diff --git a/.travis.yml b/.travis.yml
index 4efa23b8d..3f323a854 100644
--- a/.travis.yml
+++ b/.travis.yml
@@ -117,7 +117,7 @@ matrix:
     - <<: *test-alpine
       env:
         - TARGET_BOX=LINUX64_MUSL
-        - BTYPE="BINARY=64 NO_AFFINITY=1 USE_OPENMP=0 NO_LAPACK=0 TARGET=core2"
+        - BTYPE="BINARY=64 NO_AFFINITY=1 USE_OPENMP=0 NO_LAPACK=0 TARGET=CORE2"
 
     - &test-cmake
       os: linux
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 296113941..24c169afe 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -6,7 +6,7 @@ cmake_minimum_required(VERSION 2.8.5)
 project(OpenBLAS C ASM)
 set(OpenBLAS_MAJOR_VERSION 0)
 set(OpenBLAS_MINOR_VERSION 3)
-set(OpenBLAS_PATCH_VERSION 4)
+set(OpenBLAS_PATCH_VERSION 5.dev)
 set(OpenBLAS_VERSION "${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.${OpenBLAS_PATCH_VERSION}")
 
 # Adhere to GNU filesystem layout conventions
diff --git a/Changelog.txt b/Changelog.txt
index 0dd17a558..49b26873a 100644
--- a/Changelog.txt
+++ b/Changelog.txt
@@ -1,4 +1,36 @@
 OpenBLAS ChangeLog
+====================================================================
+Version 0.3.5
+31-Dec-2018
+
+common:
+	* loop unrolling in TRMV has been enabled again.
+	* A domain error in the thread workload distribution for SYRK
+	  has been fixed.
+	* gmake builds will now automatically add -fPIC to the build
+	  options if the platform requires it.
+	* a pthreads key leakage (and associate crash on dlclose) in
+	  the USE_TLS codepath was fixed.
+	* building of the utest cases on systems that do not provide
+	  an implementation of complex.h was fixed.
+	  
+x86_64:
+	* the SkylakeX code was changed to compile on OSX.
+	* unwanted application of the -march=skylake-avx512 option
+	  to the common code parts of a DYNAMIC_ARCH build was fixed.
+	* improved performance of SGEMM for small workloads on Skylake X.
+	* performance of SGEMM and DGEMM was improved on Haswell.
+
+ARMV8:
+	* a configuration error that broke the CNRM2 kernel was corrected.
+	* compilation of the GEMM kernels with CMAKE was fixed.
+	* DYNAMIC_ARCH builds are now available with CMAKE as well.
+	* using CMAKE for cross-compilation to the new cpu TARGETs
+	  introduced in 0.3.4 now works.
+	  
+POWER:
+	* a problem in cpu autodetection for AIX has been corrected.
+	
 ====================================================================
 Version 0.3.4
 02-Dec-2018
diff --git a/Makefile b/Makefile
index d42f9b8c3..21096f893 100644
--- a/Makefile
+++ b/Makefile
@@ -131,7 +131,7 @@ endif
 endif
 
 libs :
-ifeq ($(CORE), UNKOWN)
+ifeq ($(CORE), UNKNOWN)
 	$(error OpenBLAS: Detecting CPU failed. Please set TARGET explicitly, e.g. make TARGET=your_cpu_target. Please read README for the detail.)
 endif
 ifeq ($(NOFORTRAN), 1)
diff --git a/Makefile.arm64 b/Makefile.arm64
index a529fab80..cd16dbfae 100644
--- a/Makefile.arm64
+++ b/Makefile.arm64
@@ -30,8 +30,8 @@ FCOMMON_OPT += -march=armv8-a -mtune=thunderx
 endif
 
 ifeq ($(CORE), FALKOR)
-CCOMMON_OPT += -march=armv8.1-a -mtune=falkor
-FCOMMON_OPT += -march=armv8.1-a -mtune=falkor
+CCOMMON_OPT += -march=armv8-a -mtune=falkor
+FCOMMON_OPT += -march=armv8-a -mtune=falkor
 endif
 
 ifeq ($(CORE), THUNDERX2T99)
diff --git a/Makefile.rule b/Makefile.rule
index f3086a01b..0d5b83b39 100644
--- a/Makefile.rule
+++ b/Makefile.rule
@@ -3,7 +3,7 @@
 #
 
 # This library's version
-VERSION = 0.3.4
+VERSION = 0.3.5.dev
 
 # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a
 # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library
diff --git a/Makefile.system b/Makefile.system
index 22fe24337..fb8e7ea41 100644
--- a/Makefile.system
+++ b/Makefile.system
@@ -12,6 +12,12 @@ endif
 # Catch conflicting usage of ARCH in some BSD environments
 ifeq ($(ARCH), amd64)
 override ARCH=x86_64
+else ifeq ($(ARCH), powerpc64)
+override ARCH=power
+else ifeq ($(ARCH), i386)
+override ARCH=x86
+else ifeq ($(ARCH), aarch64)
+override ARCH=arm64
 endif
 
 NETLIB_LAPACK_DIR = $(TOPDIR)/lapack-netlib
@@ -1148,8 +1154,6 @@ ifndef FCOMMON_OPT
 FCOMMON_OPT = -O2 -frecursive
 endif
 
-
-
 override CFLAGS     += $(COMMON_OPT) $(CCOMMON_OPT) -I$(TOPDIR)
 override PFLAGS     += $(COMMON_OPT) $(CCOMMON_OPT) -I$(TOPDIR) -DPROFILE $(COMMON_PROF)
 
@@ -1157,6 +1161,12 @@ override FFLAGS     += $(COMMON_OPT) $(FCOMMON_OPT)
 override FPFLAGS    += $(FCOMMON_OPT) $(COMMON_PROF)
 #MAKEOVERRIDES =
 
+ifdef NEED_PIC
+ifeq (,$(findstring PIC,$(FFLAGS)))
+override FFLAGS += -fPIC
+endif
+endif
+
 #For LAPACK Fortran codes.
 #Disable -fopenmp for LAPACK Fortran codes on Windows.
 ifdef OS_WINDOWS
diff --git a/Makefile.x86_64 b/Makefile.x86_64
index f2647fb7d..1b7fe3ef4 100644
--- a/Makefile.x86_64
+++ b/Makefile.x86_64
@@ -9,6 +9,7 @@ endif
 endif
 
 ifeq ($(CORE), SKYLAKEX)
+ifndef DYNAMIC_ARCH
 ifndef NO_AVX512
 CCOMMON_OPT += -march=skylake-avx512
 FCOMMON_OPT += -march=skylake-avx512
@@ -22,6 +23,18 @@ endif
 endif
 endif
 endif
+endif
+
+ifeq ($(CORE), HASWELL)
+ifndef DYNAMIC_ARCH
+ifndef NO_AVX2
+CCOMMON_OPT += -mavx2
+FCOMMON_OPT += -mavx2
+endif
+endif
+endif
+
+
 
 ifeq ($(OSNAME), Interix)
 ARFLAGS		= -m x64
diff --git a/README.md b/README.md
index 9ed9be337..26055c745 100644
--- a/README.md
+++ b/README.md
@@ -201,7 +201,7 @@ Please see Changelog.txt to view the differences between OpenBLAS and GotoBLAS2
 * Please use GCC version 4.6 and above to compile Sandy Bridge AVX kernels on Linux/MinGW/BSD.
 * Please use Clang version 3.1 and above to compile the library on Sandy Bridge microarchitecture.
   Clang 3.0 will generate the wrong AVX binary code.
-* Please use GCC version 6 or LLVM version 6 and above to compile Skyalke AVX512 kernels.
+* Please use GCC version 6 or LLVM version 6 and above to compile Skylake AVX512 kernels.
 * The number of CPUs/cores should less than or equal to 256. On Linux `x86_64` (`amd64`),
   there is experimental support for up to 1024 CPUs/cores and 128 numa nodes if you build
   the library with `BIGNUMA=1`.
diff --git a/cmake/arch.cmake b/cmake/arch.cmake
index 52fb64eaa..63fb86fa2 100644
--- a/cmake/arch.cmake
+++ b/cmake/arch.cmake
@@ -44,6 +44,10 @@ endif ()
 
 
 if (DYNAMIC_ARCH)
+  if (ARM64)
+    set(DYNAMIC_CORE ARMV8 CORTEXA53 CORTEXA57 CORTEXA72 CORTEXA73 FALKOR THUNDERX THUNDERX2T99)
+  endif ()
+  
   if (X86)
     set(DYNAMIC_CORE KATMAI COPPERMINE NORTHWOOD PRESCOTT BANIAS CORE2 PENRYN DUNNINGTON NEHALEM ATHLON OPTERON OPTERON_SSE3 BARCELONA BOBCAT ATOM NANO)
   endif ()
diff --git a/cmake/prebuild.cmake b/cmake/prebuild.cmake
index f29bc3a75..6ed99e807 100644
--- a/cmake/prebuild.cmake
+++ b/cmake/prebuild.cmake
@@ -116,18 +116,19 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
       "#define L2_LINESIZE\t64\n"
       "#define DTB_DEFAULT_ENTRIES\t64\n"
       "#define DTB_SIZE\t4096\n"
-      "#define L2_ASSOCIATIVE\t32\n")
+      "#define L2_ASSOCIATIVE\t32\n"
+      "#define ARMV8\n")
     set(SGEMM_UNROLL_M 4)
     set(SGEMM_UNROLL_N 4)
-  elseif ("${CORE}" STREQUAL "CORTEXA57")
+  elseif ("${CORE}" STREQUAL "CORTEXA57" OR "${CORE}" STREQUAL "CORTEXA53")
     file(APPEND ${TARGET_CONF_TEMP}
-      "#define L1_CODE_SIZE\t49152\n"
+      "#define L1_CODE_SIZE\t32768\n"
       "#define L1_CODE_LINESIZE\t64\n"
       "#define L1_CODE_ASSOCIATIVE\t3\n"
       "#define L1_DATA_SIZE\t32768\n"
       "#define L1_DATA_LINESIZE\t64\n"
       "#define L1_DATA_ASSOCIATIVE\t2\n"
-      "#define L2_SIZE\t2097152\n"
+      "#define L2_SIZE\t262144\n"
       "#define L2_LINESIZE\t64\n"
       "#define L2_ASSOCIATIVE\t16\n"
       "#define DTB_DEFAULT_ENTRIES\t64\n"
@@ -135,7 +136,8 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
       "#define HAVE_VFPV4\n"
       "#define HAVE_VFPV3\n"
       "#define HAVE_VFP\n"
-      "#define HAVE_NEON\n")
+      "#define HAVE_NEON\n"
+      "#define ARMV8\n")
     set(SGEMM_UNROLL_M 16)
     set(SGEMM_UNROLL_N 4)
     set(DGEMM_UNROLL_M 8)
@@ -144,6 +146,109 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
     set(CGEMM_UNROLL_N 4)
     set(ZGEMM_UNROLL_M 8)
     set(ZGEMM_UNROLL_N 4)
+  elseif ("${CORE}" STREQUAL "CORTEXA72" OR "${CORE}" STREQUAL "CORTEXA73")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_CODE_SIZE\t49152\n"
+      "#define L1_CODE_LINESIZE\t64\n"
+      "#define L1_CODE_ASSOCIATIVE\t3\n"
+      "#define L1_DATA_SIZE\t32768\n"
+      "#define L1_DATA_LINESIZE\t64\n"
+      "#define L1_DATA_ASSOCIATIVE\t2\n"
+      "#define L2_SIZE\t524288\n"
+      "#define L2_LINESIZE\t64\n"
+      "#define L2_ASSOCIATIVE\t16\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define HAVE_VFPV4\n"
+      "#define HAVE_VFPV3\n"
+      "#define HAVE_VFP\n"
+      "#define HAVE_NEON\n"
+      "#define ARMV8\n")
+    set(SGEMM_UNROLL_M 16)
+    set(SGEMM_UNROLL_N 4)
+    set(DGEMM_UNROLL_M 8)
+    set(DGEMM_UNROLL_N 4)
+    set(CGEMM_UNROLL_M 8)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 8)
+    set(ZGEMM_UNROLL_N 4)
+  elseif ("${CORE}" STREQUAL "FALKOR")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_CODE_SIZE\t65536\n"
+      "#define L1_CODE_LINESIZE\t64\n"
+      "#define L1_CODE_ASSOCIATIVE\t3\n"
+      "#define L1_DATA_SIZE\t32768\n"
+      "#define L1_DATA_LINESIZE\t128\n"
+      "#define L1_DATA_ASSOCIATIVE\t2\n"
+      "#define L2_SIZE\t524288\n"
+      "#define L2_LINESIZE\t64\n"
+      "#define L2_ASSOCIATIVE\t16\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define HAVE_VFPV4\n"
+      "#define HAVE_VFPV3\n"
+      "#define HAVE_VFP\n"
+      "#define HAVE_NEON\n"
+      "#define ARMV8\n")
+    set(SGEMM_UNROLL_M 16)
+    set(SGEMM_UNROLL_N 4)
+    set(DGEMM_UNROLL_M 8)
+    set(DGEMM_UNROLL_N 4)
+    set(CGEMM_UNROLL_M 8)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 8)
+    set(ZGEMM_UNROLL_N 4)
+  elseif ("${CORE}" STREQUAL "THUNDERX)
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_CODE_SIZE\t32768\n"
+      "#define L1_CODE_LINESIZE\t64\n"
+      "#define L1_CODE_ASSOCIATIVE\t3\n"
+      "#define L1_DATA_SIZE\t32768\n"
+      "#define L1_DATA_LINESIZE\t128\n"
+      "#define L1_DATA_ASSOCIATIVE\t2\n"
+      "#define L2_SIZE\t167772164\n"
+      "#define L2_LINESIZE\t128\n"
+      "#define L2_ASSOCIATIVE\t16\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define HAVE_VFPV4\n"
+      "#define HAVE_VFPV3\n"
+      "#define HAVE_VFP\n"
+      "#define HAVE_NEON\n"
+      "#define ARMV8\n")
+    set(SGEMM_UNROLL_M 4)
+    set(SGEMM_UNROLL_N 4)
+    set(DGEMM_UNROLL_M 2)
+    set(DGEMM_UNROLL_N 2)
+    set(CGEMM_UNROLL_M 2)
+    set(CGEMM_UNROLL_N 2)
+    set(ZGEMM_UNROLL_M 2)
+    set(ZGEMM_UNROLL_N 2)
+  elseif ("${CORE}" STREQUAL "THUNDERX2T99)
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_CODE_SIZE\t32768\n"
+      "#define L1_CODE_LINESIZE\t64\n"
+      "#define L1_CODE_ASSOCIATIVE\t8\n"
+      "#define L1_DATA_SIZE\t32768\n"
+      "#define L1_DATA_LINESIZE\t64\n"
+      "#define L1_DATA_ASSOCIATIVE\t8\n"
+      "#define L2_SIZE\t262144\n"
+      "#define L2_LINESIZE\t64\n"
+      "#define L2_ASSOCIATIVE\t8\n"
+      "#define L3_SIZE\t33554432\n"
+      "#define L3_LINESIZE\t64\n"
+      "#define L3_ASSOCIATIVE\t32\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define VULCAN\n")
+    set(SGEMM_UNROLL_M 16)
+    set(SGEMM_UNROLL_N 4)
+    set(DGEMM_UNROLL_M 8)
+    set(DGEMM_UNROLL_N 4)
+    set(CGEMM_UNROLL_M 8)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 4)
+    set(ZGEMM_UNROLL_N 4)
   endif()
 
   # Or should this actually be NUM_CORES?
@@ -163,6 +268,7 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
   file(APPEND ${TARGET_CONF_TEMP}
     "#define GEMM_MULTITHREAD_THRESHOLD\t${GEMM_MULTITHREAD_THRESHOLD}\n")
   # Move to where gen_config_h would place it
+  file(MAKE_DIRECTORY ${TARGET_CONF_DIR})
   file(RENAME ${TARGET_CONF_TEMP} "${TARGET_CONF_DIR}/${TARGET_CONF}")  
 
 else(NOT CMAKE_CROSSCOMPILING)
diff --git a/cmake/system.cmake b/cmake/system.cmake
index d803bb9eb..a060d98cb 100644
--- a/cmake/system.cmake
+++ b/cmake/system.cmake
@@ -42,9 +42,19 @@ if (DEFINED BINARY AND DEFINED TARGET AND BINARY EQUAL 32)
 endif ()
 
 if (DEFINED TARGET)
-if (${TARGET} STREQUAL "SKYLAKEX" AND NOT NO_AVX512)
-  set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
-endif()
+  if (${TARGET} STREQUAL "SKYLAKEX" AND NOT NO_AVX512)
+    set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
+  endif()
+  if (${TARGET} STREQUAL "HASWELL" AND NOT NO_AVX2)
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
+      execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
+      if (${GCC_VERSION} VERSION_GREATER 4.7 OR ${GCC_VERSION} VERSION_EQUAL 4.7)
+        set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx2")
+      endif()
+    elseif (${CMAKE_C_COMPILER_ID} STREQUAL "CLANG")
+      set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx2")
+    endif()
+  endif()
 endif()
 
 if (DEFINED TARGET)
diff --git a/common_level3.h b/common_level3.h
index 1f5490baa..6fa902be8 100644
--- a/common_level3.h
+++ b/common_level3.h
@@ -47,6 +47,14 @@ __global__ void cuda_dgemm_kernel(int, int, int, double *, double *, double *);
 extern "C" {
 #endif
 
+extern void sgemm_kernel_direct(BLASLONG M, BLASLONG N, BLASLONG K,
+	float * A, BLASLONG strideA,
+	float * B, BLASLONG strideB,
+	float * R, BLASLONG strideR);
+
+extern int sgemm_kernel_direct_performant(BLASLONG M, BLASLONG N, BLASLONG K);
+
+
 int sgemm_beta(BLASLONG, BLASLONG, BLASLONG, float,
 	       float  *, BLASLONG, float   *, BLASLONG, float  *, BLASLONG);
 int dgemm_beta(BLASLONG, BLASLONG, BLASLONG, double,
diff --git a/cpuid_arm.c b/cpuid_arm.c
index 2f8959242..19aa90718 100644
--- a/cpuid_arm.c
+++ b/cpuid_arm.c
@@ -34,7 +34,7 @@
 #define CPU_CORTEXA15       	4
 
 static char *cpuname[] = {
-  "UNKOWN",
+  "UNKNOWN",
   "ARMV6",
   "ARMV7",
   "CORTEXA9",
diff --git a/cpuid_arm64.c b/cpuid_arm64.c
index c914fbc2b..5077d7b11 100644
--- a/cpuid_arm64.c
+++ b/cpuid_arm64.c
@@ -270,7 +270,7 @@ void get_cpuconfig(void)
 			break;
 
 		case CPU_THUNDERX2T99:
-			printf("#define VULCAN                        \n");
+			printf("#define THUNDERX2T99                  \n");
 			printf("#define L1_CODE_SIZE         32768    \n");
 			printf("#define L1_CODE_LINESIZE     64       \n");
 			printf("#define L1_CODE_ASSOCIATIVE  8        \n");
diff --git a/cpuid_mips.c b/cpuid_mips.c
index c09902936..6f2932c94 100644
--- a/cpuid_mips.c
+++ b/cpuid_mips.c
@@ -75,7 +75,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CPU_1004K	2
 
 static char *cpuname[] = {
-  "UNKOWN",
+  "UNKNOWN",
   "P5600",
   "1004K"
 };
diff --git a/cpuid_mips64.c b/cpuid_mips64.c
index dcb559a7c..0e32bfc0b 100644
--- a/cpuid_mips64.c
+++ b/cpuid_mips64.c
@@ -79,7 +79,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CPU_I6500       6
 
 static char *cpuname[] = {
-  "UNKOWN",
+  "UNKNOWN",
   "SICORTEX",
   "LOONGSON3A",
   "LOONGSON3B",
diff --git a/cpuid_power.c b/cpuid_power.c
index 23e98ebb0..82a3f4aac 100644
--- a/cpuid_power.c
+++ b/cpuid_power.c
@@ -136,7 +136,7 @@ int detect(void){
   char buffer[512], *p;
 
   p = (char *)NULL;
-  infile = popen("prtconf|grep 'Processor Type'");
+  infile = popen("prtconf|grep 'Processor Type'", "r");
   while (fgets(buffer, sizeof(buffer), infile)){
     if (!strncmp("Pro", buffer, 3)){
 	p = strchr(buffer, ':') + 2;
diff --git a/cpuid_x86.c b/cpuid_x86.c
index 8e4a7cb84..eb986b6b6 100644
--- a/cpuid_x86.c
+++ b/cpuid_x86.c
@@ -1649,7 +1649,7 @@ static char *lowercpuname[] = {
 };
 
 static char *corename[] = {
-  "UNKOWN",
+  "UNKNOWN",
   "80486",
   "P5",
   "P6",
diff --git a/driver/level2/trmv_U.c b/driver/level2/trmv_U.c
index 7f8895e7f..90ffb7370 100644
--- a/driver/level2/trmv_U.c
+++ b/driver/level2/trmv_U.c
@@ -54,16 +54,12 @@ int CNAME(BLASLONG m, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, FLOAT *bu
     COPY_K(m, b, incb, buffer, 1);
   }
 
-/*FIXME the GEMV unrolling performed here was found to be broken, see issue 1332 */
-/* Multiplying DTB size by 100 is just a quick-and-dirty hack to disable it for now[B */
+  for (is = 0; is < m; is += DTB_ENTRIES){
 
-  for (is = 0; is < m; is += DTB_ENTRIES * 100){
-
-    min_i = MIN(m - is, DTB_ENTRIES * 100);
+    min_i = MIN(m - is, DTB_ENTRIES);
 
 #ifndef TRANSA
-    if (is > 0){
-fprintf(stderr,"WARNING unrolling of the trmv_U loop may give wrong results\n");    
+    if (is > 0){ 
       GEMV_N(is, min_i, 0, dp1,
 	     a + is * lda,  lda,
 	     B + is, 1,
diff --git a/driver/level3/syrk_thread.c b/driver/level3/syrk_thread.c
index 5f40853dc..b26d363c4 100644
--- a/driver/level3/syrk_thread.c
+++ b/driver/level3/syrk_thread.c
@@ -48,7 +48,7 @@ int CNAME(int mode, blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n, int (
 
   BLASLONG width, i;
   BLASLONG n_from, n_to;
-  double dnum, nf, nt, di;
+  double dnum, nf, nt, di, dinum;
 
   int num_cpu;
   int mask = 0;
@@ -109,7 +109,11 @@ int CNAME(int mode, blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n, int (
       if (nthreads - num_cpu > 1) {
 
 	di = (double)i;
-	width = (BLASLONG)(( sqrt(di * di + dnum) - di + mask)/(mask+1)) * (mask+1);
+	dinum = di * di +dnum;
+	if (dinum <0)
+	  width = (BLASLONG)(( - di + mask)/(mask+1)) * (mask+1);
+	else
+	  width = (BLASLONG)(( sqrt(dinum) - di + mask)/(mask+1)) * (mask+1);
 
 	if ((width <= 0) || (width > n_to - i)) width = n_to - i;
 
@@ -136,9 +140,7 @@ int CNAME(int mode, blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n, int (
 
     nf = (double)(arg -> n - n_from);
     nt = (double)(arg -> n - n_to);
-
     dnum = (nt * nt - nf * nf) / (double)nthreads;
-
     num_cpu  = 0;
 
     range[0] = n_from;
@@ -149,8 +151,11 @@ int CNAME(int mode, blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n, int (
       if (nthreads - num_cpu > 1) {
 
 	di = (double)(arg -> n - i);
-	width = ((BLASLONG)((-sqrt(di * di + dnum) + di) + mask)/(mask+1)) * (mask+1);
-
+	dinum = di * di + dnum;
+	if (dinum<0)
+	  width = ((BLASLONG)(di + mask)/(mask+1)) * (mask+1);
+	else
+	  width = ((BLASLONG)((-sqrt(dinum) + di) + mask)/(mask+1)) * (mask+1);
 	if ((width <= 0) || (width > n_to - i)) width = n_to - i;
 
       } else {
diff --git a/driver/others/CMakeLists.txt b/driver/others/CMakeLists.txt
index e20b14e79..a07e00b3b 100644
--- a/driver/others/CMakeLists.txt
+++ b/driver/others/CMakeLists.txt
@@ -47,7 +47,11 @@ GenerateNamedObjects("abs.c" "DOUBLE" "z_abs" 0 "" "" 1)
 GenerateNamedObjects("openblas_get_config.c;openblas_get_parallel.c" "" "" 0 "" "" 1)
 
 if (DYNAMIC_ARCH)
-  list(APPEND COMMON_SOURCES dynamic.c)
+  if (ARM64)
+    list(APPEND COMMON_SOURCES dynamic_arm64.c)
+  else ()  
+    list(APPEND COMMON_SOURCES dynamic.c)
+  endif ()  
 else ()
   list(APPEND COMMON_SOURCES parameter.c)
 endif ()
diff --git a/driver/others/memory.c b/driver/others/memory.c
index 36815a39c..6f7a7db82 100644
--- a/driver/others/memory.c
+++ b/driver/others/memory.c
@@ -1073,6 +1073,11 @@ static volatile int memory_initialized = 0;
     }
     free(table);
   }
+#if defined(OS_WINDOWS)
+  TlsFree(local_storage_key);
+#else
+  pthread_key_delete(local_storage_key);
+#endif		
 }
 
 static void blas_memory_init(){
diff --git a/interface/gemm.c b/interface/gemm.c
index a3bac5984..97e71bc85 100644
--- a/interface/gemm.c
+++ b/interface/gemm.c
@@ -271,6 +271,14 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
 
   PRINT_DEBUG_CNAME;
 
+#if !defined(COMPLEX) && !defined(DOUBLE) && defined(USE_SGEMM_KERNEL_DIRECT)
+  if (beta == 0 && alpha == 1.0 && order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans && sgemm_kernel_direct_performant(m,n,k)) {
+	sgemm_kernel_direct(m, n, k, a, lda, b, ldb, c, ldc);
+	return;
+  }
+
+#endif
+
 #ifndef COMPLEX
   args.alpha = (void *)&alpha;
   args.beta  = (void *)&beta;
diff --git a/kernel/CMakeLists.txt b/kernel/CMakeLists.txt
index 947114ebe..2a330df4e 100644
--- a/kernel/CMakeLists.txt
+++ b/kernel/CMakeLists.txt
@@ -125,10 +125,13 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
       set(USE_TRMM true)
     endif ()
 
-    foreach (float_type ${FLOAT_TYPES})
+    foreach (float_type SINGLE DOUBLE)
       string(SUBSTRING ${float_type} 0 1 float_char)
       GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMKERNEL}" "" "gemm_kernel" false "" "" false ${float_type})
+    endforeach()
 
+    foreach (float_type ${FLOAT_TYPES})
+      string(SUBSTRING ${float_type} 0 1 float_char)
       if (${float_char}GEMMINCOPY)
         GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMINCOPY}" "${float_type}" "${${float_char}GEMMINCOPYOBJ}" false "" "" true ${float_type})
       endif ()
diff --git a/kernel/Makefile b/kernel/Makefile
index 923ffc363..e81225075 100644
--- a/kernel/Makefile
+++ b/kernel/Makefile
@@ -5,8 +5,43 @@ endif
 TOPDIR	= ..
 include $(TOPDIR)/Makefile.system
 
+AVX2OPT = 
+ifeq ($(C_COMPILER), GCC)
+# AVX2 support was added in 4.7.0
+  GCCVERSIONGTEQ4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 4)
+  GCCMINORVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 7)
+  ifeq ($(GCCVERSIONGTEQ4)$(GCCMINORVERSIONGTEQ7), 11)
+   AVX2OPT = -mavx2
+  endif
+endif
+ifeq ($(C_COMPILER), CLANG)
+# Any clang posing as gcc 4.2 should be new enough (3.4 or later)
+  GCCVERSIONGTEQ4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 4)
+  GCCMINORVERSIONGTEQ2 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 2)
+  ifeq ($(GCCVERSIONGTEQ4)$(GCCMINORVERSIONGTEQ2), 11)
+   AVX2OPT = -mavx2
+  endif
+endif
+ifdef NO_AVX2
+ AVX2OPT=
+endif
+
 ifdef TARGET_CORE
-override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE)
+ifeq ($(TARGET_CORE), SKYLAKEX)
+ override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE) -march=skylake-avx512
+ ifeq ($(OSNAME), CYGWIN_NT)
+  override CFLAGS += -fno-asynchronous-unwind-tables
+ endif
+ ifeq ($(OSNAME), WINNT)
+  ifeq ($(C_COMPILER), GCC)
+   override CFLAGS += -fno-asynchronous-unwind-tables
+  endif
+ endif
+else ifeq ($(TARGET_CORE), HASWELL)
+ override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE) $(AVX2OPT)
+else
+ override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE)
+endif
 BUILD_KERNEL = 1
 KDIR =
 TSUFFIX = _$(TARGET_CORE)
diff --git a/kernel/arm64/KERNEL.ARMV8 b/kernel/arm64/KERNEL.ARMV8
index 5c70390dc..a2a435738 100644
--- a/kernel/arm64/KERNEL.ARMV8
+++ b/kernel/arm64/KERNEL.ARMV8
@@ -93,8 +93,8 @@ IZAMAXKERNEL   = izamax.S
 
 ifneq ($(OS_DARWIN)$(CROSS),11)
 SNRM2KERNEL    = nrm2.S
-CNRM2KERNEL    = nrm2.S
-DNRM2KERNEL    = znrm2.S
+DNRM2KERNEL    = nrm2.S
+CNRM2KERNEL    = znrm2.S
 ZNRM2KERNEL    = znrm2.S
 endif
 
@@ -104,8 +104,38 @@ CDOTKERNEL     = zdot.S
 ZDOTKERNEL     = zdot.S
 DSDOTKERNEL    = dot.S
 
-ifneq ($(OS_DARWIN)$(CROSS),11)
+ifeq ($(OS_DARWIN)$(CROSS),11)
 
+STRMMKERNEL	= ../generic/trmmkernel_2x2.c
+DTRMMKERNEL	= ../generic/trmmkernel_2x2.c
+CTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
+ZTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
+
+SGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
+SGEMMONCOPY    =  ../generic/gemm_ncopy_2.c
+SGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+DGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
+DGEMMONCOPY    = ../generic/gemm_ncopy_2.c
+DGEMMOTCOPY    = ../generic/gemm_tcopy_2.c
+DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+CGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
+CGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
+CGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+ZGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
+ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
+ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+else
 SGEMMKERNEL    =  sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
 STRMMKERNEL    =  strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
 ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
@@ -173,35 +203,4 @@ ZGEMMOTCOPY    =  ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c
 ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
 ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
-else
-
-STRMMKERNEL	= ../generic/trmmkernel_2x2.c
-DTRMMKERNEL	= ../generic/trmmkernel_2x2.c
-CTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
-ZTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
-
-SGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
-SGEMMONCOPY    =  ../generic/gemm_ncopy_2.c
-SGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c
-SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
-SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
-
-DGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
-DGEMMONCOPY    = ../generic/gemm_ncopy_2.c
-DGEMMOTCOPY    = ../generic/gemm_tcopy_2.c
-DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
-DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
-
-CGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
-CGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
-CGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
-CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
-CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
-
-ZGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
-ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
-ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
-ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
-ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
-
 endif
diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index 848de38df..f98728a41 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -33,9 +33,10 @@ ZAXPYKERNEL = zaxpy.c
 
 STRMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
+SGEMM_BETA     =  sgemm_beta_skylakex.c
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
 SGEMMITCOPY    =  ../generic/gemm_tcopy_16.c
-SGEMMONCOPY    =  ../generic/gemm_ncopy_4.c
+SGEMMONCOPY    =  sgemm_ncopy_4_skylakex.c
 SGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
 SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
 SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
@@ -44,9 +45,10 @@ SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
 DTRMMKERNEL    =  dtrmm_kernel_4x8_haswell.c
 DGEMMKERNEL    =  dgemm_kernel_4x8_haswell.S
+DGEMM_BETA     =  dgemm_beta_skylakex.c
 DGEMMINCOPY    =  ../generic/gemm_ncopy_4.c
 DGEMMITCOPY    =  ../generic/gemm_tcopy_4.c
-DGEMMONCOPY    =  ../generic/gemm_ncopy_8.c
+DGEMMONCOPY    =  dgemm_ncopy_8_skylakex.c
 DGEMMOTCOPY    =  ../generic/gemm_tcopy_8.c
 DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
 DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
diff --git a/kernel/x86_64/dgemm_beta_skylakex.c b/kernel/x86_64/dgemm_beta_skylakex.c
index 6a824c9b5..5cd001920 100644
--- a/kernel/x86_64/dgemm_beta_skylakex.c
+++ b/kernel/x86_64/dgemm_beta_skylakex.c
@@ -50,7 +50,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
   FLOAT ctemp5, ctemp6, ctemp7, ctemp8;
 
   /* fast path.. just zero the whole matrix */
-  if (m == ldc && (unsigned long)beta == (unsigned long)ZERO) {
+  if (m == ldc && beta == ZERO) {
 	memset(c, 0, m * n * sizeof(FLOAT));
 	return 0;
   }
@@ -61,17 +61,17 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
   c_offset = c;
 
   if (beta == ZERO){
-    __m512d z_zero;
 
-    z_zero = _mm512_setzero_pd();
     j = n;
     do {
       c_offset1 = c_offset;
       c_offset += ldc;
 
       i = m;
-
+#ifdef __AVX2__
+#ifdef __AVX512CD__
       while (i >= 32) {
+	  __m512d z_zero = _mm512_setzero_pd();
 	  _mm512_storeu_pd(c_offset1, z_zero);
 	  _mm512_storeu_pd(c_offset1 + 8, z_zero);
 	  _mm512_storeu_pd(c_offset1 + 16, z_zero);
@@ -79,12 +79,20 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
 	  c_offset1 += 32;
 	  i -= 32;
       }
+#endif
       while (i >= 8) {
+#ifdef __AVX512CD__
+	  __m512d z_zero = _mm512_setzero_pd();
 	  _mm512_storeu_pd(c_offset1, z_zero);
+#else
+	 __m256d y_zero = _mm256_setzero_pd();
+	 _mm256_storeu_pd(c_offset1, y_zero);
+	 _mm256_storeu_pd(c_offset1 + 4, y_zero);
+#endif
 	  c_offset1 += 8;
 	  i -= 8;
       }
-
+#endif
       while (i > 0) {
 	  *c_offset1 = ZERO;
 	  c_offset1 ++;
diff --git a/kernel/x86_64/dgemm_kernel_4x8_skylakex.c b/kernel/x86_64/dgemm_kernel_4x8_skylakex.c
index a83ca98fa..6257e569e 100644
--- a/kernel/x86_64/dgemm_kernel_4x8_skylakex.c
+++ b/kernel/x86_64/dgemm_kernel_4x8_skylakex.c
@@ -869,7 +869,7 @@ CNAME(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double * __restrict__ A,
 			"vmovapd %%zmm1, %%zmm27\n"
 			"vmovapd %%zmm1, %%zmm28\n"
 			"jmp .label24\n"
-			".align 32\n"
+			".p2align 5\n"
 			/* Inner math loop */
 			".label24:\n"
 			"vmovupd     -128(%[AO]),%%zmm0\n"
@@ -1037,7 +1037,7 @@ CNAME(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double * __restrict__ A,
 			"vmovapd %%zmm1, %%zmm17\n"
 			"vmovapd %%zmm1, %%zmm18\n"
 			"jmp .label16\n"
-			".align 32\n"
+			".p2align 5\n"
 			/* Inner math loop */
 			".label16:\n"
 			"vmovupd     -128(%[AO]),%%zmm0\n"
@@ -1165,7 +1165,7 @@ CNAME(BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double * __restrict__ A,
 			"vmovapd %%zmm1, %%zmm8\n"
 			"vbroadcastsd (%[alpha]), %%zmm9\n"
 			"jmp .label1\n"
-			".align 32\n"
+			".p2align 5\n"
 			/* Inner math loop */
 			".label1:\n"
 			"vmovupd     -128(%[AO]),%%zmm0\n"
diff --git a/kernel/x86_64/sgemm_beta_skylakex.c b/kernel/x86_64/sgemm_beta_skylakex.c
index 498c46f0d..1c29c1168 100644
--- a/kernel/x86_64/sgemm_beta_skylakex.c
+++ b/kernel/x86_64/sgemm_beta_skylakex.c
@@ -50,7 +50,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
   FLOAT ctemp5, ctemp6, ctemp7, ctemp8;
 
   /* fast path.. just zero the whole matrix */
-  if (m == ldc && (unsigned long)beta == (unsigned long)ZERO) {
+  if (m == ldc && beta == ZERO) {
 	memset(c, 0, m * n * sizeof(FLOAT));
 	return 0;
   }
@@ -61,30 +61,36 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
   c_offset = c;
 
   if (beta == ZERO){
-    __m512 z_zero;
-    __m256 y_zero;
 
-    z_zero = _mm512_setzero_ps();
-    y_zero = _mm256_setzero_ps();
     j = n;
     do {
       c_offset1 = c_offset;
       c_offset += ldc;
 
       i = m;
-
+#ifdef __AVX2__
       while (i >= 32) {
+#ifdef __AVX512CD__
+	  __m512 z_zero = _mm512_setzero_ps();
 	  _mm512_storeu_ps(c_offset1, z_zero);
 	  _mm512_storeu_ps(c_offset1 + 16, z_zero);
+#else
+	  __m256 y_zero = _mm256_setzero_ps();
+	  _mm256_storeu_ps(c_offset1, y_zero);
+	  _mm256_storeu_ps(c_offset1 + 8, y_zero);
+	  _mm256_storeu_ps(c_offset1 + 16, y_zero);
+	  _mm256_storeu_ps(c_offset1 + 24, y_zero);
+#endif
 	  c_offset1 += 32;
 	  i -= 32;
       }
       while (i >= 8) {
+	    __m256 y_zero = _mm256_setzero_ps();
 	  _mm256_storeu_ps(c_offset1, y_zero);
 	  c_offset1 += 8;
 	  i -= 8;
       }
-
+#endif
       while (i > 0) {
 	  *c_offset1 = ZERO;
 	  c_offset1 ++;
diff --git a/kernel/x86_64/sgemm_kernel_16x4_skylakex.c b/kernel/x86_64/sgemm_kernel_16x4_skylakex.c
index 10d3d22ed..3246e681f 100644
--- a/kernel/x86_64/sgemm_kernel_16x4_skylakex.c
+++ b/kernel/x86_64/sgemm_kernel_16x4_skylakex.c
@@ -760,7 +760,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *************************************************************************************/
 
 int __attribute__ ((noinline))
-CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, float * __restrict__ B, float * __restrict__ C, BLASLONG ldc)
+CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict A, float * __restrict B, float * __restrict C, BLASLONG ldc)
 {
 	unsigned long M = m, N = n, K = k;
 	if (M == 0)
@@ -1175,3 +1175,468 @@ CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, f
 
 	return 0;
 }
+
+
+/*
+ * "Direct sgemm" code. This code operates directly on the inputs and outputs
+ * of the sgemm call, avoiding the copies, memory realignments and threading,
+ * and only supports alpha = 1 and beta = 0.
+ * This is a common case and provides value for relatively small matrixes.
+ * For larger matrixes the "regular" sgemm code is superior, there the cost of
+ * copying/shuffling the B matrix really pays off.
+ */
+
+
+
+#define DECLARE_RESULT_512(N,M) __m512 result##N##M = _mm512_setzero_ps()
+#define BROADCAST_LOAD_A_512(N,M) __m512 Aval##M = _mm512_broadcastss_ps(_mm_load_ss(&A[k  + strideA * (i+M)]))
+#define LOAD_B_512(N,M)  __m512 Bval##N = _mm512_loadu_ps(&B[strideB * k + j + (N*16)])
+#define MATMUL_512(N,M)  result##N##M = _mm512_fmadd_ps(Aval##M, Bval##N , result##N##M)
+#define STORE_512(N,M) _mm512_storeu_ps(&R[(i+M) * strideR + j+(N*16)], result##N##M)
+
+
+#define DECLARE_RESULT_256(N,M) __m256 result##N##M = _mm256_setzero_ps()
+#define BROADCAST_LOAD_A_256(N,M) __m256 Aval##M = _mm256_broadcastss_ps(_mm_load_ss(&A[k  + strideA * (i+M)]))
+#define LOAD_B_256(N,M)  __m256 Bval##N = _mm256_loadu_ps(&B[strideB * k + j + (N*8)])
+#define MATMUL_256(N,M)  result##N##M = _mm256_fmadd_ps(Aval##M, Bval##N , result##N##M)
+#define STORE_256(N,M) _mm256_storeu_ps(&R[(i+M) * strideR + j+(N*8)], result##N##M)
+
+#define DECLARE_RESULT_128(N,M) __m128 result##N##M = _mm_setzero_ps()
+#define BROADCAST_LOAD_A_128(N,M) __m128 Aval##M = _mm_broadcastss_ps(_mm_load_ss(&A[k  + strideA * (i+M)]))
+#define LOAD_B_128(N,M)  __m128 Bval##N = _mm_loadu_ps(&B[strideB * k + j + (N*4)])
+#define MATMUL_128(N,M)  result##N##M = _mm_fmadd_ps(Aval##M, Bval##N , result##N##M)
+#define STORE_128(N,M) _mm_storeu_ps(&R[(i+M) * strideR + j+(N*4)], result##N##M)
+
+#define DECLARE_RESULT_SCALAR(N,M) float result##N##M = 0;
+#define BROADCAST_LOAD_A_SCALAR(N,M) float Aval##M = A[k + strideA * (i + M)];
+#define LOAD_B_SCALAR(N,M)  float Bval##N  = B[k * strideB + j + N];
+#define MATMUL_SCALAR(N,M) result##N##M +=  Aval##M * Bval##N;
+#define STORE_SCALAR(N,M)  R[(i+M) * strideR + j + N] = result##N##M;
+
+int sgemm_kernel_direct_performant(BLASLONG M, BLASLONG N, BLASLONG K)
+{
+	int mnk = M * N * K;
+	/* large matrixes -> not performant */
+	if (mnk >= 28 * 512 * 512)
+		return 0;
+
+	/*
+	 * if the B matrix is not a nice multiple if 4 we get many unaligned accesses,
+	 * and the regular sgemm copy/realignment of data pays off much quicker
+	 */
+	if ((N & 3) != 0 && (mnk >= 8 * 512 * 512))
+		return 0;
+
+#ifdef SMP
+	/* if we can run multithreaded, the threading changes the based threshold */
+	if (mnk > 2 * 350 * 512 && num_cpu_avail(3)> 1)
+		return 0;
+#endif
+
+	return 1;
+}
+
+
+
+void sgemm_kernel_direct (BLASLONG M, BLASLONG N, BLASLONG K, float * __restrict A, BLASLONG strideA, float * __restrict B, BLASLONG strideB , float * __restrict R, BLASLONG strideR)
+{
+	int i, j, k;
+
+        int m4 = M & ~3;
+	int m2 = M & ~1;
+
+	int n64 = N & ~63;
+	int n32 = N & ~31;
+	int n16 = N & ~15;
+	int n8 = N & ~7;
+	int n4 = N & ~3;
+	int n2 = N & ~1;
+
+	i = 0;
+
+	for (i = 0; i < m4; i+=4) {
+
+		for (j = 0; j < n64; j+= 64) {
+			k = 0;
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);    			DECLARE_RESULT_512(2, 0);    DECLARE_RESULT_512(3, 0);
+			DECLARE_RESULT_512(0, 1);    DECLARE_RESULT_512(1, 1);    			DECLARE_RESULT_512(2, 1);    DECLARE_RESULT_512(3, 1);
+			DECLARE_RESULT_512(0, 2);    DECLARE_RESULT_512(1, 2);    			DECLARE_RESULT_512(2, 2);    DECLARE_RESULT_512(3, 2);
+			DECLARE_RESULT_512(0, 3);    DECLARE_RESULT_512(1, 3);    			DECLARE_RESULT_512(2, 3);    DECLARE_RESULT_512(3, 3);
+
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+				BROADCAST_LOAD_A_512(x, 2);
+				BROADCAST_LOAD_A_512(x, 3);
+
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);			LOAD_B_512(2, x);		LOAD_B_512(3, x);
+
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);			MATMUL_512(2, 0);		MATMUL_512(3, 0);
+				MATMUL_512(0, 1);		MATMUL_512(1, 1);			MATMUL_512(2, 1);		MATMUL_512(3, 1);
+				MATMUL_512(0, 2);		MATMUL_512(1, 2);			MATMUL_512(2, 2);		MATMUL_512(3, 2);
+				MATMUL_512(0, 3);		MATMUL_512(1, 3);			MATMUL_512(2, 3);		MATMUL_512(3, 3);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);			STORE_512(2, 0);		STORE_512(3, 0);
+			STORE_512(0, 1);		STORE_512(1, 1);			STORE_512(2, 1);		STORE_512(3, 1);
+			STORE_512(0, 2);		STORE_512(1, 2);			STORE_512(2, 2);		STORE_512(3, 2);
+			STORE_512(0, 3);		STORE_512(1, 3);			STORE_512(2, 3);		STORE_512(3, 3);
+		}
+
+		for (; j < n32; j+= 32) {
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);
+			DECLARE_RESULT_512(0, 1);    DECLARE_RESULT_512(1, 1);
+			DECLARE_RESULT_512(0, 2);    DECLARE_RESULT_512(1, 2);
+			DECLARE_RESULT_512(0, 3);    DECLARE_RESULT_512(1, 3);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+				BROADCAST_LOAD_A_512(x, 2);
+				BROADCAST_LOAD_A_512(x, 3);
+
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);
+
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);
+				MATMUL_512(0, 1);		MATMUL_512(1, 1);
+				MATMUL_512(0, 2);		MATMUL_512(1, 2);
+				MATMUL_512(0, 3);		MATMUL_512(1, 3);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);
+			STORE_512(0, 1);		STORE_512(1, 1);
+			STORE_512(0, 2);		STORE_512(1, 2);
+			STORE_512(0, 3);		STORE_512(1, 3);
+		}
+
+		for (; j < n16; j+= 16) {
+			DECLARE_RESULT_512(0, 0);
+			DECLARE_RESULT_512(0, 1);
+			DECLARE_RESULT_512(0, 2);
+			DECLARE_RESULT_512(0, 3);
+
+		 	for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+				BROADCAST_LOAD_A_512(x, 2);
+				BROADCAST_LOAD_A_512(x, 3);
+
+				LOAD_B_512(0, x);
+
+				MATMUL_512(0, 0);
+				MATMUL_512(0, 1);
+				MATMUL_512(0, 2);
+				MATMUL_512(0, 3);
+			}
+			STORE_512(0, 0);
+			STORE_512(0, 1);
+			STORE_512(0, 2);
+			STORE_512(0, 3);
+		}
+
+		for (; j < n8; j+= 8) {
+			DECLARE_RESULT_256(0, 0);
+			DECLARE_RESULT_256(0, 1);
+			DECLARE_RESULT_256(0, 2);
+			DECLARE_RESULT_256(0, 3);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_256(x, 0);
+				BROADCAST_LOAD_A_256(x, 1);
+				BROADCAST_LOAD_A_256(x, 2);
+				BROADCAST_LOAD_A_256(x, 3);
+
+				LOAD_B_256(0, x);
+
+				MATMUL_256(0, 0);
+				MATMUL_256(0, 1);
+				MATMUL_256(0, 2);
+				MATMUL_256(0, 3);
+			}
+			STORE_256(0, 0);
+			STORE_256(0, 1);
+			STORE_256(0, 2);
+			STORE_256(0, 3);
+		}
+
+		for (; j < n4; j+= 4) {
+			DECLARE_RESULT_128(0, 0);
+			DECLARE_RESULT_128(0, 1);
+			DECLARE_RESULT_128(0, 2);
+			DECLARE_RESULT_128(0, 3);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_128(x, 0);
+				BROADCAST_LOAD_A_128(x, 1);
+				BROADCAST_LOAD_A_128(x, 2);
+				BROADCAST_LOAD_A_128(x, 3);
+
+				LOAD_B_128(0, x);
+
+				MATMUL_128(0, 0);
+				MATMUL_128(0, 1);
+				MATMUL_128(0, 2);
+				MATMUL_128(0, 3);
+			}
+			STORE_128(0, 0);
+			STORE_128(0, 1);
+			STORE_128(0, 2);
+			STORE_128(0, 3);
+		}
+
+		for (; j < n2; j+= 2) {
+			DECLARE_RESULT_SCALAR(0, 0);	DECLARE_RESULT_SCALAR(1, 0);
+			DECLARE_RESULT_SCALAR(0, 1);	DECLARE_RESULT_SCALAR(1, 1);
+			DECLARE_RESULT_SCALAR(0, 2);	DECLARE_RESULT_SCALAR(1, 2);
+			DECLARE_RESULT_SCALAR(0, 3);	DECLARE_RESULT_SCALAR(1, 3);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(x, 0);
+				BROADCAST_LOAD_A_SCALAR(x, 1);
+				BROADCAST_LOAD_A_SCALAR(x, 2);
+				BROADCAST_LOAD_A_SCALAR(x, 3);
+
+				LOAD_B_SCALAR(0, x);	LOAD_B_SCALAR(1, x);
+
+				MATMUL_SCALAR(0, 0);	MATMUL_SCALAR(1, 0);
+				MATMUL_SCALAR(0, 1);	MATMUL_SCALAR(1, 1);
+				MATMUL_SCALAR(0, 2);	MATMUL_SCALAR(1, 2);
+				MATMUL_SCALAR(0, 3);	MATMUL_SCALAR(1, 3);
+			}
+			STORE_SCALAR(0, 0);	STORE_SCALAR(1, 0);
+			STORE_SCALAR(0, 1);	STORE_SCALAR(1, 1);
+			STORE_SCALAR(0, 2);	STORE_SCALAR(1, 2);
+			STORE_SCALAR(0, 3);	STORE_SCALAR(1, 3);
+		}
+
+		for (; j < N; j++) {
+			DECLARE_RESULT_SCALAR(0, 0)
+			DECLARE_RESULT_SCALAR(0, 1)
+			DECLARE_RESULT_SCALAR(0, 2)
+			DECLARE_RESULT_SCALAR(0, 3)
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(0, 0);
+				BROADCAST_LOAD_A_SCALAR(0, 1);
+				BROADCAST_LOAD_A_SCALAR(0, 2);
+				BROADCAST_LOAD_A_SCALAR(0, 3);
+
+				LOAD_B_SCALAR(0, 0);
+
+				MATMUL_SCALAR(0, 0);
+				MATMUL_SCALAR(0, 1);
+				MATMUL_SCALAR(0, 2);
+				MATMUL_SCALAR(0, 3);
+			}
+			STORE_SCALAR(0, 0);
+			STORE_SCALAR(0, 1);
+			STORE_SCALAR(0, 2);
+			STORE_SCALAR(0, 3);
+		}
+	}
+
+	for (; i < m2; i+=2) {
+		j = 0;
+
+		for (; j < n64; j+= 64) {
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);    			DECLARE_RESULT_512(2, 0);    DECLARE_RESULT_512(3, 0);
+			DECLARE_RESULT_512(0, 1);    DECLARE_RESULT_512(1, 1);    			DECLARE_RESULT_512(2, 1);    DECLARE_RESULT_512(3, 1);
+
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);			LOAD_B_512(2, x);		LOAD_B_512(3, x);
+
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);			MATMUL_512(2, 0);		MATMUL_512(3, 0);
+				MATMUL_512(0, 1);		MATMUL_512(1, 1);			MATMUL_512(2, 1);		MATMUL_512(3, 1);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);			STORE_512(2, 0);		STORE_512(3, 0);
+			STORE_512(0, 1);		STORE_512(1, 1);			STORE_512(2, 1);		STORE_512(3, 1);
+		}
+
+		for (; j < n32; j+= 32) {
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);
+			DECLARE_RESULT_512(0, 1);    DECLARE_RESULT_512(1, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);
+
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);
+				MATMUL_512(0, 1);		MATMUL_512(1, 1);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);
+			STORE_512(0, 1);		STORE_512(1, 1);
+		}
+
+
+		for (; j < n16; j+= 16) {
+			DECLARE_RESULT_512(0, 0);
+			DECLARE_RESULT_512(0, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+
+				LOAD_B_512(0, x);
+
+				MATMUL_512(0, 0);
+				MATMUL_512(0, 1);
+			}
+			STORE_512(0, 0);
+			STORE_512(0, 1);
+		}
+
+		for (; j < n8; j+= 8) {
+			DECLARE_RESULT_256(0, 0);
+			DECLARE_RESULT_256(0, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_256(x, 0);
+				BROADCAST_LOAD_A_256(x, 1);
+
+				LOAD_B_256(0, x);
+
+				MATMUL_256(0, 0);
+				MATMUL_256(0, 1);
+			}
+			STORE_256(0, 0);
+			STORE_256(0, 1);
+		}
+
+		for (; j < n4; j+= 4) {
+			DECLARE_RESULT_128(0, 0);
+			DECLARE_RESULT_128(0, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_128(x, 0);
+				BROADCAST_LOAD_A_128(x, 1);
+
+				LOAD_B_128(0, x);
+
+				MATMUL_128(0, 0);
+				MATMUL_128(0, 1);
+			}
+			STORE_128(0, 0);
+			STORE_128(0, 1);
+		}
+		for (; j < n2; j+= 2) {
+			DECLARE_RESULT_SCALAR(0, 0);	DECLARE_RESULT_SCALAR(1, 0);
+			DECLARE_RESULT_SCALAR(0, 1);	DECLARE_RESULT_SCALAR(1, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(x, 0);
+				BROADCAST_LOAD_A_SCALAR(x, 1);
+
+				LOAD_B_SCALAR(0, x);	LOAD_B_SCALAR(1, x);
+
+				MATMUL_SCALAR(0, 0);	MATMUL_SCALAR(1, 0);
+				MATMUL_SCALAR(0, 1);	MATMUL_SCALAR(1, 1);
+			}
+			STORE_SCALAR(0, 0);	STORE_SCALAR(1, 0);
+			STORE_SCALAR(0, 1);	STORE_SCALAR(1, 1);
+		}
+
+		for (; j < N; j++) {
+			DECLARE_RESULT_SCALAR(0, 0);
+			DECLARE_RESULT_SCALAR(0, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(0, 0);
+				BROADCAST_LOAD_A_SCALAR(0, 1);
+
+				LOAD_B_SCALAR(0, 0);
+
+				MATMUL_SCALAR(0, 0);
+				MATMUL_SCALAR(0, 1);
+			}
+			STORE_SCALAR(0, 0);
+			STORE_SCALAR(0, 1);
+		}
+	}
+
+	for (; i < M; i+=1) {
+		j = 0;
+		for (; j < n64; j+= 64) {
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);    			DECLARE_RESULT_512(2, 0);    DECLARE_RESULT_512(3, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);			LOAD_B_512(2, x);		LOAD_B_512(3, x);
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);			MATMUL_512(2, 0);		MATMUL_512(3, 0);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);			STORE_512(2, 0);		STORE_512(3, 0);
+		}
+		for (; j < n32; j+= 32) {
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);
+		}
+
+
+		for (; j < n16; j+= 16) {
+			DECLARE_RESULT_512(0, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+
+				LOAD_B_512(0, x);
+
+				MATMUL_512(0, 0);
+			}
+			STORE_512(0, 0);
+		}
+
+		for (; j < n8; j+= 8) {
+			DECLARE_RESULT_256(0, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_256(x, 0);
+				LOAD_B_256(0, x);
+				MATMUL_256(0, 0);
+			}
+			STORE_256(0, 0);
+		}
+
+		for (; j < n4; j+= 4) {
+			DECLARE_RESULT_128(0, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_128(x, 0);
+				LOAD_B_128(0, x);
+				MATMUL_128(0, 0);
+			}
+			STORE_128(0, 0);
+		}
+
+		for (; j < n2; j+= 2) {
+			DECLARE_RESULT_SCALAR(0, 0);	DECLARE_RESULT_SCALAR(1, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(x, 0);
+				LOAD_B_SCALAR(0, 0);	LOAD_B_SCALAR(1, 0);
+				MATMUL_SCALAR(0, 0);	MATMUL_SCALAR(1, 0);
+			}
+			STORE_SCALAR(0, 0);	STORE_SCALAR(1, 0);
+		}
+
+		for (; j < N; j++) {
+			DECLARE_RESULT_SCALAR(0, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(0, 0);
+				LOAD_B_SCALAR(0, 0);
+				MATMUL_SCALAR(0, 0);
+			}
+			STORE_SCALAR(0, 0);
+		}
+	}
+}
\ No newline at end of file
diff --git a/kernel/x86_64/sgemm_ncopy_4_skylakex.c b/kernel/x86_64/sgemm_ncopy_4_skylakex.c
index 8577e3b38..6b2b0f5b1 100644
--- a/kernel/x86_64/sgemm_ncopy_4_skylakex.c
+++ b/kernel/x86_64/sgemm_ncopy_4_skylakex.c
@@ -49,8 +49,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT * __restrict a, BLASLONG lda, FLOAT * __
   FLOAT *b_offset;
   FLOAT  ctemp1,  ctemp2,  ctemp3,  ctemp4;
   FLOAT  ctemp5,  ctemp6,  ctemp7,  ctemp8;
-  FLOAT  ctemp9, ctemp10, ctemp11, ctemp12;
-  FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
+  FLOAT  ctemp9,  ctemp13;
 
   a_offset = a;
   b_offset = b;
diff --git a/param.h b/param.h
index 8f56cdaaa..fa6730208 100644
--- a/param.h
+++ b/param.h
@@ -1508,6 +1508,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SYMV_P  8
 
 #define SWITCH_RATIO	32
+#define GEMM_PREFERED_SIZE	16
 
 #ifdef ARCH_X86
 
@@ -1628,6 +1629,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define SWITCH_RATIO	32
 #define GEMM_PREFERED_SIZE	32
+#define USE_SGEMM_KERNEL_DIRECT 1
 
 #ifdef ARCH_X86
 
diff --git a/utest/test_dotu.c b/utest/test_dotu.c
index ef04dd9a8..918541848 100644
--- a/utest/test_dotu.c
+++ b/utest/test_dotu.c
@@ -32,7 +32,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 **********************************************************************************/
 
 #include "openblas_utest.h"
-#include <complex.h>
 
 CTEST( zdotu,zdotu_n_1)
 {