Fix #686. Merge branch 'ashwinyes-develop' into develop

.gitignore

@@ -68,3 +68,4 @@ test/zblat2
 test/zblat3
 build
 build.*
+*.swp

Makefile.arm64

@@ -4,4 +4,8 @@ CCOMMON_OPT += -march=armv8-a
 FCOMMON_OPT += -march=armv8-a
 endif
 
+ifeq ($(CORE), CORTEXA57)
+CCOMMON_OPT += -march=armv8-a+crc+crypto+fp+simd -mtune=cortex-a57
+FCOMMON_OPT += -march=armv8-a+crc+crypto+fp+simd -mtune=cortex-a57
+endif
 

TargetList.txt

@@ -74,3 +74,5 @@ ARMV5
 
 7.ARM 64-bit CPU:
 ARMV8
+CORTEXA57
+

benchmark/gemm.c

@@ -172,7 +172,7 @@ int main(int argc, char *argv[]){
   srandom(getpid());
 #endif
  
-	for(j = 0; j < m; j++){
+	for(j = 0; j < to; j++){
       		for(i = 0; i < to * COMPSIZE; i++){
 			a[i + j * to * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
 			b[i + j * to * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;

common.h

@@ -86,6 +86,7 @@ extern "C" {
 #if !defined(_MSC_VER)
 #include <unistd.h>
 #endif
+#include <time.h>
 
 #ifdef OS_LINUX
 #include <malloc.h>

common_arm64.h

@@ -89,8 +89,10 @@ static inline int blas_quickdivide(blasint x, blasint y){
 #if defined(ASSEMBLER) && !defined(NEEDPARAM)
 
 #define PROLOGUE \
+	.text ;\
+	.align	4 ;\
 	.global	REALNAME ;\
-	.func	REALNAME  ;\
+	.type	REALNAME, %function ;\
 REALNAME:
 
 #define EPILOGUE
@@ -107,7 +109,11 @@ REALNAME:
 #endif
 #define HUGE_PAGESIZE   ( 4 << 20)
 
+#if defined(CORTEXA57)
+#define BUFFER_SIZE     (40 << 20)
+#else
 #define BUFFER_SIZE     (16 << 20)
+#endif
 
 
 #define BASE_ADDRESS (START_ADDRESS - BUFFER_SIZE * MAX_CPU_NUMBER)

cpuid_arm64.c

@@ -29,12 +29,19 @@
 
 #define CPU_UNKNOWN     	0
 #define CPU_ARMV8       	1
+#define CPU_CORTEXA57       	2
 
 static char *cpuname[] = {
-  "UNKOWN",
-  "ARMV8"
+  "UNKNOWN",
+  "ARMV8" ,
+  "CORTEXA57"
 };
 
+static char *cpuname_lower[] = {
+  "unknown",
+  "armv8" ,
+  "cortexa57"
+};
 
 int get_feature(char *search)
 {
@@ -53,13 +60,13 @@ int get_feature(char *search)
 		{
 			p = strchr(buffer, ':') + 2;
 			break;
-      		}
-  	}
+		}
+	}
 
-  	fclose(infile);
+	fclose(infile);
 
 
-	if( p == NULL ) return;
+	if( p == NULL ) return 0;
 
 	t = strtok(p," ");
 	while( t = strtok(NULL," "))
@@ -82,11 +89,30 @@ int detect(void)
   	p = (char *) NULL ;
 
   	infile = fopen("/proc/cpuinfo", "r");
+	while (fgets(buffer, sizeof(buffer), infile))
+	{
 
+		if (!strncmp("CPU part", buffer, 8))
+		{
+			p = strchr(buffer, ':') + 2;
+			break;
+		}
+	}
+
+	fclose(infile);
+	if(p != NULL) {
+	  if (strstr(p, "0xd07")) {
+	    return CPU_CORTEXA57;
+	  }
+	}
+
+	p = (char *) NULL ;
+	infile = fopen("/proc/cpuinfo", "r");
 	while (fgets(buffer, sizeof(buffer), infile))
 	{
 
-		if ((!strncmp("model name", buffer, 10)) || (!strncmp("Processor", buffer, 9)))
+		if ((!strncmp("model name", buffer, 10)) || (!strncmp("Processor", buffer, 9)) ||
+		    (!strncmp("CPU architecture", buffer, 16)))
 		{
 			p = strchr(buffer, ':') + 2;
 			break;
@@ -100,7 +126,7 @@ int detect(void)
 
 		if (strstr(p, "AArch64"))
 		{
-			 	return CPU_ARMV8;
+			return CPU_ARMV8;
 
 		}
 
@@ -118,23 +144,13 @@ char *get_corename(void)
 
 void get_architecture(void)
 {
-	printf("ARM");
+	printf("ARM64");
 }
 
 void get_subarchitecture(void)
 {
 	int d = detect();
-	switch (d)
-	{
-
-		case CPU_ARMV8:
-			printf("ARMV8");
-			break;
-
-		default:
-			printf("UNKNOWN");
-			break;
-	}
+	printf("%s", cpuname[d]);
 }
 
 void get_subdirname(void)
@@ -160,26 +176,32 @@ void get_cpuconfig(void)
     			printf("#define L2_ASSOCIATIVE 4\n");
 			break;
 
-
+		case CPU_CORTEXA57:
+			printf("#define CORTEXA57\n");
+			printf("#define HAVE_VFP\n");
+			printf("#define HAVE_VFPV3\n");
+			printf("#define HAVE_NEON\n");
+			printf("#define HAVE_VFPV4\n");
+			printf("#define L1_CODE_SIZE 49152\n");
+			printf("#define L1_CODE_LINESIZE 64\n");
+			printf("#define L1_CODE_ASSOCIATIVE 3\n");
+			printf("#define L1_DATA_SIZE 32768\n");
+			printf("#define L1_DATA_LINESIZE 64\n");
+			printf("#define L1_DATA_ASSOCIATIVE 2\n");
+			printf("#define L2_SIZE 2097152\n");
+			printf("#define L2_LINESIZE 64\n");
+			printf("#define L2_ASSOCIATIVE 16\n");
+			break;
 	}
 }
 
 
 void get_libname(void)
 {
-
 	int d = detect();
-	switch (d)
-	{
-
-		case CPU_ARMV8:
-    			printf("armv8\n");
-			break;
-
-	}
+	printf("%s", cpuname_lower[d]);
 }
 
-
 void get_features(void)
 {
 

driver/level2/spmv_thread.c

@@ -55,7 +55,7 @@
 static int spmv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *dummy1, FLOAT *buffer, BLASLONG pos){
 
   FLOAT *a, *x, *y;
-  BLASLONG incx, incy;
+  BLASLONG incx;
   BLASLONG m_from, m_to, i;
 #ifndef COMPLEX
   FLOAT result;
@@ -68,7 +68,6 @@ static int spmv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, F
   y = (FLOAT *)args -> c;
 
   incx = args -> ldb;
-  incy = args -> ldc;
 
   m_from = 0;
   m_to   = args -> m;

driver/level2/spr2_thread.c

@@ -43,7 +43,7 @@
 static int syr_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *dummy1, FLOAT *buffer, BLASLONG pos){
 
   FLOAT *a, *x, *y;
-  BLASLONG lda, incx, incy;
+  BLASLONG incx, incy;
   BLASLONG i, m_from, m_to;
   FLOAT alpha_r;
 #ifdef COMPLEX
@@ -56,7 +56,6 @@ static int syr_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FL
 
   incx = args -> lda;
   incy = args -> ldb;
-  lda  = args -> ldc;
 
   alpha_r = *((FLOAT *)args -> alpha + 0);
 #ifdef COMPLEX

driver/level2/spr_thread.c

@@ -46,7 +46,7 @@ static int syr_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FL
   BLASLONG incx;
   BLASLONG i, m_from, m_to;
   FLOAT alpha_r;
-#if defined(COMPLEX) && !defined(HER) && !defined(HERREV)
+#if defined(COMPLEX) && !defined(HEMV) && !defined(HEMVREV)
   FLOAT alpha_i;
 #endif
 
@@ -56,7 +56,7 @@ static int syr_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FL
   incx = args -> lda;
 
   alpha_r = *((FLOAT *)args -> alpha + 0);
-#if defined(COMPLEX) && !defined(HER) && !defined(HERREV)
+#if defined(COMPLEX) && !defined(HEMV) && !defined(HEMVREV)
   alpha_i = *((FLOAT *)args -> alpha + 1);
 #endif
 

driver/level2/symv_thread.c

@@ -55,7 +55,7 @@
 static int symv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *dummy1, FLOAT *buffer, BLASLONG pos){
 
   FLOAT *a, *x, *y;
-  BLASLONG lda, incx, incy;
+  BLASLONG lda, incx;
   BLASLONG m_from, m_to;
 
   a = (FLOAT *)args -> a;
@@ -64,7 +64,6 @@ static int symv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, F
 
   lda  = args -> lda;
   incx = args -> ldb;
-  incy = args -> ldc;
 
   m_from = 0;
   m_to   = args -> m;

driver/level2/tbmv_L.c

@@ -45,13 +45,11 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/tbmv_U.c

@@ -45,13 +45,11 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/tbsv_L.c

@@ -45,13 +45,11 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/tbsv_U.c

@@ -45,13 +45,11 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/tpsv_L.c

@@ -43,12 +43,10 @@
 int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level2/tpsv_U.c

@@ -43,12 +43,10 @@
 int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level2/ztbmv_L.c

@@ -45,7 +45,6 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 #if (TRANSA == 2) || (TRANSA == 4)
@@ -57,7 +56,6 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG inc
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) * COMPSIZE+ 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/ztbmv_U.c

@@ -45,7 +45,6 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 #if (TRANSA == 2) || (TRANSA == 4)
@@ -57,7 +56,6 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG inc
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) * COMPSIZE + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/ztbsv_L.c

@@ -45,7 +45,6 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 #if (TRANSA == 2) || (TRANSA == 4)
@@ -57,7 +56,6 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG inc
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) * COMPSIZE + 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/ztbsv_U.c

@@ -45,7 +45,6 @@ const static FLOAT dp1 = 1.;
 int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, void *buffer){
 
   BLASLONG i;
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
   BLASLONG length;
 #if (TRANSA == 2) || (TRANSA == 4)
@@ -57,7 +56,6 @@ int CNAME(BLASLONG n, BLASLONG k, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG inc
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + n * sizeof(FLOAT) * COMPSIZE+ 4095) & ~4095);
     COPY_K(n, b, incb, buffer, 1);
   }
 

driver/level2/ztpmv_L.c

@@ -49,12 +49,10 @@ int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 #ifndef UNIT
   FLOAT atemp1, atemp2, btemp1, btemp2;
 #endif
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) * 2 + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level2/ztpmv_U.c

@@ -49,12 +49,10 @@ int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 #ifndef UNIT
   FLOAT atemp1, atemp2, btemp1, btemp2;
 #endif
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) * 2 + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level2/ztpsv_L.c

@@ -51,12 +51,10 @@ int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 #ifndef UNIT
   FLOAT ar, ai, br, bi, ratio, den;
 #endif
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) * 2 + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level2/ztpsv_U.c

@@ -49,12 +49,10 @@ int CNAME(BLASLONG m, FLOAT *a, FLOAT *b, BLASLONG incb, void *buffer){
 #ifndef UNIT
   FLOAT ar, ai, br, bi, ratio, den;
 #endif
-  FLOAT *gemvbuffer = (FLOAT *)buffer;
   FLOAT *B = b;
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) * 2 + 4095) & ~4095);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level3/gemm_thread_mn.c

@@ -65,7 +65,7 @@ int CNAME(int mode, blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n, int (
   blas_queue_t queue[MAX_CPU_NUMBER];
 
   BLASLONG range_M[MAX_CPU_NUMBER + 1], range_N[MAX_CPU_NUMBER + 1];
-  BLASLONG procs, total_procs, num_cpu_m, num_cpu_n;
+  BLASLONG procs, num_cpu_m, num_cpu_n;
 
   BLASLONG width, i, j;
   BLASLONG divM, divN;

driver/level3/level3_thread.c

@@ -230,7 +230,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
   BLASLONG is, min_i, div_n;
 
   BLASLONG i, current;
-  BLASLONG l1stride, l2size;
+  BLASLONG l1stride;
 
 #ifdef TIMING
   BLASULONG rpcc_counter;
@@ -298,8 +298,6 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 #endif
       ) return 0;
 
-  l2size = GEMM_P * GEMM_Q;
-
 #if 0
   fprintf(stderr, "Thread[%ld]  m_from : %ld m_to : %ld n_from : %ld n_to : %ld N_from : %ld N_to : %ld\n",
 	  mypos, m_from, m_to, n_from, n_to, N_from, N_to);
@@ -706,7 +704,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
     n = n_to - n_from;
   }
 
-  if ((args -> m < nthreads * SWITCH_RATIO) || (args -> n < nthreads * SWITCH_RATIO)) {
+  if ((m < nthreads * SWITCH_RATIO) || (n < nthreads * SWITCH_RATIO)) {
     GEMM_LOCAL(args, range_m, range_n, sa, sb, 0);
     return 0;
   }

driver/others/memory.c

@@ -914,7 +914,6 @@ static volatile struct {
 } memory[NUM_BUFFERS];
 
 static int memory_initialized = 0;
-static void gotoblas_memory_init(void);
 
 /*       Memory allocation routine           */
 /* procpos ... indicates where it comes from */

getarch.c

@@ -819,10 +819,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
        "-DL2_SIZE=262144 -DL2_LINESIZE=64 " \
        "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=32 " 
 #define LIBNAME   "armv8"
-#define CORENAME  "XGENE1"
-#else
+#define CORENAME  "ARMV8"
 #endif
 
+#ifdef FORCE_CORTEXA57
+#define FORCE
+#define ARCHITECTURE    "ARM64"
+#define SUBARCHITECTURE "ARMV8"
+#define SUBDIRNAME      "arm64"
+#define ARCHCONFIG   "-DCORTEXA57 " \
+       "-DL1_CODE_SIZE=49152 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=3 " \
+       "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 -DL1_DATA_ASSOCIATIVE=2 " \
+       "-DL2_SIZE=2097152 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=16 " \
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
+       "-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON"
+#define LIBNAME   "cortexa57"
+#define CORENAME  "CORTEXA57"
+#else
+#endif
 
 #ifndef FORCE
 

interface/symm.c

@@ -91,6 +91,27 @@
 #endif
 #endif
 
+
+#ifdef SMP
+#ifndef COMPLEX
+#ifdef XDOUBLE
+#define MODE	(BLAS_XDOUBLE | BLAS_REAL)
+#elif defined(DOUBLE)
+#define MODE	(BLAS_DOUBLE  | BLAS_REAL)
+#else
+#define MODE	(BLAS_SINGLE  | BLAS_REAL)
+#endif
+#else
+#ifdef XDOUBLE
+#define MODE	(BLAS_XDOUBLE | BLAS_COMPLEX)
+#elif defined(DOUBLE)
+#define MODE	(BLAS_DOUBLE  | BLAS_COMPLEX)
+#else
+#define MODE	(BLAS_SINGLE  | BLAS_COMPLEX)
+#endif
+#endif
+#endif
+
 static int (*symm[])(blas_arg_t *, BLASLONG *, BLASLONG *, FLOAT *, FLOAT *, BLASLONG) = {
 #ifndef GEMM3M
 #ifndef HEMM
@@ -135,26 +156,6 @@ void NAME(char *SIDE, char *UPLO,
   FLOAT *buffer;
   FLOAT *sa, *sb;
 
-#ifdef SMP
-#ifndef COMPLEX
-#ifdef XDOUBLE
-  int mode  =  BLAS_XDOUBLE | BLAS_REAL;
-#elif defined(DOUBLE)
-  int mode  =  BLAS_DOUBLE  | BLAS_REAL;
-#else
-  int mode  =  BLAS_SINGLE  | BLAS_REAL;
-#endif
-#else
-#ifdef XDOUBLE
-  int mode  =  BLAS_XDOUBLE | BLAS_COMPLEX;
-#elif defined(DOUBLE)
-  int mode  =  BLAS_DOUBLE  | BLAS_COMPLEX;
-#else
-  int mode  =  BLAS_SINGLE  | BLAS_COMPLEX;
-#endif
-#endif
-#endif
-
 #if defined(SMP) && !defined(NO_AFFINITY)
   int nodes;
 #endif
@@ -246,26 +247,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo,
   FLOAT *buffer;
   FLOAT *sa, *sb;
 
-#ifdef SMP
-#ifndef COMPLEX
-#ifdef XDOUBLE
-  int mode  =  BLAS_XDOUBLE | BLAS_REAL;
-#elif defined(DOUBLE)
-  int mode  =  BLAS_DOUBLE  | BLAS_REAL;
-#else
-  int mode  =  BLAS_SINGLE  | BLAS_REAL;
-#endif
-#else
-#ifdef XDOUBLE
-  int mode  =  BLAS_XDOUBLE | BLAS_COMPLEX;
-#elif defined(DOUBLE)
-  int mode  =  BLAS_DOUBLE  | BLAS_COMPLEX;
-#else
-  int mode  =  BLAS_SINGLE  | BLAS_COMPLEX;
-#endif
-#endif
-#endif
-
 #if defined(SMP) && !defined(NO_AFFINITY)
   int nodes;
 #endif
@@ -407,7 +388,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo,
 
       args.nthreads /= nodes;
 
-      gemm_thread_mn(mode, &args, NULL, NULL,
+      gemm_thread_mn(MODE, &args, NULL, NULL,
 		     symm[4 | (side << 1) | uplo ], sa, sb, nodes);
 
     } else {
@@ -419,7 +400,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo,
 
 #else
 
-      GEMM_THREAD(mode, &args, NULL, NULL, symm[(side << 1) | uplo ], sa, sb, args.nthreads);
+      GEMM_THREAD(MODE, &args, NULL, NULL, symm[(side << 1) | uplo ], sa, sb, args.nthreads);
 
 #endif
 

interface/syr.c

@@ -116,7 +116,7 @@ void NAME(char *UPLO, blasint *N, FLOAT  *ALPHA,
 void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT alpha, FLOAT *x, blasint incx, FLOAT *a, blasint lda) {
 
   FLOAT *buffer;
-  int trans, uplo;
+  int uplo;
   blasint info;
 #ifdef SMP
   int nthreads;
@@ -124,7 +124,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT alpha,
 
   PRINT_DEBUG_CNAME;
 
-  trans = -1;
   uplo  = -1;
   info  =  0;
 

interface/syr2.c

@@ -118,7 +118,7 @@ void NAME(char *UPLO, blasint *N, FLOAT  *ALPHA,
 void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT alpha, FLOAT *x, blasint incx, FLOAT *y, blasint incy, FLOAT *a, blasint lda) {
 
   FLOAT *buffer;
-  int trans, uplo;
+  int uplo;
   blasint info;
 #ifdef SMP
   int nthreads;
@@ -126,7 +126,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT alpha,
 
   PRINT_DEBUG_CNAME;
 
-  trans = -1;
   uplo  = -1;
   info  =  0;
 

interface/zhemv.c

@@ -117,7 +117,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT *ALPHA
   FLOAT beta_i	= BETA[1];
 
   FLOAT *buffer;
-  int trans, uplo;
+  int uplo;
   blasint info;
 #ifdef SMP
   int nthreads;
@@ -135,7 +135,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT *ALPHA
 
   PRINT_DEBUG_CNAME;
 
-  trans = -1;
   uplo  = -1;
   info  =  0;
 

interface/zher.c

@@ -116,7 +116,7 @@ void NAME(char *UPLO, blasint *N, FLOAT  *ALPHA,
 void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT alpha, FLOAT *x, blasint incx, FLOAT *a, blasint lda) {
 
   FLOAT *buffer;
-  int trans, uplo;
+  int uplo;
   blasint info;
 #ifdef SMP
   int nthreads;
@@ -124,7 +124,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT alpha,
 
   PRINT_DEBUG_CNAME;
 
-  trans = -1;
   uplo  = -1;
   info  =  0;
 

interface/zher2.c

@@ -121,7 +121,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT *ALPHA
   FLOAT alpha_r	= ALPHA[0];
   FLOAT alpha_i	= ALPHA[1];
   FLOAT *buffer;
-  int trans, uplo;
+  int uplo;
   blasint info;
 #ifdef SMP
   int nthreads;
@@ -129,7 +129,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT *ALPHA
 
   PRINT_DEBUG_CNAME;
 
-  trans = -1;
   uplo  = -1;
   info  =  0;
 

kernel/Makefile.L1

@@ -637,49 +637,49 @@ $(KDIR)xasum_k$(TSUFFIX).$(SUFFIX)  $(KDIR)xasum_k$(TPSUFFIX).$(PSUFFIX)  : $(KE
 	$(CC) -c $(CFLAGS) -DCOMPLEX -DXDOUBLE $< -o $@
 
 $(KDIR)saxpy_k$(TSUFFIX).$(SUFFIX)  $(KDIR)saxpy_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(SAXPYKERNEL)
-	$(CC) -c $(CFLAGS) -UCOMPLEX -UCOMPLEX -UDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE $< -o $@
 
 $(KDIR)daxpy_k$(TSUFFIX).$(SUFFIX)  $(KDIR)daxpy_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(DAXPYKERNEL)
-	$(CC) -c $(CFLAGS) -UCOMPLEX -UCOMPLEX -DDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -UCOMPLEX -DDOUBLE $< -o $@
 
 $(KDIR)qaxpy_k$(TSUFFIX).$(SUFFIX)  $(KDIR)qaxpy_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(QAXPYKERNEL)
-	$(CC) -c $(CFLAGS) -UCOMPLEX -UCOMPLEX -DXDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -UCOMPLEX -DXDOUBLE $< -o $@
 
 $(KDIR)caxpy_k$(TSUFFIX).$(SUFFIX)  $(KDIR)caxpy_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(CAXPYKERNEL)
 	$(CC) -c $(CFLAGS) -DCOMPLEX -DCOMPLEX -UCONJ -UDOUBLE $< -o $@
 
 $(KDIR)zaxpy_k$(TSUFFIX).$(SUFFIX)  $(KDIR)zaxpy_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(ZAXPYKERNEL)
-	$(CC) -c $(CFLAGS) -DCOMPLEX -DCOMPLEX -UCONJ -DDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -DCOMPLEX -UCONJ -DDOUBLE $< -o $@
 
 $(KDIR)xaxpy_k$(TSUFFIX).$(SUFFIX)  $(KDIR)xaxpy_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(XAXPYKERNEL)
-	$(CC) -c $(CFLAGS) -DCOMPLEX -DCOMPLEX -UCONJ -DXDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -DCOMPLEX -UCONJ -DXDOUBLE $< -o $@
 
 $(KDIR)caxpyc_k$(TSUFFIX).$(SUFFIX)  $(KDIR)caxpyc_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(CAXPYKERNEL)
-	$(CC) -c $(CFLAGS) -DCOMPLEX -DCOMPLEX -DCONJ -UDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -DCOMPLEX -DCONJ -UDOUBLE $< -o $@
 
 $(KDIR)zaxpyc_k$(TSUFFIX).$(SUFFIX)  $(KDIR)zaxpyc_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(ZAXPYKERNEL)
-	$(CC) -c $(CFLAGS) -DCOMPLEX -DCOMPLEX -DCONJ -DDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -DCOMPLEX -DCONJ -DDOUBLE $< -o $@
 
 $(KDIR)xaxpyc_k$(TSUFFIX).$(SUFFIX)  $(KDIR)xaxpyc_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(XAXPYKERNEL)
-	$(CC) -c $(CFLAGS) -DCOMPLEX -DCOMPLEX -DCONJ -DXDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -DCOMPLEX -DCONJ -DXDOUBLE $< -o $@
 
 $(KDIR)scopy_k$(TSUFFIX).$(SUFFIX) $(KDIR)scopy_k$(TPSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SCOPYKERNEL)
-	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE -UCOMPLEX -DC_INTERFACE $< -o $@
+	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE -DC_INTERFACE $< -o $@
 
 $(KDIR)dcopy_k$(TSUFFIX).$(SUFFIX) $(KDIR)dcopy_k$(TPSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(DCOPYKERNEL)
-	$(CC) -c $(CFLAGS) -UCOMPLEX -DDOUBLE -UCOMPLEX -DC_INTERFACE $< -o $@
+	$(CC) -c $(CFLAGS) -UCOMPLEX -DDOUBLE -DC_INTERFACE $< -o $@
 
 $(KDIR)qcopy_k$(TSUFFIX).$(SUFFIX) $(KDIR)qcopy_k$(TPSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(QCOPYKERNEL)
-	$(CC) -c $(CFLAGS) -UCOMPLEX -DXDOUBLE -UCOMPLEX -DC_INTERFACE $< -o $@
+	$(CC) -c $(CFLAGS) -UCOMPLEX -DXDOUBLE -DC_INTERFACE $< -o $@
 
 $(KDIR)ccopy_k$(TSUFFIX).$(SUFFIX) $(KDIR)ccopy_k$(TPSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(CCOPYKERNEL)
-	$(CC) -c $(CFLAGS) -DCOMPLEX -UDOUBLE -DCOMPLEX -DC_INTERFACE $< -o $@
+	$(CC) -c $(CFLAGS) -DCOMPLEX -UDOUBLE -DC_INTERFACE $< -o $@
 
 $(KDIR)zcopy_k$(TSUFFIX).$(SUFFIX) $(KDIR)zcopy_k$(TPSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(ZCOPYKERNEL)
-	$(CC) -c $(CFLAGS) -DCOMPLEX -DDOUBLE -DCOMPLEX -DC_INTERFACE $< -o $@
+	$(CC) -c $(CFLAGS) -DCOMPLEX -DDOUBLE -DC_INTERFACE $< -o $@
 
 $(KDIR)xcopy_k$(TSUFFIX).$(SUFFIX) $(KDIR)xcopy_k$(TPSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(XCOPYKERNEL)
-	$(CC) -c $(CFLAGS) -DCOMPLEX -DXDOUBLE -DCOMPLEX -DC_INTERFACE $< -o $@
+	$(CC) -c $(CFLAGS) -DCOMPLEX -DXDOUBLE -DC_INTERFACE $< -o $@
 
 $(KDIR)ddot_k$(TSUFFIX).$(SUFFIX) $(KDIR)ddot_k$(TPSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(DDOTKERNEL)
 	$(CC) -c $(CFLAGS) -UCOMPLEX -DDOUBLE $< -o $@
@@ -799,15 +799,15 @@ $(KDIR)xswap_k$(TSUFFIX).$(SUFFIX)  $(KDIR)xswap_k$(TPSUFFIX).$(PSUFFIX)  : $(KE
 	$(CC) -c $(CFLAGS) -DCOMPLEX -DXDOUBLE $< -o $@
 
 $(KDIR)saxpby_k$(TSUFFIX).$(SUFFIX)  $(KDIR)saxpby_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(SAXPBYKERNEL)
-	$(CC) -c $(CFLAGS) -UCOMPLEX -UCOMPLEX -UDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE $< -o $@
 
 $(KDIR)daxpby_k$(TSUFFIX).$(SUFFIX)  $(KDIR)daxpby_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(DAXPBYKERNEL)
-	$(CC) -c $(CFLAGS) -UCOMPLEX -UCOMPLEX -DDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -UCOMPLEX -DDOUBLE $< -o $@
 
 $(KDIR)caxpby_k$(TSUFFIX).$(SUFFIX)  $(KDIR)caxpby_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(CAXPBYKERNEL)
-	$(CC) -c $(CFLAGS) -DCOMPLEX -DCOMPLEX -UCONJ -UDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -DCOMPLEX -UCONJ -UDOUBLE $< -o $@
 
 $(KDIR)zaxpby_k$(TSUFFIX).$(SUFFIX)  $(KDIR)zaxpby_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(ZAXPBYKERNEL)
-	$(CC) -c $(CFLAGS) -DCOMPLEX -DCOMPLEX -UCONJ -DDOUBLE $< -o $@
+	$(CC) -c $(CFLAGS) -DCOMPLEX -UCONJ -DDOUBLE $< -o $@
 
 

kernel/arm/amax.c

@@ -54,13 +54,15 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG ix=0;
 	FLOAT maxf=0.0;
 
-	if (n < 0 || inc_x < 1 ) return(maxf);
+	if (n <= 0 || inc_x <= 0) return(maxf);
 
 	maxf=ABS(x[0]);
+	ix += inc_x;
+	i++;
 
 	while(i < n)
 	{
-		if( ABS(x[ix]) > ABS(maxf) )
+		if( ABS(x[ix]) > maxf )
 		{
 			maxf = ABS(x[ix]);
 		}

kernel/arm/amin.c

@@ -54,13 +54,15 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG ix=0;
 	FLOAT minf=0.0;
 
-	if (n < 0 || inc_x < 1 ) return(minf);
+	if (n <= 0 || inc_x <= 0) return(minf);
 
 	minf=ABS(x[0]);
+	ix += inc_x;
+	i++;
 
 	while(i < n)
 	{
-		if( ABS(x[ix]) < ABS(minf) )
+		if( ABS(x[ix]) < minf )
 		{
 			minf = ABS(x[ix]);
 		}

kernel/arm/asum.c

@@ -53,7 +53,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 {
 	BLASLONG i=0;
 	FLOAT sumf = 0.0;
-	if (n < 0 || inc_x < 1 ) return(sumf);
+	if (n <= 0 || inc_x <= 0) return(sumf);
 
 	n *= inc_x;
 	while(i < n)

kernel/arm/iamax.c

@@ -55,13 +55,15 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	FLOAT maxf=0.0;
 	BLASLONG max=0;
 
-	if (n < 0 || inc_x < 1 ) return(max);
+	if (n <= 0 || inc_x <= 0) return(max);
 
 	maxf=ABS(x[0]);
+	ix += inc_x;
+	i++;
 
 	while(i < n)
 	{
-		if( ABS(x[ix]) > ABS(maxf) )
+		if( ABS(x[ix]) > maxf )
 		{
 			max = i;
 			maxf = ABS(x[ix]);

kernel/arm/iamin.c

@@ -55,9 +55,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	FLOAT minf=0.0;
 	BLASLONG min=0;
 
-	if (n < 0 || inc_x < 1 ) return(min);
+	if (n <= 0 || inc_x <= 0) return(min);
 
 	minf=ABS(x[0]);
+	ix += inc_x;
+	i++;
 
 	while(i < n)
 	{

kernel/arm/imax.c

@@ -47,9 +47,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	FLOAT maxf=0.0;
 	BLASLONG max=0;
 
-	if (n < 0 || inc_x < 1 ) return(max);
+	if (n <= 0 || inc_x <= 0) return(max);
 
 	maxf=x[0];
+	ix += inc_x;
+	i++;
 
 	while(i < n)
 	{

kernel/arm/imin.c

@@ -45,9 +45,11 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	FLOAT minf=0.0;
 	BLASLONG min=0;
 
-	if (n < 0 || inc_x < 1 ) return(min);
+	if (n <= 0 || inc_x <= 0) return(min);
 
 	minf=x[0];
+	ix += inc_x;
+	i++;
 
 	while(i < n)
 	{

kernel/arm/izamax.c

@@ -53,24 +53,24 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 {
 	BLASLONG i=0;
 	BLASLONG ix=0;
-	FLOAT maxf[2];
+	FLOAT maxf;
 	BLASLONG max=0;
 	BLASLONG inc_x2;
 
-	if (n < 0 || inc_x < 1 ) return(max);
+	if (n <= 0 || inc_x <= 0) return(max);
 
 	inc_x2 = 2 * inc_x;
 
-	maxf[0] = ABS(x[ix]);
-	maxf[1] = ABS(x[ix+1]);
+	maxf = CABS1(x,0);
+	ix += inc_x2;
+	i++;
 
 	while(i < n)
 	{
-		if( CABS1(x,ix) > CABS1(maxf,0) )
+		if( CABS1(x,ix) > maxf )
 		{
 			max = i;
-			maxf[0] = ABS(x[ix]);
-			maxf[1] = ABS(x[ix+1]);
+			maxf = CABS1(x,ix);
 		}
 		ix += inc_x2;
 		i++;

kernel/arm/izamin.c

@@ -53,24 +53,24 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 {
 	BLASLONG i=0;
 	BLASLONG ix=0;
-	FLOAT minf[2];
+	FLOAT minf;
 	BLASLONG min=0;
 	BLASLONG inc_x2;
 
-	if (n < 0 || inc_x < 1 ) return(min);
+	if (n <= 0 || inc_x <= 0) return(min);
 
 	inc_x2 = 2 * inc_x;
 
-	minf[0] = ABS(x[ix]);
-	minf[1] = ABS(x[ix+1]);
+	minf = CABS1(x,0);
+	ix += inc_x2;
+	i++;
 
 	while(i < n)
 	{
-		if( CABS1(x,ix) < CABS1(minf,0) )
+		if( CABS1(x,ix) < minf )
 		{
 			min = i;
-			minf[0] = ABS(x[ix]);
-			minf[1] = ABS(x[ix+1]);
+			minf = CABS1(x,ix);
 		}
 		ix += inc_x2;
 		i++;

kernel/arm/max.c

@@ -44,9 +44,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG ix=0;
 	FLOAT maxf=0.0;
 
-	if (n < 0 || inc_x < 1 ) return(maxf);
+	if (n <= 0 || inc_x <= 0) return(maxf);
 
 	maxf=x[0];
+	ix += inc_x;
+	i++;
 
 	while(i < n)
 	{

kernel/arm/min.c

@@ -44,9 +44,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG ix=0;
 	FLOAT minf=0.0;
 
-	if (n < 0 || inc_x < 1 ) return(minf);
+	if (n <= 0 || inc_x <= 0) return(minf);
 
 	minf=x[0];
+	ix += inc_x;
+	i++;
 
 	while(i < n)
 	{

kernel/arm/nrm2.c

@@ -57,7 +57,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	FLOAT absxi = 0.0;
 
 
-	if (n < 0 || inc_x < 1 ) return(0.0);
+	if (n <= 0 || inc_x <= 0) return(0.0);
 	if ( n == 1 ) return( ABS(x[0]) );
 
 	n *= inc_x;

kernel/arm/zamax.c

@@ -53,29 +53,27 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 {
 	BLASLONG i=0;
 	BLASLONG ix=0;
-	FLOAT maxf[2];
-	BLASLONG max=0;
+	FLOAT maxf;
 	BLASLONG inc_x2;
 
-	if (n < 0 || inc_x < 1 ) return(0.0);
+	if (n <= 0 || inc_x <= 0) return(0.0);
 
 	inc_x2 = 2 * inc_x;
 
-	maxf[0] = ABS(x[ix]);
-	maxf[1] = ABS(x[ix+1]);
+	maxf = CABS1(x,0);
+	ix += inc_x2;
+	i++;
 
 	while(i < n)
 	{
-		if( CABS1(x,ix) > CABS1(maxf,0) )
+		if( CABS1(x,ix) > maxf )
 		{
-			max = i;
-			maxf[0] = ABS(x[ix]);
-			maxf[1] = ABS(x[ix+1]);
+			maxf = CABS1(x,ix);
 		}
 		ix += inc_x2;
 		i++;
 	}
-	return(CABS1(maxf,0));
+	return(maxf);
 }
 
 

kernel/arm/zamin.c

@@ -53,29 +53,27 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 {
 	BLASLONG i=0;
 	BLASLONG ix=0;
-	FLOAT minf[2];
-	BLASLONG min=0;
+	FLOAT minf;
 	BLASLONG inc_x2;
 
-	if (n < 0 || inc_x < 1 ) return(0.0);
+	if (n <= 0 || inc_x <= 0) return(0.0);
 
 	inc_x2 = 2 * inc_x;
 
-	minf[0] = ABS(x[ix]);
-	minf[1] = ABS(x[ix+1]);
+	minf = CABS1(x,0);
+	ix += inc_x2;
+	i++;
 
 	while(i < n)
 	{
-		if( CABS1(x,ix) < CABS1(minf,0) )
+		if( CABS1(x,ix) < minf )
 		{
-			min = i;
-			minf[0] = ABS(x[ix]);
-			minf[1] = ABS(x[ix+1]);
+			minf = CABS1(x,ix);
 		}
 		ix += inc_x2;
 		i++;
 	}
-	return(CABS1(minf,0));
+	return(minf);
 }
 
 

kernel/arm/zasum.c

@@ -55,7 +55,8 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG i=0;
 	FLOAT sumf = 0.0;
 	BLASLONG inc_x2;
-	if (n < 0 || inc_x < 1 ) return(sumf);
+
+	if (n <= 0 || inc_x <= 0) return(sumf);
 
 	inc_x2 = 2 * inc_x;
 

kernel/arm/zaxpby.c

@@ -37,11 +37,9 @@ int CNAME(BLASLONG n, FLOAT alpha_r, FLOAT alpha_i, FLOAT *x, BLASLONG inc_x, FL
 	BLASLONG i=0;
 	BLASLONG ix,iy;
 	FLOAT temp;
+	BLASLONG inc_x2, inc_y2;
 
-	BLASLONG inc_x2;
-	BLASLONG inc_y2;
-
-	if ( n < 0     )  return(0);
+	if ( n <= 0     )  return(0);
 
 	ix = 0;
 	iy = 0;

kernel/arm/znrm2.c

@@ -57,7 +57,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	BLASLONG inc_x2;
 	FLOAT temp;
 
-	if (n < 0 || inc_x < 1 ) return(0.0);
+	if (n <= 0 || inc_x <= 0) return(0.0);
 
 	inc_x2 = 2 * inc_x;
 

kernel/arm64/KERNEL.CORTEXA57

@@ -0,0 +1,91 @@
+include $(KERNELDIR)/KERNEL.ARMV8
+
+SAMAXKERNEL  = amax.S
+DAMAXKERNEL  = amax.S
+CAMAXKERNEL  = zamax.S
+ZAMAXKERNEL  = zamax.S
+
+ISAMAXKERNEL = isamax.S
+IDAMAXKERNEL = idamax.S
+ICAMAXKERNEL = izamax.S
+IZAMAXKERNEL = izamax.S
+
+SASUMKERNEL  = asum.S
+DASUMKERNEL  = asum.S
+CASUMKERNEL  = casum.S
+ZASUMKERNEL  = zasum.S
+
+SAXPYKERNEL  = axpy.S
+DAXPYKERNEL  = axpy.S
+CAXPYKERNEL  = zaxpy.S
+ZAXPYKERNEL  = zaxpy.S
+
+SCOPYKERNEL  = copy.S
+DCOPYKERNEL  = copy.S
+CCOPYKERNEL  = copy.S
+ZCOPYKERNEL  = copy.S
+
+DOTKERNEL    = dot.S
+DDOTKERNEL   = dot.S
+CDOTKERNEL   = zdot.S
+ZDOTKERNEL   = zdot.S
+
+SNRM2KERNEL  = snrm2.S
+DNRM2KERNEL  = dnrm2.S
+CNRM2KERNEL  = znrm2.S
+ZNRM2KERNEL  = znrm2.S
+
+SROTKERNEL   = rot.S
+DROTKERNEL   = rot.S
+CROTKERNEL   = zrot.S
+ZROTKERNEL   = zrot.S
+
+SCALKERNEL  = scal.S
+DSCALKERNEL  = scal.S
+CSCALKERNEL  = zscal.S
+ZSCALKERNEL  = zscal.S
+
+SSWAPKERNEL  = swap.S
+DSWAPKERNEL  = swap.S
+CSWAPKERNEL  = swap.S
+ZSWAPKERNEL  = swap.S
+
+SGEMVNKERNEL = gemv_n.S
+DGEMVNKERNEL = gemv_n.S
+CGEMVNKERNEL = zgemv_n.S
+ZGEMVNKERNEL = zgemv_n.S
+
+SGEMVTKERNEL = gemv_t.S
+DGEMVTKERNEL = gemv_t.S
+CGEMVTKERNEL = zgemv_t.S
+ZGEMVTKERNEL = zgemv_t.S
+
+STRMMKERNEL = strmm_kernel_4x4.S
+DTRMMKERNEL = dtrmm_kernel_4x4.S
+CTRMMKERNEL = ctrmm_kernel_4x4.S
+ZTRMMKERNEL = ztrmm_kernel_4x4.S
+
+SGEMMKERNEL    =  sgemm_kernel_4x4.S
+SGEMMONCOPY    =  ../generic/gemm_ncopy_4.c
+SGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
+SGEMMONCOPYOBJ =  sgemm_oncopy.o
+SGEMMOTCOPYOBJ =  sgemm_otcopy.o
+
+DGEMMKERNEL    =  dgemm_kernel_4x4.S
+DGEMMONCOPY    =  ../generic/gemm_ncopy_4.c
+DGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
+DGEMMONCOPYOBJ =  dgemm_oncopy.o
+DGEMMOTCOPYOBJ =  dgemm_otcopy.o
+
+CGEMMKERNEL    =  cgemm_kernel_4x4.S
+CGEMMONCOPY    =  ../generic/zgemm_ncopy_4.c
+CGEMMOTCOPY    =  ../generic/zgemm_tcopy_4.c
+CGEMMONCOPYOBJ =  cgemm_oncopy.o
+CGEMMOTCOPYOBJ =  cgemm_otcopy.o
+
+ZGEMMKERNEL    =  zgemm_kernel_4x4.S
+ZGEMMONCOPY    =  ../generic/zgemm_ncopy_4.c
+ZGEMMOTCOPY    =  ../generic/zgemm_tcopy_4.c
+ZGEMMONCOPYOBJ =  zgemm_oncopy.o
+ZGEMMOTCOPYOBJ =  zgemm_otcopy.o
+

kernel/arm64/amax.S

@@ -0,0 +1,249 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if defined(USE_MIN)
+#define COND	le
+#else
+#define COND	ge
+#endif
+
+#if !defined(DOUBLE)
+#define REG0	wzr
+#define MAXF	s0
+#define TMPF	s1
+#define TMPVF	{v1.s}[0]
+#define SZ	4
+#else
+#define REG0	xzr
+#define MAXF	d0
+#define TMPF	d1
+#define TMPVF	{v1.d}[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro INIT_F1
+	ldr	MAXF, [X], #SZ
+#if defined(USE_ABS)
+	fabs	MAXF, MAXF
+#endif
+.endm
+
+.macro KERNEL_F1
+	ldr	TMPF, [X], #SZ
+#if defined(USE_ABS)
+	fabs	TMPF, TMPF
+#endif
+	fcmp	MAXF, TMPF
+	fcsel	MAXF, MAXF, TMPF, COND
+.endm
+
+.macro INIT_F4
+#if !defined(DOUBLE)
+	ld1	{v0.4s}, [X], #16
+#if defined(USE_ABS)
+	fabs	v0.4s, v0.4s
+#endif
+#if defined(USE_MIN)
+	fminv	MAXF, v0.4s
+#else
+	fmaxv	MAXF, v0.4s
+#endif
+#else // DOUBLE
+	ld2	{v0.2d,v1.2d}, [X], #32
+#if defined(USE_ABS)
+	fabs	v0.2d, v0.2d
+	fabs	v1.2d, v1.2d
+#endif
+#if defined(USE_MIN)
+	fmin	v0.2d, v0.2d, v1.2d
+	fminp	MAXF, v0.2d
+#else
+	fmax	v0.2d, v0.2d, v1.2d
+	fmaxp	MAXF, v0.2d
+#endif
+#endif
+.endm
+
+.macro KERNEL_F4
+#if !defined(DOUBLE)
+	ld1	{v1.4s}, [X], #16
+#if defined(USE_ABS)
+	fabs	v1.4s, v1.4s
+#endif
+#if defined(USE_MIN)
+	fminv	TMPF, v1.4s
+#else
+	fmaxv	TMPF, v1.4s
+#endif
+#else // DOUBLE
+	ld2	{v1.2d,v2.2d}, [X], #32
+#if defined(USE_ABS)
+	fabs	v1.2d, v1.2d
+	fabs	v2.2d, v2.2d
+#endif
+#if defined(USE_MIN)
+	fmin	v1.2d, v1.2d, v2.2d
+	fminp	TMPF, v1.2d
+#else
+	fmax	v1.2d, v1.2d, v2.2d
+	fmaxp	TMPF, v1.2d
+#endif
+#endif
+	fcmp	MAXF, TMPF
+	fcsel	MAXF, MAXF, TMPF, COND
+.endm
+
+.macro INIT_S
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #2
+	ld1	{v0.s}[0], [X], INC_X
+#else
+	lsl	INC_X, INC_X, #3
+	ld1	{v0.d}[0], [X], INC_X
+#endif
+#if defined(USE_ABS)
+	fabs	MAXF, MAXF
+#endif
+.endm
+
+.macro KERNEL_S1
+	ld1	TMPVF, [X], INC_X
+#if defined(USE_ABS)
+	fabs	TMPF, TMPF
+#endif
+	fcmp	MAXF, TMPF
+	fcsel	MAXF, MAXF, TMPF, COND
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	amax_kernel_zero
+	cmp	INC_X, xzr
+	ble	amax_kernel_zero
+
+	cmp	INC_X, #1
+	bne	amax_kernel_S_BEGIN
+
+amax_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	amax_kernel_F1_INIT
+
+	INIT_F4
+	subs	I, I, #1
+	beq	amax_kernel_F1
+
+amax_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	amax_kernel_F4
+
+amax_kernel_F1:
+
+	ands	I, N, #3
+	ble	amax_kernel_L999
+
+amax_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     amax_kernel_F10
+
+	ret
+
+amax_kernel_F1_INIT:
+
+	INIT_F1
+	subs	N, N, #1
+	b	amax_kernel_F1
+
+amax_kernel_S_BEGIN:
+
+	INIT_S
+
+	subs	N, N, #1
+	ble	amax_kernel_L999
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	amax_kernel_S1
+
+amax_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	amax_kernel_S4
+
+amax_kernel_S1:
+
+	ands	I, N, #3
+	ble	amax_kernel_L999
+
+amax_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     amax_kernel_S10
+
+amax_kernel_L999:
+
+	ret
+
+amax_kernel_zero:
+
+	fmov	MAXF, REG0
+	ret
+
+	EPILOGUE

kernel/arm64/asum.S

@@ -0,0 +1,194 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define REG0	wzr
+#define SUMF	s0
+#define TMPF	s1
+#define TMPVF	{v1.s}[0]
+#define SZ	4
+#else
+#define REG0	xzr
+#define SUMF	d0
+#define TMPF	d1
+#define TMPVF	{v1.d}[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+	ldr	TMPF, [X], #SZ
+	fabs	TMPF, TMPF
+	fadd	SUMF, SUMF, TMPF
+.endm
+
+.macro KERNEL_F8
+#if !defined(DOUBLE)
+	ld1	{v1.4s, v2.4s}, [X], #32	// Load [X3, X2, X1, X0]
+	fabs	v1.4s, v1.4s			// ABS() each value
+	fabs	v2.4s, v2.4s			// ABS() each value
+	fadd	v1.4s, v1.4s, v2.4s		// [X3+X1, X2+X0]
+	fadd	v0.4s, v0.4s, v1.4s		// [X3+X1, X2+X0]
+	PRFM	PLDL1KEEP, [X, #1024]
+#else // DOUBLE
+	ld1	{v2.2d, v3.2d, v4.2d, v5.2d}, [X]
+	add	X, X, #64
+	fabs	v2.2d, v2.2d
+	fabs	v3.2d, v3.2d
+	fabs	v4.2d, v4.2d
+	fabs	v5.2d, v5.2d
+
+	PRFM	PLDL1KEEP, [X, #1024]
+
+	fadd	v2.2d, v2.2d, v3.2d
+	fadd	v4.2d, v4.2d, v5.2d
+	fadd	v0.2d, v0.2d, v2.2d
+	fadd	v0.2d, v0.2d, v4.2d
+#endif
+.endm
+
+.macro KERNEL_F8_FINALIZE
+#if !defined(DOUBLE)
+	ext	v1.16b, v0.16b, v0.16b, #8
+	fadd	v0.2s, v0.2s, v1.2s
+	faddp	SUMF, v0.2s
+#else
+	faddp	SUMF, v0.2d
+#endif
+.endm
+
+.macro INIT_S
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #2
+#else
+	lsl	INC_X, INC_X, #3
+#endif
+.endm
+
+.macro KERNEL_S1
+	ld1	TMPVF, [X], INC_X
+	fabs	TMPF, TMPF
+	fadd	SUMF, SUMF, TMPF
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	fmov	SUMF, REG0
+#if !defined(DOUBLE)
+	fmov	s1, SUMF
+#else
+	fmov	d1, SUMF
+#endif
+
+	cmp	N, xzr
+	ble	asum_kernel_L999
+	cmp	INC_X, xzr
+	ble	asum_kernel_L999
+
+	cmp	INC_X, #1
+	bne	asum_kernel_S_BEGIN
+
+asum_kernel_F_BEGIN:
+
+	asr	I, N, #3
+	cmp	I, xzr
+	beq	asum_kernel_F1
+
+asum_kernel_F8:
+
+	KERNEL_F8
+
+	subs	I, I, #1
+	bne	asum_kernel_F8
+
+	KERNEL_F8_FINALIZE
+
+asum_kernel_F1:
+
+	ands	I, N, #7
+	ble	asum_kernel_L999
+
+asum_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     asum_kernel_F10
+
+asum_kernel_L999:
+	ret
+
+asum_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	asum_kernel_S1
+
+asum_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	asum_kernel_S4
+
+asum_kernel_S1:
+
+	ands	I, N, #3
+	ble	asum_kernel_L999
+
+asum_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     asum_kernel_S10
+
+	ret
+
+	EPILOGUE

kernel/arm64/axpy.S

@@ -0,0 +1,209 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x3	/* X vector address */
+#define	INC_X	x4	/* X stride */
+#define	Y	x5	/* Y vector address */
+#define	INC_Y	x6	/* Y stride */
+#define I	x1	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define	DA	s0	/* scale input value */
+#define TMPX	s1
+#define TMPVX	{v1.s}[0]
+#define TMPY	s2
+#define TMPVY	{v2.s}[0]
+#define SZ	4
+#else
+#define	DA	d0	/* scale input value */
+#define TMPX	d1
+#define TMPVX	{v1.d}[0]
+#define TMPY	d2
+#define TMPVY	{v2.d}[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+
+	ldr	TMPX, [X], #SZ
+	ldr	TMPY, [Y]
+	fmadd	TMPY, TMPX, DA, TMPY
+	str	TMPY, [Y], #SZ
+
+.endm
+
+.macro KERNEL_F4
+
+#if !defined(DOUBLE)
+	ld1	{v1.4s}, [X], #16
+	ld1	{v2.4s}, [Y]
+	fmla	v2.4s, v1.4s, v0.s[0]
+	st1	{v2.4s}, [Y], #16
+#else // DOUBLE
+	ld1	{v1.2d, v2.2d}, [X], #32
+	ld1	{v3.2d, v4.2d}, [Y]
+	fmla	v3.2d, v1.2d, v0.d[0]
+	fmla	v4.2d, v2.2d, v0.d[0]
+	st1	{v3.2d, v4.2d}, [Y], #32
+#endif
+
+.endm
+
+.macro KERNEL_F8
+#if !defined(DOUBLE)
+	ld1	{v1.4s, v2.4s}, [X], #32
+	ld1	{v3.4s, v4.4s}, [Y]
+
+	fmla	v3.4s, v1.4s, v0.s[0]
+	fmla	v4.4s, v2.4s, v0.s[0]
+
+	st1	{v3.4s, v4.4s}, [Y], #32
+#else // DOUBLE
+	ld1	{v1.2d, v2.2d, v3.2d, v4.2d}, [X], #64
+	ld1	{v16.2d, v17.2d, v18.2d, v19.2d}, [Y]
+
+	fmla	v16.2d, v1.2d, v0.d[0]
+	fmla	v17.2d, v2.2d, v0.d[0]
+	fmla	v18.2d, v3.2d, v0.d[0]
+	fmla	v19.2d, v4.2d, v0.d[0]
+
+	st1	{v16.2d, v17.2d, v18.2d, v19.2d}, [Y], #64
+#endif
+	PRFM	PLDL1KEEP, [X, #512]
+	PRFM	PLDL1KEEP, [Y, #512]
+.endm
+
+.macro INIT_S
+
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #2
+	lsl	INC_Y, INC_Y, #2
+#else
+	lsl	INC_X, INC_X, #3
+	lsl	INC_Y, INC_Y, #3
+#endif
+
+.endm
+
+.macro KERNEL_S1
+
+	ld1	TMPVX, [X], INC_X
+	ldr	TMPY, [Y]
+	fmadd	TMPY, TMPX, DA, TMPY
+	st1	TMPVY, [Y], INC_Y
+
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	axpy_kernel_L999
+
+	fcmp	DA, #0.0
+	beq	axpy_kernel_L999
+
+	cmp	INC_X, #1
+	bne	axpy_kernel_S_BEGIN
+	cmp	INC_Y, #1
+	bne	axpy_kernel_S_BEGIN
+
+axpy_kernel_F_BEGIN:
+
+	asr	I, N, #3
+	cmp	I, xzr
+	beq	axpy_kernel_F1
+
+axpy_kernel_F8:
+
+	KERNEL_F8
+
+	subs	I, I, #1
+	bne	axpy_kernel_F8
+
+axpy_kernel_F1:
+
+	ands	I, N, #7
+	ble	axpy_kernel_L999
+
+axpy_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     axpy_kernel_F10
+
+	mov	w0, wzr
+	ret
+
+axpy_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	axpy_kernel_S1
+
+axpy_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	axpy_kernel_S4
+
+axpy_kernel_S1:
+
+	ands	I, N, #3
+	ble	axpy_kernel_L999
+
+axpy_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     axpy_kernel_S10
+
+axpy_kernel_L999:
+
+	mov	w0, wzr
+	ret

kernel/arm64/casum.S

@@ -0,0 +1,170 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#define REG0	wzr
+#define SUMF	s0
+#define TMPF	s1
+#define TMPVF	{v1.s}[0]
+#define SZ	4
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+	ld1	{v1.2s}, [X], #8
+	fabs	v1.2s, v1.2s
+	ext	v2.8b, v1.8b, v1.8b, #4
+	fadd	TMPF, TMPF, s2
+	fadd	SUMF, SUMF, TMPF
+.endm
+
+.macro KERNEL_F8
+	ld1	{v1.4s, v2.4s, v3.4s, v4.4s}, [X]
+	add	X, X, #64
+	fabs	v1.4s, v1.4s
+	fabs	v2.4s, v2.4s
+	fabs	v3.4s, v3.4s
+	fabs	v4.4s, v4.4s
+
+	PRFM	PLDL1KEEP, [X, #1024]
+
+	fadd	v1.4s, v1.4s, v2.4s
+	fadd	v3.4s, v3.4s, v4.4s
+	fadd	v0.4s, v0.4s, v1.4s
+	fadd	v0.4s, v0.4s, v3.4s
+.endm
+
+.macro KERNEL_F8_FINALIZE
+	ext	v1.16b, v0.16b, v0.16b, #8
+	fadd	v0.2s, v0.2s, v1.2s
+	faddp	SUMF, v0.2s
+.endm
+
+.macro INIT_S
+	lsl	INC_X, INC_X, #3
+.endm
+
+.macro KERNEL_S1
+	ld1	{v1.2s}, [X], INC_X
+	fabs	v1.2s, v1.2s
+	ext	v2.8b, v1.8b, v1.8b, #4
+	fadd	TMPF, TMPF, s2
+	fadd	SUMF, SUMF, TMPF
+
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	fmov	SUMF, REG0
+	fmov	s1, SUMF
+
+	cmp	N, xzr
+	ble	asum_kernel_L999
+	cmp	INC_X, xzr
+	ble	asum_kernel_L999
+
+	cmp	INC_X, #1
+	bne	asum_kernel_S_BEGIN
+
+asum_kernel_F_BEGIN:
+
+	asr	I, N, #3
+	cmp	I, xzr
+	beq	asum_kernel_F1
+
+asum_kernel_F8:
+
+	KERNEL_F8
+
+	subs	I, I, #1
+	bne	asum_kernel_F8
+
+	KERNEL_F8_FINALIZE
+
+asum_kernel_F1:
+
+	ands	I, N, #7
+	ble	asum_kernel_L999
+
+asum_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     asum_kernel_F10
+
+asum_kernel_L999:
+	ret
+
+asum_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	asum_kernel_S1
+
+asum_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	asum_kernel_S4
+
+asum_kernel_S1:
+
+	ands	I, N, #3
+	ble	asum_kernel_L999
+
+asum_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     asum_kernel_S10
+
+	ret
+
+	EPILOGUE

kernel/arm64/copy.S

@@ -0,0 +1,232 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define	Y	x3	/* Y vector address */
+#define	INC_Y	x4	/* Y stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define TMPF	s0
+#define TMPVF	{v0.s}[0]
+#define SZ	4
+#else
+#define TMPF	d0
+#define TMPVF	{v0.d}[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+
+#if !defined(COMPLEX)
+	ldr	TMPF, [X], #SZ
+	str	TMPF, [Y], #SZ
+#else
+#if !defined(DOUBLE)
+	ld1	{v0.2s}, [X], #8
+	st1	{v0.2s}, [Y], #8
+#else
+	ld1	{v0.2d}, [X], #16
+	st1	{v0.2d}, [Y], #16
+#endif
+#endif
+
+.endm
+
+.macro KERNEL_F4
+
+#if !defined(COMPLEX)
+#if !defined(DOUBLE)
+	ld1	{v0.4s}, [X], #16
+	st1	{v0.4s}, [Y], #16
+#else // DOUBLE
+	ld1	{v0.4s}, [X], #16
+	ld1	{v1.4s}, [X], #16
+	st1	{v0.4s}, [Y], #16
+	st1	{v1.4s}, [Y], #16
+#endif
+#else // COMPLEX
+#if !defined(DOUBLE)
+	ld1	{v0.4s}, [X], #16
+	ld1	{v1.4s}, [X], #16
+	st1	{v0.4s}, [Y], #16
+	st1	{v1.4s}, [Y], #16
+#else // DOUBLE
+	ld1	{v0.4s}, [X], #16
+	ld1	{v1.4s}, [X], #16
+	ld1	{v2.4s}, [X], #16
+	ld1	{v3.4s}, [X], #16
+	st1	{v0.4s}, [Y], #16
+	st1	{v1.4s}, [Y], #16
+	st1	{v2.4s}, [Y], #16
+	st1	{v3.4s}, [Y], #16
+#endif
+#endif
+
+.endm
+
+.macro INIT_S
+
+#if !defined(COMPLEX)
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #2
+	lsl	INC_Y, INC_Y, #2
+#else
+	lsl	INC_X, INC_X, #3
+	lsl	INC_Y, INC_Y, #3
+#endif
+#else
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #3
+	lsl	INC_Y, INC_Y, #3
+#else
+	lsl	INC_X, INC_X, #4
+	lsl	INC_Y, INC_Y, #4
+#endif
+#endif
+
+.endm
+
+.macro KERNEL_S1
+
+#if !defined(COMPLEX)
+#if !defined(DOUBLE)
+	ldr	w10, [X]
+	add	X, X, INC_X
+	str	w10, [Y]
+	add	Y, Y, INC_Y
+#else
+	ldr	x10, [X]
+	add	X, X, INC_X
+	str	x10, [Y]
+	add	Y, Y, INC_Y
+#endif
+#else
+#if !defined(DOUBLE)
+	ld1	{v0.2s}, [X]
+	add	X, X, INC_X
+	st1	{v0.2s}, [Y]
+	add	Y, Y, INC_Y
+#else
+	ld1	{v0.2d}, [X]
+	add	X, X, INC_X
+	st1	{v0.2d}, [Y]
+	add	Y, Y, INC_Y
+#endif
+#endif
+
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	copy_kernel_L999
+
+	cmp	INC_X, #1
+	bne	copy_kernel_S_BEGIN
+	cmp	INC_Y, #1
+	bne	copy_kernel_S_BEGIN
+
+copy_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	copy_kernel_F1
+
+copy_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	copy_kernel_F4
+
+copy_kernel_F1:
+
+	ands	I, N, #3
+	ble	copy_kernel_L999
+
+copy_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     copy_kernel_F10
+
+	mov	w0, wzr
+	ret
+
+copy_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	copy_kernel_S1
+
+copy_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	copy_kernel_S4
+
+copy_kernel_S1:
+
+	ands	I, N, #3
+	ble	copy_kernel_L999
+
+copy_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     copy_kernel_S10
+
+copy_kernel_L999:
+
+	mov	w0, wzr
+	ret
+
+	EPILOGUE

kernel/arm64/dnrm2.S

@@ -0,0 +1,169 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#define TMPF	d6
+#define SSQ	d0
+#define TMPVF	{v6.d}[0]
+#define SZ	8
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+	ldr	TMPF, [X], #SZ
+	fmul	TMPF, TMPF, TMPF
+	fadd	SSQ, SSQ, TMPF
+.endm
+
+.macro KERNEL_F8
+	ld1	{v1.2d, v2.2d}, [X], #32
+	fmla	v0.2d, v1.2d, v1.2d
+	fmla	v5.2d, v2.2d, v2.2d
+	ld1	{v3.2d, v4.2d}, [X], #32
+	fmla	v0.2d, v3.2d, v3.2d
+	fmla	v5.2d, v4.2d, v4.2d
+	PRFM	PLDL1KEEP, [X, #1024]
+.endm
+
+.macro nrm2_kernel_F8_FINALIZE
+	fadd	v0.2d, v0.2d, v5.2d
+	faddp	SSQ, v0.2d
+.endm
+
+.macro INIT_S
+	lsl	INC_X, INC_X, #3
+	ld1	TMPVF, [X], INC_X
+	fmul	SSQ, TMPF, TMPF
+.endm
+
+.macro KERNEL_S1
+	ld1	TMPVF, [X], INC_X
+	fmul	TMPF, TMPF, TMPF
+	fadd	SSQ, SSQ, TMPF
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	fmov	SSQ, xzr
+	fmov	d5, SSQ
+
+	cmp	N, xzr
+	ble	nrm2_kernel_zero
+	cmp	INC_X, xzr
+	ble	nrm2_kernel_zero
+	cmp	INC_X, #1
+	bne	nrm2_kernel_S_BEGIN
+
+nrm2_kernel_F_BEGIN:
+
+	asr	I, N, #3
+	cmp	I, xzr
+	beq	nrm2_kernel_F1_INIT
+
+nrm2_kernel_F8:
+
+	KERNEL_F8
+
+	subs	I, I, #1
+	bne	nrm2_kernel_F8
+
+	nrm2_kernel_F8_FINALIZE
+
+nrm2_kernel_F1:
+
+	ands	I, N, #7
+	ble	nrm2_kernel_L999
+
+nrm2_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     nrm2_kernel_F10
+
+	b	nrm2_kernel_L999
+
+nrm2_kernel_F1_INIT:
+
+	b	nrm2_kernel_F1
+
+nrm2_kernel_S_BEGIN:
+
+	INIT_S
+
+	subs	N, N, #1
+	ble	nrm2_kernel_L999
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	nrm2_kernel_S1
+
+nrm2_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	nrm2_kernel_S4
+
+nrm2_kernel_S1:
+
+	ands	I, N, #3
+	ble	nrm2_kernel_L999
+
+nrm2_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+	bne     nrm2_kernel_S10
+
+nrm2_kernel_L999:
+	fsqrt	SSQ, SSQ
+	ret
+
+nrm2_kernel_zero:
+	ret
+
+	EPILOGUE

kernel/arm64/dot.S

@@ -0,0 +1,227 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define	Y	x3	/* Y vector address */
+#define	INC_Y	x4	/* Y stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#if !defined(DSDOT)
+#define REG0	wzr
+#define DOTF	s0
+#else // DSDOT
+#define REG0	xzr
+#define DOTF	d0
+#endif
+#define DOTI	s1
+#define TMPX	s2
+#define LD1VX	{v2.s}[0]
+#define TMPY	s3
+#define LD1VY	{v3.s}[0]
+#define TMPVY	v3.s[0]
+#define SZ	4
+#else
+#define REG0	xzr
+#define DOTF	d0
+#define DOTI	d1
+#define TMPX	d2
+#define LD1VX	{v2.d}[0]
+#define TMPY	d3
+#define LD1VY	{v3.d}[0]
+#define TMPVY	v3.d[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+	ldr	TMPX, [X], #SZ
+	ldr	TMPY, [Y], #SZ
+#if !defined(DSDOT)
+	fmadd	DOTF, TMPX, TMPY, DOTF
+#else // DSDOT
+	fmul	TMPX, TMPX, TMPY
+	fcvt	d2, TMPX
+	fadd	DOTF, DOTF, d2
+#endif
+.endm
+
+.macro KERNEL_F4
+#if !defined(DOUBLE)
+	ld1	{v2.4s}, [X], #16
+	ld1	{v3.4s}, [Y], #16
+#if !defined(DSDOT)
+	fmla	v0.4s, v2.4s, v3.4s
+#else
+	fmul	v2.4s, v2.4s, v3.4s
+	ext	v3.16b, v2.16b, v2.16b, #8
+	fcvtl	v2.2d, v2.2s
+	fcvtl	v3.2d, v3.2s
+	fadd	v0.2d, v0.2d, v2.2d
+	fadd	v0.2d, v0.2d, v3.2d
+#endif
+#else //DOUBLE
+	ld1	{v2.2d, v3.2d}, [X], #32
+	ld1	{v4.2d, v5.2d}, [Y], #32
+	fmul	v2.2d, v2.2d, v4.2d
+	fmul	v3.2d, v3.2d, v5.2d
+	fadd	v0.2d, v0.2d, v2.2d
+	fadd	v0.2d, v0.2d, v3.2d
+#endif
+	PRFM	PLDL1KEEP, [X, #1024]
+	PRFM	PLDL1KEEP, [Y, #1024]
+.endm
+
+.macro KERNEL_F4_FINALIZE
+#if !defined(DOUBLE)
+#if !defined(DSDOT)
+	ext	v1.16b, v0.16b, v0.16b, #8
+	fadd	v0.2s, v0.2s, v1.2s
+	faddp	DOTF, v0.2s
+#else
+	faddp	DOTF, v0.2d
+#endif
+#else //DOUBLE
+	faddp	DOTF, v0.2d
+#endif
+.endm
+
+.macro INIT_S
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #2
+	lsl	INC_Y, INC_Y, #2
+#else
+	lsl	INC_X, INC_X, #3
+	lsl	INC_Y, INC_Y, #3
+#endif
+.endm
+
+.macro KERNEL_S1
+	ld1	LD1VX, [X], INC_X
+	ld1	LD1VY, [Y], INC_Y
+#if !defined(DSDOT)
+	fmadd	DOTF, TMPX, TMPY, DOTF
+#else // DSDOT
+	fmul	TMPX, TMPX, TMPY
+	fcvt	d2, TMPX
+	fadd	DOTF, DOTF, d2
+#endif
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	fmov	DOTF, REG0
+#if defined(DOUBLE)
+	fmov	d6, DOTF
+#endif
+
+	cmp	N, xzr
+	ble	dot_kernel_L999
+
+	cmp	INC_X, #1
+	bne	dot_kernel_S_BEGIN
+	cmp	INC_Y, #1
+	bne	dot_kernel_S_BEGIN
+
+dot_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	dot_kernel_F1
+
+dot_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	dot_kernel_F4
+
+	KERNEL_F4_FINALIZE
+
+dot_kernel_F1:
+
+	ands	I, N, #3
+	ble	dot_kernel_L999
+
+dot_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     dot_kernel_F10
+
+	ret
+
+dot_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	dot_kernel_S1
+
+dot_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	dot_kernel_S4
+
+dot_kernel_S1:
+
+	ands	I, N, #3
+	ble	dot_kernel_L999
+
+dot_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     dot_kernel_S10
+
+dot_kernel_L999:
+
+	ret
+
+	EPILOGUE

kernel/arm64/gemv_n.S

@@ -0,0 +1,320 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	M	x0	/* Y vector length */
+#define	N	x1	/* X vector length */
+#define	A	x3	/* A vector address */
+#define	LDA	x4	/* A stride */
+#define	X	x5	/* X vector address */
+#define	INC_X	x6	/* X stride */
+#define	Y	x7	/* Y vector address */
+#define	INC_Y	x2	/* Y stride */
+#define	A_PTR	x9	/* loop A vector address */
+#define	Y_IPTR	x10	/* loop Y vector address */
+#define J	x11	/* loop variable */
+#define I	x12	/* loop variable */
+#define	Y_OPTR	x13	/* loop Y vector address */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define ALPHA	s0
+#define TEMP	s1
+#define TEMPV	{v1.s}[0]
+#define TMP1	s2
+#define TMPV1	{v2.s}[0]
+#define TMP2	s3
+#define TMPV2	{v3.s}[0]
+#define SZ	4
+#define SHZ	2
+#else
+#define ALPHA	d0
+#define TEMP	d1
+#define TEMPV	{v1.d}[0]
+#define TMP1	d2
+#define TMPV1	{v2.d}[0]
+#define TMP2	d3
+#define TMPV2	{v3.d}[0]
+#define SZ	8
+#define SHZ	3
+#endif
+
+/******************************************************************************/
+
+.macro SAVE_REGS
+	add	sp, sp, #-(11 * 16)
+	stp	d8, d9, [sp, #(0 * 16)]
+	stp	d10, d11, [sp, #(1 * 16)]
+	stp	d12, d13, [sp, #(2 * 16)]
+	stp	d14, d15, [sp, #(3 * 16)]
+	stp	d16, d17, [sp, #(4 * 16)]
+	stp	x18, x19, [sp, #(5 * 16)]
+	stp	x20, x21, [sp, #(6 * 16)]
+	stp	x22, x23, [sp, #(7 * 16)]
+	stp	x24, x25, [sp, #(8 * 16)]
+	stp	x26, x27, [sp, #(9 * 16)]
+	str	x28, [sp, #(10 * 16)]
+.endm
+
+.macro RESTORE_REGS
+	ldp	d8, d9, [sp, #(0 * 16)]
+	ldp	d10, d11, [sp, #(1 * 16)]
+	ldp	d12, d13, [sp, #(2 * 16)]
+	ldp	d14, d15, [sp, #(3 * 16)]
+	ldp	d16, d17, [sp, #(4 * 16)]
+	ldp	x18, x19, [sp, #(5 * 16)]
+	ldp	x20, x21, [sp, #(6 * 16)]
+	ldp	x22, x23, [sp, #(7 * 16)]
+	ldp	x24, x25, [sp, #(8 * 16)]
+	ldp	x26, x27, [sp, #(9 * 16)]
+	ldr	x28, [sp, #(10 * 16)]
+	add	sp, sp, #(11*16)
+.endm
+
+.macro KERNEL_F16
+#if !defined(DOUBLE)
+	ld1	{v2.4s, v3.4s}, [A_PTR], #32
+	ld1	{v4.4s, v5.4s}, [Y_IPTR], #32
+	fmla	v4.4s, v1.4s, v2.4s
+	fmla	v5.4s, v1.4s, v3.4s
+	st1	{v4.4s, v5.4s}, [Y_OPTR], #32
+
+	ld1	{v6.4s, v7.4s}, [A_PTR], #32
+	ld1	{v8.4s, v9.4s}, [Y_IPTR], #32
+	fmla	v8.4s, v1.4s, v6.4s
+	fmla	v9.4s, v1.4s, v7.4s
+	st1	{v8.4s, v9.4s}, [Y_OPTR], #32
+#else //DOUBLE
+	ld1	{v2.2d, v3.2d}, [A_PTR], #32
+	ld1	{v4.2d, v5.2d}, [Y_IPTR], #32
+	fmla	v4.2d, v1.2d, v2.2d
+	fmla	v5.2d, v1.2d, v3.2d
+	st1	{v4.2d, v5.2d}, [Y_OPTR], #32
+
+	ld1	{v6.2d, v7.2d}, [A_PTR], #32
+	ld1	{v8.2d, v9.2d}, [Y_IPTR], #32
+	fmla	v8.2d, v1.2d, v6.2d
+	fmla	v9.2d, v1.2d, v7.2d
+	st1	{v8.2d, v9.2d}, [Y_OPTR], #32
+
+	ld1	{v10.2d, v11.2d}, [A_PTR], #32
+	ld1	{v12.2d, v13.2d}, [Y_IPTR], #32
+	fmla	v12.2d, v1.2d, v10.2d
+	fmla	v13.2d, v1.2d, v11.2d
+	st1	{v12.2d, v13.2d}, [Y_OPTR], #32
+
+	ld1	{v14.2d, v15.2d}, [A_PTR], #32
+	ld1	{v16.2d, v17.2d}, [Y_IPTR], #32
+	fmla	v16.2d, v1.2d, v14.2d
+	fmla	v17.2d, v1.2d, v15.2d
+	st1	{v16.2d, v17.2d}, [Y_OPTR], #32
+#endif
+.endm
+
+.macro KERNEL_F4
+#if !defined(DOUBLE)
+	ld1	{v2.4s}, [A_PTR], #16
+	ld1	{v3.4s}, [Y_IPTR], #16
+	fmla	v3.4s, v1.4s, v2.4s
+	st1	{v3.4s}, [Y_OPTR], #16
+#else
+	ld1	{v2.2d}, [A_PTR], #16
+	ld1	{v3.2d}, [Y_IPTR], #16
+	fmla	v3.2d, v1.2d, v2.2d
+	st1	{v3.2d}, [Y_OPTR], #16
+
+	ld1	{v4.2d}, [A_PTR], #16
+	ld1	{v5.2d}, [Y_IPTR], #16
+	fmla	v5.2d, v1.2d, v4.2d
+	st1	{v5.2d}, [Y_OPTR], #16
+#endif
+.endm
+
+.macro KERNEL_F1
+
+	ld1	TMPV1, [A_PTR], #SZ
+	ld1	TMPV2, [Y_IPTR]
+	fmadd	TMP2, TEMP, TMP1, TMP2
+	st1	TMPV2, [Y_IPTR], #SZ
+
+.endm
+
+.macro INIT_S
+
+	lsl	INC_Y, INC_Y, #SHZ
+
+.endm
+
+.macro KERNEL_S1
+
+	ld1	TMPV1, [A_PTR], #SZ
+	ld1	TMPV2, [Y_IPTR]
+	fmadd	TMP2, TEMP, TMP1, TMP2
+	st1	TMPV2, [Y_IPTR], INC_Y
+
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	ldr	INC_Y, [sp]
+
+	SAVE_REGS
+
+	cmp	N, xzr
+	ble	gemv_n_kernel_L999
+	cmp	M, xzr
+	ble	gemv_n_kernel_L999
+
+	lsl	LDA, LDA, #SHZ
+	lsl	INC_X, INC_X, #SHZ
+	mov	J, N
+
+	cmp	INC_Y, #1
+	bne	gemv_n_kernel_S_BEGIN
+
+gemv_n_kernel_F_LOOP:
+
+	ld1	TEMPV, [X], INC_X
+	fmul	TEMP, ALPHA, TEMP
+#if !defined(DOUBLE)
+	ins	v1.s[1], v1.s[0]
+	ins	v1.s[2], v1.s[0]
+	ins	v1.s[3], v1.s[0]
+#else
+	ins	v1.d[1], v1.d[0]
+#endif
+	mov	A_PTR, A
+	mov	Y_IPTR, Y
+	mov	Y_OPTR, Y
+
+gemv_n_kernel_F32:
+
+	asr	I, M, #5
+	cmp	I, xzr
+	beq	gemv_n_kernel_F4
+
+gemv_n_kernel_F320:
+
+	KERNEL_F16
+	KERNEL_F16
+
+	subs	I, I, #1
+	bne	gemv_n_kernel_F320
+
+gemv_n_kernel_F4:
+	ands	I, M, #31
+	asr	I, I, #2
+	cmp	I, xzr
+	beq	gemv_n_kernel_F1
+
+gemv_n_kernel_F40:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	gemv_n_kernel_F40
+
+gemv_n_kernel_F1:
+	ands	I, M, #3
+	ble	gemv_n_kernel_F_END
+
+gemv_n_kernel_F10:
+
+	KERNEL_F1
+
+	subs	I, I, #1
+	bne	gemv_n_kernel_F10
+
+gemv_n_kernel_F_END:
+
+	add	A, A, LDA
+	subs	J, J, #1
+	bne	gemv_n_kernel_F_LOOP
+
+	b	gemv_n_kernel_L999
+
+gemv_n_kernel_S_BEGIN:
+
+	INIT_S
+
+gemv_n_kernel_S_LOOP:
+
+	ld1	TEMPV, [X], INC_X
+	fmul	TEMP, ALPHA, TEMP
+	mov	A_PTR, A
+	mov	Y_IPTR, Y
+
+	asr	I, M, #2
+	cmp	I, xzr
+	ble	gemv_n_kernel_S1
+
+gemv_n_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	gemv_n_kernel_S4
+
+gemv_n_kernel_S1:
+
+	ands	I, M, #3
+	ble	gemv_n_kernel_S_END
+
+gemv_n_kernel_S10:
+
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	gemv_n_kernel_S10
+
+gemv_n_kernel_S_END:
+
+	add	A, A, LDA
+	subs	J, J, #1
+	bne	gemv_n_kernel_S_LOOP
+
+gemv_n_kernel_L999:
+
+	mov	w0, wzr
+
+	RESTORE_REGS
+
+	ret
+
+	EPILOGUE

kernel/arm64/gemv_t.S

@@ -0,0 +1,347 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	M	x0	/* Y vector length */
+#define	N	x1	/* X vector length */
+#define	A	x3	/* A vector address */
+#define	LDA	x4	/* A stride */
+#define	X	x5	/* X vector address */
+#define	INC_X	x6	/* X stride */
+#define	Y	x7	/* Y vector address */
+#define	INC_Y	x2	/* Y stride */
+#define	A_PTR	x9	/* loop A vector address */
+#define	X_PTR	x10	/* loop X vector address */
+#define J	x11	/* loop variable */
+#define I	x12	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define REG0	wzr
+#define ALPHA	s0
+#define TEMP	s1
+#define TEMP1	s2
+#define TEMP2	s3
+#define TEMP3	s4
+#define TEMPV	{v1.s}[0]
+#define TMP1	s2
+#define TMPV1	{v2.s}[0]
+#define TMP2	s3
+#define TMPV2	{v3.s}[0]
+#define SZ	4
+#define SHZ	2
+#else
+#define REG0	xzr
+#define ALPHA	d0
+#define TEMP	d1
+#define TEMP1	d2
+#define TEMP2	d3
+#define TEMP3	d4
+#define TEMPV	{v1.d}[0]
+#define TMP1	d2
+#define TMPV1	{v2.d}[0]
+#define TMP2	d3
+#define TMPV2	{v3.d}[0]
+#define SZ	8
+#define SHZ	3
+#endif
+
+/******************************************************************************/
+
+.macro SAVE_REGS
+	add	sp, sp, #-(11 * 16)
+	stp	d8, d9, [sp, #(0 * 16)]
+	stp	d10, d11, [sp, #(1 * 16)]
+	stp	d12, d13, [sp, #(2 * 16)]
+	stp	d14, d15, [sp, #(3 * 16)]
+	stp	d16, d17, [sp, #(4 * 16)]
+	stp	x18, x19, [sp, #(5 * 16)]
+	stp	x20, x21, [sp, #(6 * 16)]
+	stp	x22, x23, [sp, #(7 * 16)]
+	stp	x24, x25, [sp, #(8 * 16)]
+	stp	x26, x27, [sp, #(9 * 16)]
+	str	x28, [sp, #(10 * 16)]
+.endm
+
+.macro RESTORE_REGS
+	ldp	d8, d9, [sp, #(0 * 16)]
+	ldp	d10, d11, [sp, #(1 * 16)]
+	ldp	d12, d13, [sp, #(2 * 16)]
+	ldp	d14, d15, [sp, #(3 * 16)]
+	ldp	d16, d17, [sp, #(4 * 16)]
+	ldp	x18, x19, [sp, #(5 * 16)]
+	ldp	x20, x21, [sp, #(6 * 16)]
+	ldp	x22, x23, [sp, #(7 * 16)]
+	ldp	x24, x25, [sp, #(8 * 16)]
+	ldp	x26, x27, [sp, #(9 * 16)]
+	ldr	x28, [sp, #(10 * 16)]
+	add	sp, sp, #(11*16)
+.endm
+
+.macro KERNEL_F32
+#if !defined(DOUBLE)
+	ld1	{v5.4s, v6.4s, v7.4s, v8.4s}, [A_PTR], #64
+	ld1	{v9.4s, v10.4s, v11.4s, v12.4s}, [X_PTR], #64
+	fmla	v1.4s, v5.4s, v9.4s
+	fmla	v2.4s, v6.4s, v10.4s
+	fmla	v3.4s, v7.4s, v11.4s
+	fmla	v4.4s, v8.4s, v12.4s
+
+	ld1	{v13.4s, v14.4s, v15.4s, v16.4s}, [A_PTR], #64
+	ld1	{v17.4s, v18.4s, v19.4s, v20.4s}, [X_PTR], #64
+	fmla	v1.4s, v13.4s, v17.4s
+	fmla	v2.4s, v14.4s, v18.4s
+	fmla	v3.4s, v15.4s, v19.4s
+	fmla	v4.4s, v16.4s, v20.4s
+#else
+	ld1	{v5.2d, v6.2d, v7.2d, v8.2d}, [A_PTR], #64
+	ld1	{v9.2d, v10.2d, v11.2d, v12.2d}, [X_PTR], #64
+	fmla	v1.2d, v5.2d, v9.2d
+	fmla	v2.2d, v6.2d, v10.2d
+	fmla	v3.2d, v7.2d, v11.2d
+	fmla	v4.2d, v8.2d, v12.2d
+
+	ld1	{v13.2d, v14.2d, v15.2d, v16.2d}, [A_PTR], #64
+	ld1	{v17.2d, v18.2d, v19.2d, v20.2d}, [X_PTR], #64
+	fmla	v1.2d, v13.2d, v17.2d
+	fmla	v2.2d, v14.2d, v18.2d
+	fmla	v3.2d, v15.2d, v19.2d
+	fmla	v4.2d, v16.2d, v20.2d
+
+	ld1	{v5.2d, v6.2d, v7.2d, v8.2d}, [A_PTR], #64
+	ld1	{v9.2d, v10.2d, v11.2d, v12.2d}, [X_PTR], #64
+	fmla	v1.2d, v5.2d, v9.2d
+	fmla	v2.2d, v6.2d, v10.2d
+	fmla	v3.2d, v7.2d, v11.2d
+	fmla	v4.2d, v8.2d, v12.2d
+
+	ld1	{v13.2d, v14.2d, v15.2d, v16.2d}, [A_PTR], #64
+	ld1	{v17.2d, v18.2d, v19.2d, v20.2d}, [X_PTR], #64
+	fmla	v1.2d, v13.2d, v17.2d
+	fmla	v2.2d, v14.2d, v18.2d
+	fmla	v3.2d, v15.2d, v19.2d
+	fmla	v4.2d, v16.2d, v20.2d
+#endif
+.endm
+
+.macro KERNEL_F32_FINALIZE
+#if !defined(DOUBLE)
+	fadd	v1.4s, v1.4s, v2.4s
+	fadd	v1.4s, v1.4s, v3.4s
+	fadd	v1.4s, v1.4s, v4.4s
+#else
+	fadd	v1.2d, v1.2d, v2.2d
+	fadd	v1.2d, v1.2d, v3.2d
+	fadd	v1.2d, v1.2d, v4.2d
+#endif
+.endm
+
+.macro KERNEL_F4
+#if !defined(DOUBLE)
+	ld1	{v2.4s}, [A_PTR], #16
+	ld1	{v3.4s}, [X_PTR], #16
+	fmla	v1.4s, v2.4s, v3.4s
+#else
+	ld1	{v2.2d}, [A_PTR], #16
+	ld1	{v3.2d}, [X_PTR], #16
+	fmla	v1.2d, v2.2d, v3.2d
+
+	ld1	{v4.2d}, [A_PTR], #16
+	ld1	{v5.2d}, [X_PTR], #16
+	fmla	v1.2d, v4.2d, v5.2d
+#endif
+.endm
+
+.macro KERNEL_F4_FINALIZE
+#if !defined(DOUBLE)
+	ext	v2.16b, v1.16b, v1.16b, #8
+	fadd	v1.2s, v1.2s, v2.2s
+	faddp	TEMP, v1.2s
+#else
+	faddp	TEMP, v1.2d
+#endif
+.endm
+
+.macro KERNEL_F1
+	ld1	TMPV1, [A_PTR], #SZ
+	ld1	TMPV2, [X_PTR], #SZ
+	fmadd	TEMP, TMP1, TMP2, TEMP
+.endm
+
+.macro INIT_S
+	lsl	INC_X, INC_X, #SHZ
+.endm
+
+.macro KERNEL_S1
+	ld1	TMPV1, [A_PTR], #SZ
+	ld1	TMPV2, [X_PTR], INC_X
+	fmadd	TEMP, TMP1, TMP2, TEMP
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	ldr	INC_Y, [sp]
+
+	SAVE_REGS
+
+	cmp	N, xzr
+	ble	gemv_t_kernel_L999
+	cmp	M, xzr
+	ble	gemv_t_kernel_L999
+
+	lsl	LDA, LDA, #SHZ
+	lsl	INC_Y, INC_Y, #SHZ
+	mov	J, N
+
+	cmp	INC_X, #1
+	bne	gemv_t_kernel_S_BEGIN
+
+gemv_t_kernel_F_LOOP:
+
+	fmov	TEMP, REG0
+	fmov	TEMP1, REG0
+	fmov	TEMP2, REG0
+	fmov	TEMP3, REG0
+
+	mov	A_PTR, A
+	mov	X_PTR, X
+
+gemv_t_kernel_F32:
+
+	asr	I, M, #5
+	cmp	I, xzr
+	beq	gemv_t_kernel_F4
+
+gemv_t_kernel_F320:
+
+	KERNEL_F32
+
+	subs	I, I, #1
+	bne	gemv_t_kernel_F320
+
+	KERNEL_F32_FINALIZE
+
+gemv_t_kernel_F4:
+	ands	I, M, #31
+	asr	I, I, #2
+	cmp	I, xzr
+	beq	gemv_t_kernel_F1
+
+gemv_t_kernel_F40:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	gemv_t_kernel_F40
+
+gemv_t_kernel_F1:
+
+	KERNEL_F4_FINALIZE
+
+	ands	I, M, #3
+	ble	gemv_t_kernel_F_END
+
+gemv_t_kernel_F10:
+
+	KERNEL_F1
+
+	subs	I, I, #1
+	bne	gemv_t_kernel_F10
+
+gemv_t_kernel_F_END:
+
+	ld1	TMPV1, [Y]
+	add	A, A, LDA
+	subs	J, J, #1
+	fmadd	TMP1, ALPHA, TEMP, TMP1
+	st1	TMPV1, [Y], INC_Y
+	bne	gemv_t_kernel_F_LOOP
+
+	b	gemv_t_kernel_L999
+
+gemv_t_kernel_S_BEGIN:
+
+	INIT_S
+
+gemv_t_kernel_S_LOOP:
+
+	fmov	TEMP, REG0
+	mov	A_PTR, A
+	mov	X_PTR, X
+
+	asr	I, M, #2
+	cmp	I, xzr
+	ble	gemv_t_kernel_S1
+
+gemv_t_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	gemv_t_kernel_S4
+
+gemv_t_kernel_S1:
+
+	ands	I, M, #3
+	ble	gemv_t_kernel_S_END
+
+gemv_t_kernel_S10:
+
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	gemv_t_kernel_S10
+
+gemv_t_kernel_S_END:
+
+	ld1	TMPV1, [Y]
+	add	A, A, LDA
+	subs    J, J, #1
+	fmadd	TMP1, ALPHA, TEMP, TMP1
+	st1	TMPV1, [Y], INC_Y
+        bne     gemv_t_kernel_S_LOOP
+
+gemv_t_kernel_L999:
+
+	RESTORE_REGS
+
+	mov	w0, wzr
+	ret
+
+	EPILOGUE

kernel/arm64/idamax.S

@@ -0,0 +1,124 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define INDEX	x3	/* index of max/min value */
+#define Z	x4	/* vector index */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if defined(USE_MIN)
+#define COND	le
+#else
+#define COND	ge
+#endif
+
+#define MAXF	d0
+#define TMPF	d1
+#define TMPVF	{v1.d}[0]
+#define SZ	8
+
+/******************************************************************************/
+
+.macro INIT_S
+	lsl	INC_X, INC_X, #3
+	ld1	{v0.d}[0], [X], INC_X
+	mov	Z, #1
+	mov	INDEX, Z
+	fabs	MAXF, MAXF
+.endm
+
+.macro KERNEL_S1
+	ld1	TMPVF, [X], INC_X
+	add	Z, Z, #1
+	fabs	TMPF, TMPF
+	fcmp	MAXF, TMPF
+	fcsel	MAXF, MAXF, TMPF, COND
+	csel	INDEX, INDEX, Z, COND
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	iamax_kernel_zero
+	cmp	INC_X, xzr
+	ble	iamax_kernel_zero
+
+	INIT_S
+
+	subs	N, N, #1
+	ble	iamax_kernel_L999
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	iamax_kernel_S1
+
+iamax_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	iamax_kernel_S4
+
+iamax_kernel_S1:
+
+	ands	I, N, #3
+	ble	iamax_kernel_L999
+
+iamax_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     iamax_kernel_S10
+
+iamax_kernel_L999:
+
+	mov	x0, INDEX
+	ret
+
+iamax_kernel_zero:
+
+	mov	x0, xzr
+	ret
+
+	EPILOGUE

kernel/arm64/isamax.S

@@ -0,0 +1,213 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define INDEX	x3	/* index of max/min value */
+#define Z	x4	/* vector index */
+#define I	x5	/* loop variable */
+#define X_COPY	x6	/* copy of X address */
+#define MAXF_Z	x7
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#define MAXF	s5
+#define TMPF	s6
+#define TMPVF	{v6.s}[0]
+#define SZ	4
+
+/******************************************************************************/
+
+.macro INIT_F1
+	ldr	MAXF, [X], #SZ
+	mov	Z, #1
+	mov	INDEX, Z
+	fabs	MAXF, MAXF
+.endm
+
+.macro KERNEL_F1
+	ldr	TMPF, [X], #SZ
+	add	Z, Z, #1
+	fabs	TMPF, TMPF
+	fcmp	TMPF, MAXF
+	fcsel	MAXF, MAXF, TMPF, le
+	csel	INDEX, INDEX, Z, le
+.endm
+
+.macro INIT_F4
+	ld1	{v0.4s}, [X], #16
+	fabs	v0.4s, v0.4s
+	fmaxv	MAXF, v0.4s
+	mov	Z, #5
+	mov	MAXF_Z, #1
+.endm
+
+.macro KERNEL_F4
+	ld1	{v0.4s}, [X], #16
+	fabs	v0.4s, v0.4s
+	fmaxv	TMPF, v0.4s
+	PRFM	PLDL1KEEP, [X, #512]
+	fcmp	TMPF, MAXF
+	fcsel	MAXF, MAXF, TMPF, le
+	csel	MAXF_Z, MAXF_Z, Z, le
+	add	Z, Z, #4
+.endm
+
+
+.macro KERNEL_F4_FINALIZE
+	mov	INDEX, MAXF_Z
+	sub	MAXF_Z, MAXF_Z, #1
+	lsl	MAXF_Z, MAXF_Z, #2
+	add	X_COPY, X_COPY, MAXF_Z
+	ldr	TMPF, [X_COPY], #SZ
+	fabs	TMPF, TMPF
+	fcmp	TMPF, MAXF
+	beq	KERNEL_F4_FINALIZE_DONE
+	add	INDEX, INDEX, #1
+	ldr	TMPF, [X_COPY], #SZ
+	fabs	TMPF, TMPF
+	fcmp	TMPF, MAXF
+	beq	KERNEL_F4_FINALIZE_DONE
+	add	INDEX, INDEX, #1
+	ldr	TMPF, [X_COPY], #SZ
+	fabs	TMPF, TMPF
+	fcmp	TMPF, MAXF
+	beq	KERNEL_F4_FINALIZE_DONE
+	add	INDEX, INDEX, #1
+KERNEL_F4_FINALIZE_DONE:
+.endm
+
+
+.macro INIT_S
+	lsl	INC_X, INC_X, #2
+	ld1	TMPVF, [X], INC_X
+	mov	Z, #1
+	mov	INDEX, Z
+	fabs	MAXF, TMPF
+.endm
+
+.macro KERNEL_S1
+	ld1	TMPVF, [X], INC_X
+	add	Z, Z, #1
+	fabs	TMPF, TMPF
+	fcmp	TMPF, MAXF
+	fcsel	MAXF, MAXF, TMPF, le
+	csel	INDEX, INDEX, Z, le
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	iamax_kernel_zero
+	cmp	INC_X, xzr
+	ble	iamax_kernel_zero
+
+	PRFM	PLDL1KEEP, [X]
+	mov	X_COPY, X
+
+	cmp	INC_X, #1
+	bne	iamax_kernel_S_BEGIN
+
+iamax_kernel_F_BEGIN:
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	iamax_kernel_F1_INIT
+
+	INIT_F4
+	subs	I, I, #1
+	beq	iamax_kernel_F4_FINALIZE
+
+iamax_kernel_F4:
+	KERNEL_F4
+	subs	I, I, #1
+	bne	iamax_kernel_F4
+
+iamax_kernel_F4_FINALIZE:
+	KERNEL_F4_FINALIZE
+
+iamax_kernel_F1:
+	ands	I, N, #3
+	ble	iamax_kernel_L999
+
+iamax_kernel_F10:
+	KERNEL_F1
+	subs	I, I, #1
+	bne	iamax_kernel_F10
+	b	iamax_kernel_L999
+
+iamax_kernel_F1_INIT:
+	INIT_F1
+	subs	N, N, #1
+	b	iamax_kernel_F1
+
+iamax_kernel_S_BEGIN:
+	INIT_S
+
+	subs	N, N, #1
+	ble	iamax_kernel_L999
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	iamax_kernel_S1
+
+iamax_kernel_S4:
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	iamax_kernel_S4
+
+iamax_kernel_S1:
+	ands	I, N, #3
+	ble	iamax_kernel_L999
+
+iamax_kernel_S10:
+	KERNEL_S1
+	subs	I, I, #1
+	bne	iamax_kernel_S10
+
+iamax_kernel_L999:
+	mov	x0, INDEX
+	ret
+
+iamax_kernel_zero:
+	mov	x0, xzr
+	ret
+
+	EPILOGUE

kernel/arm64/izamax.S

@@ -0,0 +1,151 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define INDEX	x3	/* index of max/min value */
+#define Z	x4	/* vector index */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if defined(USE_MIN)
+#define COND	le
+#else
+#define COND	ge
+#endif
+
+#if !defined(DOUBLE)
+#define MAXF	s0
+#define TMPF	s1
+#define TMPVF	{v1.s}[0]
+#define SZ	4
+#else
+#define MAXF	d0
+#define TMPF	d1
+#define TMPVF	{v1.d}[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro INIT_S
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #3
+	ld1	{v0.2s}, [X], INC_X
+	mov	Z, #1
+	mov	INDEX, Z
+	fabs	v0.2s, v0.2s
+	ext	v1.8b, v0.8b, v0.8b, #4
+	fadd	MAXF, s0, s1
+#else
+	lsl	INC_X, INC_X, #4
+	ld1	{v0.2d}, [X], INC_X
+	mov	Z, #1
+	mov	INDEX, Z
+	fabs	v0.2d, v0.2d
+	faddp	MAXF, v0.2d
+#endif
+.endm
+
+.macro KERNEL_S1
+#if !defined(DOUBLE)
+	ld1	{v1.2s}, [X], INC_X
+	add	Z, Z, #1
+	fabs	v1.2s, v1.2s
+	ext	v2.8b, v1.8b, v1.8b, #4
+	fadd	TMPF, s1, s2
+#else
+	ld1	{v1.2d}, [X], INC_X
+	add	Z, Z, #1
+	fabs	v1.2d, v1.2d
+	faddp	TMPF, v1.2d
+#endif
+	fcmp	MAXF, TMPF
+	fcsel	MAXF, MAXF, TMPF, COND
+	csel	INDEX, INDEX, Z, COND
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	iamax_kernel_zero
+	cmp	INC_X, xzr
+	ble	iamax_kernel_zero
+
+	INIT_S
+
+	subs	N, N, #1
+	ble	iamax_kernel_L999
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	iamax_kernel_S1
+
+iamax_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	iamax_kernel_S4
+
+iamax_kernel_S1:
+
+	ands	I, N, #3
+	ble	iamax_kernel_L999
+
+iamax_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     iamax_kernel_S10
+
+iamax_kernel_L999:
+
+	mov	x0, INDEX
+	ret
+
+iamax_kernel_zero:
+
+	mov	x0, xzr
+	ret
+
+	EPILOGUE

kernel/arm64/rot.S

@@ -0,0 +1,243 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define	Y	x3	/* Y vector address */
+#define	INC_Y	x4	/* Y stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define	C	s0	/* scale input value */
+#define	S	s1	/* scale input value */
+#else
+#define	C	d0	/* scale input value */
+#define	S	d1	/* scale input value */
+#endif
+
+/******************************************************************************/
+
+.macro INIT
+#if !defined(DOUBLE)
+	ins	v0.s[1], v0.s[0]		// [C, C]
+#else
+	ins	v0.d[1], v0.d[0]		// [C, C]
+#endif
+.endm
+
+.macro INIT_F1
+#if !defined(DOUBLE)
+	fneg	s2, S
+	ins	v1.s[1], v2.s[0]		// [-S, S]
+#else
+	fneg	d2, S
+	ins	v1.d[1], v2.d[0]		// [-S, S]
+#endif
+.endm
+
+.macro KERNEL_F1
+#if !defined(DOUBLE)
+	ld1	{v2.s}[0], [X]
+	ld1	{v2.s}[1], [Y]			// [Y, X]
+	ext	v3.8b, v2.8b, v2.8b, #4		// [X, Y]
+	fmul	v4.2s, v2.2s, v0.2s		// [C*Y, C*X]
+	fmla	v4.2s, v3.2s, v1.2s		// [C*Y - S*X, C*X + S*Y]
+	st1	{v4.s}[0], [X], #4
+	st1	{v4.s}[1], [Y], #4
+#else
+	ld1	{v2.d}[0], [X]
+	ld1	{v2.d}[1], [Y]			// [Y, X]
+	ext	v3.16b, v2.16b, v2.16b, #8	// [X, Y]
+	fmul	v4.2d, v2.2d, v0.2d		// [C*Y, C*X]
+	fmla	v4.2d, v3.2d, v1.2d		// [C*Y - S*X, C*X + S*Y]
+	st1	{v4.d}[0], [X], #8
+	st1	{v4.d}[1], [Y], #8
+#endif
+.endm
+
+.macro KERNEL_INIT_F4
+#if !defined(DOUBLE)
+	ins	v0.d[1], v0.d[0]		// [C, C, C, C]
+	ins	v1.s[1], v1.s[0]
+	ins	v1.d[1], v1.d[0]		// [S, S, S, S]
+#else
+	ins	v1.d[1], v1.d[0]		// [S, S]
+#endif
+.endm
+
+.macro KERNEL_F4
+#if !defined(DOUBLE)
+	ld1	{v2.4s}, [X]
+	fmul	v4.4s, v0.4s, v2.4s		// C*X3, C*X2, C*X1, C*X0
+	ld1	{v3.4s}, [Y]
+	fmla	v4.4s, v1.4s, v3.4s		// C*X3+S*Y3, ..., C*X0+S*Y0
+	st1	{v4.4s}, [X], #16
+	fmul	v5.4s, v0.4s, v3.4s		// C*Y3, C*Y2, C*Y1, C*Y0
+	fmls	v5.4s, v1.4s, v2.4s		// C*Y3-S*X3, ..., C*Y0-S*X0
+	st1	{v5.4s}, [Y], #16
+#else // DOUBLE
+	ld1	{v2.2d, v3.2d}, [X]
+	fmul	v6.2d, v0.2d, v2.2d		// C*X1, C*X0
+	fmul	v7.2d, v0.2d, v3.2d		// C*X3, C*X2
+	ld1	{v4.2d, v5.2d}, [Y]
+	fmla	v6.2d, v1.2d, v4.2d		// C*X1+S*Y1, C*X0+S*Y0
+	fmla	v7.2d, v1.2d, v5.2d		// C*X3+S*Y3, C*X2+S*Y2
+	st1	{v6.2d, v7.2d}, [X], #32
+	fmul	v16.2d, v0.2d, v4.2d		// C*Y1, C*Y0
+	fmul	v17.2d, v0.2d, v5.2d		// C*Y3, C*Y2
+	fmls	v16.2d, v1.2d, v2.2d		// C*Y1-S*X1, C*Y0-S*X0
+	fmls	v17.2d, v1.2d, v3.2d		// C*Y3-S*X3, C*Y2-S*X2
+	st1	{v16.2d, v17.2d}, [Y], #32
+	PRFM	PLDL1KEEP, [X, #512]
+	PRFM	PLDL1KEEP, [Y, #512]
+#endif
+.endm
+
+.macro INIT_S
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #2
+	lsl	INC_Y, INC_Y, #2
+#else
+	lsl	INC_X, INC_X, #3
+	lsl	INC_Y, INC_Y, #3
+#endif
+.endm
+
+.macro KERNEL_S1
+#if !defined(DOUBLE)
+	ld1	{v2.s}[0], [X]
+	ld1	{v2.s}[1], [Y]			// [Y, X]
+	ext	v3.8b, v2.8b, v2.8b, #4		// [X, Y]
+	fmul	v4.2s, v2.2s, v0.2s		// [C*Y, C*X]
+	fmla	v4.2s, v3.2s, v1.2s		// [C*Y - S*X, C*X + S*Y]
+	st1	{v4.s}[0], [X], INC_X
+	st1	{v4.s}[1], [Y], INC_Y
+#else
+	ld1	{v2.d}[0], [X]
+	ld1	{v2.d}[1], [Y]			// [Y, X]
+	ext	v3.16b, v2.16b, v2.16b, #8	// [X, Y]
+	fmul	v4.2d, v2.2d, v0.2d		// [C*Y, C*X]
+	fmla	v4.2d, v3.2d, v1.2d		// [C*Y - S*X, C*X + S*Y]
+	st1	{v4.d}[0], [X], INC_X
+	st1	{v4.d}[1], [Y], INC_Y
+#endif
+
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	rot_kernel_L999
+
+	INIT
+
+	cmp	INC_X, #1
+	bne	rot_kernel_S_BEGIN
+	cmp	INC_Y, #1
+	bne	rot_kernel_S_BEGIN
+
+rot_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	rot_kernel_F1
+
+	KERNEL_INIT_F4
+
+rot_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	rot_kernel_F4
+
+rot_kernel_F1:
+
+	ands	I, N, #3
+	ble	rot_kernel_L999
+
+	INIT_F1
+
+rot_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     rot_kernel_F10
+
+	mov	w0, wzr
+	ret
+
+rot_kernel_S_BEGIN:
+
+	INIT_S
+	INIT_F1
+
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	rot_kernel_S1
+
+rot_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	rot_kernel_S4
+
+rot_kernel_S1:
+
+	ands	I, N, #3
+	ble	rot_kernel_L999
+
+
+rot_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     rot_kernel_S10
+
+rot_kernel_L999:
+
+	mov	w0, wzr
+	ret

kernel/arm64/scal.S

@@ -0,0 +1,253 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x3	/* X vector address */
+#define	X_COPY	x5	/* X vector address */
+#define	INC_X	x4	/* X stride */
+#define I	x1	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define	DA	s0	/* scale input value */
+#define	DAV	{v0.s}[0]
+#define TMPF	s1
+#define TMPVF	{v1.s}[0]
+#define SZ	4
+#else
+#define	DA	d0	/* scale input value */
+#define	DAV	{v0.d}[0]
+#define TMPF	d1
+#define TMPVF	{v1.d}[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+
+	ldr	TMPF, [X]
+	fmul	TMPF, TMPF, DA
+	str	TMPF, [X], #SZ
+
+.endm
+
+.macro KERNEL_INIT_F8
+
+#if !defined(DOUBLE)
+	ins	v0.s[1], v0.s[0]
+	ins	v0.s[2], v0.s[0]
+	ins	v0.s[3], v0.s[0]
+#else
+	ins	v0.d[1], v0.d[0]
+#endif
+
+.endm
+
+.macro KERNEL_F8
+#if !defined(DOUBLE)
+	ld1	{v1.4s, v2.4s}, [X]
+	fmul	v1.4s, v1.4s, v0.4s
+	fmul	v2.4s, v2.4s, v0.4s
+	st1	{v1.4s, v2.4s}, [X], #32
+#else // DOUBLE
+	ld1	{v1.2d, v2.2d, v3.2d, v4.2d}, [X]
+	fmul	v1.2d, v1.2d, v0.2d
+	fmul	v2.2d, v2.2d, v0.2d
+	fmul	v3.2d, v3.2d, v0.2d
+	fmul	v4.2d, v4.2d, v0.2d
+	st1	{v1.2d, v2.2d, v3.2d, v4.2d}, [X], #64
+#endif
+	PRFM	PLDL1KEEP, [X, #1024]
+.endm
+
+.macro INIT_S
+
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #2
+#else
+	lsl	INC_X, INC_X, #3
+#endif
+
+.endm
+
+.macro KERNEL_S1
+	ldr	TMPF, [X]
+	fmul	TMPF, TMPF, DA
+	st1	TMPVF, [X], INC_X
+.endm
+
+.macro KERNEL_S4
+#if !defined(DOUBLE)
+	ldr	s1, [X]
+	add	X, X, INC_X
+	fmul	s1, s1, s0
+	str	s1, [X_COPY]
+	add	X_COPY, X_COPY, INC_X
+
+	ldr	s2, [X]
+	add	X, X, INC_X
+	fmul	s2, s2, s0
+	str	s2, [X_COPY]
+	add	X_COPY, X_COPY, INC_X
+
+	ldr	s3, [X]
+	add	X, X, INC_X
+	fmul	s3, s3, s0
+	str	s3, [X_COPY]
+	add	X_COPY, X_COPY, INC_X
+
+	ldr	s4, [X]
+	add	X, X, INC_X
+	fmul	s4, s4, s0
+	str	s4, [X_COPY]
+	add	X_COPY, X_COPY, INC_X
+#else
+	ldr	d1, [X]
+	add	X, X, INC_X
+	fmul	d1, d1, d0
+	str	d1, [X_COPY]
+	add	X_COPY, X_COPY, INC_X
+
+	ldr	d2, [X]
+	add	X, X, INC_X
+	fmul	d2, d2, d0
+	str	d2, [X_COPY]
+	add	X_COPY, X_COPY, INC_X
+
+	ldr	d3, [X]
+	add	X, X, INC_X
+	fmul	d3, d3, d0
+	str	d3, [X_COPY]
+	add	X_COPY, X_COPY, INC_X
+
+	ldr	d4, [X]
+	add	X, X, INC_X
+	fmul	d4, d4, d0
+	str	d4, [X_COPY]
+	add	X_COPY, X_COPY, INC_X
+#endif
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	scal_kernel_L999
+
+	fcmp	DA, #0.0
+	beq	scal_kernel_zero
+
+	cmp	INC_X, #1
+	bne	scal_kernel_S_BEGIN
+
+scal_kernel_F_BEGIN:
+
+	asr	I, N, #3
+	cmp	I, xzr
+	beq	scal_kernel_F1
+
+	KERNEL_INIT_F8
+
+scal_kernel_F8:
+
+	KERNEL_F8
+
+	subs	I, I, #1
+	bne	scal_kernel_F8
+
+scal_kernel_F1:
+
+	ands	I, N, #7
+	ble	scal_kernel_L999
+
+scal_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     scal_kernel_F10
+
+	mov	w0, wzr
+	ret
+
+scal_kernel_S_BEGIN:
+
+	INIT_S
+	mov	X_COPY, X
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	scal_kernel_S1
+
+scal_kernel_S4:
+
+	KERNEL_S4
+
+	subs	I, I, #1
+	bne	scal_kernel_S4
+
+scal_kernel_S1:
+
+	ands	I, N, #3
+	ble	scal_kernel_L999
+
+scal_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     scal_kernel_S10
+
+scal_kernel_L999:
+
+	mov	w0, wzr
+	ret
+
+scal_kernel_zero:
+
+	INIT_S
+
+scal_kernel_Z1:
+
+	st1	DAV, [X], INC_X
+	subs    N, N, #1
+        bne     scal_kernel_Z1
+
+	mov	w0, wzr
+	ret
+
+	EPILOGUE

kernel/arm64/snrm2.S

@@ -0,0 +1,178 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#define TMPF	s6
+#define SSQ	s0
+#define TMPVF	{v6.s}[0]
+#define SZ	4
+
+/******************************************************************************/
+
+.macro INIT_F1
+	ldr	TMPF, [X], #SZ
+	fmul	SSQ, TMPF, TMPF
+.endm
+
+.macro KERNEL_F1
+	ldr	TMPF, [X], #SZ
+	fmul	TMPF, TMPF, TMPF
+	fadd	SSQ, SSQ, TMPF
+.endm
+
+.macro INIT_F4
+	ld1	{v1.4s}, [X], #16
+	fmul	v1.4s, v1.4s, v1.4s
+	ext	v2.16b, v1.16b, v1.16b, #8
+	fadd	v2.2s, v1.2s, v2.2s
+	faddp	SSQ, v2.2s
+.endm
+
+.macro KERNEL_F4
+	ld1	{v1.4s}, [X], #16
+	fmul	v1.4s, v1.4s, v1.4s
+	ext	v2.16b, v1.16b, v1.16b, #8
+	fadd	v2.2s, v1.2s, v2.2s
+	faddp	TMPF, v2.2s
+	fadd	SSQ, SSQ, TMPF
+.endm
+
+.macro INIT_S
+	lsl	INC_X, INC_X, #2
+	ld1	TMPVF, [X], INC_X
+	fmul	SSQ, TMPF, TMPF
+.endm
+
+.macro KERNEL_S1
+	ld1	TMPVF, [X], INC_X
+	fmul	TMPF, TMPF, TMPF
+	fadd	SSQ, SSQ, TMPF
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	nrm2_kernel_zero
+	cmp	INC_X, xzr
+	ble	nrm2_kernel_zero
+	cmp	INC_X, #1
+	bne	nrm2_kernel_S_BEGIN
+
+nrm2_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	nrm2_kernel_F1_INIT
+
+	INIT_F4
+	subs	I, I, #1
+	beq	nrm2_kernel_F1
+
+nrm2_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	nrm2_kernel_F4
+
+nrm2_kernel_F1:
+
+	ands	I, N, #3
+	ble	nrm2_kernel_L999
+
+nrm2_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     nrm2_kernel_F10
+
+	b	nrm2_kernel_L999
+
+nrm2_kernel_F1_INIT:
+	INIT_F1
+	subs	N, N, #1
+	b	nrm2_kernel_F1
+
+nrm2_kernel_S_BEGIN:
+
+	INIT_S
+
+	subs	N, N, #1
+	ble	nrm2_kernel_L999
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	nrm2_kernel_S1
+
+nrm2_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	nrm2_kernel_S4
+
+nrm2_kernel_S1:
+
+	ands	I, N, #3
+	ble	nrm2_kernel_L999
+
+nrm2_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+	bne     nrm2_kernel_S10
+
+nrm2_kernel_L999:
+	fsqrt	SSQ, SSQ
+	ret
+
+nrm2_kernel_zero:
+	fmov	SSQ, wzr
+
+	ret
+
+	EPILOGUE

kernel/arm64/swap.S

@@ -0,0 +1,266 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x3	/* X vector address */
+#define	INC_X	x4	/* X stride */
+#define	Y	x5	/* Y vector address */
+#define	INC_Y	x6	/* Y stride */
+#define I	x1	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define TMP0	s0
+#define TMPV0	{v0.s}[0]
+#define TMP1	s1
+#define TMPV1	{v1.s}[0]
+#define SZ	4
+#else
+#define TMP0	d0
+#define TMPV0	{v0.d}[0]
+#define TMP1	d1
+#define TMPV1	{v1.d}[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+
+#if !defined(COMPLEX)
+	ldr	TMP0, [X]
+	ldr	TMP1, [Y]
+	str	TMP0, [Y], #SZ
+	str	TMP1, [X], #SZ
+#else
+#if !defined(DOUBLE)
+	ld1	{v0.2s}, [X]
+	ld1	{v1.2s}, [Y]
+	st1	{v0.2s}, [Y], #8
+	st1	{v1.2s}, [X], #8
+#else
+	ld1	{v0.2d}, [X]
+	ld1	{v1.2d}, [Y]
+	st1	{v0.2d}, [Y], #16
+	st1	{v1.2d}, [X], #16
+#endif
+#endif
+
+.endm
+
+.macro KERNEL_F8
+
+#if !defined(COMPLEX)
+#if !defined(DOUBLE)
+	ld1	{v0.4s, v1.4s}, [X]
+	ld1	{v2.4s, v3.4s}, [Y]
+	st1	{v0.4s, v1.4s}, [Y], #32
+	st1	{v2.4s, v3.4s}, [X], #32
+#else // DOUBLE
+	ld1	{v0.4s, v1.4s}, [X]
+	ld1	{v2.4s, v3.4s}, [Y]
+	st1	{v0.4s, v1.4s}, [Y], #32
+	st1	{v2.4s, v3.4s}, [X], #32
+	ld1	{v0.4s, v1.4s}, [X]
+	ld1	{v2.4s, v3.4s}, [Y]
+	st1	{v0.4s, v1.4s}, [Y], #32
+	st1	{v2.4s, v3.4s}, [X], #32
+#endif
+#else // COMPLEX
+#if !defined(DOUBLE)
+	ld1	{v0.4s, v1.4s}, [X]
+	ld1	{v2.4s, v3.4s}, [Y]
+	st1	{v0.4s, v1.4s}, [Y], #32
+	st1	{v2.4s, v3.4s}, [X], #32
+	ld1	{v0.4s, v1.4s}, [X]
+	ld1	{v2.4s, v3.4s}, [Y]
+	st1	{v0.4s, v1.4s}, [Y], #32
+	st1	{v2.4s, v3.4s}, [X], #32
+#else // DOUBLE
+	ld1	{v0.4s, v1.4s}, [X]
+	ld1	{v2.4s, v3.4s}, [Y]
+	st1	{v0.4s, v1.4s}, [Y], #32
+	st1	{v2.4s, v3.4s}, [X], #32
+	ld1	{v0.4s, v1.4s}, [X]
+	ld1	{v2.4s, v3.4s}, [Y]
+	st1	{v0.4s, v1.4s}, [Y], #32
+	st1	{v2.4s, v3.4s}, [X], #32
+	ld1	{v0.4s, v1.4s}, [X]
+	ld1	{v2.4s, v3.4s}, [Y]
+	st1	{v0.4s, v1.4s}, [Y], #32
+	st1	{v2.4s, v3.4s}, [X], #32
+	ld1	{v0.4s, v1.4s}, [X]
+	ld1	{v2.4s, v3.4s}, [Y]
+	st1	{v0.4s, v1.4s}, [Y], #32
+	st1	{v2.4s, v3.4s}, [X], #32
+#endif
+#endif
+
+.endm
+
+.macro INIT_S
+
+#if !defined(COMPLEX)
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #2
+	lsl	INC_Y, INC_Y, #2
+#else
+	lsl	INC_X, INC_X, #3
+	lsl	INC_Y, INC_Y, #3
+#endif
+#else
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #3
+	lsl	INC_Y, INC_Y, #3
+#else
+	lsl	INC_X, INC_X, #4
+	lsl	INC_Y, INC_Y, #4
+#endif
+#endif
+
+.endm
+
+.macro KERNEL_S1
+
+#if !defined(COMPLEX)
+#if !defined(DOUBLE)
+	ldr	w10, [X]
+	ldr	w11, [Y]
+	str	w10, [Y]
+	str	w11, [X]
+#else
+	ldr	x10, [X]
+	ldr	x11, [Y]
+	str	x10, [Y]
+	str	x11, [X]
+#endif
+#else
+#if !defined(DOUBLE)
+	ldr	x10, [X]
+	ldr	x11, [Y]
+	str	x10, [Y]
+	str	x11, [X]
+#else
+	ldr	x10, [X]
+	ldr	x11, [Y]
+	str	x10, [Y]
+	str	x11, [X]
+
+	ldr	x12, [X, #8]
+	ldr	x13, [Y, #8]
+	str	x12, [Y, #8]
+	str	x13, [X, #8]
+#endif
+#endif
+	add	Y, Y, INC_Y
+	add	X, X, INC_X
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	swap_kernel_L999
+
+	cmp	INC_X, #1
+	bne	swap_kernel_S_BEGIN
+	cmp	INC_Y, #1
+	bne	swap_kernel_S_BEGIN
+
+swap_kernel_F_BEGIN:
+
+	asr	I, N, #3
+	cmp	I, xzr
+	beq	swap_kernel_F1
+
+swap_kernel_F8:
+
+	KERNEL_F8
+
+	subs	I, I, #1
+	bne	swap_kernel_F8
+
+swap_kernel_F1:
+
+	ands	I, N, #7
+	ble	swap_kernel_L999
+
+swap_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+	bne     swap_kernel_F10
+
+	b	swap_kernel_L999
+
+
+swap_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	swap_kernel_S1
+
+swap_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	swap_kernel_S4
+
+swap_kernel_S1:
+
+	ands	I, N, #3
+	ble	swap_kernel_L999
+
+swap_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     swap_kernel_S10
+
+swap_kernel_L999:
+
+	mov	w0, wzr
+	ret
+
+	EPILOGUE

kernel/arm64/zamax.S

@@ -0,0 +1,273 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if defined(USE_MIN)
+#define COND	le
+#else
+#define COND	ge
+#endif
+
+#if !defined(DOUBLE)
+#define REG0	wzr
+#define MAXF	s0
+#define TMPF	s1
+#define TMPVF	{v1.s}[0]
+#define SZ	4
+#else
+#define REG0	xzr
+#define MAXF	d0
+#define TMPF	d1
+#define TMPVF	{v1.d}[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro INIT_F1
+#if !defined(DOUBLE)
+	ld1	{v0.2s}, [X], #8
+	fabs	v0.2s, v0.2s
+	ext	v1.8b, v0.8b, v0.8b, #4
+	fadd	MAXF, s0, s1
+#else
+	ld1	{v0.2d}, [X], #16
+	fabs	v0.2d, v0.2d
+	faddp	MAXF, v0.2d
+#endif
+.endm
+
+.macro KERNEL_F1
+#if !defined(DOUBLE)
+	ld1	{v1.2s}, [X], #8
+	fabs	v1.2s, v1.2s
+	ext	v2.8b, v1.8b, v1.8b, #4
+	fadd	TMPF, s1, s2
+#else
+	ld1	{v1.2d}, [X], #16
+	fabs	v1.2d, v1.2d
+	faddp	TMPF, v1.2d
+#endif
+	fcmp	MAXF, TMPF
+	fcsel	MAXF, MAXF, TMPF, COND
+.endm
+
+.macro INIT_F4
+#if !defined(DOUBLE)
+	ld2	{v0.4s,v1.4s}, [X], #32
+	fabs	v0.4s, v0.4s			// [X6, X4, X2, X0]
+	fabs	v1.4s, v1.4s			// [X7, X5, X3, X1]
+	fadd	v0.4s, v0.4s, v1.4s		// [X7+X6, X5+X4, X3+X2, X1+X0]
+#if defined(USE_MIN)
+	fminv	MAXF, v0.4s
+#else
+	fmaxv	MAXF, v0.4s
+#endif
+#else // DOUBLE
+	ld4	{v0.2d,v1.2d,v2.2d,v3.2d}, [X], #64
+	fabs	v0.2d, v0.2d
+	fabs	v1.2d, v1.2d
+	fabs	v2.2d, v2.2d
+	fabs	v3.2d, v3.2d
+	fadd	v0.2d, v0.2d, v1.2d
+	fadd	v2.2d, v2.2d, v3.2d
+#if defined(USE_MIN)
+	fmin	v0.2d, v0.2d, v2.2d
+	fminp	MAXF, v0.2d
+#else
+	fmax	v0.2d, v0.2d, v2.2d
+	fmaxp	MAXF, v0.2d
+#endif
+#endif
+.endm
+
+.macro KERNEL_F4
+#if !defined(DOUBLE)
+	ld2	{v1.4s,v2.4s}, [X], #32
+	fabs	v1.4s, v1.4s			// [X6, X4, X2, X0]
+	fabs	v2.4s, v2.4s			// [X7, X5, X3, X1]
+	fadd	v1.4s, v1.4s, v2.4s		// [X7+X6, X5+X4, X3+X2, X1+X0]
+#if defined(USE_MIN)
+	fminv	TMPF, v1.4s
+#else
+	fmaxv	TMPF, v1.4s
+#endif
+#else // DOUBLE
+	ld4	{v1.2d,v2.2d,v3.2d,v4.2d}, [X], #64
+	fabs	v1.2d, v1.2d
+	fabs	v2.2d, v2.2d
+	fabs	v3.2d, v3.2d
+	fabs	v4.2d, v4.2d
+	fadd	v1.2d, v1.2d, v2.2d
+	fadd	v3.2d, v3.2d, v4.2d
+#if defined(USE_MIN)
+	fmin	v1.2d, v1.2d, v3.2d
+	fminp	MAXF, v1.2d
+#else
+	fmax	v1.2d, v1.2d, v3.2d
+	fmaxp	MAXF, v1.2d
+#endif
+#endif
+	fcmp	MAXF, TMPF
+	fcsel	MAXF, MAXF, TMPF, COND
+.endm
+
+.macro INIT_S
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #3
+	ld1	{v0.2s}, [X], INC_X
+	fabs	v0.2s, v0.2s
+	ext	v1.8b, v0.8b, v0.8b, #4
+	fadd	MAXF, s0, s1
+#else
+	lsl	INC_X, INC_X, #4
+	ld1	{v0.2d}, [X], INC_X
+	fabs	v0.2d, v0.2d
+	faddp	MAXF, v0.2d
+#endif
+.endm
+
+.macro KERNEL_S1
+#if !defined(DOUBLE)
+	ld1	{v1.2s}, [X], INC_X
+	fabs	v1.2s, v1.2s
+	ext	v2.8b, v1.8b, v1.8b, #4
+	fadd	TMPF, s1, s2
+#else
+	ld1	{v1.2d}, [X], INC_X
+	fabs	v1.2d, v1.2d
+	faddp	TMPF, v1.2d
+#endif
+	fcmp	MAXF, TMPF
+	fcsel	MAXF, MAXF, TMPF, COND
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	amax_kernel_zero
+	cmp	INC_X, xzr
+	ble	amax_kernel_zero
+
+	cmp	INC_X, #1
+	bne	amax_kernel_S_BEGIN
+
+amax_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	amax_kernel_F1_INIT
+
+	INIT_F4
+	subs	I, I, #1
+	beq	amax_kernel_F1
+
+amax_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	amax_kernel_F4
+
+amax_kernel_F1:
+
+	ands	I, N, #3
+	ble	amax_kernel_L999
+
+amax_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     amax_kernel_F10
+
+	ret
+
+amax_kernel_F1_INIT:
+
+	INIT_F1
+	subs	N, N, #1
+	b	amax_kernel_F1
+
+amax_kernel_S_BEGIN:
+
+	INIT_S
+
+	subs	N, N, #1
+	ble	amax_kernel_L999
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	amax_kernel_S1
+
+amax_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	amax_kernel_S4
+
+amax_kernel_S1:
+
+	ands	I, N, #3
+	ble	amax_kernel_L999
+
+amax_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     amax_kernel_S10
+
+amax_kernel_L999:
+
+	ret
+
+amax_kernel_zero:
+
+	fmov	MAXF, REG0
+	ret
+
+	EPILOGUE

kernel/arm64/zasum.S

@@ -0,0 +1,164 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#define REG0	xzr
+#define SUMF	d0
+#define TMPF	d1
+#define TMPVF	{v1.d}[0]
+#define SZ	8
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+	ld1	{v1.2d}, [X], #16
+	fabs	v1.2d, v1.2d
+	faddp	TMPF, v1.2d
+	fadd	SUMF, SUMF, TMPF
+.endm
+
+.macro KERNEL_F4
+	ld1	{v1.2d, v2.2d, v3.2d, v4.2d}, [X], #64
+	fabs	v1.2d, v1.2d
+	fabs	v2.2d, v2.2d
+	fabs	v3.2d, v3.2d
+	fabs	v4.2d, v4.2d
+
+	fadd	v1.2d, v1.2d, v2.2d
+	fadd	v3.2d, v3.2d, v4.2d
+
+	fadd	v0.2d, v0.2d, v1.2d
+	fadd	v0.2d, v0.2d, v3.2d
+
+	PRFM	PLDL1KEEP, [X, #1024]
+.endm
+
+.macro KERNEL_F4_FINALIZE
+	faddp	SUMF, v0.2d
+.endm
+
+.macro INIT_S
+	lsl	INC_X, INC_X, #4
+.endm
+
+.macro KERNEL_S1
+	ld1	{v1.2d}, [X], INC_X
+	fabs	v1.2d, v1.2d
+	faddp	TMPF, v1.2d
+	fadd	SUMF, SUMF, TMPF
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	fmov	SUMF, REG0
+
+	cmp	N, xzr
+	ble	asum_kernel_L999
+	cmp	INC_X, xzr
+	ble	asum_kernel_L999
+
+	cmp	INC_X, #1
+	bne	asum_kernel_S_BEGIN
+
+asum_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	asum_kernel_F1
+
+asum_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	asum_kernel_F4
+
+	KERNEL_F4_FINALIZE
+
+asum_kernel_F1:
+
+	ands	I, N, #3
+	ble	asum_kernel_L999
+
+asum_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     asum_kernel_F10
+
+asum_kernel_L999:
+	ret
+
+asum_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	asum_kernel_S1
+
+asum_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	asum_kernel_S4
+
+asum_kernel_S1:
+
+	ands	I, N, #3
+	ble	asum_kernel_L999
+
+asum_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     asum_kernel_S10
+
+	ret
+
+	EPILOGUE

kernel/arm64/zaxpy.S

@@ -0,0 +1,301 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x3	/* X vector address */
+#define	INC_X	x4	/* X stride */
+#define	Y	x5	/* Y vector address */
+#define	INC_Y	x6	/* Y stride */
+#define I	x1	/* loop variable */
+#define Y_COPY	x7	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define	DA_R	s0	/* scale input value */
+#define	DA_I	s1	/* scale input value */
+#define TMPX	v2.2s
+#define TMPY	v3.2s
+#define SZ	4
+#else
+#define	DA_R	d0	/* scale input value */
+#define	DA_I	d1	/* scale input value */
+#define TMPX	v2.2d
+#define TMPY	v3.2d
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro INIT
+
+#if !defined(CONJ)
+#if !defined(DOUBLE)
+	ins	v0.s[1], v0.s[0]		// v0 = DA_R, DA_R 
+	fneg	s2, DA_I
+	ins	v1.s[1], v2.s[0]		// v1 = -DA_I, DA_I 
+	ext	v1.8b, v1.8b, v1.8b, #4		// v1 = DA_I, -DA_I
+#else
+	ins	v0.d[1], v0.d[0]		// v0 = DA_R, DA_R 
+	fneg	d2, DA_I
+	ins	v1.d[1], v2.d[0]		// v1 = -DA_I, DA_I 
+	ext	v1.16b, v1.16b, v1.16b, #8	// v1 = DA_I, -DA_I
+#endif
+#else
+#if !defined(DOUBLE)
+	fneg	s2, DA_R
+	ins	v0.s[1], v2.s[0]		// v0 = -DA_R, DA_R 
+	ins	v1.s[1], v1.s[0]		// v1 = DA_I, DA_I 
+#else
+	fneg	d2, DA_R
+	ins	v0.d[1], v2.d[0]		// v0 = -DA_R, DA_R 
+	ins	v1.d[1], v1.d[0]		// v1 = DA_I, DA_I 
+#endif
+#endif
+
+.endm
+
+.macro KERNEL_F1
+
+#if !defined(DOUBLE)
+	ld1	{v2.2s}, [X], #8		// V2 = X[ix+1], X[ix]; X += 2
+	ld1	{v3.2s}, [Y]			// V3 = Y[iy+1], Y[iy]
+	ext	v4.8b, v2.8b, v2.8b, #4		// V4 = X[ix], X[ix+1]
+	fmla	v3.2s, v0.2s, v2.2s		// Y[iy]   += DA_R * X[ix]
+						// Y[iy+1] += +-DA_R * X[ix+1]
+	fmla	v3.2s, v1.2s, v4.2s		// Y[iy]   += +-DA_I * X[ix+1]
+						// Y[iy+1] += DA_I * X[ix]
+	st1	{v3.2s}, [Y], #8
+#else
+	ld1	{v2.2d}, [X], #16		// V2 = X[ix+1], X[ix]; X += 2
+	ld1	{v3.2d}, [Y]			// V3 = Y[iy+1], Y[iy]
+	ext	v4.16b, v2.16b, v2.16b, #8	// V4 = X[ix], X[ix+1]
+	fmla	v3.2d, v0.2d, v2.2d		// Y[iy]   += DA_R * X[ix]
+						// Y[iy+1] += +-DA_R * X[ix+1]
+	fmla	v3.2d, v1.2d, v4.2d		// Y[iy]   += +-DA_I * X[ix+1]
+						// Y[iy+1] += DA_I * X[ix]
+	st1	{v3.2d}, [Y], #16
+#endif
+
+.endm
+
+.macro KERNEL_INIT_F4
+
+#if !defined(DOUBLE)
+	// Replicate the lower 2 floats into the upper 2 slots
+	ins	v0.d[1], v0.d[0]		// v0 = DA_R, DA_R, DA_R, DA_R
+	ins	v1.d[1], v1.d[0]		// v1 = DA_I, DA_I, DA_I, DA_I
+#endif
+
+.endm
+
+.macro KERNEL_F4
+
+#if !defined(DOUBLE)
+	ld1	{v2.4s,v3.4s}, [X], #32		// V2 = X[3], X[2], X[1], X[0]
+						// V3 = X[7], X[6], X[5], X[4]
+	ext	v6.8b, v2.8b, v2.8b, #4		// V6 =  -  ,  -  , X[0], X[1]
+	ins	v6.s[2], v2.s[3]		// V6 =  -  , X[3], X[0], X[1]
+	ins	v6.s[3], v2.s[2]		// V6 = X[2], X[3], X[0], X[1]
+
+	ld1	{v4.4s,v5.4s}, [Y]		// V4 = Y[3], Y[2], Y[1], Y[0]
+						// V5 = Y[7], Y[6], Y[5], Y[4]
+
+	ext	v7.8b, v3.8b, v3.8b, #4		// V7 =  -  ,  -  , X[4], X[5]
+	ins	v7.s[2], v3.s[3]		// V7 =  -  , X[7], X[4], X[5]
+	ins	v7.s[3], v3.s[2]		// V7 = X[6], X[7], X[4], X[5]
+
+	fmla	v4.4s, v0.4s, v2.4s		// Y[iy]   += DA_R * X[ix]
+						// Y[iy+1] += +-DA_R * X[ix+1]
+	fmla	v4.4s, v1.4s, v6.4s		// Y[iy]   += +-DA_I * X[ix+1]
+						// Y[iy+1] += DA_I * X[ix]
+	st1	{v4.4s}, [Y], #16
+
+	fmla	v5.4s, v0.4s, v3.4s		// Y[iy]   += DA_R * X[ix]
+	fmla	v5.4s, v1.4s, v7.4s		// Y[iy]   += +-DA_I * X[ix+1]
+						// Y[iy+1] += +-DA_R * X[ix+1]
+						// Y[iy+1] += DA_I * X[ix]
+	st1	{v5.4s}, [Y], #16
+#else // DOUBLE
+	ld1	{v2.2d,v3.2d}, [X], #32			// CX0, CX1, CX2, CX3
+	ext	v20.16b, v2.16b, v2.16b, #8		// X[ix], X[ix+1]
+	ext	v21.16b, v3.16b, v3.16b, #8		// X[ix], X[ix+1]
+
+	ld1	{v4.2d,v5.2d}, [X], #32			// CX0, CX1, CX2, CX3
+	ext	v22.16b, v4.16b, v4.16b, #8		// X[ix], X[ix+1]
+	ext	v23.16b, v5.16b, v5.16b, #8		// X[ix], X[ix+1]
+
+	ld1	{v16.2d,v17.2d}, [Y_COPY], #32	// CY0, CY1, CY2, CY3  
+
+	fmla	v16.2d, v0.2d, v2.2d
+	fmla	v17.2d, v0.2d, v3.2d
+
+	ld1	{v18.2d,v19.2d}, [Y_COPY], #32	// CY0, CY1, CY2, CY3  
+
+	fmla	v16.2d, v1.2d, v20.2d
+	fmla	v17.2d, v1.2d, v21.2d
+	st1	{v16.2d,v17.2d}, [Y], #32
+
+	fmla	v18.2d, v0.2d, v4.2d
+	fmla	v19.2d, v0.2d, v5.2d
+	fmla	v18.2d, v1.2d, v22.2d
+	fmla	v19.2d, v1.2d, v23.2d
+	st1	{v18.2d,v19.2d}, [Y], #32
+#endif
+	PRFM	PLDL1KEEP, [X, #512]
+	PRFM	PLDL1KEEP, [Y, #512]
+.endm
+
+.macro INIT_S
+
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #3
+	lsl	INC_Y, INC_Y, #3
+#else
+	lsl	INC_X, INC_X, #4
+	lsl	INC_Y, INC_Y, #4
+#endif
+
+.endm
+
+.macro KERNEL_S1
+
+#if !defined(DOUBLE)
+	ld1	{v2.2s}, [X], INC_X		// V2 = X[ix+1], X[ix]; X += 2
+	ld1	{v3.2s}, [Y]			// V3 = Y[iy+1], Y[iy]
+	ext	v4.8b, v2.8b, v2.8b, #4		// V4 = X[ix], X[ix+1]
+	fmla	v3.2s, v0.2s, v2.2s		// Y[iy]   += DA_R * X[ix]
+						// Y[iy+1] += +-DA_R * X[ix+1]
+	fmla	v3.2s, v1.2s, v4.2s		// Y[iy]   += +-DA_I * X[ix+1]
+						// Y[iy+1] += DA_I * X[ix]
+	st1	{v3.2s}, [Y], INC_Y
+#else
+	ld1	{v2.2d}, [X], INC_X		// V2 = X[ix+1], X[ix]; X += 2
+	ld1	{v3.2d}, [Y]			// V3 = Y[iy+1], Y[iy]
+	ext	v4.16b, v2.16b, v2.16b, #8	// V4 = X[ix], X[ix+1]
+	fmla	v3.2d, v0.2d, v2.2d		// Y[iy]   += DA_R * X[ix]
+						// Y[iy+1] += +-DA_R * X[ix+1]
+	fmla	v3.2d, v1.2d, v4.2d		// Y[iy]   += +-DA_I * X[ix+1]
+						// Y[iy+1] += DA_I * X[ix]
+	st1	{v3.2d}, [Y], INC_Y
+#endif
+
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	zaxpy_kernel_L999
+
+	mov	Y_COPY, Y
+
+	fcmp	DA_R, #0.0
+	bne	.L1
+	fcmp	DA_I, #0.0
+	beq	zaxpy_kernel_L999
+
+.L1:
+	INIT
+
+	cmp	INC_X, #1
+	bne	zaxpy_kernel_S_BEGIN
+	cmp	INC_Y, #1
+	bne	zaxpy_kernel_S_BEGIN
+
+zaxpy_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	zaxpy_kernel_F1
+
+	KERNEL_INIT_F4
+
+zaxpy_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	zaxpy_kernel_F4
+
+zaxpy_kernel_F1:
+
+	ands	I, N, #3
+	ble	zaxpy_kernel_L999
+
+zaxpy_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     zaxpy_kernel_F10
+
+	mov	w0, wzr
+	ret
+
+zaxpy_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	zaxpy_kernel_S1
+
+zaxpy_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	zaxpy_kernel_S4
+
+zaxpy_kernel_S1:
+
+	ands	I, N, #3
+	ble	zaxpy_kernel_L999
+
+zaxpy_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     zaxpy_kernel_S10
+
+zaxpy_kernel_L999:
+
+	mov	w0, wzr
+	ret

kernel/arm64/zdot.S

@@ -0,0 +1,302 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define	Y	x3	/* Y vector address */
+#define	INC_Y	x4	/* Y stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#if !defined(DSDOT)
+#define REG0	wzr
+#define DOTF	s0
+#else // DSDOT
+#define REG0	xzr
+#define DOTF	d0
+#endif
+#define DOTI	s1
+#define TMPX	s2
+#define LD1VX	{v2.s}[0]
+#define TMPY	s3
+#define LD1VY	{v3.s}[0]
+#define TMPVY	v3.s[0]
+#define SZ	4
+#else
+#define REG0	xzr
+#define DOTF	d0
+#define DOTI	d1
+#define TMPX	d2
+#define LD1VX	{v2.d}[0]
+#define TMPY	d3
+#define LD1VY	{v3.d}[0]
+#define TMPVY	v3.d[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+
+#if !defined(DOUBLE)
+	ld1	{v2.2s}, [X], #8		// V2 = X[ix+1], X[ix]; X += 2
+	ld1	{v3.2s}, [Y], #8		// V3 = Y[iy+1], Y[iy]; Y += 2
+	ins	v4.s[0], v2.s[1]		// V4 = X[ix+1]
+#if !defined(CONJ)
+	fmla	DOTF, s2, v3.s[0]		// dot[0] += X[ix] * Y[iy]
+	fmls	DOTF, s4, v3.s[1]		// dot[0] -= X[ix+1] * Y[iy+1]
+	fmla	DOTI, s4, v3.s[0]		// dot[1] += X[ix+1] * Y[iy]
+	fmla	DOTI, s2, v3.s[1]		// dot[1] += X[ix] * Y[iy+1]
+#else
+	fmla	DOTF, s2, v3.s[0]		// dot[0] += X[ix] * Y[iy]
+	fmla	DOTF, s4, v3.s[1]		// dot[0] += X[ix+1] * Y[iy+1]
+	fmls	DOTI, s4, v3.s[0]		// dot[1] -= X[ix+1] * Y[iy]
+	fmla	DOTI, s2, v3.s[1]		// dot[1] += X[ix] * Y[iy+1]
+#endif
+#else // DOUBLE
+	ld1	{v2.2d}, [X], #16		// V2 = X[ix+1], X[ix]; X += 2
+	ld1	{v3.2d}, [Y], #16		// V3 = Y[iy+1], Y[iy]; Y += 2
+	ins	v4.d[0], v2.d[1]		// V4 = X[ix+1]
+#if !defined(CONJ)
+	fmla	DOTF, d2, v3.d[0]		// dot[0] += X[ix] * Y[iy]
+	fmls	DOTF, d4, v3.d[1]		// dot[0] -= X[ix+1] * Y[iy+1]
+	fmla	DOTI, d4, v3.d[0]		// dot[1] += X[ix+1] * Y[iy]
+	fmla	DOTI, d2, v3.d[1]		// dot[1] += X[ix] * Y[iy+1]
+#else
+	fmla	DOTF, d2, v3.d[0]		// dot[0] += X[ix] * Y[iy]
+	fmla	DOTF, d4, v3.d[1]		// dot[0] += X[ix+1] * Y[iy+1]
+	fmls	DOTI, d4, v3.d[0]		// dot[1] -= X[ix+1] * Y[iy]
+	fmla	DOTI, d2, v3.d[1]		// dot[1] += X[ix] * Y[iy+1]
+#endif
+#endif
+
+.endm
+
+
+.macro KERNEL_F4
+
+#if !defined(DOUBLE)
+	ld2	{v2.4s, v3.4s}, [X], #32	// V2 = X[ix+1], X[ix]; X += 2
+	ld2	{v4.4s, v5.4s}, [Y], #32	// V2 = X[ix+1], X[ix]; X += 2
+
+	fmla	v0.4s, v2.4s, v4.4s		// dot[0] += X[ix] * Y[iy]
+	fmla	v1.4s, v2.4s, v5.4s		// dot[1] += X[ix] * Y[iy+1]
+	PRFM	PLDL1KEEP, [X, #1024]
+	PRFM	PLDL1KEEP, [Y, #1024]
+#if !defined(CONJ)
+	fmls	v0.4s, v3.4s, v5.4s		// dot[0] -= X[ix+1] * Y[iy+1]
+	fmla	v1.4s, v3.4s, v4.4s		// dot[1] += X[ix+1] * Y[iy]
+#else
+	fmla	v0.4s, v3.4s, v5.4s		// dot[0] += X[ix+1] * Y[iy+1]
+	fmls	v1.4s, v3.4s, v4.4s		// dot[1] -= X[ix+1] * Y[iy]
+#endif
+#else // DOUBLE
+	ld2	{v2.2d, v3.2d}, [X], #32	// V2 = X[ix+1], X[ix]; X += 2
+	ld2	{v16.2d, v17.2d}, [Y], #32
+
+	fmla	v0.2d, v2.2d, v16.2d		// dot[0] += X[ix] * Y[iy]
+	fmla	v1.2d, v2.2d, v17.2d		// dot[1] += X[ix] * Y[iy+1]
+	ld2	{v4.2d, v5.2d}, [X], #32
+	ld2	{v18.2d, v19.2d}, [Y], #32
+	fmla	v0.2d, v4.2d, v18.2d		// dot[1] += X[ix] * Y[iy+1]
+	fmla	v1.2d, v4.2d, v19.2d		// dot[1] += X[ix] * Y[iy+1]
+	PRFM	PLDL1KEEP, [X, #1024]
+	PRFM	PLDL1KEEP, [Y, #1024]
+#if !defined(CONJ)
+	fmls	v0.2d, v3.2d, v17.2d		// dot[0] -= X[ix+1] * Y[iy+1]
+	fmls	v20.2d, v5.2d, v19.2d		// dot[0] -= X[ix+1] * Y[iy+1]
+	fmla	v1.2d, v3.2d, v16.2d		// dot[1] += X[ix+1] * Y[iy]
+	fmla	v21.2d, v5.2d, v18.2d		// dot[1] += X[ix+1] * Y[iy]
+#else
+	fmla	v0.2d, v3.2d, v17.2d		// dot[0] += X[ix+1] * Y[iy+1]
+	fmla	v20.2d, v5.2d, v19.2d		// dot[0] += X[ix+1] * Y[iy+1]
+	fmls	v1.2d, v3.2d, v16.2d		// dot[1] -= X[ix+1] * Y[iy]
+	fmls	v21.2d, v5.2d, v18.2d		// dot[1] -= X[ix+1] * Y[iy]
+#endif
+#endif
+
+.endm
+
+.macro KERNEL_F4_FINALIZE
+#if !defined(DOUBLE)
+	ext	v2.16b, v0.16b, v0.16b, #8
+	fadd	v0.2s, v0.2s, v2.2s
+	faddp	DOTF, v0.2s
+	ext	v3.16b, v1.16b, v1.16b, #8
+	fadd	v1.2s, v1.2s, v3.2s
+	faddp	DOTI, v1.2s
+#else
+	fadd	v0.2d, v0.2d, v20.2d
+	faddp	DOTF, v0.2d
+	fadd	v1.2d, v1.2d, v21.2d
+	faddp	DOTI, v1.2d
+#endif
+.endm
+
+.macro INIT_S
+
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #3
+	lsl	INC_Y, INC_Y, #3
+#else
+	lsl	INC_X, INC_X, #4
+	lsl	INC_Y, INC_Y, #4
+#endif
+
+.endm
+
+.macro KERNEL_S1
+#if !defined(DOUBLE)
+	ld1	{v2.2s}, [X], INC_X		// V2 = X[ix+1], X[ix]; X += 2
+	ld1	{v3.2s}, [Y], INC_Y		// V3 = Y[iy+1], Y[iy]; Y += 2
+	ext	v4.8b, v2.8b, v2.8b, #4		// V4 = X[ix], X[ix+1]
+#if !defined(CONJ)
+	fmla	DOTF, s2, v3.s[0]		// dot[0] += X[ix] * Y[iy]
+	fmls	DOTF, s4, v3.s[1]		// dot[0] -= X[ix+1] * Y[iy+1]
+	fmla	DOTI, s4, v3.s[0]		// dot[1] += X[ix+1] * Y[iy]
+	fmla	DOTI, s2, v3.s[1]		// dot[1] += X[ix] * Y[iy+1]
+#else
+	fmla	DOTF, s2, v3.s[0]		// dot[0] += X[ix] * Y[iy]
+	fmla	DOTF, s4, v3.s[1]		// dot[0] += X[ix+1] * Y[iy+1]
+	fmls	DOTI, s4, v3.s[0]		// dot[1] -= X[ix+1] * Y[iy]
+	fmla	DOTI, s2, v3.s[1]		// dot[1] += X[ix] * Y[iy+1]
+#endif
+#else // DOUBLE
+	ld1	{v2.2d}, [X], INC_X		// V2 = X[ix+1], X[ix]; X += 2
+	ld1	{v3.2d}, [Y], INC_Y		// V3 = Y[iy+1], Y[iy]; Y += 2
+	ext	v4.16b, v2.16b, v2.16b, #8	// V4 = X[ix], X[ix+1]
+#if !defined(CONJ)
+	fmla	DOTF, d2, v3.d[0]		// dot[0] += X[ix] * Y[iy]
+	fmls	DOTF, d4, v3.d[1]		// dot[0] -= X[ix+1] * Y[iy+1]
+	fmla	DOTI, d4, v3.d[0]		// dot[1] += X[ix+1] * Y[iy]
+	fmla	DOTI, d2, v3.d[1]		// dot[1] += X[ix] * Y[iy+1]
+#else
+	fmla	DOTF, d2, v3.d[0]		// dot[0] += X[ix] * Y[iy]
+	fmla	DOTF, d4, v3.d[1]		// dot[0] += X[ix+1] * Y[iy+1]
+	fmls	DOTI, d4, v3.d[0]		// dot[1] -= X[ix+1] * Y[iy]
+	fmla	DOTI, d2, v3.d[1]		// dot[1] += X[ix] * Y[iy+1]
+#endif
+#endif
+
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	fmov	DOTF, REG0
+	fmov	DOTI, DOTF
+#if !defined(DOUBLE)
+	fmov	s20, DOTF
+	fmov	s21, DOTI
+#else
+	fmov	d20, DOTF
+	fmov	d21, DOTI
+#endif
+
+	cmp	N, xzr
+	ble	dot_kernel_L999
+
+	cmp	INC_X, #1
+	bne	dot_kernel_S_BEGIN
+	cmp	INC_Y, #1
+	bne	dot_kernel_S_BEGIN
+
+dot_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	dot_kernel_F1
+
+dot_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	dot_kernel_F4
+
+	KERNEL_F4_FINALIZE
+
+dot_kernel_F1:
+
+	ands	I, N, #3
+	ble	dot_kernel_L999
+
+dot_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     dot_kernel_F10
+
+	ret
+
+dot_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	dot_kernel_S1
+
+dot_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	dot_kernel_S4
+
+dot_kernel_S1:
+
+	ands	I, N, #3
+	ble	dot_kernel_L999
+
+dot_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     dot_kernel_S10
+
+dot_kernel_L999:
+
+	ret
+
+	EPILOGUE

kernel/arm64/zgemv_n.S

@@ -0,0 +1,514 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	M	x0	/* Y vector length */
+#define	N	x1	/* X vector length */
+#define	A	x3	/* A vector address */
+#define	LDA	x4	/* A stride */
+#define	X	x5	/* X vector address */
+#define	INC_X	x6	/* X stride */
+#define	Y	x7	/* Y vector address */
+#define	INC_Y	x2	/* Y stride */
+#define	A_PTR	x9	/* loop A vector address */
+#define	Y_IPTR	x10	/* loop Y vector address */
+#define J	x11	/* loop variable */
+#define I	x12	/* loop variable */
+#define	Y_OPTR	x13	/* loop Y vector address */
+#define	X_PTR	x14	/* loop X vector address */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define ALPHA_R		s0
+#define ALPHA_I		s1
+#define ALPHA_R_COPY	s7
+#define ALPHA_I_COPY	s8
+#define SHZ		3
+#else
+#define ALPHA_R		d0
+#define ALPHA_I		d1
+#define ALPHA_R_COPY	d7
+#define ALPHA_I_COPY	d8
+#define SHZ		4
+#endif
+
+/******************************************************************************/
+
+.macro SAVE_REGS
+	add	sp, sp, #-(11 * 16)
+	stp	d8, d9, [sp, #(0 * 16)]
+	stp	d10, d11, [sp, #(1 * 16)]
+	stp	d12, d13, [sp, #(2 * 16)]
+	stp	d14, d15, [sp, #(3 * 16)]
+	stp	d16, d17, [sp, #(4 * 16)]
+	stp	x18, x19, [sp, #(5 * 16)]
+	stp	x20, x21, [sp, #(6 * 16)]
+	stp	x22, x23, [sp, #(7 * 16)]
+	stp	x24, x25, [sp, #(8 * 16)]
+	stp	x26, x27, [sp, #(9 * 16)]
+	str	x28, [sp, #(10 * 16)]
+.endm
+
+.macro RESTORE_REGS
+	ldp	d8, d9, [sp, #(0 * 16)]
+	ldp	d10, d11, [sp, #(1 * 16)]
+	ldp	d12, d13, [sp, #(2 * 16)]
+	ldp	d14, d15, [sp, #(3 * 16)]
+	ldp	d16, d17, [sp, #(4 * 16)]
+	ldp	x18, x19, [sp, #(5 * 16)]
+	ldp	x20, x21, [sp, #(6 * 16)]
+	ldp	x22, x23, [sp, #(7 * 16)]
+	ldp	x24, x25, [sp, #(8 * 16)]
+	ldp	x26, x27, [sp, #(9 * 16)]
+	ldr	x28, [sp, #(10 * 16)]
+	add	sp, sp, #(11*16)
+.endm
+
+
+.macro INIT
+	/********** INIT FOR F4 LOOP **********/
+	fmov	ALPHA_R_COPY, ALPHA_R
+	fmov	ALPHA_I_COPY, ALPHA_I
+#if !defined(DOUBLE)
+	ins	v7.s[1], v7.s[0]		// R(ALPHA), R(ALPHA) 
+	ins	v8.s[1], v8.s[0]		// I(ALPHA), I(ALPHA) 
+	ins	v7.d[1], v7.d[0]
+	ins	v8.d[1], v8.d[0]
+#else
+	ins	v7.d[1], v7.d[0]		// R(ALPHA), R(ALPHA)
+	ins	v8.d[1], v8.d[0]		// I(ALPHA), I(ALPHA)
+#endif
+
+	/******* INIT FOR F1 AND S1 LOOP ******/
+#if !defined(DOUBLE)
+	ins	v0.s[1], v0.s[0]		// R(ALPHA), R(ALPHA) 
+	fneg	s2, ALPHA_I
+	ins	v1.s[1], v2.s[0]		// -I(ALPHA), I(ALPHA) 
+#if !defined(XCONJ)
+	ext	v1.8b, v1.8b, v1.8b, #4		// I(ALPHA), -I(ALPHA)
+#endif
+#else
+	ins	v0.d[1], v0.d[0]		// R(ALPHA), R(ALPHA) 
+	fneg	d2, ALPHA_I
+	ins	v1.d[1], v2.d[0]		// -I(ALPHA), I(ALPHA) 
+#if !defined(XCONJ)
+	ext	v1.16b, v1.16b, v1.16b, #8	// I(ALPHA), -I(ALPHA)
+#endif
+#endif
+.endm
+
+.macro INIT_LOOP
+	/********** INIT_LOOP FOR F4 LOOP **********/
+#if !defined(DOUBLE)
+	ld1	{v9.2s}, [X_PTR]		// [I(X), R(X)]
+	ins	v10.s[0], v9.s[1]
+	ins	v9.s[1], v9.s[0]		// [R(X), R(X)]
+	ins	v10.s[1], v10.s[0]		// [I(X), I(X)]
+	ins	v9.d[1], v9.d[0]
+	ins	v10.d[1], v10.d[0]
+#if !defined(CONJ)
+#if !defined(XCONJ)
+	fmul	v11.4s, v9.4s, v7.4s		// [+ R(X) * R(ALPHA)]
+	fmls	v11.4s, v10.4s, v8.4s		// [- I(X) * I(ALPHA)]
+	fmul	v12.4s, v9.4s, v8.4s		// [+ R(X) * I(ALPHA)]
+	fmla	v12.4s, v10.4s, v7.4s		// [+ I(X) * R(ALPHA)]
+#else
+	fmul	v11.4s, v9.4s, v7.4s		// [+ R(X) * R(ALPHA)]
+	fmla	v11.4s, v10.4s, v8.4s		// [+ I(X) * I(ALPHA)]
+	fmul	v12.4s, v9.4s, v8.4s		// [+ R(X) * I(ALPHA)]
+	fmls	v12.4s, v10.4s, v7.4s		// [- I(X) * R(ALPHA)]
+#endif
+#else // CONJ
+#if !defined(XCONJ)
+	fmul	v11.4s, v9.4s, v7.4s		// [+ R(X) * R(ALPHA)]
+	fmls	v11.4s, v10.4s, v8.4s		// [+ I(X) * I(ALPHA)]
+	fmul	v12.4s, v10.4s, v7.4s		// [+ I(X) * R(ALPHA)]
+	fmls	v12.4s, v9.4s, v8.4s		// [- R(X) * I(ALPHA)]
+#else
+	fmul	v11.4s, v9.4s, v7.4s		// [+ R(X) * R(ALPHA)]
+	fmls	v11.4s, v10.4s, v8.4s		// [- I(X) * I(ALPHA)]
+	fmul	v12.4s, v9.4s, v8.4s		// [R(X) * I(ALPHA)]
+	fneg	v12.4s, v12.4s			// [- R(X) * I(ALPHA)]
+	fmla	v12.4s, v10.4s, v7.4s		// [- I(X) * R(ALPHA)]
+#endif
+#endif // CONJ
+
+	/****** INIT_LOOP FOR F1 AND S1 LOOP ******/
+	ld1	{v2.2s}, [X_PTR]		// [I(X), R(X)]
+	ext	v3.8b, v2.8b, v2.8b, #4		// [R(X), I(X)]
+	fmul	v2.2s, v0.2s, v2.2s
+	fmla	v2.2s, v1.2s, v3.2s		// [I(TEMP), R(TEMP)]
+	ins	v3.s[0], v2.s[1]
+#if !defined(CONJ)
+#if !defined(XCONJ)
+	fneg	s4, s3
+	ins	v3.s[1], v4.s[0]
+	ext	v3.8b, v3.8b, v3.8b, #4		// [I(TEMP), -I(TEMP)]
+	ins	v2.s[1], v2.s[0]		// [R(TEMP), R(TEMP)]
+#else
+	fneg	s4, s3
+	ins	v3.s[1], v4.s[0]		// [-I(TEMP), I(TEMP)]
+	ins	v2.s[1], v2.s[0]		// [R(TEMP), R(TEMP)]
+#endif
+#else // CONJ
+#if !defined(XCONJ)
+	ins	v3.s[1], v3.s[0]		// [I(TEMP), I(TEMP)]
+	fneg	s4, s2
+	ins	v2.s[1], v4.s[0]		// [-R(TEMP), R(TEMP)]
+#else
+	fneg	s3, s3
+	ins	v3.s[1], v3.s[0]		// [-I(TEMP), -I(TEMP)]
+	fneg	s4, s2
+	ins	v2.s[1], v4.s[0]		// [-R(TEMP), R(TEMP)]
+#endif
+#endif // CONJ
+
+#else // DOUBLE
+
+	/********** INIT_LOOP FOR F4 LOOP **********/
+	ld1	{v9.2d}, [X_PTR]		// [I(X), R(X)]
+	ins	v10.d[0], v9.d[1]
+	ins	v9.d[1], v9.d[0]		// [R(X), R(X)]
+	ins	v10.d[1], v10.d[0]		// [I(X), I(X)]
+#if !defined(CONJ)
+#if !defined(XCONJ)
+	fmul	v11.2d, v9.2d, v7.2d		// [+ R(X) * R(ALPHA)]
+	fmls	v11.2d, v10.2d, v8.2d		// [- I(X) * I(ALPHA)]
+	fmul	v12.2d, v9.2d, v8.2d		// [+ R(X) * I(ALPHA)]
+	fmla	v12.2d, v10.2d, v7.2d		// [+ I(X) * R(ALPHA)]
+#else
+	fmul	v11.2d, v9.2d, v7.2d		// [+ R(X) * R(ALPHA)]
+	fmla	v11.2d, v10.2d, v8.2d		// [+ I(X) * I(ALPHA)]
+	fmul	v12.2d, v9.2d, v8.2d		// [+ R(X) * I(ALPHA)]
+	fmls	v12.2d, v10.2d, v7.2d		// [- I(X) * R(ALPHA)]
+#endif
+#else // CONJ
+#if !defined(XCONJ)
+	fmul	v11.2d, v9.2d, v7.2d		// [+ R(X) * R(ALPHA)]
+	fmls	v11.2d, v10.2d, v8.2d		// [+ I(X) * I(ALPHA)]
+	fmul	v12.2d, v10.2d, v7.2d		// [+ I(X) * R(ALPHA)]
+	fmls	v12.2d, v9.2d, v8.2d		// [- R(X) * I(ALPHA)]
+#else
+	fmul	v11.2d, v9.2d, v7.2d		// [+ R(X) * R(ALPHA)]
+	fmls	v11.2d, v10.2d, v8.2d		// [- I(X) * I(ALPHA)]
+	fmul	v12.2d, v9.2d, v8.2d		// [R(X) * I(ALPHA)]
+	fneg	v12.2d, v12.2d			// [- R(X) * I(ALPHA)]
+	fmla	v12.2d, v10.2d, v7.2d		// [- I(X) * R(ALPHA)]
+#endif
+#endif // CONJ
+
+	/****** INIT_LOOP FOR F1 AND S1 LOOP ******/
+	ld1	{v2.2d}, [X_PTR]		// [I(X), R(X)]
+	ext	v3.16b, v2.16b, v2.16b, #8	// [R(X), I(X)]
+	fmul	v2.2d, v0.2d, v2.2d
+	fmla	v2.2d, v1.2d, v3.2d		// [I(TEMP), R(TEMP)]
+	ins	v3.d[0], v2.d[1]		// I(TEMP)
+#if !defined(CONJ)
+#if !defined(XCONJ)
+	fneg	d4, d3				// -I(TEMP)
+	ins	v3.d[1], v4.d[0]
+	ext	v3.16b, v3.16b, v3.16b, #8	// [I(TEMP), -I(TEMP)]
+	ins	v2.d[1], v2.d[0]		// [R(TEMP), R(TEMP)]
+#else
+	fneg	d4, d3				// -I(TEMP)
+	ins	v3.d[1], v4.d[0]		// [-I(TEMP), I(TEMP)]
+	ins	v2.d[1], v2.d[0]		// [R(TEMP), R(TEMP)]
+#endif
+#else // CONJ
+#if !defined(XCONJ)
+	ins	v3.d[1], v3.d[0]		// [I(TEMP), I(TEMP)]
+	fneg	d4, d2				// -R(TEMP)
+	ins	v2.d[1], v4.d[0]		// [-R(TEMP), R(TEMP)]
+#else
+	fneg	d3, d3				// -I(TEMP)
+	ins	v3.d[1], v3.d[0]		// [-I(TEMP), -I(TEMP)]
+	fneg	d4, d2				// -R(TEMP)
+	ins	v2.d[1], v4.d[0]		// [-R(TEMP), R(TEMP)]
+#endif
+#endif // CONJ
+
+#endif // DOUBLE
+.endm
+
+.macro KERNEL_F4
+#if !defined(DOUBLE)
+
+	ld2	{v13.4s, v14.4s}, [A_PTR], #32
+	ld2	{v15.4s, v16.4s}, [Y_IPTR], #32
+#if !defined(CONJ)
+#if !defined(XCONJ)
+	fmla	v15.4s, v11.4s, v13.4s		// [+ R(ALPHA * X) * A_R]
+	fmls	v15.4s, v12.4s, v14.4s		// [- I(ALPHA * X) * A_I]
+	fmla	v16.4s, v11.4s, v14.4s		// [+ R(ALPHA * X) * A_I]
+	fmla	v16.4s, v12.4s, v13.4s		// [+ I(ALPHA * X) * A_R]
+#else
+	fmla	v15.4s, v11.4s, v13.4s		// [+ R(ALPHA * X) * A_R]
+	fmla	v15.4s, v12.4s, v14.4s		// [+ I(ALPHA * X) * A_I]
+	fmla	v16.4s, v11.4s, v14.4s		// [+ R(ALPHA * X) * A_I]
+	fmls	v16.4s, v12.4s, v13.4s		// [- I(ALPHA * X) * A_R]
+#endif
+#else // CONJ
+#if !defined(XCONJ)
+	fmla	v15.4s, v11.4s, v13.4s		// [+ R(ALPHA * X) * A_R]
+	fmla	v15.4s, v12.4s, v14.4s		// [+ I(ALPHA * X) * A_I]
+	fmls	v16.4s, v11.4s, v14.4s		// [- R(ALPHA * X) * A_I]
+	fmla	v16.4s, v12.4s, v13.4s		// [+ I(ALPHA * X) * A_R]
+#else
+	fmla	v15.4s, v11.4s, v13.4s		// [+ R(ALPHA * X) * A_R]
+	fmls	v15.4s, v12.4s, v14.4s		// [- I(ALPHA * X) * A_I]
+	fmls	v16.4s, v11.4s, v14.4s		// [- R(ALPHA * X) * A_I]
+	fmls	v16.4s, v12.4s, v13.4s		// [- I(ALPHA * X) * A_R]
+#endif
+#endif // CONJ
+	st2	{v15.4s, v16.4s}, [Y_OPTR], #32
+
+#else // DOUBLE
+
+	ld2	{v13.2d, v14.2d}, [A_PTR], #32
+	ld2	{v15.2d, v16.2d}, [Y_IPTR], #32
+#if !defined(CONJ)
+#if !defined(XCONJ)
+	fmla	v15.2d, v11.2d, v13.2d		// [+ R(ALPHA * X) * A_R]
+	fmls	v15.2d, v12.2d, v14.2d		// [- I(ALPHA * X) * A_I]
+	fmla	v16.2d, v11.2d, v14.2d		// [+ R(ALPHA * X) * A_I]
+	fmla	v16.2d, v12.2d, v13.2d		// [+ I(ALPHA * X) * A_R]
+#else
+	fmla	v15.2d, v11.2d, v13.2d		// [+ R(ALPHA * X) * A_R]
+	fmla	v15.2d, v12.2d, v14.2d		// [+ I(ALPHA * X) * A_I]
+	fmla	v16.2d, v11.2d, v14.2d		// [+ R(ALPHA * X) * A_I]
+	fmls	v16.2d, v12.2d, v13.2d		// [- I(ALPHA * X) * A_R]
+#endif
+#else // CONJ
+#if !defined(XCONJ)
+	fmla	v15.2d, v11.2d, v13.2d		// [+ R(ALPHA * X) * A_R]
+	fmla	v15.2d, v12.2d, v14.2d		// [+ I(ALPHA * X) * A_I]
+	fmls	v16.2d, v11.2d, v14.2d		// [- R(ALPHA * X) * A_I]
+	fmla	v16.2d, v12.2d, v13.2d		// [+ I(ALPHA * X) * A_R]
+#else
+	fmla	v15.2d, v11.2d, v13.2d		// [+ R(ALPHA * X) * A_R]
+	fmls	v15.2d, v12.2d, v14.2d		// [- I(ALPHA * X) * A_I]
+	fmls	v16.2d, v11.2d, v14.2d		// [- R(ALPHA * X) * A_I]
+	fmls	v16.2d, v12.2d, v13.2d		// [- I(ALPHA * X) * A_R]
+#endif
+#endif // CONJ
+	st2	{v15.2d, v16.2d}, [Y_OPTR], #32
+
+	ld2	{v17.2d, v18.2d}, [A_PTR], #32
+	ld2	{v19.2d, v20.2d}, [Y_IPTR], #32
+#if !defined(CONJ)
+#if !defined(XCONJ)
+	fmla	v19.2d, v11.2d, v17.2d		// [+ R(ALPHA * X) * A_R]
+	fmls	v19.2d, v12.2d, v18.2d		// [- I(ALPHA * X) * A_I]
+	fmla	v20.2d, v11.2d, v18.2d		// [+ R(ALPHA * X) * A_I]
+	fmla	v20.2d, v12.2d, v17.2d		// [+ I(ALPHA * X) * A_R]
+#else
+	fmla	v19.2d, v11.2d, v17.2d		// [+ R(ALPHA * X) * A_R]
+	fmla	v19.2d, v12.2d, v18.2d		// [- I(ALPHA * X) * A_I]
+	fmla	v20.2d, v11.2d, v18.2d		// [+ R(ALPHA * X) * A_I]
+	fmls	v20.2d, v12.2d, v17.2d		// [+ I(ALPHA * X) * A_R]
+#endif
+#else // CONJ
+#if !defined(XCONJ)
+	fmla	v19.2d, v11.2d, v17.2d		// [+ R(ALPHA * X) * A_R]
+	fmla	v19.2d, v12.2d, v18.2d		// [- I(ALPHA * X) * A_I]
+	fmls	v20.2d, v11.2d, v18.2d		// [+ R(ALPHA * X) * A_I]
+	fmla	v20.2d, v12.2d, v17.2d		// [+ I(ALPHA * X) * A_R]
+#else
+	fmla	v19.2d, v11.2d, v17.2d		// [+ R(ALPHA * X) * A_R]
+	fmls	v19.2d, v12.2d, v18.2d		// [- I(ALPHA * X) * A_I]
+	fmls	v20.2d, v11.2d, v18.2d		// [+ R(ALPHA * X) * A_I]
+	fmls	v20.2d, v12.2d, v17.2d		// [+ I(ALPHA * X) * A_R]
+#endif
+#endif // CONJ
+	st2	{v19.2d, v20.2d}, [Y_OPTR], #32
+
+#endif
+
+.endm
+
+.macro KERNEL_F1
+#if !defined(DOUBLE)
+	ld1	{v4.2s}, [A_PTR], #8
+	ld1	{v5.2s}, [Y_IPTR], #8
+	ext	v6.8b, v4.8b, v4.8b, #4
+	fmla	v5.2s, v2.2s, v4.2s
+	fmla	v5.2s, v3.2s, v6.2s
+	st1	{v5.2s}, [Y_OPTR], #8
+#else // DOUBLE
+	ld1	{v4.2d}, [A_PTR], #16
+	ld1	{v5.2d}, [Y_IPTR], #16
+	ext	v6.16b, v4.16b, v4.16b, #8
+	fmla	v5.2d, v2.2d, v4.2d
+	fmla	v5.2d, v3.2d, v6.2d
+	st1	{v5.2d}, [Y_OPTR], #16
+#endif
+.endm
+
+.macro INIT_S
+	lsl	INC_Y, INC_Y, #SHZ
+.endm
+
+.macro KERNEL_S1
+#if !defined(DOUBLE)
+	ld1	{v4.2s}, [A_PTR], #8
+	ld1	{v5.2s}, [Y_IPTR], INC_Y
+	ext	v6.8b, v4.8b, v4.8b, #4
+	fmla	v5.2s, v2.2s, v4.2s
+	fmla	v5.2s, v3.2s, v6.2s
+	st1	{v5.2s}, [Y_OPTR], INC_Y
+#else // DOUBLE
+	ld1	{v4.2d}, [A_PTR], #16
+	ld1	{v5.2d}, [Y_IPTR], INC_Y
+	ext	v6.16b, v4.16b, v4.16b, #8
+	fmla	v5.2d, v2.2d, v4.2d
+	fmla	v5.2d, v3.2d, v6.2d
+	st1	{v5.2d}, [Y_OPTR], INC_Y
+#endif
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	ldr	INC_Y, [sp]
+
+	SAVE_REGS
+
+	cmp	N, xzr
+	ble	zgemv_n_kernel_L999
+	cmp	M, xzr
+	ble	zgemv_n_kernel_L999
+
+	lsl	LDA, LDA, #SHZ
+	lsl	INC_X, INC_X, #SHZ
+	mov	J, N
+
+	INIT
+
+	cmp	INC_Y, #1
+	bne	zgemv_n_kernel_S_BEGIN
+
+zgemv_n_kernel_F_LOOP:
+	mov	A_PTR, A
+	mov	Y_IPTR, Y
+	mov	Y_OPTR, Y
+	mov	X_PTR, X
+	add	X, X, INC_X
+	INIT_LOOP
+
+	asr	I, M, #2
+	cmp	I, xzr
+	beq	zgemv_n_kernel_F1
+
+zgemv_n_kernel_F4:
+
+	KERNEL_F1
+	KERNEL_F1
+	KERNEL_F1
+	KERNEL_F1
+
+	subs	I, I, #1
+	bne	zgemv_n_kernel_F4
+
+zgemv_n_kernel_F1:
+
+	ands	I, M, #3
+	ble	zgemv_n_kernel_F_END
+
+zgemv_n_kernel_F10:
+
+	KERNEL_F1
+
+	subs	I, I, #1
+	bne	zgemv_n_kernel_F10
+
+zgemv_n_kernel_F_END:
+
+	add	A, A, LDA
+	subs    J, J, #1
+        bne     zgemv_n_kernel_F_LOOP
+
+	b	zgemv_n_kernel_L999
+
+zgemv_n_kernel_S_BEGIN:
+
+	INIT_S
+
+zgemv_n_kernel_S_LOOP:
+	mov	A_PTR, A
+	mov	Y_IPTR, Y
+	mov	Y_OPTR, Y
+	mov	X_PTR, X
+	add	X, X, INC_X
+	INIT_LOOP
+
+	asr	I, M, #2
+	cmp	I, xzr
+	ble	zgemv_n_kernel_S1
+
+zgemv_n_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	zgemv_n_kernel_S4
+
+zgemv_n_kernel_S1:
+
+	ands	I, M, #3
+	ble	zgemv_n_kernel_S_END
+
+zgemv_n_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     zgemv_n_kernel_S10
+
+zgemv_n_kernel_S_END:
+
+	add	A, A, LDA
+	subs    J, J, #1
+        bne     zgemv_n_kernel_S_LOOP
+
+zgemv_n_kernel_L999:
+	RESTORE_REGS
+
+	mov	w0, wzr
+	ret
+
+	EPILOGUE

kernel/arm64/zgemv_t.S

@@ -0,0 +1,448 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	M	x0	/* Y vector length */
+#define	N	x1	/* X vector length */
+#define	A	x3	/* A vector address */
+#define	LDA	x4	/* A stride */
+#define	X	x5	/* X vector address */
+#define	INC_X	x6	/* X stride */
+#define	Y	x7	/* Y vector address */
+#define	INC_Y	x2	/* Y stride */
+#define	A_PTR	x9	/* loop A vector address */
+#define	X_PTR	x10	/* loop Y vector address */
+#define J	x11	/* loop variable */
+#define I	x12	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define ALPHA_R		s0
+#define ALPHA_I		s1
+#define ALPHA_R_COPY	s7
+#define ALPHA_I_COPY	s8
+#define SHZ		3
+#else
+#define ALPHA_R		d0
+#define ALPHA_I		d1
+#define ALPHA_R_COPY	d7
+#define ALPHA_I_COPY	d8
+#define SHZ		4
+#endif
+
+/******************************************************************************/
+
+
+.macro SAVE_REGS
+	add	sp, sp, #-(11 * 16)
+	stp	d8, d9, [sp, #(0 * 16)]
+	stp	d10, d11, [sp, #(1 * 16)]
+	stp	d12, d13, [sp, #(2 * 16)]
+	stp	d14, d15, [sp, #(3 * 16)]
+	stp	d16, d17, [sp, #(4 * 16)]
+	stp	x18, x19, [sp, #(5 * 16)]
+	stp	x20, x21, [sp, #(6 * 16)]
+	stp	x22, x23, [sp, #(7 * 16)]
+	stp	x24, x25, [sp, #(8 * 16)]
+	stp	x26, x27, [sp, #(9 * 16)]
+	str	x28, [sp, #(10 * 16)]
+.endm
+
+.macro RESTORE_REGS
+	ldp	d8, d9, [sp, #(0 * 16)]
+	ldp	d10, d11, [sp, #(1 * 16)]
+	ldp	d12, d13, [sp, #(2 * 16)]
+	ldp	d14, d15, [sp, #(3 * 16)]
+	ldp	d16, d17, [sp, #(4 * 16)]
+	ldp	x18, x19, [sp, #(5 * 16)]
+	ldp	x20, x21, [sp, #(6 * 16)]
+	ldp	x22, x23, [sp, #(7 * 16)]
+	ldp	x24, x25, [sp, #(8 * 16)]
+	ldp	x26, x27, [sp, #(9 * 16)]
+	ldr	x28, [sp, #(10 * 16)]
+	add	sp, sp, #(11*16)
+.endm
+
+.macro INIT
+#if !defined(XCONJ)
+#if !defined(DOUBLE)
+	ins	v0.s[1], v0.s[0]		// v0 = ALPHA_R, ALPHA_R 
+	fneg	s2, ALPHA_I
+	ins	v1.s[1], v2.s[0]
+	ext	v1.8b, v1.8b, v1.8b, #4		// v1 = ALPHA_I, -ALPHA_I 
+#else
+	ins	v0.d[1], v0.d[0]		// v0 = ALPHA_R, ALPHA_R 
+	fneg	d2, ALPHA_I
+	ins	v1.d[1], v2.d[0]
+	ext	v1.16b, v1.16b, v1.16b, #8	// v1 = ALPHA_I, -ALPHA_I 
+#endif
+#else // XCONJ
+#if !defined(DOUBLE)
+	fneg	s2, ALPHA_R
+	ins	v0.s[1], v2.s[0]		// v0 = -ALPHA_R, ALPHA_R 
+	ins	v1.s[1], v1.s[0]		// v1 = ALPHA_I, ALPHA_I 
+#else
+	fneg	d2, ALPHA_R
+	ins	v0.d[1], v2.d[0]		// v0 = -ALPHA_R, ALPHA_R 
+	ins	v1.d[1], v1.d[0]		// v1 = ALPHA_I, ALPHA_I 
+#endif
+#endif
+.endm
+
+.macro INIT_LOOP
+	fmov	d9, xzr				// TEMP_R = [0, 0]
+	fmov	d10, xzr			// TEMP_I = [0, 0]
+#if !defined(DOUBLE)
+#else
+	fmov	d15, xzr			// TEMP_R = [0, 0]
+	fmov	d16, xzr			// TEMP_I = [0, 0]
+#endif
+
+	fmov	d2, xzr				// TEMP = [0, 0]
+.endm
+
+.macro KERNEL_F4
+#if !defined(DOUBLE)
+
+	ld2	{v11.4s, v12.4s}, [X_PTR], #32
+	ld2	{v13.4s, v14.4s}, [A_PTR], #32
+
+#if !defined(CONJ)
+#if !defined(XCONJ)
+	fmla	v9.4s,  v11.4s, v13.4s		// [+ R(X) * A_R]
+	fmls	v9.4s,  v12.4s, v14.4s		// [- I(X) * A_I]
+	fmla	v10.4s, v11.4s, v14.4s		// [+ R(X) * A_I]
+	fmla	v10.4s, v12.4s, v13.4s		// [+ I(X) * A_R]
+#else
+	fmla	v9.4s,  v11.4s, v13.4s		// [+ R(X) * A_R]
+	fmla	v9.4s,  v12.4s, v14.4s		// [+ I(X) * A_I]
+	fmla	v10.4s, v11.4s, v14.4s		// [+ R(X) * A_I]
+	fmls	v10.4s, v12.4s, v13.4s		// [- I(X) * A_R]
+#endif
+#else // CONJ
+#if !defined(XCONJ)
+	fmla	v9.4s,  v11.4s, v13.4s		// [+ R(X) * A_R]
+	fmla	v9.4s,  v12.4s, v14.4s		// [+ I(X) * A_I]
+	fmls	v10.4s, v11.4s, v14.4s		// [- R(X) * A_I]
+	fmla	v10.4s, v12.4s, v13.4s		// [+ I(X) * A_R]
+#else
+	fmla	v9.4s,  v11.4s, v13.4s		// [+ R(X) * A_R]
+	fmls	v9.4s,  v12.4s, v14.4s		// [- I(X) * A_I]
+	fmls	v10.4s, v11.4s, v14.4s		// [- R(X) * A_I]
+	fmls	v10.4s, v12.4s, v13.4s		// [- I(X) * A_R]
+#endif
+#endif // CONJ
+
+#else // DOUBLE
+	ld2	{v11.2d, v12.2d}, [X_PTR], #32
+	ld2	{v13.2d, v14.2d}, [A_PTR], #32
+	prfm    PLDL1STRM, [X_PTR, #512]
+#if !defined(CONJ)
+#if !defined(XCONJ)
+	fmla	v9.2d,  v11.2d, v13.2d		// [+ R(X) * A_R]
+	fmls	v9.2d,  v12.2d, v14.2d		// [- I(X) * A_I]
+	fmla	v10.2d, v11.2d, v14.2d		// [+ R(X) * A_I]
+	fmla	v10.2d, v12.2d, v13.2d		// [+ I(X) * A_R]
+#else
+	fmla	v9.2d,  v11.2d, v13.2d		// [+ R(X) * A_R]
+	fmla	v9.2d,  v12.2d, v14.2d		// [+ I(X) * A_I]
+	fmla	v10.2d, v11.2d, v14.2d		// [+ R(X) * A_I]
+	fmls	v10.2d, v12.2d, v13.2d		// [- I(X) * A_R]
+#endif
+#else // CONJ
+#if !defined(XCONJ)
+	fmla	v9.2d,  v11.2d, v13.2d		// [+ R(X) * A_R]
+	fmla	v9.2d,  v12.2d, v14.2d		// [+ I(X) * A_I]
+	fmls	v10.2d, v11.2d, v14.2d		// [- R(X) * A_I]
+	fmla	v10.2d, v12.2d, v13.2d		// [+ I(X) * A_R]
+#else
+	fmla	v9.2d,  v11.2d, v13.2d		// [+ R(X) * A_R]
+	fmls	v9.2d,  v12.2d, v14.2d		// [- I(X) * A_I]
+	fmls	v10.2d, v11.2d, v14.2d		// [- R(X) * A_I]
+	fmls	v10.2d, v12.2d, v13.2d		// [- I(X) * A_R]
+#endif
+#endif // CONJ
+	ld2	{v17.2d, v18.2d}, [X_PTR], #32
+	ld2	{v19.2d, v20.2d}, [A_PTR], #32
+	prfm    PLDL1STRM, [A_PTR, #512]
+#if !defined(CONJ)
+#if !defined(XCONJ)
+	fmla	v15.2d, v17.2d, v19.2d		// [+ R(X) * A_R]
+	fmls	v15.2d, v18.2d, v20.2d		// [- I(X) * A_I]
+	fmla	v16.2d, v17.2d, v20.2d		// [+ R(X) * A_I]
+	fmla	v16.2d, v18.2d, v19.2d		// [+ I(X) * A_R]
+#else
+	fmla	v15.2d, v17.2d, v19.2d		// [+ R(X) * A_R]
+	fmla	v15.2d, v18.2d, v20.2d		// [- I(X) * A_I]
+	fmla	v16.2d, v17.2d, v20.2d		// [+ R(X) * A_I]
+	fmls	v16.2d, v18.2d, v19.2d		// [+ I(X) * A_R]
+#endif
+#else // CONJ
+#if !defined(XCONJ)
+	fmla	v15.2d, v17.2d, v19.2d		// [+ R(X) * A_R]
+	fmla	v15.2d, v18.2d, v20.2d		// [- I(X) * A_I]
+	fmls	v16.2d, v17.2d, v20.2d		// [+ R(X) * A_I]
+	fmla	v16.2d, v18.2d, v19.2d		// [+ I(X) * A_R]
+#else
+	fmla	v15.2d, v17.2d, v19.2d		// [+ R(X) * A_R]
+	fmls	v15.2d, v18.2d, v20.2d		// [- I(X) * A_I]
+	fmls	v16.2d, v17.2d, v20.2d		// [+ R(X) * A_I]
+	fmls	v16.2d, v18.2d, v19.2d		// [+ I(X) * A_R]
+#endif
+#endif // CONJ
+#endif //DOUBLE
+.endm
+
+.macro KERNEL_F4_FINALIZE
+#if !defined(DOUBLE)
+	ext	v21.16b, v9.16b, v9.16b, #8
+	fadd	v9.2s, v9.2s, v21.2s
+	faddp	s9, v9.2s
+
+	ext	v21.16b, v10.16b, v10.16b, #8
+	fadd	v10.2s, v10.2s, v21.2s
+	faddp	s10, v10.2s
+
+	ins	v2.s[0], v9.s[0]
+	ins	v2.s[1], v10.s[0]
+#else
+	fadd	v9.2d, v9.2d, v15.2d
+	fadd	v10.2d, v10.2d, v16.2d
+
+	faddp	d9, v9.2d
+	faddp	d10, v10.2d
+
+	ins	v2.d[0], v9.d[0]
+	ins	v2.d[1], v10.d[0]
+#endif
+.endm
+
+
+.macro KERNEL_F1
+#if !defined(DOUBLE)
+	ld1r	{v4.2s}, [A_PTR], #4		// [A0, A0]
+	ld1	{v5.s}[0], [A_PTR], #4		// A1
+	ld1	{v6.2s}, [X_PTR], #8		// [X1, X0]
+	fneg	s16, s5
+	ins	v5.s[1], v16.s[0]		// [-A1, A1]
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+	ext	v5.8b, v5.8b, v5.8b, #4		// [A1, -A1]
+#endif
+	ext	v7.8b, v6.8b, v6.8b, #4		// [X0, X1]
+	fmla	v2.2s, v4.2s, v6.2s
+	fmla	v2.2s, v5.2s, v7.2s
+#else // DOUBLE
+	ld1r	{v4.2d}, [A_PTR], #8		// [A0, A0]
+	ld1	{v5.d}[0], [A_PTR], #8		// A1
+	ld1	{v6.2d}, [X_PTR], #16		// [X1, X0]
+	fneg	d16, d5
+	ins	v5.d[1], v16.d[0]		// [-A1, A1]
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+	ext	v5.16b, v5.16b, v5.16b, #8	// [A1, -A1]
+#endif
+	ext	v7.16b, v6.16b, v6.16b, #8	// [X0, X1]
+	fmla	v2.2d, v4.2d, v6.2d
+	fmla	v2.2d, v5.2d, v7.2d
+#endif
+.endm
+
+.macro INIT_S
+	lsl	INC_X, INC_X, #SHZ
+.endm
+
+.macro KERNEL_S1
+#if !defined(DOUBLE)
+	ld1r	{v4.2s}, [A_PTR], #4		// [A0, A0]
+	ld1	{v5.s}[0], [A_PTR], #4		// A1
+	ld1	{v6.2s}, [X_PTR], INC_X		// [X1, X0]
+	fneg	s16, s5
+	ins	v5.s[1], v16.s[0]		// [-A1, A1]
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+	ext	v5.8b, v5.8b, v5.8b, #4		// [A1, -A1]
+#endif
+	ext	v7.8b, v6.8b, v6.8b, #4		// [X0, X1]
+	fmla	v2.2s, v4.2s, v6.2s
+	fmla	v2.2s, v5.2s, v7.2s
+#else // DOUBLE
+	ld1r	{v4.2d}, [A_PTR], #8		// [A0, A0]
+	ld1	{v5.d}[0], [A_PTR], #8		// A1
+	ld1	{v6.2d}, [X_PTR], INC_X		// [X1, X0]
+	fneg	d16, d5
+	ins	v5.d[1], v16.d[0]		// [-A1, A1]
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+	ext	v5.16b, v5.16b, v5.16b, #8	// [A1, -A1]
+#endif
+	ext	v7.16b, v6.16b, v6.16b, #8	// [X0, X1]
+	fmla	v2.2d, v4.2d, v6.2d
+	fmla	v2.2d, v5.2d, v7.2d
+#endif
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	ldr	INC_Y, [sp]
+	SAVE_REGS
+
+	cmp	N, xzr
+	ble	zgemv_t_kernel_L999
+	cmp	M, xzr
+	ble	zgemv_t_kernel_L999
+
+	lsl	LDA, LDA, #SHZ
+	lsl	INC_Y, INC_Y, #SHZ
+	mov	J, N
+
+	INIT
+
+	cmp	INC_X, #1
+	bne	zgemv_t_kernel_S_BEGIN
+
+zgemv_t_kernel_F_LOOP:
+
+	mov	A_PTR, A
+	mov	X_PTR, X
+
+	INIT_LOOP
+
+	asr	I, M, #2
+	cmp	I, xzr
+	beq	zgemv_t_kernel_F1
+
+zgemv_t_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	zgemv_t_kernel_F4
+
+	KERNEL_F4_FINALIZE
+
+zgemv_t_kernel_F1:
+
+	ands	I, M, #3
+	ble	zgemv_t_kernel_F_END
+
+zgemv_t_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     zgemv_t_kernel_F10
+
+zgemv_t_kernel_F_END:
+
+#if !defined(DOUBLE)
+	ld1	{v4.2s}, [Y]
+	ext	v3.8b, v2.8b, v2.8b, #4		// [TEMP_R, TEMP_I]
+	fmla	v4.2s, v0.2s, v2.2s
+	fmla	v4.2s, v1.2s, v3.2s
+	st1	{v4.2s}, [Y], INC_Y
+#else // DOUBLE
+	ld1	{v4.2d}, [Y]
+	ext	v3.16b, v2.16b, v2.16b, #8	// [TEMP_R, TEMP_I]
+	fmla	v4.2d, v0.2d, v2.2d
+	fmla	v4.2d, v1.2d, v3.2d
+	st1	{v4.2d}, [Y], INC_Y
+#endif
+
+	add	A, A, LDA
+	subs    J, J, #1
+        bne     zgemv_t_kernel_F_LOOP
+
+	b	zgemv_t_kernel_L999
+
+zgemv_t_kernel_S_BEGIN:
+
+	INIT_S
+
+zgemv_t_kernel_S_LOOP:
+
+	mov	A_PTR, A
+	mov	X_PTR, X
+	INIT_LOOP
+
+	asr	I, M, #2
+	cmp	I, xzr
+	ble	zgemv_t_kernel_S1
+
+zgemv_t_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	zgemv_t_kernel_S4
+
+zgemv_t_kernel_S1:
+
+	ands	I, M, #3
+	ble	zgemv_t_kernel_S_END
+
+zgemv_t_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     zgemv_t_kernel_S10
+
+zgemv_t_kernel_S_END:
+
+#if !defined(DOUBLE)
+	ld1	{v4.2s}, [Y]
+	ext	v3.8b, v2.8b, v2.8b, #4		// [TEMP_R, TEMP_I]
+	fmla	v4.2s, v0.2s, v2.2s
+	fmla	v4.2s, v1.2s, v3.2s
+	st1	{v4.2s}, [Y], INC_Y
+#else // DOUBLE
+	ld1	{v4.2d}, [Y]
+	ext	v3.16b, v2.16b, v2.16b, #8	// [TEMP_R, TEMP_I]
+	fmla	v4.2d, v0.2d, v2.2d
+	fmla	v4.2d, v1.2d, v3.2d
+	st1	{v4.2d}, [Y], INC_Y
+#endif
+
+	add	A, A, LDA
+	subs    J, J, #1
+        bne     zgemv_t_kernel_S_LOOP
+
+zgemv_t_kernel_L999:
+	RESTORE_REGS
+	mov	w0, wzr
+	ret
+
+	EPILOGUE

kernel/arm64/znrm2.S

@@ -0,0 +1,228 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define TMPF	s6
+#define SSQ	s0
+#define TMPVF	{v6.s}[0]
+#define SZ	4
+#else
+#define TMPF	d6
+#define SSQ	d0
+#define TMPVF	{v6.d}[0]
+#define SZ	8
+#endif
+
+/******************************************************************************/
+
+.macro KERNEL_F1
+#if !defined(DOUBLE)
+	ld1	{v1.2s}, [X], #8
+	fmul	v1.2s, v1.2s, v1.2s
+	faddp	TMPF, v1.2s
+	fadd	SSQ, SSQ, TMPF
+#else
+	ld1	{v1.2d}, [X], #16
+	fmul	v1.2d, v1.2d, v1.2d
+	faddp	TMPF, v1.2d
+	fadd	SSQ, SSQ, TMPF
+#endif
+.endm
+
+.macro KERNEL_F8
+#if !defined(DOUBLE)
+	ld1	{v1.4s, v2.4s}, [X], #32
+	fmla	v0.4s, v1.4s, v1.4s
+	fmla	v5.4s, v2.4s, v2.4s
+	ld1	{v3.4s,v4.4s}, [X], #32
+	fmla	v0.4s, v3.4s, v3.4s
+	fmla	v5.4s, v4.4s, v4.4s
+	PRFM	PLDL1KEEP, [X, #1024]
+#else // DOUBLE
+	ld1	{v1.2d, v2.2d}, [X], #32
+	fmla	v0.2d, v1.2d, v1.2d
+	fmla	v5.2d, v2.2d, v2.2d
+	ld1	{v3.2d, v4.2d}, [X], #32
+	fmla	v0.2d, v3.2d, v3.2d
+	fmla	v5.2d, v4.2d, v4.2d
+
+	ld1	{v16.2d, v17.2d}, [X], #32
+	fmla	v0.2d, v16.2d, v16.2d
+	fmla	v5.2d, v17.2d, v17.2d
+	ld1	{v18.2d, v19.2d}, [X], #32
+	fmla	v0.2d, v18.2d, v18.2d
+	fmla	v5.2d, v19.2d, v19.2d
+#endif
+.endm
+
+.macro nrm2_kernel_F8_FINALIZE
+#if !defined(DOUBLE)
+	fadd	v0.4s, v0.4s, v5.4s
+	ext	v1.16b, v0.16b, v0.16b, #8
+	fadd	v0.2s, v0.2s,  v1.2s
+	faddp	SSQ, v0.2s
+#else
+	fadd	v0.2d, v0.2d, v5.2d
+	faddp	SSQ, v0.2d
+#endif
+.endm
+
+.macro INIT_S
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #3
+	ld1	{v1.2s}, [X], INC_X
+	fmul	v1.2s, v1.2s, v1.2s
+	faddp	SSQ, v1.2s
+#else
+	lsl	INC_X, INC_X, #4
+	ld1	{v1.2d}, [X], INC_X
+	fmul	v1.2d, v1.2d, v1.2d
+	faddp	SSQ, v1.2d
+#endif
+.endm
+
+.macro KERNEL_S1
+#if !defined(DOUBLE)
+	ld1	{v1.2s}, [X], INC_X
+	fmul	v1.2s, v1.2s, v1.2s
+	faddp	TMPF, v1.2s
+	fadd	SSQ, SSQ, TMPF
+#else
+	ld1	{v1.2d}, [X], INC_X
+	fmul	v1.2d, v1.2d, v1.2d
+	faddp	TMPF, v1.2d
+	fadd	SSQ, SSQ, TMPF
+#endif
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+#if !defined(DOUBLE)
+	fmov	SSQ, wzr
+	fmov	s5, SSQ
+#else
+	fmov	SSQ, xzr
+	fmov	d5, SSQ
+#endif
+
+	cmp	N, xzr
+	ble	nrm2_kernel_zero
+	cmp	INC_X, xzr
+	ble	nrm2_kernel_zero
+	cmp	INC_X, #1
+	bne	nrm2_kernel_S_BEGIN
+
+nrm2_kernel_F_BEGIN:
+
+	asr	I, N, #3
+	cmp	I, xzr
+	beq	nrm2_kernel_F1_INIT
+
+nrm2_kernel_F8:
+
+	KERNEL_F8
+
+	subs	I, I, #1
+	bne	nrm2_kernel_F8
+
+	nrm2_kernel_F8_FINALIZE
+
+nrm2_kernel_F1:
+
+	ands	I, N, #7
+	ble	nrm2_kernel_L999
+
+nrm2_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     nrm2_kernel_F10
+
+	b	nrm2_kernel_L999
+
+nrm2_kernel_F1_INIT:
+
+	b	nrm2_kernel_F1
+
+nrm2_kernel_S_BEGIN:
+
+	INIT_S
+
+	subs	N, N, #1
+	ble	nrm2_kernel_L999
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	nrm2_kernel_S1
+
+nrm2_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	nrm2_kernel_S4
+
+nrm2_kernel_S1:
+
+	ands	I, N, #3
+	ble	nrm2_kernel_L999
+
+nrm2_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+	bne     nrm2_kernel_S10
+
+nrm2_kernel_L999:
+	fsqrt	SSQ, SSQ
+	ret
+
+nrm2_kernel_zero:
+	ret
+
+	EPILOGUE

kernel/arm64/zrot.S

@@ -0,0 +1,256 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x1	/* X vector address */
+#define	INC_X	x2	/* X stride */
+#define	Y	x3	/* Y vector address */
+#define	INC_Y	x4	/* Y stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define	C	s0	/* scale input value */
+#define	S	s1	/* scale input value */
+#else
+#define	C	d0	/* scale input value */
+#define	S	d1	/* scale input value */
+#endif
+
+/******************************************************************************/
+
+.macro INIT
+
+#if !defined(DOUBLE)
+	ins	v0.s[1], v0.s[0]		// [C, C]
+	ins	v1.s[1], v1.s[0]		// [S, S]
+#else
+	ins	v0.d[1], v0.d[0]		// [C, C]
+	ins	v1.d[1], v1.d[0]		// [S, S]
+#endif
+
+.endm
+
+.macro KERNEL_F1
+
+#if !defined(DOUBLE)
+	ld1	{v2.2s}, [X]
+	ld1	{v3.2s}, [Y]
+	fmul	v4.2s, v0.2s, v2.2s		// [C*X1, C*X0]
+	fmla	v4.2s, v1.2s, v3.2s		// [C*X1 + S*Y1, C*X0 + S*Y0]
+	fmul	v5.2s, v0.2s, v3.2s		// [C*Y1, C*Y0]
+	fmls	v5.2s, v1.2s, v2.2s		// [C*Y1 - S*X1, C*Y0 - S*X0]
+	st1	{v4.2s}, [X], #8
+	st1	{v5.2s}, [Y], #8
+#else
+	ld1	{v2.2d}, [X]
+	ld1	{v3.2d}, [Y]
+	fmul	v4.2d, v0.2d, v2.2d		// [C*X1, C*X0]
+	fmla	v4.2d, v1.2d, v3.2d		// [C*X1 + S*Y1, C*X0 + S*Y0]
+	fmul	v5.2d, v0.2d, v3.2d		// [C*Y1, C*Y0]
+	fmls	v5.2d, v1.2d, v2.2d		// [C*Y1 - S*X1, C*Y0 - S*X0]
+	st1	{v4.2d}, [X], #16
+	st1	{v5.2d}, [Y], #16
+#endif
+
+.endm
+
+.macro KERNEL_INIT_F4
+
+#if !defined(DOUBLE)
+	ins	v0.d[1], v0.d[0]		// [C, C, C, C]
+	ins	v1.d[1], v1.d[0]		// [S, S, S, S]
+#endif
+
+.endm
+
+.macro KERNEL_F4
+
+#if !defined(DOUBLE)
+	ld1	{v2.4s, v3.4s}, [X]
+	ld1	{v4.4s, v5.4s}, [Y]
+	fmul	v6.4s, v0.4s, v2.4s		// C*X3, C*X2, C*X1, C*X0
+	fmul	v7.4s, v0.4s, v3.4s		// C*X7, C*X6, C*X5, C*X4
+	fmla	v6.4s, v1.4s, v4.4s		// C*X3+S*Y3, ..., C*X0+S*Y0
+	fmla	v7.4s, v1.4s, v5.4s		// C*X7+S*Y7, ..., C*X4+S*Y4
+	fmul	v16.4s, v0.4s, v4.4s		// C*Y3, C*Y2, C*Y1, C*Y0
+	fmul	v17.4s, v0.4s, v5.4s		// C*Y7, C*Y6, C*Y5, C*Y4
+	fmls	v16.4s, v1.4s, v2.4s		// C*Y3-S*X3, ..., C*Y0-S*X0
+	fmls	v17.4s, v1.4s, v3.4s		// C*Y7-S*X7, ..., C*Y4-S*X4
+	st1	{v6.4s,v7.4s}, [X], #32
+	st1	{v16.4s,v17.4s}, [Y], #32
+#else // DOUBLE
+	ld1	{v2.2d, v3.2d}, [X]
+	ld1	{v4.2d, v5.2d}, [Y]
+	fmul	v6.2d, v0.2d, v2.2d		// C*X3, C*X2, C*X1, C*X0
+	fmul	v7.2d, v0.2d, v3.2d		// C*X7, C*X6, C*X5, C*X4
+	fmla	v6.2d, v1.2d, v4.2d		// C*X3+S*Y3, ..., C*X0+S*Y0
+	fmla	v7.2d, v1.2d, v5.2d		// C*X7+S*Y7, ..., C*X4+S*Y4
+	fmul	v16.2d, v0.2d, v4.2d		// C*Y3, C*Y2, C*Y1, C*Y0
+	fmul	v17.2d, v0.2d, v5.2d		// C*Y7, C*Y6, C*Y5, C*Y4
+	fmls	v16.2d, v1.2d, v2.2d		// C*Y3-S*X3, ..., C*Y0-S*X0
+	fmls	v17.2d, v1.2d, v3.2d		// C*Y7-S*X7, ..., C*Y4-S*X4
+	st1	{v6.2d,v7.2d}, [X], #32
+	st1	{v16.2d,v17.2d}, [Y], #32
+	ld1	{v2.2d, v3.2d}, [X]
+	ld1	{v4.2d, v5.2d}, [Y]
+	fmul	v6.2d, v0.2d, v2.2d		// C*X3, C*X2, C*X1, C*X0
+	fmul	v7.2d, v0.2d, v3.2d		// C*X7, C*X6, C*X5, C*X4
+	fmla	v6.2d, v1.2d, v4.2d		// C*X3+S*Y3, ..., C*X0+S*Y0
+	fmla	v7.2d, v1.2d, v5.2d		// C*X7+S*Y7, ..., C*X4+S*Y4
+	fmul	v16.2d, v0.2d, v4.2d		// C*Y3, C*Y2, C*Y1, C*Y0
+	fmul	v17.2d, v0.2d, v5.2d		// C*Y7, C*Y6, C*Y5, C*Y4
+	fmls	v16.2d, v1.2d, v2.2d		// C*Y3-S*X3, ..., C*Y0-S*X0
+	fmls	v17.2d, v1.2d, v3.2d		// C*Y7-S*X7, ..., C*Y4-S*X4
+	st1	{v6.2d,v7.2d}, [X], #32
+	st1	{v16.2d,v17.2d}, [Y], #32
+#endif
+
+.endm
+
+.macro INIT_S
+
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #3
+	lsl	INC_Y, INC_Y, #3
+#else
+	lsl	INC_X, INC_X, #4
+	lsl	INC_Y, INC_Y, #4
+#endif
+
+.endm
+
+.macro KERNEL_S1
+
+#if !defined(DOUBLE)
+	ld1	{v2.2s}, [X]
+	ld1	{v3.2s}, [Y]
+	fmul	v4.2s, v0.2s, v2.2s		// [C*X1, C*X0]
+	fmla	v4.2s, v1.2s, v3.2s		// [C*X1 + S*Y1, C*X0 + S*Y0]
+	fmul	v5.2s, v0.2s, v3.2s		// [C*Y1, C*Y0]
+	fmls	v5.2s, v1.2s, v2.2s		// [C*Y1 - S*X1, C*Y0 - S*X0]
+	st1	{v4.2s}, [X], INC_X
+	st1	{v5.2s}, [Y], INC_Y
+#else
+	ld1	{v2.2d}, [X]
+	ld1	{v3.2d}, [Y]
+	fmul	v4.2d, v0.2d, v2.2d		// [C*X1, C*X0]
+	fmla	v4.2d, v1.2d, v3.2d		// [C*X1 + S*Y1, C*X0 + S*Y0]
+	fmul	v5.2d, v0.2d, v3.2d		// [C*Y1, C*Y0]
+	fmls	v5.2d, v1.2d, v2.2d		// [C*Y1 - S*X1, C*Y0 - S*X0]
+	st1	{v4.2d}, [X], INC_X
+	st1	{v5.2d}, [Y], INC_Y
+#endif
+
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	rot_kernel_L999
+
+	INIT
+
+	cmp	INC_X, #1
+	bne	rot_kernel_S_BEGIN
+	cmp	INC_Y, #1
+	bne	rot_kernel_S_BEGIN
+
+rot_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	rot_kernel_F1
+
+	KERNEL_INIT_F4
+
+rot_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	rot_kernel_F4
+
+rot_kernel_F1:
+
+	ands	I, N, #3
+	ble	rot_kernel_L999
+
+rot_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     rot_kernel_F10
+
+	mov	w0, wzr
+	ret
+
+rot_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	rot_kernel_S1
+
+rot_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	rot_kernel_S4
+
+rot_kernel_S1:
+
+	ands	I, N, #3
+	ble	rot_kernel_L999
+
+rot_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     rot_kernel_S10
+
+rot_kernel_L999:
+
+	mov	w0, wzr
+	ret

kernel/arm64/zscal.S

@@ -0,0 +1,274 @@
+/*******************************************************************************
+Copyright (c) 2015, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*******************************************************************************/
+
+#define ASSEMBLER
+#include "common.h"
+
+#define	N	x0	/* vector length */
+#define	X	x3	/* X vector address */
+#define	INC_X	x4	/* X stride */
+#define I	x5	/* loop variable */
+
+/*******************************************************************************
+* Macro definitions
+*******************************************************************************/
+
+#if !defined(DOUBLE)
+#define	DA_R	s0	/* real scale input value */
+#define	DA_I	s1	/* imaginary scale input value */
+#else
+#define	DA_R	d0	/* real scale input value */
+#define	DA_I	d1	/* imaginary scale input value */
+#endif
+
+/******************************************************************************/
+
+.macro INIT
+
+#if !defined(DOUBLE)
+	ins	v0.s[1], v0.s[0]		// v0 = DA_R, DA_R 
+	fneg	s2, DA_I
+	ins	v1.s[1], v2.s[0]		// v1 = -DA_I, DA_I 
+	ext	v1.8b, v1.8b, v1.8b, #4		// v1 = DA_I, -DA_I
+#else
+	ins	v0.d[1], v0.d[0]		// v0 = DA_R, DA_R 
+	fneg	d2, DA_I
+	ins	v1.d[1], v2.d[0]		// v1 = DA_I, DA_I 
+	ext	v1.16b, v1.16b, v1.16b, #8	// v1 = DA_I, -DA_I
+#endif
+
+.endm
+
+.macro KERNEL_F1
+
+#if !defined(DOUBLE)
+	ld1	{v2.2s}, [X]			// X1, X0
+	ext	v3.8b, v2.8b, v2.8b, #4		// X0, X1
+	fmul	v2.2s, v2.2s, v0.2s		// DA_R*X1, DA_R*X0
+	fmla	v2.2s, v3.2s, v1.2s		// DA_R*X1+DA_I*X0, DA_R*X0-DA_I*X1
+	st1	{v2.2s}, [X], #8
+#else
+	ld1	{v2.2d}, [X]			// X1, X0
+	ext	v3.16b, v2.16b, v2.16b, #8	// X0, X1
+	fmul	v2.2d, v2.2d, v0.2d		// DA_R*X1, DA_R*X0
+	fmla	v2.2d, v3.2d, v1.2d		// DA_R*X1+DA_I*X0, DA_R*X0-DA_I*X1
+	st1	{v2.2d}, [X], #16
+#endif
+
+.endm
+
+.macro KERNEL_INIT_F4
+
+#if !defined(DOUBLE)
+	// Replicate the lower 2 floats into the upper 2 slots
+	ins	v0.d[1], v0.d[0]		// v0 = DA_R, DA_R, DA_R, DA_R
+	ins	v1.d[1], v1.d[0]		// v1 = DA_I, DA_I, DA_I, DA_I
+#endif
+
+.endm
+
+.macro KERNEL_F4
+
+#if !defined(DOUBLE)
+	ld1	{v2.4s,v3.4s}, [X]		// V2 = X[3], X[2], X[1], X[0]
+						// V3 = X[7], X[6], X[5], X[4]
+
+	ext	v6.8b, v2.8b, v2.8b, #4		// V6 =  -  ,  -  , X[0], X[1]
+	ins	v6.s[2], v2.s[3]		// V6 =  -  , X[3], X[0], X[1]
+	ins	v6.s[3], v2.s[2]		// V6 = X[2], X[3], X[0], X[1]
+	fmul	v2.4s, v0.4s, v2.4s		// X'[ix]   += DA_R * X[ix]
+						// X'[ix+1] += DA_R * X[ix+1]
+	fmla	v2.4s, v1.4s, v6.4s		// X'[ix]   += -DA_I * X[ix+1]
+						// X'[ix+1] += DA_I * X[ix]
+
+	ext	v7.8b, v3.8b, v3.8b, #4		// V7 =  -  ,  -  , X[4], X[5]
+	ins	v7.s[2], v3.s[3]		// V7 =  -  , X[7], X[4], X[5]
+	ins	v7.s[3], v3.s[2]		// V7 = X[6], X[7], X[4], X[5]
+	fmul	v3.4s, v0.4s, v3.4s		// X'[ix]   += DA_R * X[ix]
+						// X'[ix+1] += DA_R * X[ix+1]
+	fmla	v3.4s, v1.4s, v7.4s		// X'[ix]   += -DA_I * X[ix+1]
+						// X'[ix+1] += DA_I * X[ix]
+
+	st1	{v2.4s,v3.4s}, [X], #32
+#else // DOUBLE
+	ld1	{v2.2d,v3.2d,v4.2d,v5.2d}, [X]	// CX0, CX1, CX2, CX3
+	ext	v20.16b, v2.16b, v2.16b, #8	// X[ix], X[ix+1]
+	ext	v21.16b, v3.16b, v3.16b, #8	// X[ix], X[ix+1]
+	ext	v22.16b, v4.16b, v4.16b, #8	// X[ix], X[ix+1]
+	ext	v23.16b, v5.16b, v5.16b, #8	// X[ix], X[ix+1]
+
+	fmul	v2.2d, v0.2d, v2.2d
+	fmla	v2.2d, v1.2d, v20.2d
+
+	fmul	v3.2d, v0.2d, v3.2d
+	fmla	v3.2d, v1.2d, v21.2d
+	st1	{v2.2d,v3.2d}, [X], #32
+
+	fmul	v4.2d, v0.2d, v4.2d
+	fmla	v4.2d, v1.2d, v22.2d
+
+	fmul	v5.2d, v0.2d, v5.2d
+	fmla	v5.2d, v1.2d, v23.2d
+	st1	{v4.2d,v5.2d}, [X], #32
+#endif
+	PRFM	PLDL1KEEP, [X, #1024]
+.endm
+
+.macro INIT_S
+
+#if !defined(DOUBLE)
+	lsl	INC_X, INC_X, #3
+#else
+	lsl	INC_X, INC_X, #4
+#endif
+
+.endm
+
+.macro KERNEL_S1
+
+#if !defined(DOUBLE)
+	ld1	{v2.2s}, [X]			// X1, X0
+	ext	v3.8b, v2.8b, v2.8b, #4		// X0, X1
+	fmul	v2.2s, v2.2s, v0.2s		// DA_R*X1, DA_R*X0
+	fmla	v2.2s, v3.2s, v1.2s		// DA_R*X1+DA_I*X0, DA_R*X0-DA_I*X1
+	st1	{v2.2s}, [X], INC_X
+#else
+	ld1	{v2.2d}, [X]			// X1, X0
+	ext	v3.16b, v2.16b, v2.16b, #8	// X0, X1
+	fmul	v2.2d, v2.2d, v0.2d		// DA_R*X1, DA_R*X0
+	fmla	v2.2d, v3.2d, v1.2d		// DA_R*X1+DA_I*X0, DA_R*X0-DA_I*X1
+	st1	{v2.2d}, [X], INC_X
+#endif
+
+.endm
+
+/*******************************************************************************
+* End of macro definitions
+*******************************************************************************/
+
+	PROLOGUE
+
+	cmp	N, xzr
+	ble	zscal_kernel_L999
+
+	fcmp	DA_R, #0.0
+	bne	zscal_kernel_1
+
+	fcmp	DA_I, #0.0
+	beq	zscal_kernel_zero
+
+	// TODO: special case DA_R == 0 && DA_I != 0
+
+zscal_kernel_1:
+
+	// TODO: special case DA_R != 0 && DA_I == 0
+
+	INIT
+
+	cmp	INC_X, #1
+	bne	zscal_kernel_S_BEGIN
+
+zscal_kernel_F_BEGIN:
+
+	asr	I, N, #2
+	cmp	I, xzr
+	beq	zscal_kernel_F1
+
+	KERNEL_INIT_F4
+
+zscal_kernel_F4:
+
+	KERNEL_F4
+
+	subs	I, I, #1
+	bne	zscal_kernel_F4
+
+zscal_kernel_F1:
+
+	ands	I, N, #3
+	ble	zscal_kernel_L999
+
+zscal_kernel_F10:
+
+	KERNEL_F1
+
+	subs    I, I, #1
+        bne     zscal_kernel_F10
+
+	mov	w0, wzr
+	ret
+
+zscal_kernel_S_BEGIN:
+
+	INIT_S
+
+	asr	I, N, #2
+	cmp	I, xzr
+	ble	zscal_kernel_S1
+
+zscal_kernel_S4:
+
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+	KERNEL_S1
+
+	subs	I, I, #1
+	bne	zscal_kernel_S4
+
+zscal_kernel_S1:
+
+	ands	I, N, #3
+	ble	zscal_kernel_L999
+
+zscal_kernel_S10:
+
+	KERNEL_S1
+
+	subs    I, I, #1
+        bne     zscal_kernel_S10
+
+zscal_kernel_L999:
+
+	mov	w0, wzr
+	ret
+
+zscal_kernel_zero:
+
+	INIT_S
+
+zscal_kernel_Z1:
+
+	stp	DA_R, DA_I, [X]
+	add	X, X, INC_X
+	subs    N, N, #1
+        bne     zscal_kernel_Z1
+
+	mov	w0, wzr
+	ret
+
+	EPILOGUE

kernel/generic/ztrmmkernel_4x4.c

@@ -0,0 +1,883 @@
+#include "common.h"
+
+#define MADD_ALPHA_N_STORE(C, res, alpha) \
+	C[0] = res ## _r * alpha ## _r - res ## _i * alpha ## _i; \
+	C[1] = res ## _r * alpha ## _i + res ## _i * alpha ## _r;
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT)
+#define MADD(res, op1, op2) \
+	res ## _r += op1 ## _r * op2 ## _r; \
+	res ## _r -= op1 ## _i * op2 ## _i; \
+	res ## _i += op1 ## _r * op2 ## _i; \
+	res ## _i += op1 ## _i * op2 ## _r;
+#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
+#define MADD(res, op1, op2) \
+	res ## _r += op1 ## _r * op2 ## _r; \
+	res ## _r += op1 ## _i * op2 ## _i; \
+	res ## _i -= op1 ## _r * op2 ## _i; \
+	res ## _i += op1 ## _i * op2 ## _r;
+#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
+#define MADD(res, op1, op2) \
+	res ## _r += op1 ## _r * op2 ## _r; \
+	res ## _r += op1 ## _i * op2 ## _i; \
+	res ## _i += op1 ## _r * op2 ## _i; \
+	res ## _i -= op1 ## _i * op2 ## _r;
+#elif defined(RR) || defined(RC) || defined(CR) || defined(CC)
+#define MADD(res, op1, op2) \
+	res ## _r += op1 ## _r * op2 ## _r; \
+	res ## _r -= op1 ## _i * op2 ## _i; \
+	res ## _i -= op1 ## _r * op2 ## _i; \
+	res ## _i -= op1 ## _i * op2 ## _r;
+#endif
+
+int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha_r, FLOAT alpha_i,FLOAT* ba,FLOAT* bb,FLOAT* C,BLASLONG ldc
+		, BLASLONG offset
+		)
+{
+
+	BLASLONG i,j,k;
+	FLOAT *C0,*C1,*C2,*C3,*ptrba,*ptrbb;
+	FLOAT res00_r, res01_r, res02_r, res03_r;
+	FLOAT res00_i, res01_i, res02_i, res03_i;
+	FLOAT res10_r, res11_r, res12_r, res13_r;
+	FLOAT res10_i, res11_i, res12_i, res13_i;
+	FLOAT res20_r, res21_r, res22_r, res23_r;
+	FLOAT res20_i, res21_i, res22_i, res23_i;
+	FLOAT res30_r, res31_r, res32_r, res33_r;
+	FLOAT res30_i, res31_i, res32_i, res33_i;
+	FLOAT a0_r, a1_r;
+	FLOAT a0_i, a1_i;
+	FLOAT b0_r, b1_r, b2_r, b3_r;
+	FLOAT b0_i, b1_i, b2_i, b3_i;
+	BLASLONG off, temp;
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+	off = -offset;
+#endif
+
+	for (j=0; j<bn/4; j+=1) // do blocks of the Mx4 loops 
+	{
+		C0 = C;
+		C1 = C0+2*ldc;
+		C2 = C1+2*ldc;
+		C3 = C2+2*ldc;
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+		off = offset;
+#endif
+
+		ptrba = ba;
+
+		for (i=0; i<bm/4; i+=1) // do blocks of 4x4
+		{
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			ptrbb = bb;
+#else
+
+			ptrba += off*4*2; // number of values in A
+			ptrbb = bb + off*4*2; // number of values in B
+#endif
+
+			res00_r = 0;
+			res00_i = 0;
+			res01_r = 0;
+			res01_i = 0;
+			res02_r = 0;
+			res02_i = 0;
+			res03_r = 0;
+			res03_i = 0;
+
+			res10_r = 0;
+			res10_i = 0;
+			res11_r = 0;
+			res11_i = 0;
+			res12_r = 0;
+			res12_i = 0;
+			res13_r = 0;
+			res13_i = 0;
+
+			res20_r = 0;
+			res20_i = 0;
+			res21_r = 0;
+			res21_i = 0;
+			res22_r = 0;
+			res22_i = 0;
+			res23_r = 0;
+			res23_i = 0;
+
+			res30_r = 0;
+			res30_i = 0;
+			res31_r = 0;
+			res31_i = 0;
+			res32_r = 0;
+			res32_i = 0;
+			res33_r = 0;
+			res33_i = 0;
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+			temp = bk - off;
+#elif defined(LEFT)
+			temp = off + 4;
+#else
+			temp = off + 4;
+#endif
+
+			for (k=0; k<temp; k++)
+			{
+				b0_r = ptrbb[2*0+0]; b0_i = ptrbb[2*0+1];
+				b1_r = ptrbb[2*1+0]; b1_i = ptrbb[2*1+1];
+				b2_r = ptrbb[2*2+0]; b2_i = ptrbb[2*2+1];
+				b3_r = ptrbb[2*3+0]; b3_i = ptrbb[2*3+1];
+
+				a0_r = ptrba[2*0+0]; a0_i = ptrba[2*0+1];
+				MADD(res00, a0, b0);
+				MADD(res10, a0, b1);
+				MADD(res20, a0, b2);
+				MADD(res30, a0, b3);
+
+				a1_r = ptrba[2*1+0]; a1_i = ptrba[2*1+1];
+				MADD(res01, a1, b0);
+				MADD(res11, a1, b1);
+				MADD(res21, a1, b2);
+				MADD(res31, a1, b3);
+
+				a0_r = ptrba[2*2+0]; a0_i = ptrba[2*2+1];
+				MADD(res02, a0, b0);
+				MADD(res12, a0, b1);
+				MADD(res22, a0, b2);
+				MADD(res32, a0, b3);
+
+
+				a1_r = ptrba[2*3+0]; a1_i = ptrba[2*3+1];
+				MADD(res03, a1, b0);
+				MADD(res13, a1, b1);
+				MADD(res23, a1, b2);
+				MADD(res33, a1, b3);
+
+				ptrba = ptrba+8;
+				ptrbb = ptrbb+8;
+			}
+
+			MADD_ALPHA_N_STORE(C0, res00, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res01, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res02, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res03, alpha);
+			C0 = C0 + 2;
+
+			MADD_ALPHA_N_STORE(C1, res10, alpha);
+			C1 = C1 + 2;
+			MADD_ALPHA_N_STORE(C1, res11, alpha);
+			C1 = C1 + 2;
+			MADD_ALPHA_N_STORE(C1, res12, alpha);
+			C1 = C1 + 2;
+			MADD_ALPHA_N_STORE(C1, res13, alpha);
+			C1 = C1 + 2;
+
+			MADD_ALPHA_N_STORE(C2, res20, alpha);
+			C2 = C2 + 2;
+			MADD_ALPHA_N_STORE(C2, res21, alpha);
+			C2 = C2 + 2;
+			MADD_ALPHA_N_STORE(C2, res22, alpha);
+			C2 = C2 + 2;
+			MADD_ALPHA_N_STORE(C2, res23, alpha);
+			C2 = C2 + 2;
+
+			MADD_ALPHA_N_STORE(C3, res30, alpha);
+			C3 = C3 + 2;
+			MADD_ALPHA_N_STORE(C3, res31, alpha);
+			C3 = C3 + 2;
+			MADD_ALPHA_N_STORE(C3, res32, alpha);
+			C3 = C3 + 2;
+			MADD_ALPHA_N_STORE(C3, res33, alpha);
+			C3 = C3 + 2;
+
+
+#if ( defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			temp = bk-off;
+#if defined(LEFT)
+			temp = temp - 4;
+#else
+			temp = temp - 4;
+#endif
+			ptrba += temp*4*2; // number of values in A
+			ptrbb += temp*4*2; // number of values in B
+#endif
+#ifdef LEFT
+			off += 4; // number of values in A
+#endif
+
+
+		}
+
+		if ( bm & 2 ) // do any 2x4 loop
+		{
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			ptrbb = bb;
+#else
+			ptrba += off*2*2;
+			ptrbb = bb + off*4*2;
+#endif
+
+
+			res00_r = 0;
+			res00_i = 0;
+			res01_r = 0;
+			res01_i = 0;
+
+			res10_r = 0;
+			res10_i = 0;
+			res11_r = 0;
+			res11_i = 0;
+
+			res20_r = 0;
+			res20_i = 0;
+			res21_r = 0;
+			res21_i = 0;
+
+			res30_r = 0;
+			res30_i = 0;
+			res31_r = 0;
+			res31_i = 0;
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+			temp = bk-off;
+#elif defined(LEFT)
+			temp = off+2;	// number of values in A
+#else
+			temp = off+4;	// number of values in B
+#endif
+
+			for (k=0; k<temp; k++)
+			{
+				b0_r = ptrbb[2*0+0]; b0_i = ptrbb[2*0+1];
+				b1_r = ptrbb[2*1+0]; b1_i = ptrbb[2*1+1];
+				b2_r = ptrbb[2*2+0]; b2_i = ptrbb[2*2+1];
+				b3_r = ptrbb[2*3+0]; b3_i = ptrbb[2*3+1];
+
+				a0_r = ptrba[2*0+0]; a0_i = ptrba[2*0+1];
+				MADD(res00, a0, b0);
+				MADD(res10, a0, b1);
+				MADD(res20, a0, b2);
+				MADD(res30, a0, b3);
+
+				a1_r = ptrba[2*1+0]; a1_i = ptrba[2*1+1];
+				MADD(res01, a1, b0);
+				MADD(res11, a1, b1);
+				MADD(res21, a1, b2);
+				MADD(res31, a1, b3);
+
+
+				ptrba = ptrba+4;
+				ptrbb = ptrbb+8;
+			}
+
+			MADD_ALPHA_N_STORE(C0, res00, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res01, alpha);
+			C0 = C0 + 2;
+
+			MADD_ALPHA_N_STORE(C1, res10, alpha);
+			C1 = C1 + 2;
+			MADD_ALPHA_N_STORE(C1, res11, alpha);
+			C1 = C1 + 2;
+
+			MADD_ALPHA_N_STORE(C2, res20, alpha);
+			C2 = C2 + 2;
+			MADD_ALPHA_N_STORE(C2, res21, alpha);
+			C2 = C2 + 2;
+
+			MADD_ALPHA_N_STORE(C3, res30, alpha);
+			C3 = C3 + 2;
+			MADD_ALPHA_N_STORE(C3, res31, alpha);
+			C3 = C3 + 2;
+
+
+
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			temp = bk - off;
+#ifdef LEFT
+			temp -= 2; // number of values in A
+#else
+			temp -= 4; // number of values in B
+#endif
+			ptrba += temp*2*2;
+			ptrbb += temp*4*2;
+#endif
+
+#ifdef LEFT
+			off += 2; // number of values in A
+#endif
+
+
+		}
+
+		if ( bm & 1 ) // do any 1x4 loop
+		{
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			ptrbb = bb;
+#else
+			ptrba += off*1*2;
+			ptrbb = bb + off*4*2;
+#endif
+
+			res00_r = 0;
+			res00_i = 0;
+			res10_r = 0;
+			res10_i = 0;
+			res20_r = 0;
+			res20_i = 0;
+			res30_r = 0;
+			res30_i = 0;
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+			temp = bk-off;
+#elif defined(LEFT)
+			temp = off+1;	// number of values in A
+#else
+			temp = off+4;	// number of values in B
+#endif
+
+			for (k=0; k<temp; k++)
+			{
+				b0_r = ptrbb[2*0+0]; b0_i = ptrbb[2*0+1];
+				b1_r = ptrbb[2*1+0]; b1_i = ptrbb[2*1+1];
+				b2_r = ptrbb[2*2+0]; b2_i = ptrbb[2*2+1];
+				b3_r = ptrbb[2*3+0]; b3_i = ptrbb[2*3+1];
+
+				a0_r = ptrba[2*0+0]; a0_i = ptrba[2*0+1];
+				MADD(res00, a0, b0);
+				MADD(res10, a0, b1);
+				MADD(res20, a0, b2);
+				MADD(res30, a0, b3);
+
+
+				ptrba = ptrba+2;
+				ptrbb = ptrbb+8;
+			}
+
+			MADD_ALPHA_N_STORE(C0, res00, alpha);
+			C0 = C0 + 2;
+
+			MADD_ALPHA_N_STORE(C1, res10, alpha);
+			C1 = C1 + 2;
+
+			MADD_ALPHA_N_STORE(C2, res20, alpha);
+			C2 = C2 + 2;
+
+			MADD_ALPHA_N_STORE(C3, res30, alpha);
+			C3 = C3 + 2;
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			temp = bk - off;
+#ifdef LEFT
+			temp -= 1; // number of values in A
+#else
+			temp -= 4; // number of values in B
+#endif
+			ptrba += temp*1*2;
+			ptrbb += temp*4*2;
+#endif
+
+#ifdef LEFT
+			off += 1; // number of values in A
+#endif
+
+
+		}
+
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+		off += 4;
+#endif
+
+		k = (bk<<3);
+		bb = bb+k;
+		i = (ldc<<3);
+		C = C+i;
+	}
+
+	for (j=0; j<(bn&2); j+=2) // do the Mx2 loops 
+	{
+		C0 = C;
+		C1 = C0+ldc*2;
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+		off = offset;
+#endif
+
+
+		ptrba = ba;
+
+		for (i=0; i<bm/4; i+=1) // do blocks of 4x2
+		{
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			ptrbb = bb;
+#else
+			ptrba += off*4*2;
+			ptrbb = bb + off*2*2;
+#endif
+
+			res00_r = 0;
+			res00_i = 0;
+			res01_r = 0;
+			res01_i = 0;
+			res02_r = 0;
+			res02_i = 0;
+			res03_r = 0;
+			res03_i = 0;
+
+			res10_r = 0;
+			res10_i = 0;
+			res11_r = 0;
+			res11_i = 0;
+			res12_r = 0;
+			res12_i = 0;
+			res13_r = 0;
+			res13_i = 0;
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+			temp = bk-off;
+#elif defined(LEFT)
+			temp = off+4;	// number of values in A
+#else
+			temp = off+2;	// number of values in B
+#endif
+
+			for (k=0; k<temp; k++)
+			{
+				b0_r = ptrbb[2*0+0]; b0_i = ptrbb[2*0+1];
+				b1_r = ptrbb[2*1+0]; b1_i = ptrbb[2*1+1];
+
+				a0_r = ptrba[2*0+0]; a0_i = ptrba[2*0+1];
+				MADD(res00, a0, b0);
+				MADD(res10, a0, b1);
+
+				a1_r = ptrba[2*1+0]; a1_i = ptrba[2*1+1];
+				MADD(res01, a1, b0);
+				MADD(res11, a1, b1);
+
+				a0_r = ptrba[2*2+0]; a0_i = ptrba[2*2+1];
+				MADD(res02, a0, b0);
+				MADD(res12, a0, b1);
+
+				a1_r = ptrba[2*3+0]; a1_i = ptrba[2*3+1];
+				MADD(res03, a1, b0);
+				MADD(res13, a1, b1);
+
+				ptrba = ptrba+8;
+				ptrbb = ptrbb+4;
+			}
+
+			MADD_ALPHA_N_STORE(C0, res00, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res01, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res02, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res03, alpha);
+			C0 = C0 + 2;
+
+			MADD_ALPHA_N_STORE(C1, res10, alpha);
+			C1 = C1 + 2;
+			MADD_ALPHA_N_STORE(C1, res11, alpha);
+			C1 = C1 + 2;
+			MADD_ALPHA_N_STORE(C1, res12, alpha);
+			C1 = C1 + 2;
+			MADD_ALPHA_N_STORE(C1, res13, alpha);
+			C1 = C1 + 2;
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			temp = bk - off;
+#ifdef LEFT
+			temp -= 4; // number of values in A
+#else
+			temp -= 2; // number of values in B
+#endif
+			ptrba += temp*4*2;
+			ptrbb += temp*2*2;
+#endif
+
+#ifdef LEFT
+			off += 4; // number of values in A
+#endif
+
+		}
+
+		if ( bm & 2 ) // do any 2x2 loop
+		{
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			ptrbb = bb;
+#else
+			ptrba += off*2*2;
+			ptrbb = bb + off*2*2;
+#endif
+
+			res00_r = 0;
+			res00_i = 0;
+			res01_r = 0;
+			res01_i = 0;
+
+			res10_r = 0;
+			res10_i = 0;
+			res11_r = 0;
+			res11_i = 0;
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+			temp = bk-off;
+#elif defined(LEFT)
+			temp = off+2;	// number of values in A
+#else
+			temp = off+2;	// number of values in B
+#endif
+
+			for (k=0; k<temp; k++)
+			{
+				b0_r = ptrbb[2*0+0]; b0_i = ptrbb[2*0+1];
+				b1_r = ptrbb[2*1+0]; b1_i = ptrbb[2*1+1];
+
+				a0_r = ptrba[2*0+0]; a0_i = ptrba[2*0+1];
+				MADD(res00, a0, b0);
+				MADD(res10, a0, b1);
+
+				a1_r = ptrba[2*1+0]; a1_i = ptrba[2*1+1];
+				MADD(res01, a1, b0);
+				MADD(res11, a1, b1);
+
+
+				ptrba = ptrba+4;
+				ptrbb = ptrbb+4;
+			}
+
+			MADD_ALPHA_N_STORE(C0, res00, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res01, alpha);
+			C0 = C0 + 2;
+
+			MADD_ALPHA_N_STORE(C1, res10, alpha);
+			C1 = C1 + 2;
+			MADD_ALPHA_N_STORE(C1, res11, alpha);
+			C1 = C1 + 2;
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			temp = bk - off;
+#ifdef LEFT
+			temp -= 2; // number of values in A
+#else
+			temp -= 2; // number of values in B
+#endif
+			ptrba += temp*2*2;
+			ptrbb += temp*2*2;
+#endif
+
+#ifdef LEFT
+			off += 2; // number of values in A
+#endif
+
+		}
+
+		if ( bm & 1 ) // do any 1x2 loop
+		{
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			ptrbb = bb;
+#else
+			ptrba += off*1*2;
+			ptrbb = bb + off*2*2;
+#endif
+
+			res00_r = 0;
+			res00_i = 0;
+
+			res10_r = 0;
+			res10_i = 0;
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+			temp = bk-off;
+#elif defined(LEFT)
+			temp = off+1;	// number of values in A
+#else
+			temp = off+2;	// number of values in B
+#endif
+
+			for (k=0; k<temp; k++)
+			{
+				b0_r = ptrbb[2*0+0]; b0_i = ptrbb[2*0+1];
+				b1_r = ptrbb[2*1+0]; b1_i = ptrbb[2*1+1];
+
+				a0_r = ptrba[2*0+0]; a0_i = ptrba[2*0+1];
+				MADD(res00, a0, b0);
+				MADD(res10, a0, b1);
+
+				ptrba = ptrba+2;
+				ptrbb = ptrbb+4;
+			}
+
+			MADD_ALPHA_N_STORE(C0, res00, alpha);
+			C0 = C0 + 2;
+
+			MADD_ALPHA_N_STORE(C1, res10, alpha);
+			C1 = C1 + 2;
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			temp = bk - off;
+#ifdef LEFT
+			temp -= 1; // number of values in A
+#else
+			temp -= 2; // number of values in B
+#endif
+			ptrba += temp*1*2;
+			ptrbb += temp*2*2;
+#endif
+
+#ifdef LEFT
+			off += 1; // number of values in A
+#endif
+
+		}
+
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+		off += 2;
+#endif
+
+		k = (bk<<2);
+		bb = bb+k;
+		i = (ldc<<2);
+		C = C+i;
+	}
+
+
+
+
+
+
+
+	for (j=0; j<(bn&1); j+=1) // do the Mx1 loops
+	{
+		C0 = C;
+
+#if defined(TRMMKERNEL) &&  defined(LEFT)
+		off = offset;
+#endif
+
+		ptrba = ba;
+
+		for (i=0; i<bm/4; i+=1) // do blocks of 4x1 loops
+		{
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			ptrbb = bb;
+#else
+			ptrba += off*4*2;
+			ptrbb = bb + off*1*2;
+#endif
+
+			res00_r = 0;
+			res00_i = 0;
+			res01_r = 0;
+			res01_i = 0;
+			res02_r = 0;
+			res02_i = 0;
+			res03_r = 0;
+			res03_i = 0;
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+			temp = bk-off;
+#elif defined(LEFT)
+			temp = off+4;	// number of values in A
+#else
+			temp = off+1;	// number of values in B
+#endif
+
+			for (k=0; k<temp; k++)
+			{
+				b0_r = ptrbb[2*0+0]; b0_i = ptrbb[2*0+1];
+
+				a0_r = ptrba[2*0+0]; a0_i = ptrba[2*0+1];
+				MADD(res00, a0, b0);
+
+				a1_r = ptrba[2*1+0]; a1_i = ptrba[2*1+1];
+				MADD(res01, a1, b0);
+
+				a0_r = ptrba[2*2+0]; a0_i = ptrba[2*2+1];
+				MADD(res02, a0, b0);
+
+				a1_r = ptrba[2*3+0]; a1_i = ptrba[2*3+1];
+				MADD(res03, a1, b0);
+
+				ptrba = ptrba+8;
+				ptrbb = ptrbb+2;
+			}
+
+			MADD_ALPHA_N_STORE(C0, res00, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res01, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res02, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res03, alpha);
+			C0 = C0 + 2;
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			temp = bk - off;
+#ifdef LEFT
+			temp -= 4; // number of values in A
+#else
+			temp -= 1; // number of values in B
+#endif
+			ptrba += temp*4*2;
+			ptrbb += temp*1*2;
+#endif
+
+#ifdef LEFT
+			off += 4; // number of values in A
+#endif
+
+		}
+
+		if ( bm & 2 ) // do any 2x1 loop
+		{
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			ptrbb = bb;
+#else
+			ptrba += off*2*2;
+			ptrbb = bb + off*1*2;
+#endif
+
+			res00_r = 0;
+			res00_i = 0;
+			res01_r = 0;
+			res01_i = 0;
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+			temp = bk-off;
+#elif defined(LEFT)
+			temp = off+2;	// number of values in A
+#else
+			temp = off+1;	// number of values in B
+#endif
+
+			for (k=0; k<temp; k++)
+			{
+				b0_r = ptrbb[2*0+0]; b0_i = ptrbb[2*0+1];
+
+				a0_r = ptrba[2*0+0]; a0_i = ptrba[2*0+1];
+				MADD(res00, a0, b0);
+
+				a1_r = ptrba[2*1+0]; a1_i = ptrba[2*1+1];
+				MADD(res01, a1, b0);
+
+
+				ptrba = ptrba+4;
+				ptrbb = ptrbb+2;
+			}
+
+			MADD_ALPHA_N_STORE(C0, res00, alpha);
+			C0 = C0 + 2;
+			MADD_ALPHA_N_STORE(C0, res01, alpha);
+			C0 = C0 + 2;
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			temp = bk - off;
+#ifdef LEFT
+			temp -= 2; // number of values in A
+#else
+			temp -= 1; // number of values in B
+#endif
+			ptrba += temp*2*2;
+			ptrbb += temp*1*2;
+#endif
+
+#ifdef LEFT
+			off += 2; // number of values in A
+#endif
+
+		}
+
+		if ( bm & 1 ) // do any 1x1 loop
+		{
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			ptrbb = bb;
+#else
+			ptrba += off*1*2;
+			ptrbb = bb + off*1*2;
+#endif
+
+			res00_r = 0;
+			res00_i = 0;
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+			temp = bk-off;
+#elif defined(LEFT)
+			temp = off+1;	// number of values in A
+#else
+			temp = off+1;	// number of values in B
+#endif
+
+			for (k=0; k<temp; k++)
+			{
+				b0_r = ptrbb[2*0+0]; b0_i = ptrbb[2*0+1];
+
+				a0_r = ptrba[2*0+0]; a0_i = ptrba[2*0+1];
+				MADD(res00, a0, b0);
+
+				ptrba = ptrba+2;
+				ptrbb = ptrbb+2;
+			}
+
+			MADD_ALPHA_N_STORE(C0, res00, alpha);
+			C0 = C0 + 2;
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+			temp = bk - off;
+#ifdef LEFT
+			temp -= 1; // number of values in A
+#else
+			temp -= 1; // number of values in B
+#endif
+			ptrba += temp*1*2;
+			ptrbb += temp*1*2;
+#endif
+
+#ifdef LEFT
+			off += 1; // number of values in A
+#endif
+
+		}
+
+
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+		off += 1;
+#endif
+
+		k = (bk<<1);
+		bb = bb+k;
+		i = (ldc<<1);
+		C = C+i;
+	}
+	return 0;
+}

param.h

@@ -2214,6 +2214,46 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ZGEMM_DEFAULT_R 4096
 
 
+#define SYMV_P	16
+#endif
+
+
+#if defined(CORTEXA57)
+#define SNUMOPT		2
+#define DNUMOPT		2
+
+#define GEMM_DEFAULT_OFFSET_A 0
+#define GEMM_DEFAULT_OFFSET_B 0
+#define GEMM_DEFAULT_ALIGN 0x03fffUL
+
+#define SGEMM_DEFAULT_UNROLL_M  4
+#define SGEMM_DEFAULT_UNROLL_N  4
+
+#define DGEMM_DEFAULT_UNROLL_M  4
+#define DGEMM_DEFAULT_UNROLL_N  4
+
+#define CGEMM_DEFAULT_UNROLL_M  4
+#define CGEMM_DEFAULT_UNROLL_N  4
+
+#define ZGEMM_DEFAULT_UNROLL_M  4
+#define ZGEMM_DEFAULT_UNROLL_N  4
+
+#define SGEMM_DEFAULT_P	128
+#define DGEMM_DEFAULT_P	256
+#define CGEMM_DEFAULT_P 256
+#define ZGEMM_DEFAULT_P 128
+
+#define SGEMM_DEFAULT_Q 240
+#define DGEMM_DEFAULT_Q 1024
+#define CGEMM_DEFAULT_Q 1024
+#define ZGEMM_DEFAULT_Q 512
+
+#define SGEMM_DEFAULT_R 12288
+#define DGEMM_DEFAULT_R 4096
+#define CGEMM_DEFAULT_R 4096
+#define ZGEMM_DEFAULT_R 2048
+
+
 #define SYMV_P	16
 #endif