(meshes[i].material);
+		mesh.first_vertex = LittleLong(meshes[i].first_vertex);
+		mesh.num_vertexes = LittleLong(meshes[i].num_vertexes);
+		mesh.first_triangle = LittleLong(meshes[i].first_triangle);
+		mesh.num_triangles = LittleLong(meshes[i].num_triangles);
+
+		loadmodel->sortedmodelsurfaces[i] = i;
+		surface = loadmodel->data_surfaces + i;
+		surface->texture = loadmodel->data_textures + i;
+		surface->num_firsttriangle = mesh.first_triangle;
+		surface->num_triangles = mesh.num_triangles;
+		surface->num_firstvertex = mesh.first_vertex;
+		surface->num_vertices = mesh.num_vertexes;
+
+		Mod_BuildAliasSkinsFromSkinFiles(loadmodel->data_textures + i, skinfiles, &text[mesh.name], &text[mesh.material]);
+	}
+
 	Mod_FreeSkinFiles(skinfiles);
-	Mem_Free(animfilebuffer);
 	Mod_MakeSortedSurfaces(loadmodel);
 
 	// compute all the mesh information that was not loaded from the file
-	// TODO: honor smoothing groups somehow?
 	if (loadmodel->surfmesh.data_element3s)
 		for (i = 0;i < loadmodel->surfmesh.num_triangles*3;i++)
 			loadmodel->surfmesh.data_element3s[i] = loadmodel->surfmesh.data_element3i[i];
-	Mod_ValidateElements(loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles, 0, loadmodel->surfmesh.num_vertices, __FILE__, __LINE__);
-	Mod_BuildBaseBonePoses();
-	Mod_BuildNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_normal3f, true);
-	Mod_BuildTextureVectorsFromNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_texcoordtexture2f, loadmodel->surfmesh.data_normal3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_svector3f, loadmodel->surfmesh.data_tvector3f, true);
-	Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
-	Mod_Alias_CalculateBoundingBox();
+	if (!vnormal)
+		Mod_BuildNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_normal3f, r_smoothnormals_areaweighting.integer != 0);
+	if (!vnormal || !vtangent)
+		Mod_BuildTextureVectorsFromNormals(0, loadmodel->surfmesh.num_vertices, loadmodel->surfmesh.num_triangles, loadmodel->surfmesh.data_vertex3f, loadmodel->surfmesh.data_texcoordtexture2f, loadmodel->surfmesh.data_normal3f, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.data_svector3f, loadmodel->surfmesh.data_tvector3f, r_smoothnormals_areaweighting.integer != 0);
+	if (!header.ofs_neighbors && loadmodel->surfmesh.data_neighbor3i)
+		Mod_BuildTriangleNeighbors(loadmodel->surfmesh.data_neighbor3i, loadmodel->surfmesh.data_element3i, loadmodel->surfmesh.num_triangles);
+	if (!header.ofs_bounds)
+		Mod_Alias_CalculateBoundingBox();
+
+	// Always make a BIH for the first frame, we can use it where possible.
+	Mod_MakeCollisionBIH(loadmodel, true, &loadmodel->collision_bih);
+	if (!loadmodel->surfmesh.isanimated)
+	{
+		loadmodel->TraceBox = Mod_CollisionBIH_TraceBox;
+		loadmodel->TraceBrush = Mod_CollisionBIH_TraceBrush;
+		loadmodel->TraceLine = Mod_CollisionBIH_TraceLine;
+		loadmodel->TracePoint = Mod_CollisionBIH_TracePoint_Mesh;
+		loadmodel->PointSuperContents = Mod_CollisionBIH_PointSuperContents_Mesh;
+	}
 
-	loadmodel->surfmesh.isanimated = loadmodel->numframes > 1 || loadmodel->animscenes[0].framecount > 1;
+	if (joint        ) Mem_Free(joint        );joint         = NULL;
+	if (joint1       ) Mem_Free(joint1       );joint1        = NULL;
+	if (pose         ) Mem_Free(pose         );pose          = NULL;
+	if (pose1        ) Mem_Free(pose1        );pose1         = NULL;
+
+	// because shaders can do somewhat unexpected things, check for unusual features now
+	for (i = 0;i < loadmodel->num_textures;i++)
+	{
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_SKY))
+			mod->DrawSky = R_Q1BSP_DrawSky;
+		if (loadmodel->data_textures[i].basematerialflags & (MATERIALFLAG_WATERSHADER | MATERIALFLAG_REFRACTION | MATERIALFLAG_REFLECTION | MATERIALFLAG_CAMERA))
+			mod->DrawAddWaterPlanes = R_Q1BSP_DrawAddWaterPlanes;
+	}
 }