Squashed 'third_party/blasfeo/' content from commit 2a828ca

Change-Id: If1c3caa4799b2d4eb287ef83fa17043587ef07a3
git-subtree-dir: third_party/blasfeo
git-subtree-split: 2a828ca5442108c4c58e4b42b061a0469043f6ea
diff --git a/auxiliary/c99/Makefile b/auxiliary/c99/Makefile
new file mode 100644
index 0000000..6e9ea7b
--- /dev/null
+++ b/auxiliary/c99/Makefile
@@ -0,0 +1,77 @@
+###################################################################################################
+#                                                                                                 #
+# This file is part of BLASFEO.                                                                   #
+#                                                                                                 #
+# BLASFEO -- BLAS For Embedded Optimization.                                                      #
+# Copyright (C) 2016-2017 by Gianluca Frison.                                                     #
+# Developed at IMTEK (University of Freiburg) under the supervision of Moritz Diehl.              #
+# All rights reserved.                                                                            #
+#                                                                                                 #
+# HPMPC is free software; you can redistribute it and/or                                          #
+# modify it under the terms of the GNU Lesser General Public                                      #
+# License as published by the Free Software Foundation; either                                    #
+# version 2.1 of the License, or (at your option) any later version.                              #
+#                                                                                                 #
+# HPMPC is distributed in the hope that it will be useful,                                        #
+# but WITHOUT ANY WARRANTY; without even the implied warranty of                                  #
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.                                            #
+# See the GNU Lesser General Public License for more details.                                     #
+#                                                                                                 #
+# You should have received a copy of the GNU Lesser General Public                                #
+# License along with HPMPC; if not, write to the Free Software                                    #
+# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA                  #
+#                                                                                                 #
+# Author: Gianluca Frison, giaf (at) dtu.dk                                                       #
+#                          gianluca.frison (at) imtek.uni-freiburg.de                             #
+#                                                                                                 #
+###################################################################################################
+
+include ../../Makefile.rule
+
+OBJS = 
+
+ifeq ($(LA), HIGH_PERFORMANCE)
+
+ifeq ($(TARGET), X64_INTEL_HASWELL)
+OBJS += 
+OBJS += kernel_sgetr_lib4.o
+endif
+
+ifeq ($(TARGET), X64_INTEL_SANDY_BRIDGE)
+OBJS += 
+OBJS += kernel_sgetr_lib4.o
+endif
+
+ifeq ($(TARGET), X64_INTEL_CORE)
+OBJS += kernel_dgecp_lib4.o kernel_dgetr_lib4.o
+OBJS += kernel_sgetr_lib4.o
+endif
+
+ifeq ($(TARGET), X64_AMD_BULLDOZER)
+OBJS += kernel_dgecp_lib4.o kernel_dgetr_lib4.o
+OBJS += kernel_sgetr_lib4.o
+endif
+
+ifeq ($(TARGET), ARMV8A_ARM_CORTEX_A57)
+OBJS += kernel_dgecp_lib4.o kernel_dgetr_lib4.o
+OBJS += kernel_sgetr_lib4.o
+endif
+
+ifeq ($(TARGET), ARMV7A_ARM_CORTEX_A15)
+OBJS += kernel_dgecp_lib4.o kernel_dgetr_lib4.o
+OBJS += kernel_sgetr_lib4.o
+endif
+
+ifeq ($(TARGET), GENERIC)
+OBJS += kernel_dgecp_lib4.o kernel_dgetr_lib4.o
+OBJS += kernel_sgetr_lib4.o
+endif
+
+else # LA_REFERENCE | LA_BLAS
+
+endif # LA choice
+
+obj: $(OBJS)
+
+clean:
+	rm -f *.o
diff --git a/auxiliary/c99/kernel_dgecp_lib4.c b/auxiliary/c99/kernel_dgecp_lib4.c
new file mode 100644
index 0000000..e883072
--- /dev/null
+++ b/auxiliary/c99/kernel_dgecp_lib4.c
@@ -0,0 +1,1261 @@
+/**************************************************************************************************
+*                                                                                                 *
+* This file is part of BLASFEO.                                                                   *
+*                                                                                                 *
+* BLASFEO -- BLAS For Embedded Optimization.                                                      *
+* Copyright (C) 2016-2017 by Gianluca Frison.                                                     *
+* Developed at IMTEK (University of Freiburg) under the supervision of Moritz Diehl.              *
+* All rights reserved.                                                                            *
+*                                                                                                 *
+* HPMPC is free software; you can redistribute it and/or                                          *
+* modify it under the terms of the GNU Lesser General Public                                      *
+* License as published by the Free Software Foundation; either                                    *
+* version 2.1 of the License, or (at your option) any later version.                              *
+*                                                                                                 *
+* HPMPC is distributed in the hope that it will be useful,                                        *
+* but WITHOUT ANY WARRANTY; without even the implied warranty of                                  *
+* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.                                            *
+* See the GNU Lesser General Public License for more details.                                     *
+*                                                                                                 *
+* You should have received a copy of the GNU Lesser General Public                                *
+* License along with HPMPC; if not, write to the Free Software                                    *
+* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA                  *
+*                                                                                                 *
+* Author: Gianluca Frison, giaf (at) dtu.dk                                                       *
+*                          gianluca.frison (at) imtek.uni-freiburg.de                             *
+*                                                                                                 *
+**************************************************************************************************/
+
+
+
+// both A and B are aligned to 256-bit boundaries
+void kernel_dgecp_4_0_lib4(int tri, int kmax, double alpha, double *A, double *B)
+	{
+
+	if(tri==1)
+		{
+		// A and C are lower triangular
+		// kmax+1 4-wide + end 3x3 triangle
+
+		kmax += 1;
+		}
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+		B[0+bs*0] = alpha*A[0+bs*0];
+		B[1+bs*0] = alpha*A[1+bs*0];
+		B[2+bs*0] = alpha*A[2+bs*0];
+		B[3+bs*0] = alpha*A[3+bs*0];
+
+		B[0+bs*1] = alpha*A[0+bs*1];
+		B[1+bs*1] = alpha*A[1+bs*1];
+		B[2+bs*1] = alpha*A[2+bs*1];
+		B[3+bs*1] = alpha*A[3+bs*1];
+
+		B[0+bs*2] = alpha*A[0+bs*2];
+		B[1+bs*2] = alpha*A[1+bs*2];
+		B[2+bs*2] = alpha*A[2+bs*2];
+		B[3+bs*2] = alpha*A[3+bs*2];
+
+		B[0+bs*3] = alpha*A[0+bs*3];
+		B[1+bs*3] = alpha*A[1+bs*3];
+		B[2+bs*3] = alpha*A[2+bs*3];
+		B[3+bs*3] = alpha*A[3+bs*3];
+
+		A += 16;
+		B += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] = alpha*A[0+bs*0];
+		B[1+bs*0] = alpha*A[1+bs*0];
+		B[2+bs*0] = alpha*A[2+bs*0];
+		B[3+bs*0] = alpha*A[3+bs*0];
+
+		A += 4;
+		B += 4;
+
+		}
+	
+	if(tri==1)
+		{
+		// 3x3 triangle
+
+		B[1+bs*0] = alpha*A[1+bs*0];
+		B[2+bs*0] = alpha*A[2+bs*0];
+		B[3+bs*0] = alpha*A[3+bs*0];
+
+		B[2+bs*1] = alpha*A[2+bs*1];
+		B[3+bs*1] = alpha*A[3+bs*1];
+
+		B[3+bs*2] = alpha*A[3+bs*2];
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 256-bit boundaries, 1 element of A must be skipped
+void kernel_dgecp_4_1_lib4(int tri, int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(tri==1)
+		{
+		// A and C are lower triangular
+		// kmax+1 4-wide + end 3x3 triangle
+
+		kmax += 1;
+		}
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] = alpha*A0[1+bs*0];
+		B[1+bs*0] = alpha*A0[2+bs*0];
+		B[2+bs*0] = alpha*A0[3+bs*0];
+		B[3+bs*0] = alpha*A1[0+bs*0];
+
+		B[0+bs*1] = alpha*A0[1+bs*1];
+		B[1+bs*1] = alpha*A0[2+bs*1];
+		B[2+bs*1] = alpha*A0[3+bs*1];
+		B[3+bs*1] = alpha*A1[0+bs*1];
+
+		B[0+bs*2] = alpha*A0[1+bs*2];
+		B[1+bs*2] = alpha*A0[2+bs*2];
+		B[2+bs*2] = alpha*A0[3+bs*2];
+		B[3+bs*2] = alpha*A1[0+bs*2];
+
+		B[0+bs*3] = alpha*A0[1+bs*3];
+		B[1+bs*3] = alpha*A0[2+bs*3];
+		B[2+bs*3] = alpha*A0[3+bs*3];
+		B[3+bs*3] = alpha*A1[0+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] = alpha*A0[1+bs*0];
+		B[1+bs*0] = alpha*A0[2+bs*0];
+		B[2+bs*0] = alpha*A0[3+bs*0];
+		B[3+bs*0] = alpha*A1[0+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+	
+	if(tri==1)
+		{
+		// 3x3 triangle
+
+		B[1+0*bs] = alpha*A0[2+0*bs];
+		B[2+0*bs] = alpha*A0[3+0*bs];
+		B[3+0*bs] = alpha*A1[0+0*bs];
+
+		B[2+1*bs] = alpha*A0[3+1*bs];
+		B[3+1*bs] = alpha*A1[0+1*bs];
+
+		B[3+2*bs] = alpha*A1[0+2*bs];
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 256-bit boundaries, 2 elements of A must be skipped
+void kernel_dgecp_4_2_lib4(int tri, int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(tri==1)
+		{
+		// A and C are lower triangular
+		// kmax+1 4-wide + end 3x3 triangle
+
+		kmax += 1;
+		}
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] = alpha*A0[2+bs*0];
+		B[1+bs*0] = alpha*A0[3+bs*0];
+		B[2+bs*0] = alpha*A1[0+bs*0];
+		B[3+bs*0] = alpha*A1[1+bs*0];
+
+		B[0+bs*1] = alpha*A0[2+bs*1];
+		B[1+bs*1] = alpha*A0[3+bs*1];
+		B[2+bs*1] = alpha*A1[0+bs*1];
+		B[3+bs*1] = alpha*A1[1+bs*1];
+
+		B[0+bs*2] = alpha*A0[2+bs*2];
+		B[1+bs*2] = alpha*A0[3+bs*2];
+		B[2+bs*2] = alpha*A1[0+bs*2];
+		B[3+bs*2] = alpha*A1[1+bs*2];
+
+		B[0+bs*3] = alpha*A0[2+bs*3];
+		B[1+bs*3] = alpha*A0[3+bs*3];
+		B[2+bs*3] = alpha*A1[0+bs*3];
+		B[3+bs*3] = alpha*A1[1+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] = alpha*A0[2+bs*0];
+		B[1+bs*0] = alpha*A0[3+bs*0];
+		B[2+bs*0] = alpha*A1[0+bs*0];
+		B[3+bs*0] = alpha*A1[1+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+	
+	if(tri==1)
+		{
+		// 3x3 triangle}
+
+		B[1+bs*0] = alpha*A0[3+bs*0];
+		B[2+bs*0] = alpha*A1[0+bs*0];
+		B[3+bs*0] = alpha*A1[1+bs*0];
+
+		B[2+bs*1] = alpha*A1[0+bs*1];
+		B[3+bs*1] = alpha*A1[1+bs*1];
+
+		B[3+bs*2] = alpha*A1[1+bs*2];
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 256-bit boundaries, 3 elements of A must be skipped
+void kernel_dgecp_4_3_lib4(int tri, int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(tri==1)
+		{
+		// A and C are lower triangular
+		// kmax+1 4-wide + end 3x3 triangle
+
+		kmax += 1;
+		}
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] = alpha*A0[3+bs*0];
+		B[1+bs*0] = alpha*A1[0+bs*0];
+		B[2+bs*0] = alpha*A1[1+bs*0];
+		B[3+bs*0] = alpha*A1[2+bs*0];
+
+		B[0+bs*1] = alpha*A0[3+bs*1];
+		B[1+bs*1] = alpha*A1[0+bs*1];
+		B[2+bs*1] = alpha*A1[1+bs*1];
+		B[3+bs*1] = alpha*A1[2+bs*1];
+
+		B[0+bs*2] = alpha*A0[3+bs*2];
+		B[1+bs*2] = alpha*A1[0+bs*2];
+		B[2+bs*2] = alpha*A1[1+bs*2];
+		B[3+bs*2] = alpha*A1[2+bs*2];
+
+		B[0+bs*3] = alpha*A0[3+bs*3];
+		B[1+bs*3] = alpha*A1[0+bs*3];
+		B[2+bs*3] = alpha*A1[1+bs*3];
+		B[3+bs*3] = alpha*A1[2+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] = alpha*A0[3+bs*0];
+		B[1+bs*0] = alpha*A1[0+bs*0];
+		B[2+bs*0] = alpha*A1[1+bs*0];
+		B[3+bs*0] = alpha*A1[2+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+	
+	if(tri==1)
+		{
+		// 3x3 triangle
+
+		B[1+bs*0] = alpha*A1[0+bs*0];
+		B[2+bs*0] = alpha*A1[1+bs*0];
+		B[3+bs*0] = alpha*A1[2+bs*0];
+
+		B[2+bs*1] = alpha*A1[1+bs*1];
+		B[3+bs*1] = alpha*A1[2+bs*1];
+
+		B[3+bs*2] = alpha*A1[2+bs*2];
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 64-bit boundaries
+void kernel_dgecp_3_0_lib4(int tri, int kmax, double alpha, double *A, double *B)
+	{
+
+	if(tri==1)
+		{
+		// A and C are lower triangular
+		// kmax+1 3-wide + end 2x2 triangle
+
+		kmax += 1;
+		}
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+		B[0+bs*0] = alpha*A[0+bs*0];
+		B[1+bs*0] = alpha*A[1+bs*0];
+		B[2+bs*0] = alpha*A[2+bs*0];
+
+		B[0+bs*1] = alpha*A[0+bs*1];
+		B[1+bs*1] = alpha*A[1+bs*1];
+		B[2+bs*1] = alpha*A[2+bs*1];
+
+		B[0+bs*2] = alpha*A[0+bs*2];
+		B[1+bs*2] = alpha*A[1+bs*2];
+		B[2+bs*2] = alpha*A[2+bs*2];
+
+		B[0+bs*3] = alpha*A[0+bs*3];
+		B[1+bs*3] = alpha*A[1+bs*3];
+		B[2+bs*3] = alpha*A[2+bs*3];
+
+		A += 16;
+		B += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] = alpha*A[0+bs*0];
+		B[1+bs*0] = alpha*A[1+bs*0];
+		B[2+bs*0] = alpha*A[2+bs*0];
+
+		A += 4;
+		B += 4;
+
+		}
+	
+	if(tri==1)
+		{
+		// 2x2 triangle
+
+		B[1+bs*0] = alpha*A[1+bs*0];
+		B[2+bs*0] = alpha*A[2+bs*0];
+
+		B[2+bs*1] = alpha*A[2+bs*1];
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 256-bit boundaries, 2 elements of A must be skipped
+void kernel_dgecp_3_2_lib4(int tri, int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(tri==1)
+		{
+		// A and C are lower triangular
+		// kmax+1 3-wide + end 2x2 triangle
+
+		kmax += 1;
+		}
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] = alpha*A0[2+bs*0];
+		B[1+bs*0] = alpha*A0[3+bs*0];
+		B[2+bs*0] = alpha*A1[0+bs*0];
+
+		B[0+bs*1] = alpha*A0[2+bs*1];
+		B[1+bs*1] = alpha*A0[3+bs*1];
+		B[2+bs*1] = alpha*A1[0+bs*1];
+
+		B[0+bs*2] = alpha*A0[2+bs*2];
+		B[1+bs*2] = alpha*A0[3+bs*2];
+		B[2+bs*2] = alpha*A1[0+bs*2];
+
+		B[0+bs*3] = alpha*A0[2+bs*3];
+		B[1+bs*3] = alpha*A0[3+bs*3];
+		B[2+bs*3] = alpha*A1[0+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] = alpha*A0[2+bs*0];
+		B[1+bs*0] = alpha*A0[3+bs*0];
+		B[2+bs*0] = alpha*A1[0+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+	
+	if(tri==1)
+		{
+		// 2x2 triangle
+
+		B[1+bs*0] = alpha*A0[3+bs*0];
+		B[2+bs*0] = alpha*A1[0+bs*0];
+
+		B[2+bs*1] = alpha*A1[0+bs*1];
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 256-bit boundaries, 3 elements of A must be skipped
+void kernel_dgecp_3_3_lib4(int tri, int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(tri==1)
+		{
+		// A and C are lower triangular
+		// kmax+1 3-wide + end 2x2 triangle
+
+		kmax += 1;
+		}
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] = alpha*A0[3+bs*0];
+		B[1+bs*0] = alpha*A1[0+bs*0];
+		B[2+bs*0] = alpha*A1[1+bs*0];
+
+		B[0+bs*1] = alpha*A0[3+bs*1];
+		B[1+bs*1] = alpha*A1[0+bs*1];
+		B[2+bs*1] = alpha*A1[1+bs*1];
+
+		B[0+bs*2] = alpha*A0[3+bs*2];
+		B[1+bs*2] = alpha*A1[0+bs*2];
+		B[2+bs*2] = alpha*A1[1+bs*2];
+
+		B[0+bs*3] = alpha*A0[3+bs*3];
+		B[1+bs*3] = alpha*A1[0+bs*3];
+		B[2+bs*3] = alpha*A1[1+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] = alpha*A0[3+bs*0];
+		B[1+bs*0] = alpha*A1[0+bs*0];
+		B[2+bs*0] = alpha*A1[1+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+	
+	if(tri==1)
+		{
+		// 2x2 triangle
+
+		B[1+bs*0] = alpha*A1[0+bs*0];
+		B[2+bs*0] = alpha*A1[1+bs*0];
+
+		B[2+bs*1] = alpha*A1[1+bs*1];
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 64-bit boundaries
+void kernel_dgecp_2_0_lib4(int tri, int kmax, double alpha, double *A, double *B)
+	{
+
+	if(tri==1)
+		{
+		// A and C are lower triangular
+		// kmax+1 2-wide + end 1x1 triangle
+
+		kmax += 1;
+		}
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+		B[0+bs*0] = alpha*A[0+bs*0];
+		B[1+bs*0] = alpha*A[1+bs*0];
+
+		B[0+bs*1] = alpha*A[0+bs*1];
+		B[1+bs*1] = alpha*A[1+bs*1];
+
+		B[0+bs*2] = alpha*A[0+bs*2];
+		B[1+bs*2] = alpha*A[1+bs*2];
+
+		B[0+bs*3] = alpha*A[0+bs*3];
+		B[1+bs*3] = alpha*A[1+bs*3];
+
+		A += 16;
+		B += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] = alpha*A[0+bs*0];
+		B[1+bs*0] = alpha*A[1+bs*0];
+
+		A += 4;
+		B += 4;
+
+		}
+	
+	if(tri==1)
+		{
+		// 1x1 triangle
+
+		B[1+bs*0] = alpha*A[1+bs*0];
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 128-bit boundaries, 3 elements of A must be skipped
+void kernel_dgecp_2_3_lib4(int tri, int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(tri==1)
+		{
+		// A and C are lower triangular
+		// kmax+1 2-wide + end 1x1 triangle
+
+		kmax += 1;
+		}
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] = alpha*A0[3+bs*0];
+		B[1+bs*0] = alpha*A1[0+bs*0];
+
+		B[0+bs*1] = alpha*A0[3+bs*1];
+		B[1+bs*1] = alpha*A1[0+bs*1];
+
+		B[0+bs*2] = alpha*A0[3+bs*2];
+		B[1+bs*2] = alpha*A1[0+bs*2];
+
+		B[0+bs*3] = alpha*A0[3+bs*3];
+		B[1+bs*3] = alpha*A1[0+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] = alpha*A0[3+bs*0];
+		B[1+bs*0] = alpha*A1[0+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+	
+	if(tri==1)
+		{
+		// 1x1 triangle
+
+		B[1+bs*0] = alpha*A1[0+bs*0];
+
+		}
+
+	}
+
+
+
+// both A and B are aligned 64-bit boundaries
+void kernel_dgecp_1_0_lib4(int tri, int kmax, double alpha, double *A, double *B)
+	{
+
+	if(tri==1)
+		{
+		// A and C are lower triangular
+		// kmax+1 1-wide
+
+		kmax += 1;
+		}
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+		B[0+bs*0] = alpha*A[0+bs*0];
+
+		B[0+bs*1] = alpha*A[0+bs*1];
+
+		B[0+bs*2] = alpha*A[0+bs*2];
+
+		B[0+bs*3] = alpha*A[0+bs*3];
+
+		A += 16;
+		B += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] = alpha*A[0+bs*0];
+
+		A += 4;
+		B += 4;
+
+		}
+
+	}
+
+
+
+
+// both A and B are aligned to 256-bit boundaries
+void kernel_dgead_4_0_lib4(int kmax, double alpha, double *A, double *B)
+	{
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+		B[0+bs*0] += alpha * A[0+bs*0];
+		B[1+bs*0] += alpha * A[1+bs*0];
+		B[2+bs*0] += alpha * A[2+bs*0];
+		B[3+bs*0] += alpha * A[3+bs*0];
+
+		B[0+bs*1] += alpha * A[0+bs*1];
+		B[1+bs*1] += alpha * A[1+bs*1];
+		B[2+bs*1] += alpha * A[2+bs*1];
+		B[3+bs*1] += alpha * A[3+bs*1];
+
+		B[0+bs*2] += alpha * A[0+bs*2];
+		B[1+bs*2] += alpha * A[1+bs*2];
+		B[2+bs*2] += alpha * A[2+bs*2];
+		B[3+bs*2] += alpha * A[3+bs*2];
+
+		B[0+bs*3] += alpha * A[0+bs*3];
+		B[1+bs*3] += alpha * A[1+bs*3];
+		B[2+bs*3] += alpha * A[2+bs*3];
+		B[3+bs*3] += alpha * A[3+bs*3];
+
+		A += 16;
+		B += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] += alpha * A[0+bs*0];
+		B[1+bs*0] += alpha * A[1+bs*0];
+		B[2+bs*0] += alpha * A[2+bs*0];
+		B[3+bs*0] += alpha * A[3+bs*0];
+
+		A += 4;
+		B += 4;
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 256-bit boundaries, 1 element of A must be skipped
+void kernel_dgead_4_1_lib4(int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] += alpha * A0[1+bs*0];
+		B[1+bs*0] += alpha * A0[2+bs*0];
+		B[2+bs*0] += alpha * A0[3+bs*0];
+		B[3+bs*0] += alpha * A1[0+bs*0];
+
+		B[0+bs*1] += alpha * A0[1+bs*1];
+		B[1+bs*1] += alpha * A0[2+bs*1];
+		B[2+bs*1] += alpha * A0[3+bs*1];
+		B[3+bs*1] += alpha * A1[0+bs*1];
+
+		B[0+bs*2] += alpha * A0[1+bs*2];
+		B[1+bs*2] += alpha * A0[2+bs*2];
+		B[2+bs*2] += alpha * A0[3+bs*2];
+		B[3+bs*2] += alpha * A1[0+bs*2];
+
+		B[0+bs*3] += alpha * A0[1+bs*3];
+		B[1+bs*3] += alpha * A0[2+bs*3];
+		B[2+bs*3] += alpha * A0[3+bs*3];
+		B[3+bs*3] += alpha * A1[0+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] += alpha * A0[1+bs*0];
+		B[1+bs*0] += alpha * A0[2+bs*0];
+		B[2+bs*0] += alpha * A0[3+bs*0];
+		B[3+bs*0] += alpha * A1[0+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 256-bit boundaries, 2 elements of A must be skipped
+void kernel_dgead_4_2_lib4(int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] += alpha * A0[2+bs*0];
+		B[1+bs*0] += alpha * A0[3+bs*0];
+		B[2+bs*0] += alpha * A1[0+bs*0];
+		B[3+bs*0] += alpha * A1[1+bs*0];
+
+		B[0+bs*1] += alpha * A0[2+bs*1];
+		B[1+bs*1] += alpha * A0[3+bs*1];
+		B[2+bs*1] += alpha * A1[0+bs*1];
+		B[3+bs*1] += alpha * A1[1+bs*1];
+
+		B[0+bs*2] += alpha * A0[2+bs*2];
+		B[1+bs*2] += alpha * A0[3+bs*2];
+		B[2+bs*2] += alpha * A1[0+bs*2];
+		B[3+bs*2] += alpha * A1[1+bs*2];
+
+		B[0+bs*3] += alpha * A0[2+bs*3];
+		B[1+bs*3] += alpha * A0[3+bs*3];
+		B[2+bs*3] += alpha * A1[0+bs*3];
+		B[3+bs*3] += alpha * A1[1+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] += alpha * A0[2+bs*0];
+		B[1+bs*0] += alpha * A0[3+bs*0];
+		B[2+bs*0] += alpha * A1[0+bs*0];
+		B[3+bs*0] += alpha * A1[1+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 256-bit boundaries, 3 elements of A must be skipped
+void kernel_dgead_4_3_lib4(int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] += alpha * A0[3+bs*0];
+		B[1+bs*0] += alpha * A1[0+bs*0];
+		B[2+bs*0] += alpha * A1[1+bs*0];
+		B[3+bs*0] += alpha * A1[2+bs*0];
+
+		B[0+bs*1] += alpha * A0[3+bs*1];
+		B[1+bs*1] += alpha * A1[0+bs*1];
+		B[2+bs*1] += alpha * A1[1+bs*1];
+		B[3+bs*1] += alpha * A1[2+bs*1];
+
+		B[0+bs*2] += alpha * A0[3+bs*2];
+		B[1+bs*2] += alpha * A1[0+bs*2];
+		B[2+bs*2] += alpha * A1[1+bs*2];
+		B[3+bs*2] += alpha * A1[2+bs*2];
+
+		B[0+bs*3] += alpha * A0[3+bs*3];
+		B[1+bs*3] += alpha * A1[0+bs*3];
+		B[2+bs*3] += alpha * A1[1+bs*3];
+		B[3+bs*3] += alpha * A1[2+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] += alpha * A0[3+bs*0];
+		B[1+bs*0] += alpha * A1[0+bs*0];
+		B[2+bs*0] += alpha * A1[1+bs*0];
+		B[3+bs*0] += alpha * A1[2+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 64-bit boundaries
+void kernel_dgead_3_0_lib4(int kmax, double alpha, double *A, double *B)
+	{
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+		B[0+bs*0] += alpha * A[0+bs*0];
+		B[1+bs*0] += alpha * A[1+bs*0];
+		B[2+bs*0] += alpha * A[2+bs*0];
+
+		B[0+bs*1] += alpha * A[0+bs*1];
+		B[1+bs*1] += alpha * A[1+bs*1];
+		B[2+bs*1] += alpha * A[2+bs*1];
+
+		B[0+bs*2] += alpha * A[0+bs*2];
+		B[1+bs*2] += alpha * A[1+bs*2];
+		B[2+bs*2] += alpha * A[2+bs*2];
+
+		B[0+bs*3] += alpha * A[0+bs*3];
+		B[1+bs*3] += alpha * A[1+bs*3];
+		B[2+bs*3] += alpha * A[2+bs*3];
+
+		A += 16;
+		B += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] += alpha * A[0+bs*0];
+		B[1+bs*0] += alpha * A[1+bs*0];
+		B[2+bs*0] += alpha * A[2+bs*0];
+
+		A += 4;
+		B += 4;
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 256-bit boundaries, 2 elements of A must be skipped
+void kernel_dgead_3_2_lib4(int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] += alpha * A0[2+bs*0];
+		B[1+bs*0] += alpha * A0[3+bs*0];
+		B[2+bs*0] += alpha * A1[0+bs*0];
+
+		B[0+bs*1] += alpha * A0[2+bs*1];
+		B[1+bs*1] += alpha * A0[3+bs*1];
+		B[2+bs*1] += alpha * A1[0+bs*1];
+
+		B[0+bs*2] += alpha * A0[2+bs*2];
+		B[1+bs*2] += alpha * A0[3+bs*2];
+		B[2+bs*2] += alpha * A1[0+bs*2];
+
+		B[0+bs*3] += alpha * A0[2+bs*3];
+		B[1+bs*3] += alpha * A0[3+bs*3];
+		B[2+bs*3] += alpha * A1[0+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] += alpha * A0[2+bs*0];
+		B[1+bs*0] += alpha * A0[3+bs*0];
+		B[2+bs*0] += alpha * A1[0+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 256-bit boundaries, 3 elements of A must be skipped
+void kernel_dgead_3_3_lib4(int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] += alpha * A0[3+bs*0];
+		B[1+bs*0] += alpha * A1[0+bs*0];
+		B[2+bs*0] += alpha * A1[1+bs*0];
+
+		B[0+bs*1] += alpha * A0[3+bs*1];
+		B[1+bs*1] += alpha * A1[0+bs*1];
+		B[2+bs*1] += alpha * A1[1+bs*1];
+
+		B[0+bs*2] += alpha * A0[3+bs*2];
+		B[1+bs*2] += alpha * A1[0+bs*2];
+		B[2+bs*2] += alpha * A1[1+bs*2];
+
+		B[0+bs*3] += alpha * A0[3+bs*3];
+		B[1+bs*3] += alpha * A1[0+bs*3];
+		B[2+bs*3] += alpha * A1[1+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] += alpha * A0[3+bs*0];
+		B[1+bs*0] += alpha * A1[0+bs*0];
+		B[2+bs*0] += alpha * A1[1+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 64-bit boundaries
+void kernel_dgead_2_0_lib4(int kmax, double alpha, double *A, double *B)
+	{
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+		B[0+bs*0] += alpha * A[0+bs*0];
+		B[1+bs*0] += alpha * A[1+bs*0];
+
+		B[0+bs*1] += alpha * A[0+bs*1];
+		B[1+bs*1] += alpha * A[1+bs*1];
+
+		B[0+bs*2] += alpha * A[0+bs*2];
+		B[1+bs*2] += alpha * A[1+bs*2];
+
+		B[0+bs*3] += alpha * A[0+bs*3];
+		B[1+bs*3] += alpha * A[1+bs*3];
+
+		A += 16;
+		B += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] += alpha * A[0+bs*0];
+		B[1+bs*0] += alpha * A[1+bs*0];
+
+		A += 4;
+		B += 4;
+
+		}
+
+	}
+
+
+
+// both A and B are aligned to 128-bit boundaries, 3 elements of A must be skipped
+void kernel_dgead_2_3_lib4(int kmax, double alpha, double *A0, int sda, double *B)
+	{
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	double *A1 = A0 + bs*sda;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+
+		B[0+bs*0] += alpha * A0[3+bs*0];
+		B[1+bs*0] += alpha * A1[0+bs*0];
+
+		B[0+bs*1] += alpha * A0[3+bs*1];
+		B[1+bs*1] += alpha * A1[0+bs*1];
+
+		B[0+bs*2] += alpha * A0[3+bs*2];
+		B[1+bs*2] += alpha * A1[0+bs*2];
+
+		B[0+bs*3] += alpha * A0[3+bs*3];
+		B[1+bs*3] += alpha * A1[0+bs*3];
+
+		A0 += 16;
+		A1 += 16;
+		B  += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] += alpha * A0[3+bs*0];
+		B[1+bs*0] += alpha * A1[0+bs*0];
+
+		A0 += 4;
+		A1 += 4;
+		B  += 4;
+
+		}
+
+	}
+
+
+
+// both A and B are aligned 64-bit boundaries
+void kernel_dgead_1_0_lib4(int kmax, double alpha, double *A, double *B)
+	{
+
+	if(kmax<=0)
+		return;
+
+	const int bs = 4;
+
+	int k;
+
+	for(k=0; k<kmax-3; k+=4)
+		{
+		B[0+bs*0] += alpha * A[0+bs*0];
+
+		B[0+bs*1] += alpha * A[0+bs*1];
+
+		B[0+bs*2] += alpha * A[0+bs*2];
+
+		B[0+bs*3] += alpha * A[0+bs*3];
+
+		A += 16;
+		B += 16;
+
+		}
+	for(; k<kmax; k++)
+		{
+
+		B[0+bs*0] += alpha * A[0+bs*0];
+
+		A += 4;
+		B += 4;
+
+		}
+
+	}
+
+
+
+
diff --git a/auxiliary/c99/kernel_dgetr_lib4.c b/auxiliary/c99/kernel_dgetr_lib4.c
new file mode 100644
index 0000000..7d62277
--- /dev/null
+++ b/auxiliary/c99/kernel_dgetr_lib4.c
@@ -0,0 +1,414 @@
+/**************************************************************************************************
+*                                                                                                 *
+* This file is part of BLASFEO.                                                                   *
+*                                                                                                 *
+* BLASFEO -- BLAS For Embedded Optimization.                                                      *
+* Copyright (C) 2016-2017 by Gianluca Frison.                                                     *
+* Developed at IMTEK (University of Freiburg) under the supervision of Moritz Diehl.              *
+* All rights reserved.                                                                            *
+*                                                                                                 *
+* HPMPC is free software; you can redistribute it and/or                                          *
+* modify it under the terms of the GNU Lesser General Public                                      *
+* License as published by the Free Software Foundation; either                                    *
+* version 2.1 of the License, or (at your option) any later version.                              *
+*                                                                                                 *
+* HPMPC is distributed in the hope that it will be useful,                                        *
+* but WITHOUT ANY WARRANTY; without even the implied warranty of                                  *
+* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.                                            *
+* See the GNU Lesser General Public License for more details.                                     *
+*                                                                                                 *
+* You should have received a copy of the GNU Lesser General Public                                *
+* License along with HPMPC; if not, write to the Free Software                                    *
+* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA                  *
+*                                                                                                 *
+* Author: Gianluca Frison, giaf (at) dtu.dk                                                       *
+*                          gianluca.frison (at) imtek.uni-freiburg.de                             *
+*                                                                                                 *
+**************************************************************************************************/
+
+
+
+// transposed of general matrices, read along panels, write across panels
+void kernel_dgetr_4_lib4(int tri, int kmax, int kna, double alpha, double *A, double *C, int sdc)
+	{
+
+	if(tri==1)
+		{
+		// A is lower triangular, C is upper triangular
+		// kmax+1 4-wide + end 3x3 triangle
+
+		kmax += 1;
+		}
+
+	const int bs = 4;
+	
+	int k;
+
+	k = 0;
+
+	if(kmax<kna)
+		goto cleanup_loop;
+
+	if(kna>0)
+		{
+		for( ; k<kna; k++)
+			{
+			C[0+bs*0] = alpha * A[0+bs*0];
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[0+bs*3] = alpha * A[3+bs*0];
+
+			C += 1;
+			A += bs;
+			}
+		C += bs*(sdc-1);
+		}
+	
+	for( ; k<kmax-3; k+=4)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+		C[0+bs*2] = alpha * A[2+bs*0];
+		C[0+bs*3] = alpha * A[3+bs*0];
+
+		C[1+bs*0] = alpha * A[0+bs*1];
+		C[1+bs*1] = alpha * A[1+bs*1];
+		C[1+bs*2] = alpha * A[2+bs*1];
+		C[1+bs*3] = alpha * A[3+bs*1];
+
+		C[2+bs*0] = alpha * A[0+bs*2];
+		C[2+bs*1] = alpha * A[1+bs*2];
+		C[2+bs*2] = alpha * A[2+bs*2];
+		C[2+bs*3] = alpha * A[3+bs*2];
+
+		C[3+bs*0] = alpha * A[0+bs*3];
+		C[3+bs*1] = alpha * A[1+bs*3];
+		C[3+bs*2] = alpha * A[2+bs*3];
+		C[3+bs*3] = alpha * A[3+bs*3];
+
+		C += bs*sdc;
+		A += bs*bs;
+		}
+
+	cleanup_loop:
+
+	for( ; k<kmax; k++)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+		C[0+bs*2] = alpha * A[2+bs*0];
+		C[0+bs*3] = alpha * A[3+bs*0];
+
+		C += 1;
+		A += bs;
+		}
+
+	if(tri==1)
+		{
+		// end 3x3 triangle
+		kna = (bs-(bs-kna+kmax)%bs)%bs;
+
+		if(kna==1)
+			{
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[0+bs*3] = alpha * A[3+bs*0];
+			C[1+bs*(sdc+1)] = alpha * A[2+bs*1];
+			C[1+bs*(sdc+2)] = alpha * A[3+bs*1];
+			C[2+bs*(sdc+2)] = alpha * A[3+bs*2];
+			}
+		else if(kna==2)
+			{
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[0+bs*3] = alpha * A[3+bs*0];
+			C[1+bs*2] = alpha * A[2+bs*1];
+			C[1+bs*3] = alpha * A[3+bs*1];
+			C[2+bs*(sdc+2)] = alpha * A[3+bs*2];
+			}
+		else
+			{
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[0+bs*3] = alpha * A[3+bs*0];
+			C[1+bs*2] = alpha * A[2+bs*1];
+			C[1+bs*3] = alpha * A[3+bs*1];
+			C[2+bs*3] = alpha * A[3+bs*2];
+			}
+		}
+
+	}
+
+
+
+// transposed of general matrices, read along panels, write across panels
+void kernel_dgetr_3_lib4(int tri, int kmax, int kna, double alpha, double *A, double *C, int sdc)
+	{
+
+	if(tri==1)
+		{
+		// A is lower triangular, C is upper triangular
+		// kmax+1 3-wide + end 2x2 triangle
+
+		kmax += 1;
+		}
+
+	const int bs = 4;
+	
+	int k;
+
+	k = 0;
+
+	if(kmax<kna)
+		goto cleanup_loop;
+
+	if(kna>0)
+		{
+		for( ; k<kna; k++)
+			{
+			C[0+bs*0] = alpha * A[0+bs*0];
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+
+			C += 1;
+			A += bs;
+			}
+		C += bs*(sdc-1);
+		}
+	
+	for( ; k<kmax-3; k+=4)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+		C[0+bs*2] = alpha * A[2+bs*0];
+
+		C[1+bs*0] = alpha * A[0+bs*1];
+		C[1+bs*1] = alpha * A[1+bs*1];
+		C[1+bs*2] = alpha * A[2+bs*1];
+
+		C[2+bs*0] = alpha * A[0+bs*2];
+		C[2+bs*1] = alpha * A[1+bs*2];
+		C[2+bs*2] = alpha * A[2+bs*2];
+
+		C[3+bs*0] = alpha * A[0+bs*3];
+		C[3+bs*1] = alpha * A[1+bs*3];
+		C[3+bs*2] = alpha * A[2+bs*3];
+
+		C += bs*sdc;
+		A += bs*bs;
+		}
+	
+	cleanup_loop:
+
+	for( ; k<kmax; k++)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+		C[0+bs*2] = alpha * A[2+bs*0];
+
+		C += 1;
+		A += bs;
+		}
+
+	if(tri==1)
+		{
+		// end 2x2 triangle
+		kna = (bs-(bs-kna+kmax)%bs)%bs;
+
+		if(kna==1)
+			{
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[1+bs*(sdc+1)] = alpha * A[2+bs*1];
+			}
+		else
+			{
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[1+bs*2] = alpha * A[2+bs*1];
+			}
+		}
+
+	}
+
+
+
+// transposed of general matrices, read along panels, write across panels
+void kernel_dgetr_2_lib4(int tri, int kmax, int kna, double alpha, double *A, double *C, int sdc)
+	{
+
+	if(tri==1)
+		{
+		// A is lower triangular, C is upper triangular
+		// kmax+1 2-wide + end 1x1 triangle
+
+		kmax += 1;
+		}
+
+	const int bs = 4;
+	
+	int k;
+
+	k = 0;
+
+	if(kmax<kna)
+		goto cleanup_loop;
+
+	if(kna>0)
+		{
+		for( ; k<kna; k++)
+			{
+			C[0+bs*0] = alpha * A[0+bs*0];
+			C[0+bs*1] = alpha * A[1+bs*0];
+
+			C += 1;
+			A += bs;
+			}
+		C += bs*(sdc-1);
+		}
+	
+	for( ; k<kmax-3; k+=4)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+
+		C[1+bs*0] = alpha * A[0+bs*1];
+		C[1+bs*1] = alpha * A[1+bs*1];
+
+		C[2+bs*0] = alpha * A[0+bs*2];
+		C[2+bs*1] = alpha * A[1+bs*2];
+
+		C[3+bs*0] = alpha * A[0+bs*3];
+		C[3+bs*1] = alpha * A[1+bs*3];
+
+		C += bs*sdc;
+		A += bs*bs;
+		}
+	
+	cleanup_loop:
+
+	for( ; k<kmax; k++)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+
+		C += 1;
+		A += bs;
+		}
+	
+	if(tri==1)
+		{
+		// end 1x1 triangle
+		C[0+bs*1] = alpha * A[1+bs*0];
+		}
+
+	}
+
+
+
+// transposed of general matrices, read along panels, write across panels
+void kernel_dgetr_1_lib4(int tri, int kmax, int kna, double alpha, double *A, double *C, int sdc)
+	{
+
+	if(tri==1)
+		{
+		// A is lower triangular, C is upper triangular
+		// kmax+1 1-wide
+
+		kmax += 1;
+		}
+
+	const int bs = 4;
+	
+	int k;
+
+	k = 0;
+
+	if(kmax<kna)
+		goto cleanup_loop;
+
+	if(kna>0)
+		{
+		for( ; k<kna; k++)
+			{
+			C[0+bs*0] = alpha * A[0+bs*0];
+
+			C += 1;
+			A += bs;
+			}
+		C += bs*(sdc-1);
+		}
+	
+	for( ; k<kmax-3; k+=4)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+
+		C[1+bs*0] = alpha * A[0+bs*1];
+
+		C[2+bs*0] = alpha * A[0+bs*2];
+
+		C[3+bs*0] = alpha * A[0+bs*3];
+
+		C += bs*sdc;
+		A += bs*bs;
+		}
+	
+	cleanup_loop:
+
+	for( ; k<kmax; k++)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+
+		C += 1;
+		A += bs;
+		}
+
+	}
+
+
+
+// transposed of general matrices, read across panels, write along panels
+void kernel_dgetr_4_0_lib4(int kmax, double *A, int sda, double *B)
+	{
+	const int ps = 4;
+	int k;
+	for(k=0; k<kmax-3; k+=4)
+		{
+		//
+		B[0+ps*0] = A[0+ps*0];
+		B[0+ps*1] = A[1+ps*0];
+		B[0+ps*2] = A[2+ps*0];
+		B[0+ps*3] = A[3+ps*0];
+		//
+		B[1+ps*0] = A[0+ps*1];
+		B[1+ps*1] = A[1+ps*1];
+		B[1+ps*2] = A[2+ps*1];
+		B[1+ps*3] = A[3+ps*1];
+		//
+		B[2+ps*0] = A[0+ps*2];
+		B[2+ps*1] = A[1+ps*2];
+		B[2+ps*2] = A[2+ps*2];
+		B[2+ps*3] = A[3+ps*2];
+		//
+		B[3+ps*0] = A[0+ps*3];
+		B[3+ps*1] = A[1+ps*3];
+		B[3+ps*2] = A[2+ps*3];
+		B[3+ps*3] = A[3+ps*3];
+
+		A += ps*sda;
+		B += ps*ps;
+		}
+	for( ; k<kmax; k++)
+		{
+		//
+		B[0+ps*0] = A[0+ps*0];
+		B[1+ps*0] = A[0+ps*1];
+		B[2+ps*0] = A[0+ps*2];
+		B[3+ps*0] = A[0+ps*3];
+
+		A += 1;
+		B += ps;
+		}
+	return;
+	}
+
diff --git a/auxiliary/c99/kernel_sgetr_lib4.c b/auxiliary/c99/kernel_sgetr_lib4.c
new file mode 100644
index 0000000..4cf6fa2
--- /dev/null
+++ b/auxiliary/c99/kernel_sgetr_lib4.c
@@ -0,0 +1,370 @@
+/**************************************************************************************************
+*                                                                                                 *
+* This file is part of BLASFEO.                                                                   *
+*                                                                                                 *
+* BLASFEO -- BLAS For Embedded Optimization.                                                      *
+* Copyright (C) 2016-2017 by Gianluca Frison.                                                     *
+* Developed at IMTEK (University of Freiburg) under the supervision of Moritz Diehl.              *
+* All rights reserved.                                                                            *
+*                                                                                                 *
+* HPMPC is free software; you can redistribute it and/or                                          *
+* modify it under the terms of the GNU Lesser General Public                                      *
+* License as published by the Free Software Foundation; either                                    *
+* version 2.1 of the License, or (at your option) any later version.                              *
+*                                                                                                 *
+* HPMPC is distributed in the hope that it will be useful,                                        *
+* but WITHOUT ANY WARRANTY; without even the implied warranty of                                  *
+* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.                                            *
+* See the GNU Lesser General Public License for more details.                                     *
+*                                                                                                 *
+* You should have received a copy of the GNU Lesser General Public                                *
+* License along with HPMPC; if not, write to the Free Software                                    *
+* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA                  *
+*                                                                                                 *
+* Author: Gianluca Frison, giaf (at) dtu.dk                                                       *
+*                          gianluca.frison (at) imtek.uni-freiburg.de                             *
+*                                                                                                 *
+**************************************************************************************************/
+
+
+
+// transposed of general matrices, read along panels, write across panels
+void kernel_sgetr_4_lib4(int tri, int kmax, int kna, float alpha, float *A, float *C, int sdc)
+	{
+
+	if(tri==1)
+		{
+		// A is lower triangular, C is upper triangular
+		// kmax+1 4-wide + end 3x3 triangle
+
+		kmax += 1;
+		}
+
+	const int bs = 4;
+	
+	int k;
+
+	k = 0;
+
+	if(kmax<kna)
+		goto cleanup_loop;
+
+	if(kna>0)
+		{
+		for( ; k<kna; k++)
+			{
+			C[0+bs*0] = alpha * A[0+bs*0];
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[0+bs*3] = alpha * A[3+bs*0];
+
+			C += 1;
+			A += bs;
+			}
+		C += bs*(sdc-1);
+		}
+	
+	for( ; k<kmax-3; k+=4)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+		C[0+bs*2] = alpha * A[2+bs*0];
+		C[0+bs*3] = alpha * A[3+bs*0];
+
+		C[1+bs*0] = alpha * A[0+bs*1];
+		C[1+bs*1] = alpha * A[1+bs*1];
+		C[1+bs*2] = alpha * A[2+bs*1];
+		C[1+bs*3] = alpha * A[3+bs*1];
+
+		C[2+bs*0] = alpha * A[0+bs*2];
+		C[2+bs*1] = alpha * A[1+bs*2];
+		C[2+bs*2] = alpha * A[2+bs*2];
+		C[2+bs*3] = alpha * A[3+bs*2];
+
+		C[3+bs*0] = alpha * A[0+bs*3];
+		C[3+bs*1] = alpha * A[1+bs*3];
+		C[3+bs*2] = alpha * A[2+bs*3];
+		C[3+bs*3] = alpha * A[3+bs*3];
+
+		C += bs*sdc;
+		A += bs*bs;
+		}
+
+	cleanup_loop:
+
+	for( ; k<kmax; k++)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+		C[0+bs*2] = alpha * A[2+bs*0];
+		C[0+bs*3] = alpha * A[3+bs*0];
+
+		C += 1;
+		A += bs;
+		}
+
+	if(tri==1)
+		{
+		// end 3x3 triangle
+		kna = (bs-(bs-kna+kmax)%bs)%bs;
+
+		if(kna==1)
+			{
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[0+bs*3] = alpha * A[3+bs*0];
+			C[1+bs*(sdc+1)] = alpha * A[2+bs*1];
+			C[1+bs*(sdc+2)] = alpha * A[3+bs*1];
+			C[2+bs*(sdc+2)] = alpha * A[3+bs*2];
+			}
+		else if(kna==2)
+			{
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[0+bs*3] = alpha * A[3+bs*0];
+			C[1+bs*2] = alpha * A[2+bs*1];
+			C[1+bs*3] = alpha * A[3+bs*1];
+			C[2+bs*(sdc+2)] = alpha * A[3+bs*2];
+			}
+		else
+			{
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[0+bs*3] = alpha * A[3+bs*0];
+			C[1+bs*2] = alpha * A[2+bs*1];
+			C[1+bs*3] = alpha * A[3+bs*1];
+			C[2+bs*3] = alpha * A[3+bs*2];
+			}
+		}
+
+	}
+
+
+
+// transposed of general matrices, read along panels, write across panels
+void kernel_sgetr_3_lib4(int tri, int kmax, int kna, float alpha, float *A, float *C, int sdc)
+	{
+
+	if(tri==1)
+		{
+		// A is lower triangular, C is upper triangular
+		// kmax+1 3-wide + end 2x2 triangle
+
+		kmax += 1;
+		}
+
+	const int bs = 4;
+	
+	int k;
+
+	k = 0;
+
+	if(kmax<kna)
+		goto cleanup_loop;
+
+	if(kna>0)
+		{
+		for( ; k<kna; k++)
+			{
+			C[0+bs*0] = alpha * A[0+bs*0];
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+
+			C += 1;
+			A += bs;
+			}
+		C += bs*(sdc-1);
+		}
+	
+	for( ; k<kmax-3; k+=4)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+		C[0+bs*2] = alpha * A[2+bs*0];
+
+		C[1+bs*0] = alpha * A[0+bs*1];
+		C[1+bs*1] = alpha * A[1+bs*1];
+		C[1+bs*2] = alpha * A[2+bs*1];
+
+		C[2+bs*0] = alpha * A[0+bs*2];
+		C[2+bs*1] = alpha * A[1+bs*2];
+		C[2+bs*2] = alpha * A[2+bs*2];
+
+		C[3+bs*0] = alpha * A[0+bs*3];
+		C[3+bs*1] = alpha * A[1+bs*3];
+		C[3+bs*2] = alpha * A[2+bs*3];
+
+		C += bs*sdc;
+		A += bs*bs;
+		}
+	
+	cleanup_loop:
+
+	for( ; k<kmax; k++)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+		C[0+bs*2] = alpha * A[2+bs*0];
+
+		C += 1;
+		A += bs;
+		}
+
+	if(tri==1)
+		{
+		// end 2x2 triangle
+		kna = (bs-(bs-kna+kmax)%bs)%bs;
+
+		if(kna==1)
+			{
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[1+bs*(sdc+1)] = alpha * A[2+bs*1];
+			}
+		else
+			{
+			C[0+bs*1] = alpha * A[1+bs*0];
+			C[0+bs*2] = alpha * A[2+bs*0];
+			C[1+bs*2] = alpha * A[2+bs*1];
+			}
+		}
+
+	}
+
+
+
+// transposed of general matrices, read along panels, write across panels
+void kernel_sgetr_2_lib4(int tri, int kmax, int kna, float alpha, float *A, float *C, int sdc)
+	{
+
+	if(tri==1)
+		{
+		// A is lower triangular, C is upper triangular
+		// kmax+1 2-wide + end 1x1 triangle
+
+		kmax += 1;
+		}
+
+	const int bs = 4;
+	
+	int k;
+
+	k = 0;
+
+	if(kmax<kna)
+		goto cleanup_loop;
+
+	if(kna>0)
+		{
+		for( ; k<kna; k++)
+			{
+			C[0+bs*0] = alpha * A[0+bs*0];
+			C[0+bs*1] = alpha * A[1+bs*0];
+
+			C += 1;
+			A += bs;
+			}
+		C += bs*(sdc-1);
+		}
+	
+	for( ; k<kmax-3; k+=4)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+
+		C[1+bs*0] = alpha * A[0+bs*1];
+		C[1+bs*1] = alpha * A[1+bs*1];
+
+		C[2+bs*0] = alpha * A[0+bs*2];
+		C[2+bs*1] = alpha * A[1+bs*2];
+
+		C[3+bs*0] = alpha * A[0+bs*3];
+		C[3+bs*1] = alpha * A[1+bs*3];
+
+		C += bs*sdc;
+		A += bs*bs;
+		}
+	
+	cleanup_loop:
+
+	for( ; k<kmax; k++)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+		C[0+bs*1] = alpha * A[1+bs*0];
+
+		C += 1;
+		A += bs;
+		}
+	
+	if(tri==1)
+		{
+		// end 1x1 triangle
+		C[0+bs*1] = alpha * A[1+bs*0];
+		}
+
+	}
+
+
+
+// transposed of general matrices, read along panels, write across panels
+void kernel_sgetr_1_lib4(int tri, int kmax, int kna, float alpha, float *A, float *C, int sdc)
+	{
+
+	if(tri==1)
+		{
+		// A is lower triangular, C is upper triangular
+		// kmax+1 1-wide
+
+		kmax += 1;
+		}
+
+	const int bs = 4;
+	
+	int k;
+
+	k = 0;
+
+	if(kmax<kna)
+		goto cleanup_loop;
+
+	if(kna>0)
+		{
+		for( ; k<kna; k++)
+			{
+			C[0+bs*0] = alpha * A[0+bs*0];
+
+			C += 1;
+			A += bs;
+			}
+		C += bs*(sdc-1);
+		}
+	
+	for( ; k<kmax-3; k+=4)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+
+		C[1+bs*0] = alpha * A[0+bs*1];
+
+		C[2+bs*0] = alpha * A[0+bs*2];
+
+		C[3+bs*0] = alpha * A[0+bs*3];
+
+		C += bs*sdc;
+		A += bs*bs;
+		}
+	
+	cleanup_loop:
+
+	for( ; k<kmax; k++)
+		{
+		C[0+bs*0] = alpha * A[0+bs*0];
+
+		C += 1;
+		A += bs;
+		}
+
+	}
+
+
+
+