Merge pull request #286 from dreamstalker/asmlib_update

Use SSE for some typical floating point operations

rehlds/engine/common.h

@@ -154,6 +154,9 @@ extern int loadsize;
 #define Q_atoi atoi
 #define Q_atof atof
 #define Q_sqrt M_sqrt
+#define Q_min M_min
+#define Q_max M_max
+#define Q_clamp M_clamp
 //#define Q_strtoull strtoull
 //#define Q_FileNameCmp FileNameCmp
 #define Q_vsnprintf _vsnprintf
@@ -182,6 +185,9 @@ extern int loadsize;
 #define Q_atoi atoi
 #define Q_atof atof
 #define Q_sqrt sqrt
+#define Q_min min
+#define Q_max max
+#define Q_clamp clamp
 //#define Q_strtoull strtoull
 //#define Q_FileNameCmp FileNameCmp
 #define Q_vsnprintf _vsnprintf

rehlds/engine/host.cpp

@@ -477,9 +477,9 @@ void SV_DropClient_internal(client_t *cl, qboolean crash, const char *string)
 			MSG_WriteByte(&cl->netchan.message, svc_disconnect);
 			MSG_WriteString(&cl->netchan.message, string);
 			final[0] = svc_disconnect;
-			Q_strncpy((char *)&final[1], string, min(sizeof(final) - 1, Q_strlen(string) + 1));
+			Q_strncpy((char *)&final[1], string, Q_min(sizeof(final) - 1, Q_strlen(string) + 1));
 			final[sizeof(final) - 1] = 0;
-			i = 1 + min(sizeof(final) - 1, Q_strlen(string) + 1);
+			i = 1 + Q_min(sizeof(final) - 1, Q_strlen(string) + 1);
 		}
 		if (cl->edict && cl->spawned)
 			gEntityInterface.pfnClientDisconnect(cl->edict);

rehlds/engine/mathlib_e.h

@@ -56,24 +56,57 @@ static const int nanmask = 0x7F800000;
 
 #define IS_NAN(fvar) ((*reinterpret_cast<int*>(&(fvar)) & nanmask) == nanmask)
 
-inline double M_sqrt(int value) {
-	return sqrt(value);
-}
-
 inline float M_sqrt(float value) {
 	return _mm_cvtss_f32(_mm_sqrt_ss(_mm_load_ss(&value)));
 }
 
 inline double M_sqrt(double value) {
-	double ret;
 	auto v = _mm_load_sd(&value);
-	_mm_store_sd(&ret, _mm_sqrt_sd(v, v));
-	return ret;
+	return _mm_cvtsd_f64(_mm_sqrt_sd(v, v));
 }
 
-inline double M_sqrt(long double value)
-{
-	return M_sqrt(double(value));
+template<typename T>
+inline double M_sqrt(T value) {
+	return sqrt(value);
+}
+
+inline float M_min(float a, float b) {
+	return _mm_cvtss_f32(_mm_min_ss(_mm_load_ss(&a), _mm_load_ss(&b)));
+}
+
+inline double M_min(double a, double b) {
+	return _mm_cvtsd_f64(_mm_min_sd(_mm_load_sd(&a), _mm_load_sd(&b)));
+}
+
+template<typename T>
+inline T M_min(T a, T b) {
+	return min(a, b);
+}
+
+inline float M_max(float a, float b) {
+	return _mm_cvtss_f32(_mm_max_ss(_mm_load_ss(&a), _mm_load_ss(&b)));
+}
+
+inline double M_max(double a, double b) {
+	return _mm_cvtsd_f64(_mm_max_sd(_mm_load_sd(&a), _mm_load_sd(&b)));
+}
+
+template<typename T>
+inline T M_max(T a, T b) {
+	return max(a, b);
+}
+
+inline float M_clamp(float a, float min, float max) {
+	return _mm_cvtss_f32(_mm_min_ss(_mm_max_ss(_mm_load_ss(&a), _mm_load_ss(&min)), _mm_load_ss(&max)));
+}
+
+inline double M_clamp(double a, double min, double max) {
+	return _mm_cvtsd_f64(_mm_min_sd(_mm_max_sd(_mm_load_sd(&a), _mm_load_sd(&min)), _mm_load_sd(&max)));
+}
+
+template<typename T>
+inline T M_clamp(T a, T min, T max) {
+	return clamp(a, min, max);
 }
 
 float anglemod(float a);

rehlds/engine/net_ws.cpp

@@ -1360,7 +1360,7 @@ int NET_SendLong(netsrc_t sock, int s, const char *buf, int len, int flags, cons
 
 		while (len > 0)
 		{
-			size = min(int(SPLIT_SIZE), len);
+			size = Q_min(int(SPLIT_SIZE), len);
 
 			pPacket->packetID = (packetNumber << 4) + packetCount;
 

rehlds/engine/r_studio.cpp

@@ -883,7 +883,7 @@ qboolean SV_CheckSphereIntersection(edict_t *ent, const vec_t *start, const vec_
 	pseqdesc += ent->v.sequence;
 	for (int i = 0; i < 3; i++)
 	{
-		maxDim[i] = max(fabs(pseqdesc->bbmax[i]), fabs(pseqdesc->bbmin[i]));
+		maxDim[i] = Q_max(fabs(pseqdesc->bbmax[i]), fabs(pseqdesc->bbmin[i]));
 	}
 	radiusSquared = maxDim[0] * maxDim[0] + maxDim[1] * maxDim[1] + maxDim[2] * maxDim[2];
 	return DoesSphereIntersect(ent->v.origin, radiusSquared, traceOrg, traceDir) != 0;

rehlds/engine/sv_main.cpp

@@ -719,7 +719,7 @@ void SV_StartParticle(const vec_t *org, const vec_t *dir, int color, int count)
 
 		for (int i = 0; i < 3; i++)
 		{
-			MSG_WriteChar(&g_psv.datagram, clamp((int)(dir[i] * 16.0f), -128, 127));
+			MSG_WriteChar(&g_psv.datagram, Q_clamp((int)(dir[i] * 16.0f), -128, 127));
 		}
 
 		MSG_WriteByte(&g_psv.datagram, count);
@@ -1771,7 +1771,7 @@ int SV_GetFragmentSize(void *state)
 		if (val[0] != 0)
 		{
 			size = Q_atoi( val );
-			size = clamp(size, FRAGMENT_S2C_MIN_SIZE, FRAGMENT_S2C_MAX_SIZE);
+			size = Q_clamp(size, FRAGMENT_S2C_MIN_SIZE, FRAGMENT_S2C_MAX_SIZE);
 		}
 	}
 
@@ -4937,7 +4937,7 @@ void SV_ExtractFromUserinfo(client_t *cl)
 	if (val[0] != 0)
 	{
 		i = Q_atoi(val);
-		cl->netchan.rate = clamp(float(i), MIN_RATE, MAX_RATE);
+		cl->netchan.rate = Q_clamp(float(i), MIN_RATE, MAX_RATE);
 	}
 
 	val = Info_ValueForKey(userinfo, "topcolor");

rehlds/public/utlrbtree.h

@@ -1119,7 +1119,7 @@ int CUtlRBTree<T, I>::Depth( I node ) const
 	
 	int depthright = Depth( RightChild(node) );
 	int depthleft = Depth( LeftChild(node) );
-	return max(depthright, depthleft) + 1;
+	return Q_max(depthright, depthleft) + 1;
 }