Merge branch 'develop' into betterPowerGEMVTail

.github/workflows/riscv64_vector.yml

@@ -28,6 +28,9 @@ jobs:
           - target: RISCV64_ZVL256B
             opts: TARGET=RISCV64_ZVL256B BINARY=64 ARCH=riscv64
             qemu_cpu: rv64,g=true,c=true,v=true,vext_spec=v1.0,vlen=256,elen=64
+          - target: DYNAMIC_ARCH=1
+            opts: TARGET=RISCV64_GENERIC BINARY=64 ARCH=riscv64 DYNAMIC_ARCH=1
+            qemu_cpu: rv64,g=true,c=true,v=true,vext_spec=v1.0,vlen=256,elen=64
 
     steps:
       - name: Checkout repository

Jenkinsfile.pwr

@@ -1,7 +1,7 @@
 pipeline {
     agent { 
         docker {
-            image 'osuosl/ubuntu-ppc64le'
+            image 'osuosl/ubuntu-ppc64le:18.04'
         }
     }
     stages {

Makefile.arm64

@@ -276,12 +276,19 @@ endif
 endif
 endif
 
-ifeq (1, $(filter 1,$(GCCVERSIONGTEQ11) $(ISCLANG)))
 ifeq ($(CORE), A64FX)
+ifeq (1, $(filter 1,$(GCCVERSIONGTEQ10) $(ISCLANG)))
+ifeq (1, $(filter 1,$(GCCMINORVERSIONGTEQ3) $(GCCVERSIONGTEQ11) $(ISCLANG)))
 CCOMMON_OPT += -march=armv8.2-a+sve -mtune=a64fx
 ifneq ($(F_COMPILER), NAG)
 FCOMMON_OPT += -march=armv8.2-a+sve -mtune=a64fx
 endif
+else
+CCOMMON_OPT += -march=armv8.4-a+sve -mtune=neoverse-n1
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=armv8.4-a -mtune=neoverse-n1
+endif
+endif
 endif
 endif
 

Makefile.system

@@ -268,13 +268,24 @@ SMALL_MATRIX_OPT = 1
 else ifeq ($(ARCH), power)
 SMALL_MATRIX_OPT = 1
 BUILD_BFLOAT16 = 1
+else ifeq ($(ARCH), arm64)
+SMALL_MATRIX_OPT = 1
 endif
 ifeq ($(ARCH), loongarch64)
 SMALL_MATRIX_OPT = 1
 endif
+ifeq ($(ARCH), arm64)
+GEMM_GEMV_FORWARD = 1
+endif
+
 ifeq ($(SMALL_MATRIX_OPT), 1)
 CCOMMON_OPT += -DSMALL_MATRIX_OPT
 endif
+ifeq ($(GEMM_GEMV_FORWARD), 1)
+ifneq ($(ONLY_CBLAS), 1)
+CCOMMON_OPT += -DGEMM_GEMV_FORWARD
+endif
+endif
 
 # This operation is expensive, so execution should be once.
 ifndef GOTOBLAS_MAKEFILE
@@ -689,6 +700,7 @@ ifneq ($(NO_SVE), 1)
 DYNAMIC_CORE += NEOVERSEV1
 DYNAMIC_CORE += NEOVERSEN2
 DYNAMIC_CORE += ARMV8SVE
+DYNAMIC_CORE += A64FX
 endif
 DYNAMIC_CORE += THUNDERX
 DYNAMIC_CORE += THUNDERX2T99
@@ -715,6 +727,17 @@ ifeq ($(ARCH), loongarch64)
 DYNAMIC_CORE = LOONGSON3R5 LOONGSON2K1000 LOONGSONGENERIC
 endif
 
+ifeq ($(ARCH), riscv64)
+DYNAMIC_CORE = RISCV64_GENERIC
+DYNAMIC_CORE += RISCV64_ZVL128B
+DYNAMIC_CORE += RISCV64_ZVL256B
+ifdef DYNAMIC_LIST
+override DYNAMIC_CORE = RISCV64_GENERIC $(DYNAMIC_LIST)
+XCCOMMON_OPT = -DDYNAMIC_LIST -DDYN_RISCV64_GENERIC
+XCCOMMON_OPT += $(foreach dcore,$(DYNAMIC_LIST),-DDYN_$(dcore))
+endif
+endif
+
 ifeq ($(ARCH), zarch)
 DYNAMIC_CORE = ZARCH_GENERIC
 

README.md

@@ -234,6 +234,8 @@ For **POWER**, the list encompasses POWER6, POWER8 and POWER9. POWER10 is additi
 
 on **ZARCH** it comprises Z13 and Z14 as well as generic zarch support.
 
+On **riscv64**, DYNAMIC_ARCH enables support for riscv64_zvl128b and riscv64_zvl256b in addition to generic riscv64 support.  A compiler that supports RVV 1.0 is required to build OpenBLAS for riscv64 when DYNAMIC_ARCH is enabled.
+
 The `TARGET` option can be used in conjunction with `DYNAMIC_ARCH=1` to specify which cpu model should be assumed for all the
 common code in the library, usually you will want to set this to the oldest model you expect to encounter.
 Please note that it is not possible to combine support for different architectures, so no combined 32 and 64 bit or x86_64 and arm64 in the same library.

benchmark/pybench/benchmarks/bench_blas.py

@@ -234,14 +234,10 @@ def test_gesdd(benchmark, mn, variant):
     gesdd = ow.get_func('gesdd', variant)
     u, s, vt, info = benchmark(run_gesdd, a, lwork, gesdd)
 
-    if variant != 's':
-        # On entry to SLASCL parameter number  4 had an illegal value
-        # under codspeed (cannot repro locally or on CI w/o codspeed)
-        # https://github.com/OpenMathLib/OpenBLAS/issues/4776
-        assert info == 0
-
-        atol = {'s': 1e-5, 'd': 1e-13}
-        np.testing.assert_allclose(u @ np.diag(s) @ vt, a, atol=atol[variant])
+    assert info == 0
+
+    atol = {'s': 1e-5, 'd': 1e-13}
+    np.testing.assert_allclose(u @ np.diag(s) @ vt, a, atol=atol[variant])
 
 
 # linalg.eigh

c_check

@@ -356,6 +356,9 @@ if [ "$compiler" = "GCC" ]; then
         no_avx2=0
         oldgcc=0
         data=`$compiler_name -dumpversion`
+        case "$data" in *-*)
+            data="${data%-*}"
+        esac
         case "$data" in *.*.*)
             data="${data%.*}"
         esac

cmake/arch.cmake

@@ -46,7 +46,7 @@ if (DYNAMIC_ARCH)
   if (ARM64)
 	  set(DYNAMIC_CORE ARMV8 CORTEXA53 CORTEXA57 THUNDERX THUNDERX2T99 TSV110 EMAG8180 NEOVERSEN1 THUNDERX3T110)
     if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 9.99)
-          set(DYNAMIC_CORE ${DYNAMIC_CORE} NEOVERSEV1 NEOVERSEN2 ARMV8SVE)
+	    set(DYNAMIC_CORE ${DYNAMIC_CORE} NEOVERSEV1 NEOVERSEN2 ARMV8SVE A64FX)
     endif ()
     if (DYNAMIC_LIST)
 	  set(DYNAMIC_CORE ARMV8 ${DYNAMIC_LIST})

cmake/prebuild.cmake

@@ -1218,6 +1218,37 @@ endif ()
     set(ZGEMM_UNROLL_M 4)
     set(ZGEMM_UNROLL_N 4)
     set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "A64FX")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_CODE_SIZE\t65536\n"
+      "#define L1_CODE_LINESIZE\t256\n"
+      "#define L1_CODE_ASSOCIATIVE\t8\n"
+      "#define L1_DATA_SIZE\t32768\n"
+      "#define L1_DATA_LINESIZE\t256\n"
+      "#define L1_DATA_ASSOCIATIVE\t8\n"
+      "#define L2_SIZE\t8388608\n\n"
+      "#define L2_LINESIZE\t256\n"
+      "#define L2_ASSOCIATIVE\t8\n"
+      "#define L3_SIZE\t0\n\n"
+      "#define L3_LINESIZE\t0\n\n"
+      "#define L3_ASSOCIATIVE\t0\n\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define HAVE_VFPV4\n"
+      "#define HAVE_VFPV3\n"
+      "#define HAVE_VFP\n"
+      "#define HAVE_NEON\n"
+      "#define HAVE_SVE\n"
+      "#define ARMV8\n")
+    set(SGEMM_UNROLL_M 4)
+    set(SGEMM_UNROLL_N 8)
+    set(DGEMM_UNROLL_M 2)
+    set(DGEMM_UNROLL_N 8)
+    set(CGEMM_UNROLL_M 2)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 2)
+    set(ZGEMM_UNROLL_N 4)
+    set(SYMV_P 16)
   elseif ("${TCORE}" STREQUAL "P5600")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L2_SIZE 1048576\n"

cmake/system.cmake

@@ -310,6 +310,18 @@ if (${TARGET} STREQUAL NEOVERSEV1)
       set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv8.2-a+sve")
     endif()
   endif()
+  if (${TARGET} STREQUAL A64FX)
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
+      set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -Msve-intrinsics -march=armv8.2-a+sve -mtune=a64fx")
+    else ()
+      execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
+      if (${GCC_VERSION} VERSION_GREATER 10.4 OR ${GCC_VERSION} VERSION_EQUAL 10.4)
+        set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv8.2-a+sve -mtune=a64fx")
+      else ()
+        message(FATAL_ERROR "Compiler $${CMAKE_C_COMPILER} {GCC_VERSION} does not support A64FX.")
+      endif()
+    endif()
+  endif()
 
 endif()
 
@@ -379,6 +391,13 @@ endif ()
 if (X86_64 OR ${CORE} STREQUAL POWER10)
   set(SMALL_MATRIX_OPT TRUE)
 endif ()
+if (ARM64)
+  set(GEMM_GEMV_FORWARD TRUE)
+endif ()
+
+if (GEMM_GEMV_FORWARD AND NOT ONLY_CBLAS)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DGEMM_GEMV_FORWARD")
+endif ()
 if (SMALL_MATRIX_OPT)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DSMALL_MATRIX_OPT")
 endif ()

ctest/Makefile

@@ -26,7 +26,7 @@ endif
 override CFLAGS += -DADD$(BU) -DCBLAS
 ifeq ($(F_COMPILER),GFORTRAN)
 ifneq (, $(filter $(CORE),LOONGSON3R3 LOONGSON3R4))
-	override FFLAGS = $(filter_out(-O2 -O3,$(FFLAGS)) -O0
+	override FFLAGS = $(filter_out(-O2 -O3,$(FFLAGS))) -O0
 endif
 	override FFLAGS += -fno-tree-vectorize
 endif

docs/install.md

@@ -245,7 +245,7 @@ newer installed.
     On Windows 11 with Visual Studio 2022, this would be done by invoking:
     
     ```shell
-    "c:\Program Files\Microsoft Visual Studio\2022\Preview\vc\Auxiliary\Build\vcvars64.bat"
+    "c:\Program Files\Microsoft Visual Studio\2022\Community\vc\Auxiliary\Build\vcvars64.bat"
     ```
    
     With VS2019, the command should be the same (except for the year number of course).

driver/others/Makefile

@@ -30,12 +30,16 @@ else
 ifeq ($(ARCH),loongarch64)
 COMMONOBJS += dynamic_loongarch64.$(SUFFIX)
 else
+ifeq ($(ARCH),riscv64)
+COMMONOBJS += dynamic_riscv64.$(SUFFIX) detect_riscv64.$(SUFFIX)
+else
 COMMONOBJS	+=  dynamic.$(SUFFIX)
 endif
 endif
 endif
 endif
 endif
+endif
 else
 COMMONOBJS	+=  parameter.$(SUFFIX)
 endif
@@ -106,12 +110,16 @@ else
 ifeq ($(ARCH),loongarch64)
 HPLOBJS = memory.$(SUFFIX) xerbla.$(SUFFIX) dynamic_loongarch64.$(SUFFIX)
 else
+ifeq ($(ARCH),riscv64)
+HPLOBJS = memory.$(SUFFIX) xerbla.$(SUFFIX) dynamic_riscv64.$(SUFFIX) detect_riscv64.$(SUFFIX)
+else
 HPLOBJS = memory.$(SUFFIX) xerbla.$(SUFFIX) dynamic.$(SUFFIX)
 endif
 endif
 endif
 endif
 endif
+endif
 else
 HPLOBJS = memory.$(SUFFIX) xerbla.$(SUFFIX) parameter.$(SUFFIX)
 endif
@@ -209,6 +217,9 @@ addx.$(SUFFIX) : $(ARCH)/addx.c
 mulx.$(SUFFIX) : $(ARCH)/mulx.c
 	$(CC) $(CFLAGS) -c -DXDOUBLE -UCOMPLEX $< -o $(@F)
 
+detect_riscv64.$(SUFFIX): detect_riscv64.c
+	$(CC) $(CFLAGS) -c -march=rv64imafdcv $< -o $(@F)
+
 xerbla.$(PSUFFIX) : xerbla.c
 	$(CC) $(PFLAGS) -c $< -o $(@F)
 

driver/others/detect_riscv64.c

@@ -0,0 +1,75 @@
+/*****************************************************************************
+Copyright (c) 2024, 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 COPYRIGHT OWNER 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.
+**********************************************************************************/
+
+#include <stdint.h>
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+unsigned detect_riscv64_get_vlenb(void) {
+#ifdef __riscv_v_intrinsic
+	return __riscv_vlenb();
+#else
+	return 0;
+#endif
+}
+
+/*
+ * Based on the approach taken here:
+ * https://code.videolan.org/videolan/dav1d/-/merge_requests/1629
+ *
+ * Only to be called after we've determined we have some sort of
+ * RVV support.
+ */
+
+uint64_t detect_riscv64_rvv100(void)
+{
+	uint64_t rvv10_supported;
+
+	/*
+	 * After the vsetvli statement vtype will either be a value > 0 if the
+	 * vsetvli succeeded or less than 0 if it failed.  If 0 < vtype
+	 * we're good and the function will return 1, otherwise there's no
+	 * RVV 1.0 and we return 0.
+	 */
+
+	asm volatile("vsetvli x0, x0, e8, m1, ta, ma\n\t"
+		     "csrr %0, vtype\n\t"
+		     "slt %0, x0, %0\n"
+		     : "=r" (rvv10_supported)
+		     :
+		     :);
+
+	return rvv10_supported;
+}
+

driver/others/dynamic_arm64.c

@@ -120,6 +120,11 @@ extern gotoblas_t  gotoblas_CORTEXA55;
 #else
 #define gotoblas_CORTEXA55 gotoblas_ARMV8
 #endif
+#ifdef DYN_A64FX
+extern gotoblas_t gotoblas_A64FX;
+#else
+#define gotoblas_A64FX gotoblas_ARMV8
+#endif
 #else
 extern gotoblas_t  gotoblas_CORTEXA53;
 #define gotoblas_CORTEXA55 gotoblas_CORTEXA53
@@ -136,10 +141,12 @@ extern gotoblas_t  gotoblas_NEOVERSEN1;
 extern gotoblas_t  gotoblas_NEOVERSEV1;
 extern gotoblas_t  gotoblas_NEOVERSEN2;
 extern gotoblas_t  gotoblas_ARMV8SVE;
+extern gotoblas_t  gotoblas_A64FX;
 #else
 #define gotoblas_NEOVERSEV1 gotoblas_ARMV8
 #define gotoblas_NEOVERSEN2 gotoblas_ARMV8
 #define gotoblas_ARMV8SVE   gotoblas_ARMV8
+#define gotoblas_A64FX      gotoblas_ARMV8
 #endif
 extern gotoblas_t  gotoblas_THUNDERX3T110;
 #endif
@@ -149,7 +156,7 @@ extern void openblas_warning(int verbose, const char * msg);
 #define FALLBACK_VERBOSE 1
 #define NEOVERSEN1_FALLBACK "OpenBLAS : Your OS does not support SVE instructions. OpenBLAS is using Neoverse N1 kernels as a fallback, which may give poorer performance.\n"
 
-#define NUM_CORETYPES   17
+#define NUM_CORETYPES   18
 
 /*
  * In case asm/hwcap.h is outdated on the build system, make sure
@@ -184,6 +191,7 @@ static char *corename[] = {
   "thunderx3t110",
   "cortexa55",
   "armv8sve",
+  "a64fx",
   "unknown"
 };
 
@@ -205,6 +213,7 @@ char *gotoblas_corename(void) {
   if (gotoblas == &gotoblas_THUNDERX3T110) return corename[14];
   if (gotoblas == &gotoblas_CORTEXA55)    return corename[15];
   if (gotoblas == &gotoblas_ARMV8SVE)     return corename[16];
+  if (gotoblas == &gotoblas_A64FX)        return corename[17];
   return corename[NUM_CORETYPES];
 }
 
@@ -241,6 +250,7 @@ static gotoblas_t *force_coretype(char *coretype) {
     case 14: return (&gotoblas_THUNDERX3T110);
     case 15: return (&gotoblas_CORTEXA55);
     case 16: return (&gotoblas_ARMV8SVE);
+    case 17: return (&gotoblas_A64FX);
   }
   snprintf(message, 128, "Core not found: %s\n", coretype);
   openblas_warning(1, message);
@@ -346,6 +356,15 @@ static gotoblas_t *get_coretype(void) {
           return &gotoblas_THUNDERX3T110;
       }
       break;
+    case 0x46: // Fujitsu
+      switch (part)
+      {
+#ifndef NO_SVE
+        case 0x001: // A64FX
+          return &gotoblas_A64FX;
+#endif
+      }
+      break;
     case 0x48: // HiSilicon
       switch (part)
       {

driver/others/dynamic_riscv64.c

@@ -0,0 +1,269 @@
+/*****************************************************************************
+Copyright (c) 2024, 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 COPYRIGHT OWNER 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.
+**********************************************************************************/
+
+#include <stdbool.h>
+
+#include "common.h"
+
+/*
+ * OpenBLAS contains some kernels that are optimised for RVV 1.0.  Before we
+ * can use these kernels we need to determine whether the device supports
+ * RVV 1.0 and what the device's VLEN is.  Our strategy will be as follows.
+ *
+ * First we'll invoke the hwprobe syscall to detect RVV 1.0.  In an ideal world,
+ * this is all we should need to do.  If the syscall is not implemented we
+ * should be able to deduce that RVV 1.0 is not supported (as it was added to
+ * Linux after hwprobe) and if the syscall is implemented we can use it to
+ * determine whether RVV 1.0 is supported.  However, there are some riscv64
+ * boards out there that implement RVV 1.0 but ship with a Linux kernel that
+ * predates RVV vector support and hwprobe support.  These kernels contain
+ * the backported RVV patches but not the hwprobe patches and so they
+ * advertise support for RVV via hwcap.  To cater for these boards we need
+ * to fall back to hwcap if hwprobe is not supported. Unfortunately, some
+ * boards indicate support for RVV via hwcap even though they only support
+ * RVV 0.7.1, which is incompatible with RVV 1.0.  So an additional check is
+ * required to test if the devices advertising support for RVV via hwcap really
+ * support RVV 1.0.  This test works by executing a vsetvli instruction that
+ * sets the tail agnostic and mask agnostic bits in the vtype register.
+ * These bits are not supported prior to RVV 0.9 so will cause the VIL bit to
+ * be set on the VTYPE register in CPUs supporting 0.7.1.  If this bit is set
+ * we can determine that RVV 1.0 is not supported.
+ *
+ * This approach is borrowed from
+ * VideoLan dav1d:
+ *   (https://code.videolan.org/videolan/dav1d/-/merge_requests/1629).
+ *
+ * We assume that if a kernel reports the presence of RVV via hwcap that
+ * the device supports the vsetvli instruction.
+ *
+ * For now we're just going to invoke the hwprobe syscall directly, rather than
+ * invoking it through glibc.  Support for hwprobe has been added to glibc but
+ * at the time of writing this support has not yet been included in a glibc
+ * release.  Once it has, it will be better to invoke hwprobe via glibc as doing
+ * so should take advantage of the vdso entry and be more efficient.
+ */
+
+/*
+ * This should work on Android as well but I have no way of testing.
+ */
+
+#if defined(OS_LINUX)
+#include <unistd.h>
+#include <sys/syscall.h>
+#include <stdint.h>
+#include <sys/auxv.h>
+
+#define DETECT_RISCV64_HWCAP_ISA_V (1 << ('V' - 'A'))
+
+struct riscv_hwprobe {
+	int64_t key;
+	uint64_t value;
+};
+
+/* The constants below are copied from
+ * /usr/include/riscv64-linux-gnu/asm/hwprobe.h. We duplicate the
+ *  constants as the header file from which they are copied will only
+ *  be present if we're building on a device with Linux 6.5 or greater.
+ */
+
+#define RISCV_HWPROBE_KEY_IMA_EXT_0	4
+#define		RISCV_HWPROBE_IMA_V		(1 << 2)
+
+#ifndef NR_riscv_hwprobe
+#ifndef NR_arch_specific_syscall
+#define NR_arch_specific_syscall 244
+#endif
+#define NR_riscv_hwprobe (NR_arch_specific_syscall + 14)
+#endif
+#endif // defined(OS_LINUX)
+
+unsigned detect_riscv64_get_vlenb(void);
+uint64_t detect_riscv64_rvv100(void);
+
+extern gotoblas_t gotoblas_RISCV64_GENERIC;
+#if !defined(DYNAMIC_LIST) || defined(DYN_RISCV64_ZVL256B)
+extern gotoblas_t gotoblas_RISCV64_ZVL256B;
+#endif
+#if !defined(DYNAMIC_LIST) || defined(DYN_RISCV64_ZVL128B)
+extern gotoblas_t gotoblas_RISCV64_ZVL128B;
+#endif
+
+#define CPU_GENERIC         0
+#define CPU_RISCV64_ZVL256B 1
+#define CPU_RISCV64_ZVL128B 2
+
+static char *cpuname[] = {
+	"riscv64_generic",
+	"riscv64_zvl256b",
+	"riscv64_zvl128b"
+};
+#define NUM_CORETYPES (sizeof(cpuname)/sizeof(char*))
+
+extern int openblas_verbose(void);
+extern void openblas_warning(int verbose, const char* msg);
+
+char* gotoblas_corename(void) {
+#if !defined(DYNAMIC_LIST) || defined(DYN_RISCV64_ZVL256B)
+	if (gotoblas == &gotoblas_RISCV64_ZVL256B)
+		return cpuname[CPU_RISCV64_ZVL256B];
+#endif
+#if !defined(DYNAMIC_LIST) || defined(DYN_RISCV64_ZVL128B)
+	if (gotoblas == &gotoblas_RISCV64_ZVL128B)
+		return cpuname[CPU_RISCV64_ZVL128B];
+#endif
+	if (gotoblas == &gotoblas_RISCV64_GENERIC)
+		return cpuname[CPU_GENERIC];
+
+	return "unknown";
+}
+
+static gotoblas_t* get_coretype(void) {
+	unsigned vlenb = 0;
+
+#if !defined(OS_LINUX)
+	return NULL;
+#else
+
+	/*
+	 * See the hwprobe documentation
+	 *
+	 * ( https://docs.kernel.org/arch/riscv/hwprobe.html )
+	 * for more details.
+	 */
+
+	struct riscv_hwprobe pairs[] = {
+		{ .key = RISCV_HWPROBE_KEY_IMA_EXT_0, },
+	};
+	int ret = syscall(NR_riscv_hwprobe, pairs, 1, 0, NULL, 0);
+	if (ret == 0) {
+		if (!(pairs[0].value & RISCV_HWPROBE_IMA_V))
+			return NULL;
+	} else {
+		if (!(getauxval(AT_HWCAP) & DETECT_RISCV64_HWCAP_ISA_V))
+			return NULL;
+
+		if (!detect_riscv64_rvv100())
+			return NULL;
+	}
+
+	/*
+	 * RVV 1.0 is supported.  We now just need to determine the coretype
+	 * based on the VLEN.
+	 */
+
+	vlenb = detect_riscv64_get_vlenb();
+
+	if (vlenb < 16)
+		return NULL;
+#if !defined(DYNAMIC_LIST) || defined(DYN_RISCV64_ZVL256B)
+	if (vlenb >= 32)
+		return &gotoblas_RISCV64_ZVL256B;
+#endif
+
+#if !defined(DYNAMIC_LIST) || defined(DYN_RISCV64_ZVL128B)
+	return &gotoblas_RISCV64_ZVL128B;
+#else
+	return NULL;
+#endif
+
+#endif  // !defined(OS_LINUX)
+}
+
+static gotoblas_t* force_coretype(char* coretype) {
+	size_t i;
+	char message[128];
+
+	for (i = 0; i < NUM_CORETYPES && strcasecmp(coretype, cpuname[i]); i++);
+
+	if (i == CPU_GENERIC)
+		return &gotoblas_RISCV64_GENERIC;
+
+	if (i == CPU_RISCV64_ZVL256B) {
+#if !defined(DYNAMIC_LIST) || defined(DYN_RISCV64_ZVL256B)
+		return &gotoblas_RISCV64_ZVL256B;
+#else
+		openblas_warning(1,
+				 "riscv64_zvl256b support not compiled in\n");
+		return NULL;
+#endif
+	}
+
+	if (i == CPU_RISCV64_ZVL128B) {
+#if !defined(DYNAMIC_LIST) || defined(DYN_RISCV64_ZVL128B)
+		return &gotoblas_RISCV64_ZVL128B;
+#else
+		openblas_warning(1,
+				 "riscv64_zvl128b support not compiled in\n");
+		return NULL;
+#endif
+	}
+
+	snprintf(message, sizeof(message), "Core not found: %s\n", coretype);
+	openblas_warning(1, message);
+
+	return NULL;
+}
+
+void gotoblas_dynamic_init(void) {
+
+	char coremsg[128];
+	char* p;
+
+	if (gotoblas) return;
+
+	p = getenv("OPENBLAS_CORETYPE");
+	if (p)
+		gotoblas = force_coretype(p);
+	else
+		gotoblas = get_coretype();
+
+	if (!gotoblas) {
+		snprintf(coremsg, sizeof(coremsg), "Falling back to generic riscv64 core\n");
+		openblas_warning(1, coremsg);
+		gotoblas = &gotoblas_RISCV64_GENERIC;
+	}
+
+	if (gotoblas->init) {
+		snprintf(coremsg, sizeof(coremsg), "Core: %s\n",
+			 gotoblas_corename());
+		openblas_warning(2, coremsg);
+		gotoblas->init();
+		return;
+	}
+
+	openblas_warning(0, "OpenBLAS : Architecture Initialization failed. No initialization function found.\n");
+	exit(1);
+}
+
+void gotoblas_dynamic_quit(void) {
+	gotoblas = NULL;
+}

interface/gemm.c

@@ -1,4 +1,5 @@
 /*********************************************************************/
+/* Copyright 2024 The OpenBLAS Project                               */
 /* Copyright 2009, 2010 The University of Texas at Austin.           */
 /* All rights reserved.                                              */
 /*                                                                   */
@@ -47,12 +48,16 @@
 #define SMP_THRESHOLD_MIN 65536.0
 #ifdef XDOUBLE
 #define ERROR_NAME "QGEMM "
+#define GEMV BLASFUNC(qgemv)
 #elif defined(DOUBLE)
 #define ERROR_NAME "DGEMM "
+#define GEMV BLASFUNC(dgemv)
 #elif defined(BFLOAT16)
 #define ERROR_NAME "SBGEMM "
+#define GEMV BLASFUNC(sbgemv)
 #else
 #define ERROR_NAME "SGEMM "
+#define GEMV BLASFUNC(sgemv)
 #endif
 #else
 #define SMP_THRESHOLD_MIN 8192.0
@@ -493,6 +498,52 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
 	 args.m, args.n, args.k, args.lda, args.ldb, args.ldc);
 #endif
 
+#if defined(GEMM_GEMV_FORWARD) && !defined(GEMM3M) && !defined(COMPLEX)
+  // Check if we can convert GEMM -> GEMV
+  if (args.k != 0) {
+    if (args.n == 1) {
+      blasint inc_x = 1;
+      blasint inc_y = 1;
+      // These were passed in as blasint, but the struct translates them to blaslong
+      blasint m = args.m;
+      blasint n = args.k;
+      blasint lda = args.lda;
+      // Create new transpose parameters
+      char NT = 'N';
+      if (transa & 1) {
+        NT = 'T';
+        m = args.k;
+        n = args.m;
+      }
+      if (transb & 1) {
+        inc_x = args.ldb;
+      }
+      GEMV(&NT, &m, &n, args.alpha, args.a, &lda, args.b, &inc_x, args.beta, args.c, &inc_y);
+      return;
+    }
+    if (args.m == 1) {
+      blasint inc_x = args.lda;
+      blasint inc_y = args.ldc;
+      // These were passed in as blasint, but the struct translates them to blaslong
+      blasint m = args.k;
+      blasint n = args.n;
+      blasint ldb = args.ldb;
+      // Create new transpose parameters
+      char NT = 'T';
+      if (transa & 1) {
+        inc_x = 1;
+      }
+      if (transb & 1) {
+        NT = 'N';
+        m = args.n;
+        n = args.k;
+      }
+      GEMV(&NT, &m, &n, args.alpha, args.b, &ldb, args.a, &inc_x, args.beta, args.c, &inc_y);
+      return;
+    }
+  }
+#endif
+
   IDEBUG_START;
 
   FUNCTION_PROFILE_START();

interface/scal.c

@@ -85,7 +85,7 @@ void CNAME(blasint n, FLOAT alpha, FLOAT *x, blasint incx){
   if (nthreads == 1) {
 #endif
 
-  SCAL_K(n, 0, 0, alpha, x, incx, NULL, 0, NULL, 0);
+  SCAL_K(n, 0, 0, alpha, x, incx, NULL, 0, NULL, 1);
 
 #ifdef SMP
   } else {
@@ -102,7 +102,7 @@ void CNAME(blasint n, FLOAT alpha, FLOAT *x, blasint incx){
 #else
 		       &alpha,
 #endif
-		       x, incx, NULL, 0, NULL, 0,  (int (*)(void))SCAL_K, nthreads);
+		       x, incx, NULL, 0, NULL, 1,  (int (*)(void))SCAL_K, nthreads);
 
   }
 #endif

kernel/arm/scal.c

@@ -43,9 +43,22 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 	if ( (n <= 0) || (inc_x <= 0))
 		return(0);
 	
+	if (dummy2 == 0) {
+		while(j < n)
+		{
 
-	while(j < n)
-	{
+		if ( da == 0.0 )
+			x[i]=0.0;
+		else
+			x[i] = da * x[i] ;
+
+		i += inc_x ;
+		j++;
+		}
+	} else {
+	
+		while(j < n)
+		{
 
 		if ( da == 0.0 )
 			if (!isnan(x[i]) && !isinf(x[i])) {
@@ -59,6 +72,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 		i += inc_x ;
 		j++;
 
+		}
 	}
 	return 0;
 

kernel/arm64/KERNEL.A64FX

@@ -1 +1,6 @@
 include $(KERNELDIR)/KERNEL.ARMV8SVE
+
+SGEMVNKERNEL = gemv_n_sve.c
+DGEMVNKERNEL = gemv_n_sve.c
+SGEMVTKERNEL = gemv_t_sve.c
+DGEMVTKERNEL = gemv_t_sve.c

kernel/arm64/KERNEL.ARMV8SVE

@@ -131,6 +131,16 @@ SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
 SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
 SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
+SGEMM_SMALL_M_PERMIT = gemm_small_kernel_permit_sve.c
+SGEMM_SMALL_K_NT = sgemm_small_kernel_nt_sve.c
+SGEMM_SMALL_K_B0_NT = sgemm_small_kernel_nt_sve.c
+SGEMM_SMALL_K_NN = sgemm_small_kernel_nn_sve.c
+SGEMM_SMALL_K_B0_NN = sgemm_small_kernel_nn_sve.c
+SGEMM_SMALL_K_TT = sgemm_small_kernel_tt_sve.c
+SGEMM_SMALL_K_B0_TT = sgemm_small_kernel_tt_sve.c
+SGEMM_SMALL_K_TN = sgemm_small_kernel_tn_sve.c
+SGEMM_SMALL_K_B0_TN = sgemm_small_kernel_tn_sve.c
+
 STRMMUNCOPY_M  =  trmm_uncopy_sve_v1.c
 STRMMLNCOPY_M  =  trmm_lncopy_sve_v1.c
 STRMMUTCOPY_M  =  trmm_utcopy_sve_v1.c
@@ -152,6 +162,16 @@ DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
 DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
 DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
+DGEMM_SMALL_M_PERMIT = gemm_small_kernel_permit_sve.c
+DGEMM_SMALL_K_NT = dgemm_small_kernel_nt_sve.c
+DGEMM_SMALL_K_B0_NT = dgemm_small_kernel_nt_sve.c
+DGEMM_SMALL_K_NN = dgemm_small_kernel_nn_sve.c
+DGEMM_SMALL_K_B0_NN = dgemm_small_kernel_nn_sve.c
+DGEMM_SMALL_K_TT = dgemm_small_kernel_tt_sve.c
+DGEMM_SMALL_K_B0_TT = dgemm_small_kernel_tt_sve.c
+DGEMM_SMALL_K_TN = dgemm_small_kernel_tn_sve.c
+DGEMM_SMALL_K_B0_TN = dgemm_small_kernel_tn_sve.c
+
 DTRMMUNCOPY_M  =  trmm_uncopy_sve_v1.c
 DTRMMLNCOPY_M  =  trmm_lncopy_sve_v1.c
 DTRMMUTCOPY_M  =  trmm_utcopy_sve_v1.c

kernel/arm64/KERNEL.NEOVERSEV1

@@ -1 +1,4 @@
 include $(KERNELDIR)/KERNEL.ARMV8SVE
+
+SGEMVTKERNEL = gemv_t_sve.c
+DGEMVTKERNEL = gemv_t_sve.c

kernel/arm64/dgemm_small_kernel_nn_sve.c

@@ -0,0 +1,742 @@
+/***************************************************************************
+Copyright (c) 2024, 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.
+*****************************************************************************/
+
+#include "common.h"
+
+#include <arm_neon.h>
+#include <arm_sve.h>
+#if defined(__ARM_NEON_SVE_BRIDGE) && defined(__has_include) &&                \
+  __has_include(<arm_neon_sve_bridge.h>)
+#include <arm_neon_sve_bridge.h>
+#else
+#define svdup_neonq_f32(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat32_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#define svdup_neonq_f64(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat64_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#endif
+
+#define RESET_A_POINTER() a_offset = A;
+
+#define CREATE_A_POINTER(m, scale) FLOAT* a_offset##m = a_offset + scale;
+#define UPDATE_A_POINTER(scale) a_offset = a_offset + scale;
+#define A_ELEMENT_K(m, offset_k) *(a_offset##m + (k + offset_k) * lda)
+#define A_ELEMENT(m) A_ELEMENT_K(m, 0)
+
+#define RESET_B_POINTER() b_offset = B;
+
+#define CREATE_B_POINTER(n, scale) FLOAT* b_offset##n = b_offset + scale * ldb;
+#define UPDATE_B_POINTER(scale) b_offset = b_offset + scale * ldb;
+#define B_ELEMENT_K(n, offset_k) *(b_offset##n + (k + offset_k))
+#define B_ELEMENT(n) B_ELEMENT_K(n, 0)
+
+#define CREATE_C_POINTER(n, scale) FLOAT* c_offset##n = c_offset + scale * ldc;
+#define INCR_C_POINTER(m, incr) // c_offset ## m += incr;
+#define UPDATE_C_POINTER(scale) c_offset = c_offset + scale * ldc;
+#define C_ELEMENT(m, n) *(c_offset##n + ((m * v_size) + i))
+
+// #undef C_ELEMENT
+// #define C_ELEMENT(m, n)             C[(i+(m))+(j+(n))*ldc]
+
+#define PACK_ELEMENT_K(n, offset_k) packed_b[(k + offset_k) * 4 + n]
+#define PACK_ELEMENT(n) PACK_ELEMENT_K(n, 0)
+
+// ASIMD
+#define DECLARE_RESULT_VECTOR2(m, n)                                           \
+  float64x2_t result##m##n = vdupq_n_f64(0.0);
+#define DECLARE_RESULT(m, n) float64_t result##m##n = 0.0;
+#define BROADCAST_LOAD_A2(m, offset_k)                                         \
+  float64x2_t a##m##_k##offset_k = vld1q_dup_f64(&A_ELEMENT_K(m, offset_k));
+#define LOAD_A1(m, offset_k)                                                   \
+  float64_t a##m##_k##offset_k = A_ELEMENT_K(m, offset_k);
+#define VECTOR_LOAD_B_K2(n, offset_k)                                          \
+  float64x2_t b##k##n##_k##offset_k = vld1q_f64(&B_ELEMENT_K(n, offset_k));
+#define TRANSPOSE_B2_K2(n0, n1, offset_k0, offset_k1)                          \
+  float64x2_t b##n0##_k##offset_k0 =                                           \
+    vzip1q_f64(b##k##n0##_k##offset_k0, b##k##n1##_k##offset_k0);              \
+  float64x2_t b##n0##_k##offset_k1 =                                           \
+    vzip2q_f64(b##k##n0##_k##offset_k0, b##k##n1##_k##offset_k0);
+
+#define SCALE_B2_K2(n0, offset_k0, offset_k1)                                  \
+  svfloat64_t b##s##n0##_k##offset_k0 = svdup_neonq_f64(b##n0##_k##offset_k0); \
+  svfloat64_t b##s##n0##_k##offset_k1 = svdup_neonq_f64(b##n0##_k##offset_k1);
+#define GATHER_LOAD_B2(n, offset_k)                                            \
+  float64x2_t b##n##_k##offset_k = vdupq_n_f64(B_ELEMENT_K(n, offset_k));      \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f64(B_ELEMENT_K(n + 1, offset_k), b##n##_k##offset_k, 1);
+#define VECTOR_UNPACK_B2(n, offset_k)                                          \
+  float64x2_t b##n##_k##offset_k = vld1q_f64(&PACK_ELEMENT_K(n, offset_k));
+#define VECTOR_PACK_B2(n, offset_k)                                            \
+  vst1q_f64(&PACK_ELEMENT_K(n, offset_k), b##n##_k##offset_k);
+#define PACK_B0(n, offset_k)                                                   \
+  PACK_ELEMENT_K(n, offset_k) = vget_lane_f64(b##n##_k##offset_k, 0);
+#define UPDATE_RESULT_VECTOR2(m, n, offset_k)                                  \
+  result##m##n =                                                               \
+    vfmaq_f64(result##m##n, a##m##_k##offset_k, b##n##_k##offset_k);
+#define UPDATE_RESULT(m, n, offset_k)                                          \
+  result##m##n = result##m##n + a##m##_k##offset_k * b##n##_k##offset_k;
+#ifdef B0
+#define SCATTER_STORE2(m, n)                                                   \
+  result##m##n = vmulq_f64(result##m##n, vdupq_n_f64(alpha));                  \
+  C_ELEMENT(m, n + 0) = vgetq_lane_f64(result##m##n, 0);                       \
+  C_ELEMENT(m, n + 1) = vgetq_lane_f64(result##m##n, 1);
+#else
+#define SCATTER_STORE2(m, n)                                                   \
+  result##m##n = vmulq_f64(result##m##n, vdupq_n_f64(alpha));                  \
+  C_ELEMENT(m, n + 0) =                                                        \
+    C_ELEMENT(m, n + 0) * beta + vgetq_lane_f64(result##m##n, 0);              \
+  C_ELEMENT(m, n + 1) =                                                        \
+    C_ELEMENT(m, n + 1) * beta + vgetq_lane_f64(result##m##n, 1);
+#endif
+
+// SVE
+#define DECLARE_RESULT_VECTOR(m, n) svfloat64_t result##m##n = svdup_f64(0.0);
+#define BROADCAST_LOAD_A(m, offset_k)                                          \
+  svfloat64_t a##s##m##_k##offset_k = svdup_f64(A_ELEMENT_K(m, offset_k));
+#define BROADCAST_LOAD_B(n, offset_k)                                          \
+  svfloat64_t b##s##n##_k##offset_k = svdup_f64(B_ELEMENT_K(n, offset_k));
+#define VECTOR_LOAD_A(pg, m, offset_k)                                         \
+  svfloat64_t a##s##m##_k##offset_k = svld1(pg, &A_ELEMENT_K(m, offset_k));
+#define QUADWORD_LOAD_B(n, offset_k)                                           \
+  svfloat64_t b##s##n##_k##offset_k =                                          \
+    svld1rq(pg_true, &B_ELEMENT_K(n, offset_k));
+#define PACK_B(n, offset_k)                                                    \
+  svst1(pg_first, &PACK_ELEMENT_K(n, offset_k), b##s##n##_k##offset_k);
+#define VECTOR_PACK_B(n, offset_k)                                             \
+  svst1(pg_true, &PACK_ELEMENT_K(n* v_size, offset_k), b##s##n##_k##offset_k);
+#define QUADWORD_PACK_B(n, offset_k)                                           \
+  svst1(pg_quad, &PACK_ELEMENT_K(n, offset_k), b##s##n##_k##offset_k);
+#define UNPACK_VECTOR_B(n, offset_k)                                           \
+  svfloat64_t b##s##n##_k##offset_k =                                          \
+    svld1(pg_true, &PACK_ELEMENT_K(n * v_size, offset_k));
+#define UNPACK_BROADCAST_B(n, offset_k)                                        \
+  svfloat64_t b##s##n##_k##offset_k = svdup_f64(PACK_ELEMENT_K(n, offset_k));
+#define UNPACK_QUADWORD_B(n, offset_k)                                         \
+  svfloat64_t b##s##n##_k##offset_k =                                          \
+    svld1rq(pg_true, &PACK_ELEMENT_K(n, offset_k));
+#define UPDATE_RESULT_VECTOR(pg, m, n, offset_k)                               \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, a##s##m##_k##offset_k, b##s##n##_k##offset_k);
+#define UPDATE_RESULT_VECTOR_QUADWORD(m, n, outer, lane, offset_k)             \
+  result##m##n = svmla_lane(                                                   \
+    result##m##n, a##s##m##_k##offset_k, b##s##outer##_k##offset_k, lane);
+#ifdef B0
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#else
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, svld1(pg, &C_ELEMENT(m, n)), beta_vec);          \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n = svmla_m(pg,                                                   \
+                         result##m##n,                                         \
+                         svld1_gather_index(pg, &C_ELEMENT(m, n), ldc_vec),    \
+                         beta_vec);                                            \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#endif
+
+#ifndef LIKELY
+#ifdef __GNUC__
+#define LIKELY(x) __builtin_expect(!!(x), 1)
+#else
+#define LIKELY(x) (x)
+#endif
+#endif
+
+#ifdef B0
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT* C,
+      BLASLONG ldc)
+#else
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT beta,
+      FLOAT* C,
+      BLASLONG ldc)
+#endif
+{
+  const uint64_t v_size = svcntd();
+  const uint64_t v_size2 = v_size * 2;
+  const svbool_t pg_true = svptrue_b64();
+  const svbool_t pg_quad = svwhilelt_b64(0, 2);
+  const svbool_t pg_first = svwhilelt_b64(0, 1);
+  const svfloat64_t alpha_vec = svdup_f64(alpha);
+#ifndef B0
+  const svfloat64_t beta_vec = svdup_f64(beta);
+#endif
+  const BLASLONG n4 = N & -4;
+  const BLASLONG n2 = N & -2;
+  const BLASLONG v_m2 = M & -v_size2;
+  const BLASLONG v_m1 = M & -v_size;
+  const BLASLONG k2 = K & -2;
+
+  const int pack_b = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
+  FLOAT* packed_b =
+    (pack_b) ? packed_b = (FLOAT*)malloc(K * 4 * sizeof(FLOAT)) : NULL;
+
+  FLOAT* b_offset = B;
+  FLOAT* a_offset = A;
+  FLOAT* c_offset = C;
+
+  BLASLONG j = 0;
+  for (; j < n4; j += 4) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_C_POINTER(1, 1);
+    CREATE_C_POINTER(2, 2);
+    CREATE_C_POINTER(3, 3);
+    CREATE_B_POINTER(0, 0);
+    CREATE_B_POINTER(1, 1);
+    CREATE_B_POINTER(2, 2);
+    CREATE_B_POINTER(3, 3);
+
+    BLASLONG i = 0;
+    for (; i < v_m2; i += v_size2) {
+
+      CREATE_A_POINTER(0, 0);
+      CREATE_A_POINTER(1, v_size);
+      UPDATE_A_POINTER(v_size2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+      DECLARE_RESULT_VECTOR(1, 2);
+      DECLARE_RESULT_VECTOR(1, 3);
+
+      if (LIKELY(packed_b != NULL)) {
+        if (i == 0) {
+          for (; k < k2; k += 2) {
+
+            VECTOR_LOAD_B_K2(0, 0);
+            VECTOR_LOAD_B_K2(1, 0);
+            TRANSPOSE_B2_K2(0, 1, 0, 1);
+            SCALE_B2_K2(0, 0, 1);
+            VECTOR_PACK_B2(0, 0);
+            VECTOR_PACK_B2(0, 1);
+            VECTOR_LOAD_A(pg_true, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+            VECTOR_LOAD_A(pg_true, 0, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
+            VECTOR_LOAD_B_K2(2, 0);
+            VECTOR_LOAD_B_K2(3, 0);
+            TRANSPOSE_B2_K2(2, 3, 0, 1);
+            SCALE_B2_K2(2, 0, 1);
+            VECTOR_PACK_B2(2, 0);
+            VECTOR_PACK_B2(2, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 1);
+            VECTOR_LOAD_A(pg_true, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 2, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 2, 1, 0);
+            VECTOR_LOAD_A(pg_true, 1, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 2, 0, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 2, 1, 1);
+          }
+          for (; k < K; k++) {
+
+            BROADCAST_LOAD_B(0, 0);
+            PACK_B(0, 0);
+            VECTOR_LOAD_A(pg_true, 0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+            BROADCAST_LOAD_B(1, 0);
+            PACK_B(1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+            VECTOR_LOAD_A(pg_true, 1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+            BROADCAST_LOAD_B(2, 0);
+            PACK_B(2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+            BROADCAST_LOAD_B(3, 0);
+            PACK_B(3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+          }
+        } else {
+          for (; k < K; k++) {
+
+            UNPACK_QUADWORD_B(0, 0);
+            VECTOR_LOAD_A(pg_true, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+            UNPACK_QUADWORD_B(2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+            VECTOR_LOAD_A(pg_true, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 2, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 2, 1, 0);
+          }
+        }
+      } else {
+        for (; k < k2; k += 2) {
+
+          VECTOR_LOAD_B_K2(0, 0);
+          VECTOR_LOAD_B_K2(1, 0);
+          TRANSPOSE_B2_K2(0, 1, 0, 1);
+          SCALE_B2_K2(0, 0, 1);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          VECTOR_LOAD_A(pg_true, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
+          VECTOR_LOAD_B_K2(2, 0);
+          VECTOR_LOAD_B_K2(3, 0);
+          TRANSPOSE_B2_K2(2, 3, 0, 1);
+          SCALE_B2_K2(2, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 1);
+          VECTOR_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 2, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 2, 1, 0);
+          VECTOR_LOAD_A(pg_true, 1, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 2, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 2, 1, 1);
+        }
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+          VECTOR_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+          BROADCAST_LOAD_B(2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+          BROADCAST_LOAD_B(3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      VECTOR_STORE(pg_true, 1, 2);
+      VECTOR_STORE(pg_true, 1, 3);
+      INCR_C_POINTER(0, v_size2);
+      INCR_C_POINTER(1, v_size2);
+      INCR_C_POINTER(2, v_size2);
+      INCR_C_POINTER(3, v_size2);
+    }
+    for (; i < v_m1; i += v_size) {
+
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(v_size);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      if (LIKELY(packed_b != NULL)) {
+        for (; k < K; k++) {
+
+          UNPACK_QUADWORD_B(0, 0);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UNPACK_QUADWORD_B(2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+        }
+      } else {
+        for (; k < k2; k += 2) {
+
+          VECTOR_LOAD_B_K2(0, 0);
+          VECTOR_LOAD_B_K2(1, 0);
+          TRANSPOSE_B2_K2(0, 1, 0, 1);
+          SCALE_B2_K2(0, 0, 1);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          VECTOR_LOAD_A(pg_true, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
+          VECTOR_LOAD_B_K2(2, 0);
+          VECTOR_LOAD_B_K2(3, 0);
+          TRANSPOSE_B2_K2(2, 3, 0, 1);
+          SCALE_B2_K2(2, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 1);
+        }
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+          BROADCAST_LOAD_B(2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+          BROADCAST_LOAD_B(3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      INCR_C_POINTER(0, v_size);
+      INCR_C_POINTER(1, v_size);
+      INCR_C_POINTER(2, v_size);
+      INCR_C_POINTER(3, v_size);
+    }
+    for (; i < M; i += v_size) {
+      const svbool_t pg_tail = svwhilelt_b64((uint64_t)i, (uint64_t)(M));
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(0);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      if (LIKELY(packed_b != NULL)) {
+        for (; k < K; k++) {
+
+          UNPACK_QUADWORD_B(0, 0);
+          VECTOR_LOAD_A(pg_tail, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UNPACK_QUADWORD_B(2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+        }
+      } else {
+        for (; k < k2; k += 2) {
+
+          VECTOR_LOAD_B_K2(0, 0);
+          VECTOR_LOAD_B_K2(1, 0);
+          TRANSPOSE_B2_K2(0, 1, 0, 1);
+          SCALE_B2_K2(0, 0, 1);
+          VECTOR_LOAD_A(pg_tail, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          VECTOR_LOAD_A(pg_tail, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
+          VECTOR_LOAD_B_K2(2, 0);
+          VECTOR_LOAD_B_K2(3, 0);
+          TRANSPOSE_B2_K2(2, 3, 0, 1);
+          SCALE_B2_K2(2, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 1);
+        }
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          VECTOR_LOAD_A(pg_tail, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_tail, 0, 1, 0);
+          BROADCAST_LOAD_B(2, 0);
+          UPDATE_RESULT_VECTOR(pg_tail, 0, 2, 0);
+          BROADCAST_LOAD_B(3, 0);
+          UPDATE_RESULT_VECTOR(pg_tail, 0, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      VECTOR_STORE(pg_tail, 0, 2);
+      VECTOR_STORE(pg_tail, 0, 3);
+      INCR_C_POINTER(0, 0);
+      INCR_C_POINTER(1, 0);
+      INCR_C_POINTER(2, 0);
+      INCR_C_POINTER(3, 0);
+    }
+
+    UPDATE_B_POINTER(4);
+    RESET_A_POINTER();
+    UPDATE_C_POINTER(4);
+  }
+  for (; j < n2; j += 2) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_C_POINTER(1, 1);
+    CREATE_B_POINTER(0, 0);
+    CREATE_B_POINTER(1, 1);
+
+    BLASLONG i = 0;
+    for (; i < v_m2; i += v_size2) {
+
+      CREATE_A_POINTER(0, 0);
+      CREATE_A_POINTER(1, v_size);
+      UPDATE_A_POINTER(v_size2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+
+      for (; k < k2; k += 2) {
+
+        VECTOR_LOAD_B_K2(0, 0);
+        VECTOR_LOAD_B_K2(1, 0);
+        TRANSPOSE_B2_K2(0, 1, 0, 1);
+        SCALE_B2_K2(0, 0, 1);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        VECTOR_LOAD_A(pg_true, 0, 1);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
+        VECTOR_LOAD_A(pg_true, 1, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+        VECTOR_LOAD_A(pg_true, 1, 1);
+        UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 1);
+        UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 1);
+      }
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+        BROADCAST_LOAD_B(1, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+        VECTOR_LOAD_A(pg_true, 1, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      INCR_C_POINTER(0, v_size2);
+      INCR_C_POINTER(1, v_size2);
+    }
+    for (; i < v_m1; i += v_size) {
+
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(v_size);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+
+      for (; k < k2; k += 2) {
+
+        VECTOR_LOAD_B_K2(0, 0);
+        VECTOR_LOAD_B_K2(1, 0);
+        TRANSPOSE_B2_K2(0, 1, 0, 1);
+        SCALE_B2_K2(0, 0, 1);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        VECTOR_LOAD_A(pg_true, 0, 1);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
+      }
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+        BROADCAST_LOAD_B(1, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      INCR_C_POINTER(0, v_size);
+      INCR_C_POINTER(1, v_size);
+    }
+    for (; i < M; i += v_size) {
+      const svbool_t pg_tail = svwhilelt_b64((uint64_t)i, (uint64_t)(M));
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(0);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+
+      for (; k < k2; k += 2) {
+
+        VECTOR_LOAD_B_K2(0, 0);
+        VECTOR_LOAD_B_K2(1, 0);
+        TRANSPOSE_B2_K2(0, 1, 0, 1);
+        SCALE_B2_K2(0, 0, 1);
+        VECTOR_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        VECTOR_LOAD_A(pg_tail, 0, 1);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
+      }
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+        BROADCAST_LOAD_B(1, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 1, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      INCR_C_POINTER(0, 0);
+      INCR_C_POINTER(1, 0);
+    }
+
+    UPDATE_B_POINTER(2);
+    RESET_A_POINTER();
+    UPDATE_C_POINTER(2);
+  }
+  for (; j < N; j++) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_B_POINTER(0, 0);
+
+    BLASLONG i = 0;
+    for (; i < v_m2; i += v_size2) {
+
+      CREATE_A_POINTER(0, 0);
+      CREATE_A_POINTER(1, v_size);
+      UPDATE_A_POINTER(v_size2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(1, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+        VECTOR_LOAD_A(pg_true, 1, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 1, 0);
+      INCR_C_POINTER(0, v_size2);
+    }
+    for (; i < v_m1; i += v_size) {
+
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(v_size);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      INCR_C_POINTER(0, v_size);
+    }
+    for (; i < M; i += v_size) {
+      const svbool_t pg_tail = svwhilelt_b64((uint64_t)i, (uint64_t)(M));
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(0);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      INCR_C_POINTER(0, 0);
+    }
+
+    UPDATE_B_POINTER(1);
+    RESET_A_POINTER();
+    UPDATE_C_POINTER(1);
+  }
+
+  if (pack_b)
+    free(packed_b);
+
+  return 0;
+}

kernel/arm64/dgemm_small_kernel_nt_sve.c

@@ -0,0 +1,474 @@
+/***************************************************************************
+Copyright (c) 2024, 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.
+*****************************************************************************/
+
+#include "common.h"
+
+#include <arm_neon.h>
+#include <arm_sve.h>
+#if defined(__ARM_NEON_SVE_BRIDGE) && defined(__has_include) &&                \
+  __has_include(<arm_neon_sve_bridge.h>)
+#include <arm_neon_sve_bridge.h>
+#else
+#define svdup_neonq_f32(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat32_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#define svdup_neonq_f64(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat64_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#endif
+
+#define RESET_A_POINTER() a_offset = A;
+
+#define CREATE_A_POINTER(m, scale) FLOAT* a_offset##m = a_offset + scale;
+#define UPDATE_A_POINTER(scale) a_offset = a_offset + scale;
+#define A_ELEMENT_K(m, offset_k) *(a_offset##m + (k + offset_k) * lda)
+#define A_ELEMENT(m) A_ELEMENT_K(m, 0)
+
+#define RESET_B_POINTER() b_offset = B;
+
+#define CREATE_B_POINTER(n, scale) FLOAT* b_offset##n = b_offset + scale;
+#define UPDATE_B_POINTER(scale) b_offset = b_offset + scale;
+#define B_ELEMENT_K(n, offset_k) *(b_offset##n + (k + offset_k) * ldb)
+#define B_ELEMENT(n) B_ELEMENT_K(n, 0)
+
+#define CREATE_C_POINTER(n, scale) FLOAT* c_offset##n = c_offset + scale * ldc;
+#define INCR_C_POINTER(m, incr) // c_offset ## m += incr;
+#define UPDATE_C_POINTER(scale) c_offset = c_offset + scale * ldc;
+#define C_ELEMENT(m, n) *(c_offset##n + ((m * v_size) + i))
+
+// #undef C_ELEMENT
+// #define C_ELEMENT(m, n)             C[(i+(m))+(j+(n))*ldc]
+
+#define PACK_ELEMENT_K(n, offset_k) packed_b[(k + offset_k) * 4 + n]
+#define PACK_ELEMENT(n) PACK_ELEMENT_K(n, 0)
+
+// ASIMD
+#define DECLARE_RESULT_VECTOR2(m, n)                                           \
+  float64x2_t result##m##n = vdupq_n_f64(0.0);
+#define DECLARE_RESULT(m, n) float64_t result##m##n = 0.0;
+#define BROADCAST_LOAD_A2(m, offset_k)                                         \
+  float64x2_t a##m##_k##offset_k = vld1q_dup_f64(&A_ELEMENT_K(m, offset_k));
+#define LOAD_A1(m, offset_k)                                                   \
+  float64_t a##m##_k##offset_k = A_ELEMENT_K(m, offset_k);
+#define VECTOR_LOAD_B2(n, offset_k)                                            \
+  float64x2_t b##n##_k##offset_k = vld1q_f64(&B_ELEMENT_K(n, offset_k));
+#define GATHER_LOAD_B2(n, offset_k)                                            \
+  float64x2_t b##n##_k##offset_k = vdupq_n_f64(B_ELEMENT_K(n, offset_k));      \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f64(B_ELEMENT_K(n + 1, offset_k), b##n##_k##offset_k, 1);
+#define UPDATE_RESULT_VECTOR2(m, n, offset_k)                                  \
+  result##m##n =                                                               \
+    vfmaq_f64(result##m##n, a##m##_k##offset_k, b##n##_k##offset_k);
+#define UPDATE_RESULT(m, n, offset_k)                                          \
+  result##m##n = result##m##n + a##m##_k##offset_k * b##n##_k##offset_k;
+#ifdef B0
+#define SCATTER_STORE2(m, n)                                                   \
+  result##m##n = vmulq_f64(result##m##n, vdupq_n_f64(alpha));                  \
+  C_ELEMENT(m, n + 0) = vgetq_lane_f64(result##m##n, 0);                       \
+  C_ELEMENT(m, n + 1) = vgetq_lane_f64(result##m##n, 1);
+#else
+#define SCATTER_STORE2(m, n)                                                   \
+  result##m##n = vmulq_f64(result##m##n, vdupq_n_f64(alpha));                  \
+  C_ELEMENT(m, n + 0) =                                                        \
+    C_ELEMENT(m, n + 0) * beta + vgetq_lane_f64(result##m##n, 0);              \
+  C_ELEMENT(m, n + 1) =                                                        \
+    C_ELEMENT(m, n + 1) * beta + vgetq_lane_f64(result##m##n, 1);
+#endif
+
+// SVE
+#define DECLARE_RESULT_VECTOR(m, n) svfloat64_t result##m##n = svdup_f64(0.0);
+#define BROADCAST_LOAD_A(m, offset_k)                                          \
+  svfloat64_t a##s##m##_k##offset_k = svdup_f64(A_ELEMENT_K(m, offset_k));
+#define BROADCAST_LOAD_B(n, offset_k)                                          \
+  svfloat64_t b##s##n##_k##offset_k = svdup_f64(B_ELEMENT_K(n, offset_k));
+#define VECTOR_LOAD_A(pg, m, offset_k)                                         \
+  svfloat64_t a##s##m##_k##offset_k = svld1(pg, &A_ELEMENT_K(m, offset_k));
+#define QUADWORD_LOAD_B(n, offset_k)                                           \
+  svfloat64_t b##s##n##_k##offset_k =                                          \
+    svld1rq(pg_true, &B_ELEMENT_K(n, offset_k));
+#define UPDATE_RESULT_VECTOR(pg, m, n, offset_k)                               \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, a##s##m##_k##offset_k, b##s##n##_k##offset_k);
+#define UPDATE_RESULT_VECTOR_QUADWORD(m, n, outer, lane, offset_k)             \
+  result##m##n = svmla_lane(                                                   \
+    result##m##n, a##s##m##_k##offset_k, b##s##outer##_k##offset_k, lane);
+#ifdef B0
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#else
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, svld1(pg, &C_ELEMENT(m, n)), beta_vec);          \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n = svmla_m(pg,                                                   \
+                         result##m##n,                                         \
+                         svld1_gather_index(pg, &C_ELEMENT(m, n), ldc_vec),    \
+                         beta_vec);                                            \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#endif
+
+#ifndef LIKELY
+#ifdef __GNUC__
+#define LIKELY(x) __builtin_expect(!!(x), 1)
+#else
+#define LIKELY(x) (x)
+#endif
+#endif
+
+#ifdef B0
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT* C,
+      BLASLONG ldc)
+#else
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT beta,
+      FLOAT* C,
+      BLASLONG ldc)
+#endif
+{
+  const uint64_t v_size = svcntd();
+  const uint64_t v_size2 = v_size * 2;
+  const svbool_t pg_true = svptrue_b64();
+  const svbool_t pg_quad = svwhilelt_b64(0, 2);
+  const svfloat64_t alpha_vec = svdup_f64(alpha);
+#ifndef B0
+  const svfloat64_t beta_vec = svdup_f64(beta);
+#endif
+  const BLASLONG n4 = N & -4;
+  const BLASLONG n2 = N & -2;
+  const BLASLONG v_m2 = M & -v_size2;
+  const BLASLONG v_m1 = M & -v_size;
+
+  FLOAT* b_offset = B;
+  FLOAT* a_offset = A;
+  FLOAT* c_offset = C;
+
+  BLASLONG j = 0;
+  for (; j < n4; j += 4) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_C_POINTER(1, 1);
+    CREATE_C_POINTER(2, 2);
+    CREATE_C_POINTER(3, 3);
+    CREATE_B_POINTER(0, 0);
+    CREATE_B_POINTER(1, 1);
+    CREATE_B_POINTER(2, 2);
+    CREATE_B_POINTER(3, 3);
+
+    BLASLONG i = 0;
+    for (; i < v_m2; i += v_size2) {
+
+      CREATE_A_POINTER(0, 0);
+      CREATE_A_POINTER(1, v_size);
+      UPDATE_A_POINTER(v_size2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+      DECLARE_RESULT_VECTOR(1, 2);
+      DECLARE_RESULT_VECTOR(1, 3);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        QUADWORD_LOAD_B(2, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+        VECTOR_LOAD_A(pg_true, 1, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 2, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 2, 1, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      VECTOR_STORE(pg_true, 1, 2);
+      VECTOR_STORE(pg_true, 1, 3);
+      INCR_C_POINTER(0, v_size2);
+      INCR_C_POINTER(1, v_size2);
+      INCR_C_POINTER(2, v_size2);
+      INCR_C_POINTER(3, v_size2);
+    }
+    for (; i < v_m1; i += v_size) {
+
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(v_size);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        QUADWORD_LOAD_B(2, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      INCR_C_POINTER(0, v_size);
+      INCR_C_POINTER(1, v_size);
+      INCR_C_POINTER(2, v_size);
+      INCR_C_POINTER(3, v_size);
+    }
+    for (; i < M; i += v_size) {
+      const svbool_t pg_tail = svwhilelt_b64((uint64_t)i, (uint64_t)(M));
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(0);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        QUADWORD_LOAD_B(2, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      VECTOR_STORE(pg_tail, 0, 2);
+      VECTOR_STORE(pg_tail, 0, 3);
+      INCR_C_POINTER(0, 0);
+      INCR_C_POINTER(1, 0);
+      INCR_C_POINTER(2, 0);
+      INCR_C_POINTER(3, 0);
+    }
+
+    UPDATE_B_POINTER(4);
+    RESET_A_POINTER();
+    UPDATE_C_POINTER(4);
+  }
+  for (; j < n2; j += 2) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_C_POINTER(1, 1);
+    CREATE_B_POINTER(0, 0);
+    CREATE_B_POINTER(1, 1);
+
+    BLASLONG i = 0;
+    for (; i < v_m2; i += v_size2) {
+
+      CREATE_A_POINTER(0, 0);
+      CREATE_A_POINTER(1, v_size);
+      UPDATE_A_POINTER(v_size2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        VECTOR_LOAD_A(pg_true, 1, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      INCR_C_POINTER(0, v_size2);
+      INCR_C_POINTER(1, v_size2);
+    }
+    for (; i < v_m1; i += v_size) {
+
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(v_size);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      INCR_C_POINTER(0, v_size);
+      INCR_C_POINTER(1, v_size);
+    }
+    for (; i < M; i += v_size) {
+      const svbool_t pg_tail = svwhilelt_b64((uint64_t)i, (uint64_t)(M));
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(0);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      INCR_C_POINTER(0, 0);
+      INCR_C_POINTER(1, 0);
+    }
+
+    UPDATE_B_POINTER(2);
+    RESET_A_POINTER();
+    UPDATE_C_POINTER(2);
+  }
+  for (; j < N; j++) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_B_POINTER(0, 0);
+
+    BLASLONG i = 0;
+    for (; i < v_m2; i += v_size2) {
+
+      CREATE_A_POINTER(0, 0);
+      CREATE_A_POINTER(1, v_size);
+      UPDATE_A_POINTER(v_size2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(1, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+        VECTOR_LOAD_A(pg_true, 1, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 1, 0);
+      INCR_C_POINTER(0, v_size2);
+    }
+    for (; i < v_m1; i += v_size) {
+
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(v_size);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      INCR_C_POINTER(0, v_size);
+    }
+    for (; i < M; i += v_size) {
+      const svbool_t pg_tail = svwhilelt_b64((uint64_t)i, (uint64_t)(M));
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(0);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      INCR_C_POINTER(0, 0);
+    }
+
+    UPDATE_B_POINTER(1);
+    RESET_A_POINTER();
+    UPDATE_C_POINTER(1);
+  }
+
+  return 0;
+}

kernel/arm64/dgemm_small_kernel_tn_sve.c

@@ -0,0 +1,571 @@
+/***************************************************************************
+Copyright (c) 2024, 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.
+*****************************************************************************/
+
+#include "common.h"
+
+#include <arm_neon.h>
+#include <arm_sve.h>
+#if defined(__ARM_NEON_SVE_BRIDGE) && defined(__has_include) &&                \
+  __has_include(<arm_neon_sve_bridge.h>)
+#include <arm_neon_sve_bridge.h>
+#else
+#define svdup_neonq_f32(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat32_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#define svdup_neonq_f64(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat64_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#endif
+
+#define RESET_A_POINTER() a_offset = A;
+
+#define CREATE_A_POINTER(m, scale) FLOAT* a_offset##m = a_offset + scale * lda;
+#define UPDATE_A_POINTER(scale) a_offset = a_offset + scale * lda;
+#define A_ELEMENT_K(m, offset_k) *(a_offset##m + (k + offset_k))
+#define A_ELEMENT(m) A_ELEMENT_K(m, 0)
+
+#define RESET_B_POINTER() b_offset = B;
+
+#define CREATE_B_POINTER(n, scale) FLOAT* b_offset##n = b_offset + scale * ldb;
+#define UPDATE_B_POINTER(scale) b_offset = b_offset + scale * ldb;
+#define B_ELEMENT_K(n, offset_k) *(b_offset##n + (k + offset_k))
+#define B_ELEMENT(n) B_ELEMENT_K(n, 0)
+
+#define CREATE_C_POINTER(m, scale) FLOAT* c_offset##m = c_offset + scale;
+#define INCR_C_POINTER(m, incr) // c_offset ## m += incr * ldc;
+#define UPDATE_C_POINTER(scale) c_offset += scale;
+#define C_ELEMENT(m, n)                                                        \
+  *(c_offset##m + ((j + n) * ldc)) // C[(i+(m))+(j+(n))*ldc]
+
+// #undef C_ELEMENT
+// #define C_ELEMENT(m, n)             C[(i+(m))+(j+(n))*ldc]
+
+#define PACK_ELEMENT_K(m, offset_k) packed_a[(k + offset_k) * v_size2 + m]
+#define PACK_ELEMENT(m) PACK_ELEMENT_K(m, 0)
+
+// ASIMD
+#define DECLARE_RESULT_VECTOR2(m, n)                                           \
+  float64x2_t result##m##n = vdupq_n_f64(0.0);
+#define DECLARE_RESULT(m, n) float64_t result##m##n = 0.0;
+#define BROADCAST_LOAD_A2(m, offset_k)                                         \
+  float64x2_t a##m##_k##offset_k = vld1q_dup_f64(&A_ELEMENT_K(m, offset_k));
+#define LOAD_A1(m, offset_k)                                                   \
+  float64_t a##m##_k##offset_k = A_ELEMENT_K(m, offset_k);
+#define GATHER_LOAD_B2(n, offset_k)                                            \
+  float64x2_t b##n##_k##offset_k = vdupq_n_f64(B_ELEMENT_K(n, offset_k));      \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f64(B_ELEMENT_K(n + 1, offset_k), b##n##_k##offset_k, 1);
+#define VECTOR_UNPACK_B2(n, offset_k)                                          \
+  float64x2_t b##n##_k##offset_k = vld1q_f64(&PACK_ELEMENT_K(n, offset_k));
+#define PACK_B0(n, offset_k)                                                   \
+  PACK_ELEMENT_K(n, offset_k) = vget_lane_f64(b##n##_k##offset_k, 0);
+#define UPDATE_RESULT_VECTOR2(m, n, offset_k)                                  \
+  result##m##n =                                                               \
+    vfmaq_f64(result##m##n, a##m##_k##offset_k, b##n##_k##offset_k);
+#define UPDATE_RESULT(m, n, offset_k)                                          \
+  result##m##n = result##m##n + a##m##_k##offset_k * b##n##_k##offset_k;
+#ifdef B0
+#define SCATTER_STORE2(m, n)                                                   \
+  result##m##n = vmulq_f64(result##m##n, vdupq_n_f64(alpha));                  \
+  C_ELEMENT(m, n + 0) = vgetq_lane_f64(result##m##n, 0);                       \
+  C_ELEMENT(m, n + 1) = vgetq_lane_f64(result##m##n, 1);
+#else
+#define SCATTER_STORE2(m, n)                                                   \
+  result##m##n = vmulq_f64(result##m##n, vdupq_n_f64(alpha));                  \
+  C_ELEMENT(m, n + 0) =                                                        \
+    C_ELEMENT(m, n + 0) * beta + vgetq_lane_f64(result##m##n, 0);              \
+  C_ELEMENT(m, n + 1) =                                                        \
+    C_ELEMENT(m, n + 1) * beta + vgetq_lane_f64(result##m##n, 1);
+#endif
+
+// SVE
+#define DECLARE_RESULT_VECTOR(m, n) svfloat64_t result##m##n = svdup_f64(0.0);
+#define BROADCAST_LOAD_A(m, offset_k)                                          \
+  svfloat64_t a##s##m##_k##offset_k = svdup_f64(A_ELEMENT_K(m, offset_k));
+#define BROADCAST_LOAD_B(n, offset_k)                                          \
+  svfloat64_t b##s##n##_k##offset_k = svdup_f64(B_ELEMENT_K(n, offset_k));
+#define VECTOR_LOAD_A(pg, m, offset_k)                                         \
+  svfloat64_t a##s##m##_k##offset_k = svld1(pg, &A_ELEMENT_K(m, offset_k));
+#define GATHER_LOAD_A(pg, m, offset_k)                                         \
+  svfloat64_t a##s##m##_k##offset_k =                                          \
+    svld1_gather_index(pg, &A_ELEMENT_K(m, offset_k), lda_vec);
+#define PACK_A(m, offset_k)                                                    \
+  svst1(pg_first, &PACK_ELEMENT_K(m, offset_k), a##s##m##_k##offset_k);
+#define VECTOR_PACK_A(m, offset_k)                                             \
+  svst1(pg_true, &PACK_ELEMENT_K(m* v_size, offset_k), a##s##m##_k##offset_k);
+#define QUADWORD_PACK_A(m, offset_k)                                           \
+  svst1(pg_quad, &PACK_ELEMENT_K(m, offset_k), a##s##m##_k##offset_k);
+#define UNPACK_VECTOR_A(m, offset_k)                                           \
+  svfloat64_t a##s##m##_k##offset_k =                                          \
+    svld1(pg_true, &PACK_ELEMENT_K(m * v_size, offset_k));
+#define UNPACK_BROADCAST_A(m, offset_k)                                        \
+  svfloat64_t a##s##m##_k##offset_k = svdup_f64(PACK_ELEMENT_K(m, offset_k));
+#define UNPACK_QUADWORD_A(m, offset_k)                                         \
+  svfloat64_t a##s##m##_k##offset_k =                                          \
+    svld1rq(pg_true, &PACK_ELEMENT_K(m, offset_k));
+#define UPDATE_RESULT_VECTOR(pg, m, n, offset_k)                               \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, a##s##m##_k##offset_k, b##s##n##_k##offset_k);
+#define UPDATE_RESULT_VECTOR_QUADWORD(m, n, outer, lane, offset_k)             \
+  result##m##n = svmla_lane(                                                   \
+    result##m##n, a##s##m##_k##offset_k, b##s##outer##_k##offset_k, lane);
+#ifdef B0
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#else
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, svld1(pg, &C_ELEMENT(m, n)), beta_vec);          \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n = svmla_m(pg,                                                   \
+                         result##m##n,                                         \
+                         svld1_gather_index(pg, &C_ELEMENT(m, n), ldc_vec),    \
+                         beta_vec);                                            \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#endif
+
+#ifndef LIKELY
+#ifdef __GNUC__
+#define LIKELY(x) __builtin_expect(!!(x), 1)
+#else
+#define LIKELY(x) (x)
+#endif
+#endif
+
+#ifdef B0
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT* C,
+      BLASLONG ldc)
+#else
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT beta,
+      FLOAT* C,
+      BLASLONG ldc)
+#endif
+{
+  const uint64_t v_size = svcntd();
+  const uint64_t v_size2 = v_size * 2;
+  const svbool_t pg_true = svptrue_b64();
+  const svbool_t pg_quad = svwhilelt_b64(0, 2);
+  const svbool_t pg_first = svwhilelt_b64(0, 1);
+  const svfloat64_t alpha_vec = svdup_f64(alpha);
+#ifndef B0
+  const svfloat64_t beta_vec = svdup_f64(beta);
+#endif
+  const svuint64_t lda_vec = svindex_u64(0LL, lda);
+
+  const BLASLONG v_m2 = M & -v_size2;
+  const BLASLONG v_m1 = M & -v_size;
+  const BLASLONG n4 = N & -4;
+  const BLASLONG n2 = N & -2;
+
+  const int pack_a = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
+  FLOAT* packed_a =
+    (pack_a) ? packed_a = (FLOAT*)malloc(K * v_size2 * sizeof(FLOAT)) : NULL;
+
+  FLOAT* a_offset = A;
+  FLOAT* b_offset = B;
+  FLOAT* c_offset = C;
+
+  BLASLONG i = 0;
+  for (; i < v_m2; i += v_size2) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_C_POINTER(1, v_size);
+    CREATE_A_POINTER(0, 0);
+    CREATE_A_POINTER(1, v_size);
+
+    BLASLONG j = 0;
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+      DECLARE_RESULT_VECTOR(1, 2);
+      DECLARE_RESULT_VECTOR(1, 3);
+
+      if (LIKELY(packed_a != NULL)) {
+        if (j == 0) {
+          for (; k < K; k++) {
+
+            BROADCAST_LOAD_B(0, 0);
+            GATHER_LOAD_A(pg_true, 0, 0);
+            VECTOR_PACK_A(0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+            BROADCAST_LOAD_B(1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+            GATHER_LOAD_A(pg_true, 1, 0);
+            VECTOR_PACK_A(1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+            BROADCAST_LOAD_B(2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+            BROADCAST_LOAD_B(3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+          }
+        } else {
+          for (; k < K; k++) {
+
+            BROADCAST_LOAD_B(0, 0);
+            UNPACK_VECTOR_A(0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+            BROADCAST_LOAD_B(1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+            UNPACK_VECTOR_A(1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+            BROADCAST_LOAD_B(2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+            BROADCAST_LOAD_B(3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+          }
+        }
+      } else {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+          BROADCAST_LOAD_B(2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+          BROADCAST_LOAD_B(3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      VECTOR_STORE(pg_true, 1, 2);
+      VECTOR_STORE(pg_true, 1, 3);
+      INCR_C_POINTER(0, 4);
+      INCR_C_POINTER(1, 4);
+    }
+    for (; j < n2; j += 2) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      UPDATE_B_POINTER(2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+
+      if (LIKELY(packed_a != NULL)) {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          UNPACK_VECTOR_A(0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+          UNPACK_VECTOR_A(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+        }
+      } else {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      INCR_C_POINTER(0, 2);
+      INCR_C_POINTER(1, 2);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(1, 0);
+
+      if (LIKELY(packed_a != NULL)) {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          UNPACK_VECTOR_A(0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          UNPACK_VECTOR_A(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+        }
+      } else {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 1, 0);
+      INCR_C_POINTER(0, 1);
+      INCR_C_POINTER(1, 1);
+    }
+
+    UPDATE_A_POINTER(v_size2);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(v_size2);
+  }
+  for (; i < v_m1; i += v_size) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_A_POINTER(0, 0);
+
+    BLASLONG j = 0;
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+        BROADCAST_LOAD_B(1, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+        BROADCAST_LOAD_B(2, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+        BROADCAST_LOAD_B(3, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      INCR_C_POINTER(0, 4);
+    }
+    for (; j < n2; j += 2) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      UPDATE_B_POINTER(2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+        BROADCAST_LOAD_B(1, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      INCR_C_POINTER(0, 2);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      INCR_C_POINTER(0, 1);
+    }
+
+    UPDATE_A_POINTER(v_size);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(v_size);
+  }
+  for (; i < M; i += v_size) {
+    const svbool_t pg_tail = svwhilelt_b64((uint64_t)i, (uint64_t)(M));
+    CREATE_C_POINTER(0, 0);
+    CREATE_A_POINTER(0, 0);
+
+    BLASLONG j = 0;
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+        BROADCAST_LOAD_B(1, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 1, 0);
+        BROADCAST_LOAD_B(2, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 2, 0);
+        BROADCAST_LOAD_B(3, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 3, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      VECTOR_STORE(pg_tail, 0, 2);
+      VECTOR_STORE(pg_tail, 0, 3);
+      INCR_C_POINTER(0, 4);
+    }
+    for (; j < n2; j += 2) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      UPDATE_B_POINTER(2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+        BROADCAST_LOAD_B(1, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 1, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      INCR_C_POINTER(0, 2);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      INCR_C_POINTER(0, 1);
+    }
+
+    UPDATE_A_POINTER(0);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(0);
+  }
+
+  if (pack_a)
+    free(packed_a);
+
+  return 0;
+}

kernel/arm64/dgemm_small_kernel_tt_sve.c

@@ -0,0 +1,564 @@
+/***************************************************************************
+Copyright (c) 2024, 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.
+*****************************************************************************/
+
+#include "common.h"
+
+#include <arm_neon.h>
+#include <arm_sve.h>
+#if defined(__ARM_NEON_SVE_BRIDGE) && defined(__has_include) &&                \
+  __has_include(<arm_neon_sve_bridge.h>)
+#include <arm_neon_sve_bridge.h>
+#else
+#define svdup_neonq_f32(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat32_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#define svdup_neonq_f64(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat64_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#endif
+
+#define RESET_A_POINTER() a_offset = A;
+
+#define CREATE_A_POINTER(m, scale) FLOAT* a_offset##m = a_offset + scale * lda;
+#define UPDATE_A_POINTER(scale) a_offset = a_offset + scale * lda;
+#define A_ELEMENT_K(m, offset_k) *(a_offset##m + (k + offset_k))
+#define A_ELEMENT(m) A_ELEMENT_K(m, 0)
+
+#define RESET_B_POINTER() b_offset = B;
+
+#define CREATE_B_POINTER(n, scale) FLOAT* b_offset##n = b_offset + scale;
+#define UPDATE_B_POINTER(scale) b_offset = b_offset + scale;
+#define B_ELEMENT_K(n, offset_k) *(b_offset##n + (k + offset_k) * ldb)
+#define B_ELEMENT(n) B_ELEMENT_K(n, 0)
+
+#define CREATE_C_POINTER(m, scale) FLOAT* c_offset##m = c_offset + scale;
+#define INCR_C_POINTER(m, incr) // c_offset ## m += incr * ldc;
+#define UPDATE_C_POINTER(scale) c_offset += scale;
+#define C_ELEMENT(m, n)                                                        \
+  *(c_offset##m + ((j + n) * ldc)) // C[(i+(m))+(j+(n))*ldc]
+
+// #undef C_ELEMENT
+// #define C_ELEMENT(m, n)             C[(i+(m))+(j+(n))*ldc]
+
+#define PACK_ELEMENT_K(m, offset_k) packed_a[(k + offset_k) * v_size2 + m]
+#define PACK_ELEMENT(m) PACK_ELEMENT_K(m, 0)
+
+// ASIMD
+#define DECLARE_RESULT_VECTOR2(m, n)                                           \
+  float64x2_t result##m##n = vdupq_n_f64(0.0);
+#define DECLARE_RESULT(m, n) float64_t result##m##n = 0.0;
+#define BROADCAST_LOAD_A2(m, offset_k)                                         \
+  float64x2_t a##m##_k##offset_k = vld1q_dup_f64(&A_ELEMENT_K(m, offset_k));
+#define LOAD_A1(m, offset_k)                                                   \
+  float64_t a##m##_k##offset_k = A_ELEMENT_K(m, offset_k);
+#define VECTOR_LOAD_B2(n, offset_k)                                            \
+  float64x2_t b##n##_k##offset_k = vld1q_f64(&B_ELEMENT_K(n, offset_k));
+#define GATHER_LOAD_B2(n, offset_k)                                            \
+  float64x2_t b##n##_k##offset_k = vdupq_n_f64(B_ELEMENT_K(n, offset_k));      \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f64(B_ELEMENT_K(n + 1, offset_k), b##n##_k##offset_k, 1);
+#define VECTOR_UNPACK_B2(n, offset_k)                                          \
+  float64x2_t b##n##_k##offset_k = vld1q_f64(&PACK_ELEMENT_K(n, offset_k));
+#define VECTOR_PACK_B2(n, offset_k)                                            \
+  vst1q_f64(&PACK_ELEMENT_K(n, offset_k), b##n##_k##offset_k);
+#define PACK_B0(n, offset_k)                                                   \
+  PACK_ELEMENT_K(n, offset_k) = vget_lane_f64(b##n##_k##offset_k, 0);
+#define UPDATE_RESULT_VECTOR2(m, n, offset_k)                                  \
+  result##m##n =                                                               \
+    vfmaq_f64(result##m##n, a##m##_k##offset_k, b##n##_k##offset_k);
+#define UPDATE_RESULT(m, n, offset_k)                                          \
+  result##m##n = result##m##n + a##m##_k##offset_k * b##n##_k##offset_k;
+#ifdef B0
+#define VECTOR_STORE2(m, n)                                                    \
+  vst1q_f64(&C_ELEMENT(m, n), vmulq_f64(result##m##n, vdupq_n_f64(alpha)));
+#define STORE(m, n) C_ELEMENT(m, n) = alpha * result##m##n;
+#else
+#define VECTOR_STORE2(m, n)                                                    \
+  result##m##n = vmulq_f64(result##m##n, vdupq_n_f64(alpha));                  \
+  result##m##n =                                                               \
+    vfmaq_f64(result##m##n, vld1q_f64(&C_ELEMENT(m, n)), vdupq_n_f64(beta));   \
+  vst1q_f64(&C_ELEMENT(m, n), result##m##n);
+#define STORE(m, n)                                                            \
+  C_ELEMENT(m, n) = C_ELEMENT(m, n) * beta + alpha * result##m##n;
+#endif
+
+// SVE
+#define DECLARE_RESULT_VECTOR(m, n) svfloat64_t result##m##n = svdup_f64(0.0);
+#define BROADCAST_LOAD_A(m, offset_k)                                          \
+  svfloat64_t a##s##m##_k##offset_k = svdup_f64(A_ELEMENT_K(m, offset_k));
+#define BROADCAST_LOAD_B(n, offset_k)                                          \
+  svfloat64_t b##s##n##_k##offset_k = svdup_f64(B_ELEMENT_K(n, offset_k));
+#define VECTOR_LOAD_A(pg, m, offset_k)                                         \
+  svfloat64_t a##s##m##_k##offset_k = svld1(pg, &A_ELEMENT_K(m, offset_k));
+#define QUADWORD_LOAD_B(n, offset_k)                                           \
+  svfloat64_t b##s##n##_k##offset_k =                                          \
+    svld1rq(pg_true, &B_ELEMENT_K(n, offset_k));
+#define GATHER_LOAD_A(pg, m, offset_k)                                         \
+  svfloat64_t a##s##m##_k##offset_k =                                          \
+    svld1_gather_index(pg, &A_ELEMENT_K(m, offset_k), lda_vec);
+#define PACK_A(m, offset_k)                                                    \
+  svst1(pg_first, &PACK_ELEMENT_K(m, offset_k), a##s##m##_k##offset_k);
+#define VECTOR_PACK_A(m, offset_k)                                             \
+  svst1(pg_true, &PACK_ELEMENT_K(m* v_size, offset_k), a##s##m##_k##offset_k);
+#define QUADWORD_PACK_A(m, offset_k)                                           \
+  svst1(pg_quad, &PACK_ELEMENT_K(m, offset_k), a##s##m##_k##offset_k);
+#define UNPACK_VECTOR_A(m, offset_k)                                           \
+  svfloat64_t a##s##m##_k##offset_k =                                          \
+    svld1(pg_true, &PACK_ELEMENT_K(m * v_size, offset_k));
+#define UNPACK_BROADCAST_A(m, offset_k)                                        \
+  svfloat64_t a##s##m##_k##offset_k = svdup_f64(PACK_ELEMENT_K(m, offset_k));
+#define UNPACK_QUADWORD_A(m, offset_k)                                         \
+  svfloat64_t a##s##m##_k##offset_k =                                          \
+    svld1rq(pg_true, &PACK_ELEMENT_K(m, offset_k));
+#define UPDATE_RESULT_VECTOR(pg, m, n, offset_k)                               \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, a##s##m##_k##offset_k, b##s##n##_k##offset_k);
+#define UPDATE_RESULT_VECTOR_QUADWORD(m, n, outer, lane, offset_k)             \
+  result##m##n = svmla_lane(                                                   \
+    result##m##n, a##s##m##_k##offset_k, b##s##outer##_k##offset_k, lane);
+#ifdef B0
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#else
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, svld1(pg, &C_ELEMENT(m, n)), beta_vec);          \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n = svmla_m(pg,                                                   \
+                         result##m##n,                                         \
+                         svld1_gather_index(pg, &C_ELEMENT(m, n), ldc_vec),    \
+                         beta_vec);                                            \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#endif
+
+#ifndef LIKELY
+#ifdef __GNUC__
+#define LIKELY(x) __builtin_expect(!!(x), 1)
+#else
+#define LIKELY(x) (x)
+#endif
+#endif
+
+#ifdef B0
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT* C,
+      BLASLONG ldc)
+#else
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT beta,
+      FLOAT* C,
+      BLASLONG ldc)
+#endif
+{
+  const uint64_t v_size = svcntd();
+  const uint64_t v_size2 = v_size * 2;
+  const svbool_t pg_true = svptrue_b64();
+  const svbool_t pg_quad = svwhilelt_b64(0, 2);
+  const svbool_t pg_first = svwhilelt_b64(0, 1);
+  const svfloat64_t alpha_vec = svdup_f64(alpha);
+#ifndef B0
+  const svfloat64_t beta_vec = svdup_f64(beta);
+#endif
+  const svuint64_t lda_vec = svindex_u64(0LL, lda);
+
+  const BLASLONG v_m2 = M & -v_size2;
+  const BLASLONG v_m1 = M & -v_size;
+  const BLASLONG n4 = N & -4;
+  const BLASLONG n2 = N & -2;
+
+  const int pack_a = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
+  FLOAT* packed_a =
+    (pack_a) ? packed_a = (FLOAT*)malloc(K * v_size2 * sizeof(FLOAT)) : NULL;
+
+  FLOAT* a_offset = A;
+  FLOAT* b_offset = B;
+  FLOAT* c_offset = C;
+
+  BLASLONG i = 0;
+  for (; i < v_m2; i += v_size2) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_C_POINTER(1, v_size);
+    CREATE_A_POINTER(0, 0);
+    CREATE_A_POINTER(1, v_size);
+
+    BLASLONG j = 0;
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+      DECLARE_RESULT_VECTOR(1, 2);
+      DECLARE_RESULT_VECTOR(1, 3);
+
+      if (LIKELY(packed_a != NULL)) {
+        if (j == 0) {
+          for (; k < K; k++) {
+
+            QUADWORD_LOAD_B(0, 0);
+            GATHER_LOAD_A(pg_true, 0, 0);
+            VECTOR_PACK_A(0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+            QUADWORD_LOAD_B(2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+            GATHER_LOAD_A(pg_true, 1, 0);
+            VECTOR_PACK_A(1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 2, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 2, 1, 0);
+          }
+        } else {
+          for (; k < K; k++) {
+
+            QUADWORD_LOAD_B(0, 0);
+            UNPACK_VECTOR_A(0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+            QUADWORD_LOAD_B(2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+            UNPACK_VECTOR_A(1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 2, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 2, 1, 0);
+          }
+        }
+      } else {
+        for (; k < K; k++) {
+
+          QUADWORD_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          QUADWORD_LOAD_B(2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 2, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 2, 1, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      VECTOR_STORE(pg_true, 1, 2);
+      VECTOR_STORE(pg_true, 1, 3);
+      INCR_C_POINTER(0, 4);
+      INCR_C_POINTER(1, 4);
+    }
+    for (; j < n2; j += 2) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      UPDATE_B_POINTER(2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+
+      if (LIKELY(packed_a != NULL)) {
+        for (; k < K; k++) {
+
+          QUADWORD_LOAD_B(0, 0);
+          UNPACK_VECTOR_A(0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UNPACK_VECTOR_A(1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+        }
+      } else {
+        for (; k < K; k++) {
+
+          QUADWORD_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      INCR_C_POINTER(0, 2);
+      INCR_C_POINTER(1, 2);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(1, 0);
+
+      if (LIKELY(packed_a != NULL)) {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          UNPACK_VECTOR_A(0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          UNPACK_VECTOR_A(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+        }
+      } else {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 1, 0);
+      INCR_C_POINTER(0, 1);
+      INCR_C_POINTER(1, 1);
+    }
+
+    UPDATE_A_POINTER(v_size2);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(v_size2);
+  }
+  for (; i < v_m1; i += v_size) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_A_POINTER(0, 0);
+
+    BLASLONG j = 0;
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        QUADWORD_LOAD_B(2, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      INCR_C_POINTER(0, 4);
+    }
+    for (; j < n2; j += 2) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      UPDATE_B_POINTER(2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      INCR_C_POINTER(0, 2);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      INCR_C_POINTER(0, 1);
+    }
+
+    UPDATE_A_POINTER(v_size);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(v_size);
+  }
+  for (; i < M; i += v_size) {
+    const svbool_t pg_tail = svwhilelt_b64((uint64_t)i, (uint64_t)(M));
+    CREATE_C_POINTER(0, 0);
+    CREATE_A_POINTER(0, 0);
+
+    BLASLONG j = 0;
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        QUADWORD_LOAD_B(2, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 2, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 2, 1, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      VECTOR_STORE(pg_tail, 0, 2);
+      VECTOR_STORE(pg_tail, 0, 3);
+      INCR_C_POINTER(0, 4);
+    }
+    for (; j < n2; j += 2) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      UPDATE_B_POINTER(2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      INCR_C_POINTER(0, 2);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      INCR_C_POINTER(0, 1);
+    }
+
+    UPDATE_A_POINTER(0);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(0);
+  }
+
+  if (pack_a)
+    free(packed_a);
+
+  return 0;
+}

kernel/arm64/gemm_small_kernel_permit_sve.c

@@ -0,0 +1,43 @@
+/***************************************************************************
+Copyright (c) 2024, 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.
+*****************************************************************************/
+
+#include "common.h"
+
+int CNAME(int transa, int transb, BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, FLOAT beta)
+{
+  BLASLONG MNK = M * N * K;
+
+#if defined(DOUBLE) // dgemm
+  if (MNK <= 64*64*64)
+    return 1;
+#else // sgemm
+  if (MNK <= 64*64*64)
+    return 1;
+#endif
+
+  return 0;
+}

kernel/arm64/gemv_n_sve.c

@@ -0,0 +1,92 @@
+/***************************************************************************
+Copyright (c) 2024, 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 COPYRIGHT OWNER 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.
+*****************************************************************************/
+
+#include <arm_sve.h>
+#include "common.h"
+
+#ifdef DOUBLE
+#define SV_COUNT svcntd
+#define SV_TYPE svfloat64_t
+#define SV_TRUE svptrue_b64
+#define SV_WHILE svwhilelt_b64_s64
+#define SV_DUP svdup_f64
+#else
+#define SV_COUNT svcntw
+#define SV_TYPE svfloat32_t
+#define SV_TRUE svptrue_b32
+#define SV_WHILE svwhilelt_b32_s64
+#define SV_DUP svdup_f32
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+  BLASLONG i;
+  BLASLONG ix,iy;
+  BLASLONG j;
+  FLOAT *a_ptr;
+  FLOAT temp;
+
+  ix = 0;
+  a_ptr = a;
+
+  if (inc_y == 1) {
+    uint64_t sve_size = SV_COUNT();
+    for (j = 0; j < n; j++) {
+      SV_TYPE temp_vec = SV_DUP(alpha * x[ix]);
+      i = 0;
+      svbool_t pg = SV_WHILE(i, m);
+      while (svptest_any(SV_TRUE(), pg)) {
+        SV_TYPE a_vec = svld1(pg, a_ptr + i);
+        SV_TYPE y_vec = svld1(pg, y + i);
+        y_vec = svmla_x(pg, y_vec, temp_vec, a_vec);
+        svst1(pg, y + i, y_vec);
+        i += sve_size;
+        pg = SV_WHILE(i, m);
+      }
+      a_ptr += lda;
+      ix += inc_x;
+    }
+    return(0);
+  }
+
+  for (j = 0; j < n; j++) {
+    temp = alpha * x[ix];
+    iy = 0;
+    for (i = 0; i < m; i++) {
+      y[iy] += temp * a_ptr[i];
+      iy += inc_y;
+    }
+    a_ptr += lda;
+    ix += inc_x;
+  }
+  return (0);
+}

kernel/arm64/gemv_t.S

@@ -1,5 +1,5 @@
 /*******************************************************************************
-Copyright (c) 2015, The OpenBLAS Project
+Copyright (c) 2015, 2024 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
@@ -170,39 +170,48 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro KERNEL_F32_FINALIZE
 #if !defined(DOUBLE)
-	fadd	v1.4s, v1.4s, v2.4s
+	// F8 only has 2 accumulators
+	// so add into those pairs
 	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
+	fadd	v2.4s, v2.4s, v4.4s
 #endif
 .endm
 
-.macro KERNEL_F4
+.macro KERNEL_F8
 #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
+	ld1	{v13.4s, v14.4s}, [A_PTR], #32
+	ld1	{v17.4s, v18.4s}, [X_PTR], #32
+	fmla	v1.4s, v13.4s, v17.4s
+	fmla	v2.4s, v14.4s, v18.4s
+#else	
+	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_F4_FINALIZE
+.macro KERNEL_F8_FINALIZE
 #if !defined(DOUBLE)
-	ext	v2.16b, v1.16b, v1.16b, #8
+	// Take the top two elements of v1 and 
+	// put them into the first two lanes of v3
+	ext	v3.16b, v1.16b, v1.16b, #8
+	fadd	v1.2s, v1.2s, v3.2s
+	ext	v4.16b, v2.16b, v2.16b, #8
+	fadd	v2.2s, v2.2s, v4.2s
+	// Final pair
 	fadd	v1.2s, v1.2s, v2.2s
 	faddp	TEMP, v1.2s
 #else
 	faddp	TEMP, v1.2d
+	faddp	TEMP1, v2.2d
+	faddp	TEMP2, v3.2d
+	faddp	TEMP3, v4.2d
+	fadd	TEMP, TEMP, TEMP1
+	fadd	TEMP2, TEMP2, TEMP3
+	fadd	TEMP, TEMP, TEMP2
 #endif
 .endm
 
@@ -258,7 +267,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	asr	I, M, #5
 	cmp	I, xzr
-	beq	.Lgemv_t_kernel_F4
+	beq	.Lgemv_t_kernel_F8
 
 .Lgemv_t_kernel_F320:
 
@@ -269,24 +278,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	KERNEL_F32_FINALIZE
 
-.Lgemv_t_kernel_F4:
+.Lgemv_t_kernel_F8:
 	ands	I, M, #31
-	asr	I, I, #2
+	asr	I, I, #3
 	cmp	I, xzr
 	beq	.Lgemv_t_kernel_F1
 
-.Lgemv_t_kernel_F40:
+.Lgemv_t_kernel_F80:
 
-	KERNEL_F4
+	KERNEL_F8
 
 	subs	I, I, #1
-	bne	.Lgemv_t_kernel_F40
+	bne	.Lgemv_t_kernel_F80
 
 .Lgemv_t_kernel_F1:
 
-	KERNEL_F4_FINALIZE
+	KERNEL_F8_FINALIZE
 
-	ands	I, M, #3
+	ands	I, M, #7
 	ble	.Lgemv_t_kernel_F_END
 
 .Lgemv_t_kernel_F10:

kernel/arm64/gemv_t_sve.c

@@ -0,0 +1,120 @@
+/***************************************************************************
+Copyright (c) 2024, 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 COPYRIGHT OWNER 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.
+*****************************************************************************/
+
+#include <arm_sve.h>
+#include "common.h"
+
+#ifdef DOUBLE
+#define SV_COUNT svcntd
+#define SV_TYPE svfloat64_t
+#define SV_TRUE svptrue_b64
+#define SV_WHILE svwhilelt_b64_s64
+#define SV_DUP svdup_f64
+#else
+#define SV_COUNT svcntw
+#define SV_TYPE svfloat32_t
+#define SV_TRUE svptrue_b32
+#define SV_WHILE svwhilelt_b32_s64
+#define SV_DUP svdup_f32
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+  BLASLONG i;
+  BLASLONG ix,iy;
+  BLASLONG j;
+  FLOAT *a_ptr;
+  FLOAT temp;
+
+  iy = 0;
+  a_ptr = a;
+
+  if (inc_x == 1) {
+    svbool_t pg_true = SV_TRUE();
+    uint64_t sve_size = SV_COUNT();
+    uint64_t sve_size2 = sve_size * 2;
+    BLASLONG m1 = m & -sve_size;
+    BLASLONG m2 = m & -sve_size2;
+
+    for (j = 0; j < n; j++) {
+      BLASLONG i = 0;
+
+      SV_TYPE temp_vec_v2_0 = SV_DUP(0.0);
+      SV_TYPE temp_vec_v2_1 = SV_DUP(0.0);
+      for (; i < m2; i += sve_size2) {
+        SV_TYPE a_vec0 = svld1(pg_true, a_ptr + i);
+        SV_TYPE x_vec0 = svld1(pg_true, x + i);
+        SV_TYPE a_vec1 = svld1(pg_true, a_ptr + i + sve_size);
+        SV_TYPE x_vec1 = svld1(pg_true, x + i + sve_size);
+        temp_vec_v2_0 = svmla_m(pg_true, temp_vec_v2_0, a_vec0, x_vec0);
+        temp_vec_v2_1 = svmla_m(pg_true, temp_vec_v2_1, a_vec1, x_vec1);
+      }
+
+      SV_TYPE temp_vec_v1 = SV_DUP(0.0);
+      for (; i < m1; i += sve_size) {
+        SV_TYPE a_vec0 = svld1(pg_true, a_ptr + i);
+        SV_TYPE x_vec0 = svld1(pg_true, x + i);
+        temp_vec_v1 = svmla_m(pg_true, temp_vec_v1, a_vec0, x_vec0);
+      }
+
+      SV_TYPE temp_vec = SV_DUP(0.0);
+      for (; i < m; i += sve_size) {
+        svbool_t pg = SV_WHILE(i, m);
+        SV_TYPE a_vec = svld1(pg, a_ptr + i);
+        SV_TYPE x_vec = svld1(pg, x + i);
+        temp_vec = svmla_m(pg, temp_vec, a_vec, x_vec);
+      }
+
+      y[iy] += alpha * (
+        (svaddv(SV_TRUE(), temp_vec_v2_0) + svaddv(SV_TRUE(), temp_vec)) +
+        (svaddv(SV_TRUE(), temp_vec_v2_1) + svaddv(SV_TRUE(), temp_vec_v1))
+      );
+
+      iy += inc_y;
+      a_ptr += lda;
+    }
+    return(0);
+  }
+
+  for (j = 0; j < n; j++) {
+    temp = 0.0;
+    ix = 0;
+    for (i = 0; i < m; i++) {
+      temp += a_ptr[i] * x[ix];
+      ix += inc_x;
+    }
+    y[iy] += alpha * temp;
+    iy += inc_y;
+    a_ptr += lda;
+  }
+  return (0);
+}

kernel/arm64/scal.S

@@ -33,7 +33,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define	X_COPY	x5	/* X vector address */
 #define	INC_X	x4	/* X stride */
 #define I	x1	/* loop variable */
-
+#define FLAG    x9
 /*******************************************************************************
 * Macro definitions
 *******************************************************************************/
@@ -168,9 +168,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	N, xzr
 	ble	.Lscal_kernel_L999
 
-	//fcmp	DA, #0.0
-	//beq	.Lscal_kernel_zero
+	ldr	FLAG, [sp]
+	cmp	FLAG, #1
+	beq	.Lscal_kernel_nansafe
+
+	fcmp	DA, #0.0
+	beq	.Lscal_kernel_zero
 
+.Lscal_kernel_nansafe:
 	cmp	INC_X, #1
 	bne	.Lscal_kernel_S_BEGIN
 

kernel/arm64/sgemm_small_kernel_nn_sve.c

@@ -0,0 +1,687 @@
+/***************************************************************************
+Copyright (c) 2024, 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.
+*****************************************************************************/
+
+#include "common.h"
+
+#include <arm_neon.h>
+#include <arm_sve.h>
+#if defined(__ARM_NEON_SVE_BRIDGE) && defined(__has_include) &&                \
+  __has_include(<arm_neon_sve_bridge.h>)
+#include <arm_neon_sve_bridge.h>
+#else
+#define svdup_neonq_f32(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat32_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#define svdup_neonq_f64(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat64_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#endif
+
+#define RESET_A_POINTER() a_offset = A;
+
+#define CREATE_A_POINTER(m, scale) FLOAT* a_offset##m = a_offset + scale;
+#define UPDATE_A_POINTER(scale) a_offset = a_offset + scale;
+#define A_ELEMENT_K(m, offset_k) *(a_offset##m + (k + offset_k) * lda)
+#define A_ELEMENT(m) A_ELEMENT_K(m, 0)
+
+#define RESET_B_POINTER() b_offset = B;
+
+#define CREATE_B_POINTER(n, scale) FLOAT* b_offset##n = b_offset + scale * ldb;
+#define UPDATE_B_POINTER(scale) b_offset = b_offset + scale * ldb;
+#define B_ELEMENT_K(n, offset_k) *(b_offset##n + (k + offset_k))
+#define B_ELEMENT(n) B_ELEMENT_K(n, 0)
+
+#define CREATE_C_POINTER(n, scale) FLOAT* c_offset##n = c_offset + scale * ldc;
+#define INCR_C_POINTER(m, incr) // c_offset ## m += incr;
+#define UPDATE_C_POINTER(scale) c_offset = c_offset + scale * ldc;
+#define C_ELEMENT(m, n) *(c_offset##n + ((m * v_size) + i))
+
+// #undef C_ELEMENT
+// #define C_ELEMENT(m, n)             C[(i+(m))+(j+(n))*ldc]
+
+#define PACK_ELEMENT_K(n, offset_k) packed_b[(k + offset_k) * 4 + n]
+#define PACK_ELEMENT(n) PACK_ELEMENT_K(n, 0)
+
+// ASIMD
+#define DECLARE_RESULT_VECTOR4(m, n)                                           \
+  float32x4_t result##m##n = vdupq_n_f32(0.0);
+#define DECLARE_RESULT(m, n) float32_t result##m##n = 0.0;
+#define BROADCAST_LOAD_A4(m, offset_k)                                         \
+  float32x4_t a##m##_k##offset_k = vld1q_dup_f32(&A_ELEMENT_K(m, offset_k));
+#define LOAD_A1(m, offset_k)                                                   \
+  float32_t a##m##_k##offset_k = A_ELEMENT_K(m, offset_k);
+#define VECTOR_LOAD_B_K4(n, offset_k)                                          \
+  float32x4_t b##k##n##_k##offset_k = vld1q_f32(&B_ELEMENT_K(n, offset_k));
+#define TRANSPOSE_B4_K4(                                                       \
+  n0, n1, n2, n3, offset_k0, offset_k1, offset_k2, offset_k3)                  \
+  float32x4_t b##t##n0##_k##offset_k0 =                                        \
+    vzip1q_f32(b##k##n0##_k##offset_k0, b##k##n1##_k##offset_k0);              \
+  float32x4_t b##t##n0##_k##offset_k1 =                                        \
+    vzip2q_f32(b##k##n0##_k##offset_k0, b##k##n1##_k##offset_k0);              \
+  float32x4_t b##t##n0##_k##offset_k2 =                                        \
+    vzip1q_f32(b##k##n2##_k##offset_k0, b##k##n3##_k##offset_k0);              \
+  float32x4_t b##t##n0##_k##offset_k3 =                                        \
+    vzip2q_f32(b##k##n2##_k##offset_k0, b##k##n3##_k##offset_k0);              \
+  float32x4_t b##n0##_k##offset_k0 = vreinterpretq_f32_f64(                    \
+    vzip1q_f64(vreinterpretq_f64_f32(b##t##n0##_k##offset_k0),                 \
+               vreinterpretq_f64_f32(b##t##n0##_k##offset_k2)));               \
+  float32x4_t b##n0##_k##offset_k1 = vreinterpretq_f32_f64(                    \
+    vzip2q_f64(vreinterpretq_f64_f32(b##t##n0##_k##offset_k0),                 \
+               vreinterpretq_f64_f32(b##t##n0##_k##offset_k2)));               \
+  float32x4_t b##n0##_k##offset_k2 = vreinterpretq_f32_f64(                    \
+    vzip1q_f64(vreinterpretq_f64_f32(b##t##n0##_k##offset_k1),                 \
+               vreinterpretq_f64_f32(b##t##n0##_k##offset_k3)));               \
+  float32x4_t b##n0##_k##offset_k3 = vreinterpretq_f32_f64(                    \
+    vzip2q_f64(vreinterpretq_f64_f32(b##t##n0##_k##offset_k1),                 \
+               vreinterpretq_f64_f32(b##t##n0##_k##offset_k3)));
+
+#define SCALE_B4_K4(n0, offset_k0, offset_k1, offset_k2, offset_k3)            \
+  svfloat32_t b##s##n0##_k##offset_k0 = svdup_neonq_f32(b##n0##_k##offset_k0); \
+  svfloat32_t b##s##n0##_k##offset_k1 = svdup_neonq_f32(b##n0##_k##offset_k1); \
+  svfloat32_t b##s##n0##_k##offset_k2 = svdup_neonq_f32(b##n0##_k##offset_k2); \
+  svfloat32_t b##s##n0##_k##offset_k3 = svdup_neonq_f32(b##n0##_k##offset_k3);
+#define GATHER_LOAD_B4(n, offset_k)                                            \
+  float32x4_t b##n##_k##offset_k = vdupq_n_f32(B_ELEMENT_K(n, offset_k));      \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 1, offset_k), b##n##_k##offset_k, 1);       \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 2, offset_k), b##n##_k##offset_k, 2);       \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 3, offset_k), b##n##_k##offset_k, 3);
+#define VECTOR_UNPACK_B4(n, offset_k)                                          \
+  float32x4_t b##n##_k##offset_k = vld1q_f32(&PACK_ELEMENT_K(n, offset_k));
+#define VECTOR_PACK_B4(n, offset_k)                                            \
+  vst1q_f32(&PACK_ELEMENT_K(n, offset_k), b##n##_k##offset_k);
+#define PACK_B0(n, offset_k)                                                   \
+  PACK_ELEMENT_K(n, offset_k) = vget_lane_f32(b##n##_k##offset_k, 0);
+#define UPDATE_RESULT_VECTOR4(m, n, offset_k)                                  \
+  result##m##n =                                                               \
+    vfmaq_f32(result##m##n, a##m##_k##offset_k, b##n##_k##offset_k);
+#define UPDATE_RESULT(m, n, offset_k)                                          \
+  result##m##n = result##m##n + a##m##_k##offset_k * b##n##_k##offset_k;
+#ifdef B0
+#define SCATTER_STORE4(m, n)                                                   \
+  result##m##n = vmulq_f32(result##m##n, vdupq_n_f32(alpha));                  \
+  C_ELEMENT(m, n + 0) = vgetq_lane_f32(result##m##n, 0);                       \
+  C_ELEMENT(m, n + 1) = vgetq_lane_f32(result##m##n, 1);                       \
+  C_ELEMENT(m, n + 2) = vgetq_lane_f32(result##m##n, 2);                       \
+  C_ELEMENT(m, n + 3) = vgetq_lane_f32(result##m##n, 3);
+#else
+#define SCATTER_STORE4(m, n)                                                   \
+  result##m##n = vmulq_f32(result##m##n, vdupq_n_f32(alpha));                  \
+  C_ELEMENT(m, n + 0) =                                                        \
+    C_ELEMENT(m, n + 0) * beta + vgetq_lane_f32(result##m##n, 0);              \
+  C_ELEMENT(m, n + 1) =                                                        \
+    C_ELEMENT(m, n + 1) * beta + vgetq_lane_f32(result##m##n, 1);              \
+  C_ELEMENT(m, n + 2) =                                                        \
+    C_ELEMENT(m, n + 2) * beta + vgetq_lane_f32(result##m##n, 2);              \
+  C_ELEMENT(m, n + 3) =                                                        \
+    C_ELEMENT(m, n + 3) * beta + vgetq_lane_f32(result##m##n, 3);
+#endif
+
+// SVE
+#define DECLARE_RESULT_VECTOR(m, n) svfloat32_t result##m##n = svdup_f32(0.0);
+#define BROADCAST_LOAD_A(m, offset_k)                                          \
+  svfloat32_t a##s##m##_k##offset_k = svdup_f32(A_ELEMENT_K(m, offset_k));
+#define BROADCAST_LOAD_B(n, offset_k)                                          \
+  svfloat32_t b##s##n##_k##offset_k = svdup_f32(B_ELEMENT_K(n, offset_k));
+#define VECTOR_LOAD_A(pg, m, offset_k)                                         \
+  svfloat32_t a##s##m##_k##offset_k = svld1(pg, &A_ELEMENT_K(m, offset_k));
+#define QUADWORD_LOAD_B(n, offset_k)                                           \
+  svfloat32_t b##s##n##_k##offset_k =                                          \
+    svld1rq(pg_true, &B_ELEMENT_K(n, offset_k));
+#define PACK_B(n, offset_k)                                                    \
+  svst1(pg_first, &PACK_ELEMENT_K(n, offset_k), b##s##n##_k##offset_k);
+#define VECTOR_PACK_B(n, offset_k)                                             \
+  svst1(pg_true, &PACK_ELEMENT_K(n* v_size, offset_k), b##s##n##_k##offset_k);
+#define QUADWORD_PACK_B(n, offset_k)                                           \
+  svst1(pg_quad, &PACK_ELEMENT_K(n, offset_k), b##s##n##_k##offset_k);
+#define UNPACK_VECTOR_B(n, offset_k)                                           \
+  svfloat32_t b##s##n##_k##offset_k =                                          \
+    svld1(pg_true, &PACK_ELEMENT_K(n * v_size, offset_k));
+#define UNPACK_BROADCAST_B(n, offset_k)                                        \
+  svfloat32_t b##s##n##_k##offset_k = svdup_f32(PACK_ELEMENT_K(n, offset_k));
+#define UNPACK_QUADWORD_B(n, offset_k)                                         \
+  svfloat32_t b##s##n##_k##offset_k =                                          \
+    svld1rq(pg_true, &PACK_ELEMENT_K(n, offset_k));
+#define UPDATE_RESULT_VECTOR(pg, m, n, offset_k)                               \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, a##s##m##_k##offset_k, b##s##n##_k##offset_k);
+#define UPDATE_RESULT_VECTOR_QUADWORD(m, n, outer, lane, offset_k)             \
+  result##m##n = svmla_lane(                                                   \
+    result##m##n, a##s##m##_k##offset_k, b##s##outer##_k##offset_k, lane);
+#ifdef B0
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#else
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, svld1(pg, &C_ELEMENT(m, n)), beta_vec);          \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n = svmla_m(pg,                                                   \
+                         result##m##n,                                         \
+                         svld1_gather_index(pg, &C_ELEMENT(m, n), ldc_vec),    \
+                         beta_vec);                                            \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#endif
+
+#ifndef LIKELY
+#ifdef __GNUC__
+#define LIKELY(x) __builtin_expect(!!(x), 1)
+#else
+#define LIKELY(x) (x)
+#endif
+#endif
+
+#ifdef B0
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT* C,
+      BLASLONG ldc)
+#else
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT beta,
+      FLOAT* C,
+      BLASLONG ldc)
+#endif
+{
+  const uint64_t v_size = svcntw();
+  const uint64_t v_size2 = v_size * 2;
+  const svbool_t pg_true = svptrue_b32();
+  const svbool_t pg_quad = svwhilelt_b32(0, 4);
+  const svbool_t pg_first = svwhilelt_b32(0, 1);
+  const svfloat32_t alpha_vec = svdup_f32(alpha);
+#ifndef B0
+  const svfloat32_t beta_vec = svdup_f32(beta);
+#endif
+  const BLASLONG n4 = N & -4;
+  const BLASLONG v_m2 = M & -v_size2;
+  const BLASLONG v_m1 = M & -v_size;
+  const BLASLONG k4 = K & -4;
+
+  const int pack_b = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
+  FLOAT* packed_b =
+    (pack_b) ? packed_b = (FLOAT*)malloc(K * 4 * sizeof(FLOAT)) : NULL;
+
+  FLOAT* b_offset = B;
+  FLOAT* a_offset = A;
+  FLOAT* c_offset = C;
+
+  BLASLONG j = 0;
+  for (; j < n4; j += 4) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_C_POINTER(1, 1);
+    CREATE_C_POINTER(2, 2);
+    CREATE_C_POINTER(3, 3);
+    CREATE_B_POINTER(0, 0);
+    CREATE_B_POINTER(1, 1);
+    CREATE_B_POINTER(2, 2);
+    CREATE_B_POINTER(3, 3);
+
+    BLASLONG i = 0;
+    for (; i < v_m2; i += v_size2) {
+
+      CREATE_A_POINTER(0, 0);
+      CREATE_A_POINTER(1, v_size);
+      UPDATE_A_POINTER(v_size2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+      DECLARE_RESULT_VECTOR(1, 2);
+      DECLARE_RESULT_VECTOR(1, 3);
+
+      if (LIKELY(packed_b != NULL)) {
+        if (i == 0) {
+          for (; k < k4; k += 4) {
+
+            VECTOR_LOAD_B_K4(0, 0);
+            VECTOR_LOAD_B_K4(1, 0);
+            VECTOR_LOAD_B_K4(2, 0);
+            VECTOR_LOAD_B_K4(3, 0);
+            TRANSPOSE_B4_K4(0, 1, 2, 3, 0, 1, 2, 3);
+            SCALE_B4_K4(0, 0, 1, 2, 3);
+            VECTOR_PACK_B4(0, 0);
+            VECTOR_PACK_B4(0, 1);
+            VECTOR_PACK_B4(0, 2);
+            VECTOR_PACK_B4(0, 3);
+            VECTOR_LOAD_A(pg_true, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+            VECTOR_LOAD_A(pg_true, 0, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 1);
+            VECTOR_LOAD_A(pg_true, 0, 2);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 2);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 2);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 2);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 2);
+            VECTOR_LOAD_A(pg_true, 0, 3);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 3);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 3);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 3);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 3);
+            VECTOR_LOAD_A(pg_true, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
+            VECTOR_LOAD_A(pg_true, 1, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 1);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 1);
+            VECTOR_LOAD_A(pg_true, 1, 2);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 2);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 2);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 2);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 2);
+            VECTOR_LOAD_A(pg_true, 1, 3);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 3);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 3);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 3);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 3);
+          }
+          for (; k < K; k++) {
+
+            BROADCAST_LOAD_B(0, 0);
+            PACK_B(0, 0);
+            VECTOR_LOAD_A(pg_true, 0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+            BROADCAST_LOAD_B(1, 0);
+            PACK_B(1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+            VECTOR_LOAD_A(pg_true, 1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+            BROADCAST_LOAD_B(2, 0);
+            PACK_B(2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+            BROADCAST_LOAD_B(3, 0);
+            PACK_B(3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+          }
+        } else {
+          for (; k < K; k++) {
+
+            UNPACK_QUADWORD_B(0, 0);
+            VECTOR_LOAD_A(pg_true, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+            VECTOR_LOAD_A(pg_true, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
+          }
+        }
+      } else {
+        for (; k < k4; k += 4) {
+
+          VECTOR_LOAD_B_K4(0, 0);
+          VECTOR_LOAD_B_K4(1, 0);
+          VECTOR_LOAD_B_K4(2, 0);
+          VECTOR_LOAD_B_K4(3, 0);
+          TRANSPOSE_B4_K4(0, 1, 2, 3, 0, 1, 2, 3);
+          SCALE_B4_K4(0, 0, 1, 2, 3);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+          VECTOR_LOAD_A(pg_true, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 1);
+          VECTOR_LOAD_A(pg_true, 0, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 2);
+          VECTOR_LOAD_A(pg_true, 0, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 3);
+          VECTOR_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
+          VECTOR_LOAD_A(pg_true, 1, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 1);
+          VECTOR_LOAD_A(pg_true, 1, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 2);
+          VECTOR_LOAD_A(pg_true, 1, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 3);
+        }
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+          VECTOR_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+          BROADCAST_LOAD_B(2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+          BROADCAST_LOAD_B(3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      VECTOR_STORE(pg_true, 1, 2);
+      VECTOR_STORE(pg_true, 1, 3);
+      INCR_C_POINTER(0, v_size2);
+      INCR_C_POINTER(1, v_size2);
+      INCR_C_POINTER(2, v_size2);
+      INCR_C_POINTER(3, v_size2);
+    }
+    for (; i < v_m1; i += v_size) {
+
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(v_size);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      if (LIKELY(packed_b != NULL)) {
+        for (; k < K; k++) {
+
+          UNPACK_QUADWORD_B(0, 0);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+        }
+      } else {
+        for (; k < k4; k += 4) {
+
+          VECTOR_LOAD_B_K4(0, 0);
+          VECTOR_LOAD_B_K4(1, 0);
+          VECTOR_LOAD_B_K4(2, 0);
+          VECTOR_LOAD_B_K4(3, 0);
+          TRANSPOSE_B4_K4(0, 1, 2, 3, 0, 1, 2, 3);
+          SCALE_B4_K4(0, 0, 1, 2, 3);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+          VECTOR_LOAD_A(pg_true, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 1);
+          VECTOR_LOAD_A(pg_true, 0, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 2);
+          VECTOR_LOAD_A(pg_true, 0, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 3);
+        }
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+          BROADCAST_LOAD_B(2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+          BROADCAST_LOAD_B(3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      INCR_C_POINTER(0, v_size);
+      INCR_C_POINTER(1, v_size);
+      INCR_C_POINTER(2, v_size);
+      INCR_C_POINTER(3, v_size);
+    }
+    for (; i < M; i += v_size) {
+      const svbool_t pg_tail = svwhilelt_b32((uint32_t)i, (uint32_t)(M));
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(0);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      if (LIKELY(packed_b != NULL)) {
+        for (; k < K; k++) {
+
+          UNPACK_QUADWORD_B(0, 0);
+          VECTOR_LOAD_A(pg_tail, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+        }
+      } else {
+        for (; k < k4; k += 4) {
+
+          VECTOR_LOAD_B_K4(0, 0);
+          VECTOR_LOAD_B_K4(1, 0);
+          VECTOR_LOAD_B_K4(2, 0);
+          VECTOR_LOAD_B_K4(3, 0);
+          TRANSPOSE_B4_K4(0, 1, 2, 3, 0, 1, 2, 3);
+          SCALE_B4_K4(0, 0, 1, 2, 3);
+          VECTOR_LOAD_A(pg_tail, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+          VECTOR_LOAD_A(pg_tail, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 1);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 1);
+          VECTOR_LOAD_A(pg_tail, 0, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 2);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 2);
+          VECTOR_LOAD_A(pg_tail, 0, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 3);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 3);
+        }
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          VECTOR_LOAD_A(pg_tail, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_tail, 0, 1, 0);
+          BROADCAST_LOAD_B(2, 0);
+          UPDATE_RESULT_VECTOR(pg_tail, 0, 2, 0);
+          BROADCAST_LOAD_B(3, 0);
+          UPDATE_RESULT_VECTOR(pg_tail, 0, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      VECTOR_STORE(pg_tail, 0, 2);
+      VECTOR_STORE(pg_tail, 0, 3);
+      INCR_C_POINTER(0, 0);
+      INCR_C_POINTER(1, 0);
+      INCR_C_POINTER(2, 0);
+      INCR_C_POINTER(3, 0);
+    }
+
+    UPDATE_B_POINTER(4);
+    RESET_A_POINTER();
+    UPDATE_C_POINTER(4);
+  }
+  for (; j < N; j++) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_B_POINTER(0, 0);
+
+    BLASLONG i = 0;
+    for (; i < v_m2; i += v_size2) {
+
+      CREATE_A_POINTER(0, 0);
+      CREATE_A_POINTER(1, v_size);
+      UPDATE_A_POINTER(v_size2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(1, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+        VECTOR_LOAD_A(pg_true, 1, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 1, 0);
+      INCR_C_POINTER(0, v_size2);
+    }
+    for (; i < v_m1; i += v_size) {
+
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(v_size);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      INCR_C_POINTER(0, v_size);
+    }
+    for (; i < M; i += v_size) {
+      const svbool_t pg_tail = svwhilelt_b32((uint32_t)i, (uint32_t)(M));
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(0);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      INCR_C_POINTER(0, 0);
+    }
+
+    UPDATE_B_POINTER(1);
+    RESET_A_POINTER();
+    UPDATE_C_POINTER(1);
+  }
+
+  if (pack_b)
+    free(packed_b);
+
+  return 0;
+}

kernel/arm64/sgemm_small_kernel_nt_sve.c

@@ -0,0 +1,483 @@
+/***************************************************************************
+Copyright (c) 2024, 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.
+*****************************************************************************/
+
+#include "common.h"
+
+#include <arm_neon.h>
+#include <arm_sve.h>
+#if defined(__ARM_NEON_SVE_BRIDGE) && defined(__has_include) &&                \
+  __has_include(<arm_neon_sve_bridge.h>)
+#include <arm_neon_sve_bridge.h>
+#else
+#define svdup_neonq_f32(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat32_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#define svdup_neonq_f64(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat64_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#endif
+
+#define RESET_A_POINTER() a_offset = A;
+
+#define CREATE_A_POINTER(m, scale) FLOAT* a_offset##m = a_offset + scale;
+#define UPDATE_A_POINTER(scale) a_offset = a_offset + scale;
+#define A_ELEMENT_K(m, offset_k) *(a_offset##m + (k + offset_k) * lda)
+#define A_ELEMENT(m) A_ELEMENT_K(m, 0)
+
+#define RESET_B_POINTER() b_offset = B;
+
+#define CREATE_B_POINTER(n, scale) FLOAT* b_offset##n = b_offset + scale;
+#define UPDATE_B_POINTER(scale) b_offset = b_offset + scale;
+#define B_ELEMENT_K(n, offset_k) *(b_offset##n + (k + offset_k) * ldb)
+#define B_ELEMENT(n) B_ELEMENT_K(n, 0)
+
+#define CREATE_C_POINTER(n, scale) FLOAT* c_offset##n = c_offset + scale * ldc;
+#define INCR_C_POINTER(m, incr) // c_offset ## m += incr;
+#define UPDATE_C_POINTER(scale) c_offset = c_offset + scale * ldc;
+#define C_ELEMENT(m, n) *(c_offset##n + ((m * v_size) + i))
+
+// #undef C_ELEMENT
+// #define C_ELEMENT(m, n)             C[(i+(m))+(j+(n))*ldc]
+
+#define PACK_ELEMENT_K(n, offset_k) packed_b[(k + offset_k) * 4 + n]
+#define PACK_ELEMENT(n) PACK_ELEMENT_K(n, 0)
+
+// ASIMD
+#define DECLARE_RESULT_VECTOR4(m, n)                                           \
+  float32x4_t result##m##n = vdupq_n_f32(0.0);
+#define DECLARE_RESULT(m, n) float32_t result##m##n = 0.0;
+#define BROADCAST_LOAD_A4(m, offset_k)                                         \
+  float32x4_t a##m##_k##offset_k = vld1q_dup_f32(&A_ELEMENT_K(m, offset_k));
+#define LOAD_A1(m, offset_k)                                                   \
+  float32_t a##m##_k##offset_k = A_ELEMENT_K(m, offset_k);
+#define VECTOR_LOAD_B4(n, offset_k)                                            \
+  float32x4_t b##n##_k##offset_k = vld1q_f32(&B_ELEMENT_K(n, offset_k));
+#define GATHER_LOAD_B4(n, offset_k)                                            \
+  float32x4_t b##n##_k##offset_k = vdupq_n_f32(B_ELEMENT_K(n, offset_k));      \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 1, offset_k), b##n##_k##offset_k, 1);       \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 2, offset_k), b##n##_k##offset_k, 2);       \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 3, offset_k), b##n##_k##offset_k, 3);
+#define VECTOR_UNPACK_B4(n, offset_k)                                          \
+  float32x4_t b##n##_k##offset_k = vld1q_f32(&PACK_ELEMENT_K(n, offset_k));
+#define VECTOR_PACK_B4(n, offset_k)                                            \
+  vst1q_f32(&PACK_ELEMENT_K(n, offset_k), b##n##_k##offset_k);
+#define PACK_B0(n, offset_k)                                                   \
+  PACK_ELEMENT_K(n, offset_k) = vget_lane_f32(b##n##_k##offset_k, 0);
+#define UPDATE_RESULT_VECTOR4(m, n, offset_k)                                  \
+  result##m##n =                                                               \
+    vfmaq_f32(result##m##n, a##m##_k##offset_k, b##n##_k##offset_k);
+#define UPDATE_RESULT(m, n, offset_k)                                          \
+  result##m##n = result##m##n + a##m##_k##offset_k * b##n##_k##offset_k;
+#ifdef B0
+#define SCATTER_STORE4(m, n)                                                   \
+  result##m##n = vmulq_f32(result##m##n, vdupq_n_f32(alpha));                  \
+  C_ELEMENT(m, n + 0) = vgetq_lane_f32(result##m##n, 0);                       \
+  C_ELEMENT(m, n + 1) = vgetq_lane_f32(result##m##n, 1);                       \
+  C_ELEMENT(m, n + 2) = vgetq_lane_f32(result##m##n, 2);                       \
+  C_ELEMENT(m, n + 3) = vgetq_lane_f32(result##m##n, 3);
+#else
+#define SCATTER_STORE4(m, n)                                                   \
+  result##m##n = vmulq_f32(result##m##n, vdupq_n_f32(alpha));                  \
+  C_ELEMENT(m, n + 0) =                                                        \
+    C_ELEMENT(m, n + 0) * beta + vgetq_lane_f32(result##m##n, 0);              \
+  C_ELEMENT(m, n + 1) =                                                        \
+    C_ELEMENT(m, n + 1) * beta + vgetq_lane_f32(result##m##n, 1);              \
+  C_ELEMENT(m, n + 2) =                                                        \
+    C_ELEMENT(m, n + 2) * beta + vgetq_lane_f32(result##m##n, 2);              \
+  C_ELEMENT(m, n + 3) =                                                        \
+    C_ELEMENT(m, n + 3) * beta + vgetq_lane_f32(result##m##n, 3);
+#endif
+
+// SVE
+#define DECLARE_RESULT_VECTOR(m, n) svfloat32_t result##m##n = svdup_f32(0.0);
+#define BROADCAST_LOAD_A(m, offset_k)                                          \
+  svfloat32_t a##s##m##_k##offset_k = svdup_f32(A_ELEMENT_K(m, offset_k));
+#define BROADCAST_LOAD_B(n, offset_k)                                          \
+  svfloat32_t b##s##n##_k##offset_k = svdup_f32(B_ELEMENT_K(n, offset_k));
+#define VECTOR_LOAD_A(pg, m, offset_k)                                         \
+  svfloat32_t a##s##m##_k##offset_k = svld1(pg, &A_ELEMENT_K(m, offset_k));
+#define QUADWORD_LOAD_B(n, offset_k)                                           \
+  svfloat32_t b##s##n##_k##offset_k =                                          \
+    svld1rq(pg_true, &B_ELEMENT_K(n, offset_k));
+#define PACK_B(n, offset_k)                                                    \
+  svst1(pg_first, &PACK_ELEMENT_K(n, offset_k), b##s##n##_k##offset_k);
+#define VECTOR_PACK_B(n, offset_k)                                             \
+  svst1(pg_true, &PACK_ELEMENT_K(n* v_size, offset_k), b##s##n##_k##offset_k);
+#define QUADWORD_PACK_B(n, offset_k)                                           \
+  svst1(pg_quad, &PACK_ELEMENT_K(n, offset_k), b##s##n##_k##offset_k);
+#define UNPACK_VECTOR_B(n, offset_k)                                           \
+  svfloat32_t b##s##n##_k##offset_k =                                          \
+    svld1(pg_true, &PACK_ELEMENT_K(n * v_size, offset_k));
+#define UNPACK_BROADCAST_B(n, offset_k)                                        \
+  svfloat32_t b##s##n##_k##offset_k = svdup_f32(PACK_ELEMENT_K(n, offset_k));
+#define UNPACK_QUADWORD_B(n, offset_k)                                         \
+  svfloat32_t b##s##n##_k##offset_k =                                          \
+    svld1rq(pg_true, &PACK_ELEMENT_K(n, offset_k));
+#define UPDATE_RESULT_VECTOR(pg, m, n, offset_k)                               \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, a##s##m##_k##offset_k, b##s##n##_k##offset_k);
+#define UPDATE_RESULT_VECTOR_QUADWORD(m, n, outer, lane, offset_k)             \
+  result##m##n = svmla_lane(                                                   \
+    result##m##n, a##s##m##_k##offset_k, b##s##outer##_k##offset_k, lane);
+#ifdef B0
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#else
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, svld1(pg, &C_ELEMENT(m, n)), beta_vec);          \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n = svmla_m(pg,                                                   \
+                         result##m##n,                                         \
+                         svld1_gather_index(pg, &C_ELEMENT(m, n), ldc_vec),    \
+                         beta_vec);                                            \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#endif
+
+#ifndef LIKELY
+#ifdef __GNUC__
+#define LIKELY(x) __builtin_expect(!!(x), 1)
+#else
+#define LIKELY(x) (x)
+#endif
+#endif
+
+#ifdef B0
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT* C,
+      BLASLONG ldc)
+#else
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT beta,
+      FLOAT* C,
+      BLASLONG ldc)
+#endif
+{
+  const uint64_t v_size = svcntw();
+  const uint64_t v_size2 = v_size * 2;
+  const svbool_t pg_true = svptrue_b32();
+  const svbool_t pg_quad = svwhilelt_b32(0, 4);
+  const svbool_t pg_first = svwhilelt_b32(0, 1);
+  const svfloat32_t alpha_vec = svdup_f32(alpha);
+#ifndef B0
+  const svfloat32_t beta_vec = svdup_f32(beta);
+#endif
+  const BLASLONG n4 = N & -4;
+  const BLASLONG v_m2 = M & -v_size2;
+  const BLASLONG v_m1 = M & -v_size;
+
+  const int pack_b = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
+  FLOAT* packed_b =
+    (pack_b) ? packed_b = (FLOAT*)malloc(K * 4 * sizeof(FLOAT)) : NULL;
+
+  FLOAT* b_offset = B;
+  FLOAT* a_offset = A;
+  FLOAT* c_offset = C;
+
+  BLASLONG j = 0;
+  for (; j < n4; j += 4) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_C_POINTER(1, 1);
+    CREATE_C_POINTER(2, 2);
+    CREATE_C_POINTER(3, 3);
+    CREATE_B_POINTER(0, 0);
+    CREATE_B_POINTER(1, 1);
+    CREATE_B_POINTER(2, 2);
+    CREATE_B_POINTER(3, 3);
+
+    BLASLONG i = 0;
+    for (; i < v_m2; i += v_size2) {
+
+      CREATE_A_POINTER(0, 0);
+      CREATE_A_POINTER(1, v_size);
+      UPDATE_A_POINTER(v_size2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+      DECLARE_RESULT_VECTOR(1, 2);
+      DECLARE_RESULT_VECTOR(1, 3);
+
+      if (LIKELY(packed_b != NULL)) {
+        if (i == 0) {
+          for (; k < K; k++) {
+
+            QUADWORD_LOAD_B(0, 0);
+            QUADWORD_PACK_B(0, 0);
+            VECTOR_LOAD_A(pg_true, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+            VECTOR_LOAD_A(pg_true, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
+          }
+        } else {
+          for (; k < K; k++) {
+
+            UNPACK_QUADWORD_B(0, 0);
+            VECTOR_LOAD_A(pg_true, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+            VECTOR_LOAD_A(pg_true, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
+          }
+        }
+      } else {
+        for (; k < K; k++) {
+
+          QUADWORD_LOAD_B(0, 0);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+          VECTOR_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      VECTOR_STORE(pg_true, 1, 2);
+      VECTOR_STORE(pg_true, 1, 3);
+      INCR_C_POINTER(0, v_size2);
+      INCR_C_POINTER(1, v_size2);
+      INCR_C_POINTER(2, v_size2);
+      INCR_C_POINTER(3, v_size2);
+    }
+    for (; i < v_m1; i += v_size) {
+
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(v_size);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      if (LIKELY(packed_b != NULL)) {
+        for (; k < K; k++) {
+
+          UNPACK_QUADWORD_B(0, 0);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+        }
+      } else {
+        for (; k < K; k++) {
+
+          QUADWORD_LOAD_B(0, 0);
+          VECTOR_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      INCR_C_POINTER(0, v_size);
+      INCR_C_POINTER(1, v_size);
+      INCR_C_POINTER(2, v_size);
+      INCR_C_POINTER(3, v_size);
+    }
+    for (; i < M; i += v_size) {
+      const svbool_t pg_tail = svwhilelt_b32((uint32_t)i, (uint32_t)(M));
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(0);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      if (LIKELY(packed_b != NULL)) {
+        for (; k < K; k++) {
+
+          UNPACK_QUADWORD_B(0, 0);
+          VECTOR_LOAD_A(pg_tail, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+        }
+      } else {
+        for (; k < K; k++) {
+
+          QUADWORD_LOAD_B(0, 0);
+          VECTOR_LOAD_A(pg_tail, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      VECTOR_STORE(pg_tail, 0, 2);
+      VECTOR_STORE(pg_tail, 0, 3);
+      INCR_C_POINTER(0, 0);
+      INCR_C_POINTER(1, 0);
+      INCR_C_POINTER(2, 0);
+      INCR_C_POINTER(3, 0);
+    }
+
+    UPDATE_B_POINTER(4);
+    RESET_A_POINTER();
+    UPDATE_C_POINTER(4);
+  }
+  for (; j < N; j++) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_B_POINTER(0, 0);
+
+    BLASLONG i = 0;
+    for (; i < v_m2; i += v_size2) {
+
+      CREATE_A_POINTER(0, 0);
+      CREATE_A_POINTER(1, v_size);
+      UPDATE_A_POINTER(v_size2);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(1, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+        VECTOR_LOAD_A(pg_true, 1, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 1, 0);
+      INCR_C_POINTER(0, v_size2);
+    }
+    for (; i < v_m1; i += v_size) {
+
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(v_size);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      INCR_C_POINTER(0, v_size);
+    }
+    for (; i < M; i += v_size) {
+      const svbool_t pg_tail = svwhilelt_b32((uint32_t)i, (uint32_t)(M));
+      CREATE_A_POINTER(0, 0);
+      UPDATE_A_POINTER(0);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        VECTOR_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      INCR_C_POINTER(0, 0);
+    }
+
+    UPDATE_B_POINTER(1);
+    RESET_A_POINTER();
+    UPDATE_C_POINTER(1);
+  }
+
+  if (pack_b)
+    free(packed_b);
+
+  return 0;
+}

kernel/arm64/sgemm_small_kernel_tn_sve.c

@@ -0,0 +1,719 @@
+/***************************************************************************
+Copyright (c) 2024, 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.
+*****************************************************************************/
+
+#include "common.h"
+
+#include <arm_neon.h>
+#include <arm_sve.h>
+#if defined(__ARM_NEON_SVE_BRIDGE) && defined(__has_include) &&                \
+  __has_include(<arm_neon_sve_bridge.h>)
+#include <arm_neon_sve_bridge.h>
+#else
+#define svdup_neonq_f32(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat32_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#define svdup_neonq_f64(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat64_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#endif
+
+#define RESET_A_POINTER() a_offset = A;
+
+#define CREATE_A_POINTER(m, scale) FLOAT* a_offset##m = a_offset + scale * lda;
+#define UPDATE_A_POINTER(scale) a_offset = a_offset + scale * lda;
+#define A_ELEMENT_K(m, offset_k) *(a_offset##m + (k + offset_k))
+#define A_ELEMENT(m) A_ELEMENT_K(m, 0)
+
+#define RESET_B_POINTER() b_offset = B;
+
+#define CREATE_B_POINTER(n, scale) FLOAT* b_offset##n = b_offset + scale * ldb;
+#define UPDATE_B_POINTER(scale) b_offset = b_offset + scale * ldb;
+#define B_ELEMENT_K(n, offset_k) *(b_offset##n + (k + offset_k))
+#define B_ELEMENT(n) B_ELEMENT_K(n, 0)
+
+#define CREATE_C_POINTER(m, scale) FLOAT* c_offset##m = c_offset + scale;
+#define INCR_C_POINTER(m, incr) // c_offset ## m += incr * ldc;
+#define UPDATE_C_POINTER(scale) c_offset += scale;
+#define C_ELEMENT(m, n)                                                        \
+  *(c_offset##m + ((j + n) * ldc)) // C[(i+(m))+(j+(n))*ldc]
+
+// #undef C_ELEMENT
+// #define C_ELEMENT(m, n)             C[(i+(m))+(j+(n))*ldc]
+
+#define PACK_ELEMENT_K(m, offset_k) packed_a[(k + offset_k) * v_size2 + m]
+#define PACK_ELEMENT(m) PACK_ELEMENT_K(m, 0)
+
+// ASIMD
+#define DECLARE_RESULT_VECTOR4(m, n)                                           \
+  float32x4_t result##m##n = vdupq_n_f32(0.0);
+#define DECLARE_RESULT(m, n) float32_t result##m##n = 0.0;
+#define BROADCAST_LOAD_A4(m, offset_k)                                         \
+  float32x4_t a##m##_k##offset_k = vld1q_dup_f32(&A_ELEMENT_K(m, offset_k));
+#define LOAD_A1(m, offset_k)                                                   \
+  float32_t a##m##_k##offset_k = A_ELEMENT_K(m, offset_k);
+#define GATHER_LOAD_B4(n, offset_k)                                            \
+  float32x4_t b##n##_k##offset_k = vdupq_n_f32(B_ELEMENT_K(n, offset_k));      \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 1, offset_k), b##n##_k##offset_k, 1);       \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 2, offset_k), b##n##_k##offset_k, 2);       \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 3, offset_k), b##n##_k##offset_k, 3);
+#define VECTOR_UNPACK_B4(n, offset_k)                                          \
+  float32x4_t b##n##_k##offset_k = vld1q_f32(&PACK_ELEMENT_K(n, offset_k));
+#define PACK_B0(n, offset_k)                                                   \
+  PACK_ELEMENT_K(n, offset_k) = vget_lane_f32(b##n##_k##offset_k, 0);
+#define UPDATE_RESULT_VECTOR4(m, n, offset_k)                                  \
+  result##m##n =                                                               \
+    vfmaq_f32(result##m##n, a##m##_k##offset_k, b##n##_k##offset_k);
+#define UPDATE_RESULT(m, n, offset_k)                                          \
+  result##m##n = result##m##n + a##m##_k##offset_k * b##n##_k##offset_k;
+#ifdef B0
+#define SCATTER_STORE4(m, n)                                                   \
+  result##m##n = vmulq_f32(result##m##n, vdupq_n_f32(alpha));                  \
+  C_ELEMENT(m, n + 0) = vgetq_lane_f32(result##m##n, 0);                       \
+  C_ELEMENT(m, n + 1) = vgetq_lane_f32(result##m##n, 1);                       \
+  C_ELEMENT(m, n + 2) = vgetq_lane_f32(result##m##n, 2);                       \
+  C_ELEMENT(m, n + 3) = vgetq_lane_f32(result##m##n, 3);
+#else
+#define SCATTER_STORE4(m, n)                                                   \
+  result##m##n = vmulq_f32(result##m##n, vdupq_n_f32(alpha));                  \
+  C_ELEMENT(m, n + 0) =                                                        \
+    C_ELEMENT(m, n + 0) * beta + vgetq_lane_f32(result##m##n, 0);              \
+  C_ELEMENT(m, n + 1) =                                                        \
+    C_ELEMENT(m, n + 1) * beta + vgetq_lane_f32(result##m##n, 1);              \
+  C_ELEMENT(m, n + 2) =                                                        \
+    C_ELEMENT(m, n + 2) * beta + vgetq_lane_f32(result##m##n, 2);              \
+  C_ELEMENT(m, n + 3) =                                                        \
+    C_ELEMENT(m, n + 3) * beta + vgetq_lane_f32(result##m##n, 3);
+#endif
+
+// SVE
+#define DECLARE_RESULT_VECTOR(m, n) svfloat32_t result##m##n = svdup_f32(0.0);
+#define BROADCAST_LOAD_A(m, offset_k)                                          \
+  svfloat32_t a##s##m##_k##offset_k = svdup_f32(A_ELEMENT_K(m, offset_k));
+#define BROADCAST_LOAD_B(n, offset_k)                                          \
+  svfloat32_t b##s##n##_k##offset_k = svdup_f32(B_ELEMENT_K(n, offset_k));
+#define VECTOR_LOAD_A(pg, m, offset_k)                                         \
+  svfloat32_t a##s##m##_k##offset_k = svld1(pg, &A_ELEMENT_K(m, offset_k));
+#define GATHER_LOAD_A(pg, m, offset_k)                                         \
+  svfloat32_t a##s##m##_k##offset_k =                                          \
+    svld1_gather_index(pg, &A_ELEMENT_K(m, offset_k), lda_vec);
+#define PACK_A(m, offset_k)                                                    \
+  svst1(pg_first, &PACK_ELEMENT_K(m, offset_k), a##s##m##_k##offset_k);
+#define VECTOR_PACK_A(m, offset_k)                                             \
+  svst1(pg_true, &PACK_ELEMENT_K(m* v_size, offset_k), a##s##m##_k##offset_k);
+#define QUADWORD_PACK_A(m, offset_k)                                           \
+  svst1(pg_quad, &PACK_ELEMENT_K(m, offset_k), a##s##m##_k##offset_k);
+#define UNPACK_VECTOR_A(m, offset_k)                                           \
+  svfloat32_t a##s##m##_k##offset_k =                                          \
+    svld1(pg_true, &PACK_ELEMENT_K(m * v_size, offset_k));
+#define UNPACK_BROADCAST_A(m, offset_k)                                        \
+  svfloat32_t a##s##m##_k##offset_k = svdup_f32(PACK_ELEMENT_K(m, offset_k));
+#define UNPACK_QUADWORD_A(m, offset_k)                                         \
+  svfloat32_t a##s##m##_k##offset_k =                                          \
+    svld1rq(pg_true, &PACK_ELEMENT_K(m, offset_k));
+#define UPDATE_RESULT_VECTOR(pg, m, n, offset_k)                               \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, a##s##m##_k##offset_k, b##s##n##_k##offset_k);
+#define UPDATE_RESULT_VECTOR_QUADWORD(m, n, outer, lane, offset_k)             \
+  result##m##n = svmla_lane(                                                   \
+    result##m##n, a##s##m##_k##offset_k, b##s##outer##_k##offset_k, lane);
+#ifdef B0
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#else
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, svld1(pg, &C_ELEMENT(m, n)), beta_vec);          \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n = svmla_m(pg,                                                   \
+                         result##m##n,                                         \
+                         svld1_gather_index(pg, &C_ELEMENT(m, n), ldc_vec),    \
+                         beta_vec);                                            \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#endif
+
+#ifndef LIKELY
+#ifdef __GNUC__
+#define LIKELY(x) __builtin_expect(!!(x), 1)
+#else
+#define LIKELY(x) (x)
+#endif
+#endif
+
+#ifdef B0
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT* C,
+      BLASLONG ldc)
+#else
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT beta,
+      FLOAT* C,
+      BLASLONG ldc)
+#endif
+{
+  const uint64_t v_size = svcntw();
+  const uint64_t v_size2 = v_size * 2;
+  const svbool_t pg_true = svptrue_b32();
+  const svbool_t pg_quad = svwhilelt_b32(0, 4);
+  const svbool_t pg_first = svwhilelt_b32(0, 1);
+  const svfloat32_t alpha_vec = svdup_f32(alpha);
+#ifndef B0
+  const svfloat32_t beta_vec = svdup_f32(beta);
+#endif
+  const svuint32_t lda_vec = svindex_u32(0LL, lda);
+
+  const BLASLONG v_m2 = M & -v_size2;
+  const BLASLONG v_m1 = M & -v_size;
+  const BLASLONG n8 = N & -8;
+  const BLASLONG n4 = N & -4;
+
+  const int pack_a = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
+  FLOAT* packed_a =
+    (pack_a) ? packed_a = (FLOAT*)malloc(K * v_size2 * sizeof(FLOAT)) : NULL;
+
+  FLOAT* a_offset = A;
+  FLOAT* b_offset = B;
+  FLOAT* c_offset = C;
+
+  BLASLONG i = 0;
+  for (; i < v_m2; i += v_size2) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_C_POINTER(1, v_size);
+    CREATE_A_POINTER(0, 0);
+    CREATE_A_POINTER(1, v_size);
+
+    BLASLONG j = 0;
+    for (; j < n8; j += 8) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      CREATE_B_POINTER(4, 4);
+      CREATE_B_POINTER(5, 5);
+      CREATE_B_POINTER(6, 6);
+      CREATE_B_POINTER(7, 7);
+      UPDATE_B_POINTER(8);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(0, 4);
+      DECLARE_RESULT_VECTOR(0, 5);
+      DECLARE_RESULT_VECTOR(0, 6);
+      DECLARE_RESULT_VECTOR(0, 7);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+      DECLARE_RESULT_VECTOR(1, 2);
+      DECLARE_RESULT_VECTOR(1, 3);
+      DECLARE_RESULT_VECTOR(1, 4);
+      DECLARE_RESULT_VECTOR(1, 5);
+      DECLARE_RESULT_VECTOR(1, 6);
+      DECLARE_RESULT_VECTOR(1, 7);
+
+      if (LIKELY(packed_a != NULL)) {
+        if (j == 0) {
+          for (; k < K; k++) {
+
+            BROADCAST_LOAD_B(0, 0);
+            GATHER_LOAD_A(pg_true, 0, 0);
+            VECTOR_PACK_A(0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+            BROADCAST_LOAD_B(1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+            GATHER_LOAD_A(pg_true, 1, 0);
+            VECTOR_PACK_A(1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+            BROADCAST_LOAD_B(2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+            BROADCAST_LOAD_B(3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+            BROADCAST_LOAD_B(4, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 4, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 4, 0);
+            BROADCAST_LOAD_B(5, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 5, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 5, 0);
+            BROADCAST_LOAD_B(6, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 6, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 6, 0);
+            BROADCAST_LOAD_B(7, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 7, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 7, 0);
+          }
+        } else {
+          for (; k < K; k++) {
+
+            BROADCAST_LOAD_B(0, 0);
+            UNPACK_VECTOR_A(0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+            BROADCAST_LOAD_B(1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+            UNPACK_VECTOR_A(1, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+            BROADCAST_LOAD_B(2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+            BROADCAST_LOAD_B(3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+            BROADCAST_LOAD_B(4, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 4, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 4, 0);
+            BROADCAST_LOAD_B(5, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 5, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 5, 0);
+            BROADCAST_LOAD_B(6, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 6, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 6, 0);
+            BROADCAST_LOAD_B(7, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 0, 7, 0);
+            UPDATE_RESULT_VECTOR(pg_true, 1, 7, 0);
+          }
+        }
+      } else {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+          BROADCAST_LOAD_B(2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+          BROADCAST_LOAD_B(3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+          BROADCAST_LOAD_B(4, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 4, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 4, 0);
+          BROADCAST_LOAD_B(5, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 5, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 5, 0);
+          BROADCAST_LOAD_B(6, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 6, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 6, 0);
+          BROADCAST_LOAD_B(7, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 7, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 7, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 0, 4);
+      VECTOR_STORE(pg_true, 0, 5);
+      VECTOR_STORE(pg_true, 0, 6);
+      VECTOR_STORE(pg_true, 0, 7);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      VECTOR_STORE(pg_true, 1, 2);
+      VECTOR_STORE(pg_true, 1, 3);
+      VECTOR_STORE(pg_true, 1, 4);
+      VECTOR_STORE(pg_true, 1, 5);
+      VECTOR_STORE(pg_true, 1, 6);
+      VECTOR_STORE(pg_true, 1, 7);
+      INCR_C_POINTER(0, 8);
+      INCR_C_POINTER(1, 8);
+    }
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+      DECLARE_RESULT_VECTOR(1, 2);
+      DECLARE_RESULT_VECTOR(1, 3);
+
+      if (LIKELY(packed_a != NULL)) {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          UNPACK_VECTOR_A(0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+          UNPACK_VECTOR_A(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+          BROADCAST_LOAD_B(2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+          BROADCAST_LOAD_B(3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+        }
+      } else {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          BROADCAST_LOAD_B(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
+          BROADCAST_LOAD_B(2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
+          BROADCAST_LOAD_B(3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      VECTOR_STORE(pg_true, 1, 2);
+      VECTOR_STORE(pg_true, 1, 3);
+      INCR_C_POINTER(0, 4);
+      INCR_C_POINTER(1, 4);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(1, 0);
+
+      if (LIKELY(packed_a != NULL)) {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          UNPACK_VECTOR_A(0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          UNPACK_VECTOR_A(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+        }
+      } else {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 1, 0);
+      INCR_C_POINTER(0, 1);
+      INCR_C_POINTER(1, 1);
+    }
+
+    UPDATE_A_POINTER(v_size2);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(v_size2);
+  }
+  for (; i < v_m1; i += v_size) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_A_POINTER(0, 0);
+
+    BLASLONG j = 0;
+    for (; j < n8; j += 8) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      CREATE_B_POINTER(4, 4);
+      CREATE_B_POINTER(5, 5);
+      CREATE_B_POINTER(6, 6);
+      CREATE_B_POINTER(7, 7);
+      UPDATE_B_POINTER(8);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(0, 4);
+      DECLARE_RESULT_VECTOR(0, 5);
+      DECLARE_RESULT_VECTOR(0, 6);
+      DECLARE_RESULT_VECTOR(0, 7);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+        BROADCAST_LOAD_B(1, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+        BROADCAST_LOAD_B(2, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+        BROADCAST_LOAD_B(3, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+        BROADCAST_LOAD_B(4, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 4, 0);
+        BROADCAST_LOAD_B(5, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 5, 0);
+        BROADCAST_LOAD_B(6, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 6, 0);
+        BROADCAST_LOAD_B(7, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 7, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 0, 4);
+      VECTOR_STORE(pg_true, 0, 5);
+      VECTOR_STORE(pg_true, 0, 6);
+      VECTOR_STORE(pg_true, 0, 7);
+      INCR_C_POINTER(0, 8);
+    }
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+        BROADCAST_LOAD_B(1, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
+        BROADCAST_LOAD_B(2, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
+        BROADCAST_LOAD_B(3, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      INCR_C_POINTER(0, 4);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      INCR_C_POINTER(0, 1);
+    }
+
+    UPDATE_A_POINTER(v_size);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(v_size);
+  }
+  for (; i < M; i += v_size) {
+    const svbool_t pg_tail = svwhilelt_b32((uint32_t)i, (uint32_t)(M));
+    CREATE_C_POINTER(0, 0);
+    CREATE_A_POINTER(0, 0);
+
+    BLASLONG j = 0;
+    for (; j < n8; j += 8) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      CREATE_B_POINTER(4, 4);
+      CREATE_B_POINTER(5, 5);
+      CREATE_B_POINTER(6, 6);
+      CREATE_B_POINTER(7, 7);
+      UPDATE_B_POINTER(8);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(0, 4);
+      DECLARE_RESULT_VECTOR(0, 5);
+      DECLARE_RESULT_VECTOR(0, 6);
+      DECLARE_RESULT_VECTOR(0, 7);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+        BROADCAST_LOAD_B(1, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 1, 0);
+        BROADCAST_LOAD_B(2, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 2, 0);
+        BROADCAST_LOAD_B(3, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 3, 0);
+        BROADCAST_LOAD_B(4, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 4, 0);
+        BROADCAST_LOAD_B(5, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 5, 0);
+        BROADCAST_LOAD_B(6, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 6, 0);
+        BROADCAST_LOAD_B(7, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 7, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      VECTOR_STORE(pg_tail, 0, 2);
+      VECTOR_STORE(pg_tail, 0, 3);
+      VECTOR_STORE(pg_tail, 0, 4);
+      VECTOR_STORE(pg_tail, 0, 5);
+      VECTOR_STORE(pg_tail, 0, 6);
+      VECTOR_STORE(pg_tail, 0, 7);
+      INCR_C_POINTER(0, 8);
+    }
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+        BROADCAST_LOAD_B(1, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 1, 0);
+        BROADCAST_LOAD_B(2, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 2, 0);
+        BROADCAST_LOAD_B(3, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 3, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      VECTOR_STORE(pg_tail, 0, 2);
+      VECTOR_STORE(pg_tail, 0, 3);
+      INCR_C_POINTER(0, 4);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      INCR_C_POINTER(0, 1);
+    }
+
+    UPDATE_A_POINTER(0);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(0);
+  }
+
+  if (pack_a)
+    free(packed_a);
+
+  return 0;
+}

kernel/arm64/sgemm_small_kernel_tt_sve.c

@@ -0,0 +1,678 @@
+/***************************************************************************
+Copyright (c) 2024, 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.
+*****************************************************************************/
+
+#include "common.h"
+
+#include <arm_neon.h>
+#include <arm_sve.h>
+#if defined(__ARM_NEON_SVE_BRIDGE) && defined(__has_include) &&                \
+  __has_include(<arm_neon_sve_bridge.h>)
+#include <arm_neon_sve_bridge.h>
+#else
+#define svdup_neonq_f32(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat32_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#define svdup_neonq_f64(fixed_reg)                                             \
+  ({                                                                           \
+    svfloat64_t scalable_reg;                                                  \
+    asm("mov %0.q, %q1" : "=w"(scalable_reg) : "w"(fixed_reg) :);              \
+    scalable_reg;                                                              \
+  })
+#endif
+
+#define RESET_A_POINTER() a_offset = A;
+
+#define CREATE_A_POINTER(m, scale) FLOAT* a_offset##m = a_offset + scale * lda;
+#define UPDATE_A_POINTER(scale) a_offset = a_offset + scale * lda;
+#define A_ELEMENT_K(m, offset_k) *(a_offset##m + (k + offset_k))
+#define A_ELEMENT(m) A_ELEMENT_K(m, 0)
+
+#define RESET_B_POINTER() b_offset = B;
+
+#define CREATE_B_POINTER(n, scale) FLOAT* b_offset##n = b_offset + scale;
+#define UPDATE_B_POINTER(scale) b_offset = b_offset + scale;
+#define B_ELEMENT_K(n, offset_k) *(b_offset##n + (k + offset_k) * ldb)
+#define B_ELEMENT(n) B_ELEMENT_K(n, 0)
+
+#define CREATE_C_POINTER(m, scale) FLOAT* c_offset##m = c_offset + scale;
+#define INCR_C_POINTER(m, incr) // c_offset ## m += incr * ldc;
+#define UPDATE_C_POINTER(scale) c_offset += scale;
+#define C_ELEMENT(m, n)                                                        \
+  *(c_offset##m + ((j + n) * ldc)) // C[(i+(m))+(j+(n))*ldc]
+
+// #undef C_ELEMENT
+// #define C_ELEMENT(m, n)             C[(i+(m))+(j+(n))*ldc]
+
+#define PACK_ELEMENT_K(m, offset_k) packed_a[(k + offset_k) * v_size2 + m]
+#define PACK_ELEMENT(m) PACK_ELEMENT_K(m, 0)
+
+// ASIMD
+#define DECLARE_RESULT_VECTOR4(m, n)                                           \
+  float32x4_t result##m##n = vdupq_n_f32(0.0);
+#define DECLARE_RESULT(m, n) float32_t result##m##n = 0.0;
+#define BROADCAST_LOAD_A4(m, offset_k)                                         \
+  float32x4_t a##m##_k##offset_k = vld1q_dup_f32(&A_ELEMENT_K(m, offset_k));
+#define LOAD_A1(m, offset_k)                                                   \
+  float32_t a##m##_k##offset_k = A_ELEMENT_K(m, offset_k);
+#define VECTOR_LOAD_B4(n, offset_k)                                            \
+  float32x4_t b##n##_k##offset_k = vld1q_f32(&B_ELEMENT_K(n, offset_k));
+#define GATHER_LOAD_B4(n, offset_k)                                            \
+  float32x4_t b##n##_k##offset_k = vdupq_n_f32(B_ELEMENT_K(n, offset_k));      \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 1, offset_k), b##n##_k##offset_k, 1);       \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 2, offset_k), b##n##_k##offset_k, 2);       \
+  b##n##_k##offset_k =                                                         \
+    vsetq_lane_f32(B_ELEMENT_K(n + 3, offset_k), b##n##_k##offset_k, 3);
+#define VECTOR_UNPACK_B4(n, offset_k)                                          \
+  float32x4_t b##n##_k##offset_k = vld1q_f32(&PACK_ELEMENT_K(n, offset_k));
+#define VECTOR_PACK_B4(n, offset_k)                                            \
+  vst1q_f32(&PACK_ELEMENT_K(n, offset_k), b##n##_k##offset_k);
+#define PACK_B0(n, offset_k)                                                   \
+  PACK_ELEMENT_K(n, offset_k) = vget_lane_f32(b##n##_k##offset_k, 0);
+#define UPDATE_RESULT_VECTOR4(m, n, offset_k)                                  \
+  result##m##n =                                                               \
+    vfmaq_f32(result##m##n, a##m##_k##offset_k, b##n##_k##offset_k);
+#define UPDATE_RESULT(m, n, offset_k)                                          \
+  result##m##n = result##m##n + a##m##_k##offset_k * b##n##_k##offset_k;
+#ifdef B0
+#define VECTOR_STORE4(m, n)                                                    \
+  vst1q_f32(&C_ELEMENT(m, n), vmulq_f32(result##m##n, vdupq_n_f32(alpha)));
+#define STORE(m, n) C_ELEMENT(m, n) = alpha * result##m##n;
+#else
+#define VECTOR_STORE4(m, n)                                                    \
+  result##m##n = vmulq_f32(result##m##n, vdupq_n_f32(alpha));                  \
+  result##m##n =                                                               \
+    vfmaq_f32(result##m##n, vld1q_f32(&C_ELEMENT(m, n)), vdupq_n_f32(beta));   \
+  vst1q_f32(&C_ELEMENT(m, n), result##m##n);
+#define STORE(m, n)                                                            \
+  C_ELEMENT(m, n) = C_ELEMENT(m, n) * beta + alpha * result##m##n;
+#endif
+
+// SVE
+#define DECLARE_RESULT_VECTOR(m, n) svfloat32_t result##m##n = svdup_f32(0.0);
+#define BROADCAST_LOAD_A(m, offset_k)                                          \
+  svfloat32_t a##s##m##_k##offset_k = svdup_f32(A_ELEMENT_K(m, offset_k));
+#define BROADCAST_LOAD_B(n, offset_k)                                          \
+  svfloat32_t b##s##n##_k##offset_k = svdup_f32(B_ELEMENT_K(n, offset_k));
+#define VECTOR_LOAD_A(pg, m, offset_k)                                         \
+  svfloat32_t a##s##m##_k##offset_k = svld1(pg, &A_ELEMENT_K(m, offset_k));
+#define QUADWORD_LOAD_B(n, offset_k)                                           \
+  svfloat32_t b##s##n##_k##offset_k =                                          \
+    svld1rq(pg_true, &B_ELEMENT_K(n, offset_k));
+#define GATHER_LOAD_A(pg, m, offset_k)                                         \
+  svfloat32_t a##s##m##_k##offset_k =                                          \
+    svld1_gather_index(pg, &A_ELEMENT_K(m, offset_k), lda_vec);
+#define PACK_A(m, offset_k)                                                    \
+  svst1(pg_first, &PACK_ELEMENT_K(m, offset_k), a##s##m##_k##offset_k);
+#define VECTOR_PACK_A(m, offset_k)                                             \
+  svst1(pg_true, &PACK_ELEMENT_K(m* v_size, offset_k), a##s##m##_k##offset_k);
+#define QUADWORD_PACK_A(m, offset_k)                                           \
+  svst1(pg_quad, &PACK_ELEMENT_K(m, offset_k), a##s##m##_k##offset_k);
+#define UNPACK_VECTOR_A(m, offset_k)                                           \
+  svfloat32_t a##s##m##_k##offset_k =                                          \
+    svld1(pg_true, &PACK_ELEMENT_K(m * v_size, offset_k));
+#define UNPACK_BROADCAST_A(m, offset_k)                                        \
+  svfloat32_t a##s##m##_k##offset_k = svdup_f32(PACK_ELEMENT_K(m, offset_k));
+#define UNPACK_QUADWORD_A(m, offset_k)                                         \
+  svfloat32_t a##s##m##_k##offset_k =                                          \
+    svld1rq(pg_true, &PACK_ELEMENT_K(m, offset_k));
+#define UPDATE_RESULT_VECTOR(pg, m, n, offset_k)                               \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, a##s##m##_k##offset_k, b##s##n##_k##offset_k);
+#define UPDATE_RESULT_VECTOR_QUADWORD(m, n, outer, lane, offset_k)             \
+  result##m##n = svmla_lane(                                                   \
+    result##m##n, a##s##m##_k##offset_k, b##s##outer##_k##offset_k, lane);
+#ifdef B0
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#else
+#define VECTOR_STORE(pg, m, n)                                                 \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n =                                                               \
+    svmla_m(pg, result##m##n, svld1(pg, &C_ELEMENT(m, n)), beta_vec);          \
+  svst1(pg, &C_ELEMENT(m, n), result##m##n);
+#define SCATTER_STORE(pg, m, n)                                                \
+  result##m##n = svmul_m(pg, result##m##n, alpha_vec);                         \
+  result##m##n = svmla_m(pg,                                                   \
+                         result##m##n,                                         \
+                         svld1_gather_index(pg, &C_ELEMENT(m, n), ldc_vec),    \
+                         beta_vec);                                            \
+  svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
+#endif
+
+#ifndef LIKELY
+#ifdef __GNUC__
+#define LIKELY(x) __builtin_expect(!!(x), 1)
+#else
+#define LIKELY(x) (x)
+#endif
+#endif
+
+#ifdef B0
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT* C,
+      BLASLONG ldc)
+#else
+int
+CNAME(BLASLONG M,
+      BLASLONG N,
+      BLASLONG K,
+      IFLOAT* A,
+      BLASLONG lda,
+      FLOAT alpha,
+      IFLOAT* B,
+      BLASLONG ldb,
+      FLOAT beta,
+      FLOAT* C,
+      BLASLONG ldc)
+#endif
+{
+  const uint64_t v_size = svcntw();
+  const uint64_t v_size2 = v_size * 2;
+  const svbool_t pg_true = svptrue_b32();
+  const svbool_t pg_quad = svwhilelt_b32(0, 4);
+  const svbool_t pg_first = svwhilelt_b32(0, 1);
+  const svfloat32_t alpha_vec = svdup_f32(alpha);
+#ifndef B0
+  const svfloat32_t beta_vec = svdup_f32(beta);
+#endif
+  const svuint32_t lda_vec = svindex_u32(0LL, lda);
+
+  const BLASLONG v_m2 = M & -v_size2;
+  const BLASLONG v_m1 = M & -v_size;
+  const BLASLONG n8 = N & -8;
+  const BLASLONG n4 = N & -4;
+
+  const int pack_a = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
+  FLOAT* packed_a =
+    (pack_a) ? packed_a = (FLOAT*)malloc(K * v_size2 * sizeof(FLOAT)) : NULL;
+
+  FLOAT* a_offset = A;
+  FLOAT* b_offset = B;
+  FLOAT* c_offset = C;
+
+  BLASLONG i = 0;
+  for (; i < v_m2; i += v_size2) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_C_POINTER(1, v_size);
+    CREATE_A_POINTER(0, 0);
+    CREATE_A_POINTER(1, v_size);
+
+    BLASLONG j = 0;
+    for (; j < n8; j += 8) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      CREATE_B_POINTER(4, 4);
+      CREATE_B_POINTER(5, 5);
+      CREATE_B_POINTER(6, 6);
+      CREATE_B_POINTER(7, 7);
+      UPDATE_B_POINTER(8);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(0, 4);
+      DECLARE_RESULT_VECTOR(0, 5);
+      DECLARE_RESULT_VECTOR(0, 6);
+      DECLARE_RESULT_VECTOR(0, 7);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+      DECLARE_RESULT_VECTOR(1, 2);
+      DECLARE_RESULT_VECTOR(1, 3);
+      DECLARE_RESULT_VECTOR(1, 4);
+      DECLARE_RESULT_VECTOR(1, 5);
+      DECLARE_RESULT_VECTOR(1, 6);
+      DECLARE_RESULT_VECTOR(1, 7);
+
+      if (LIKELY(packed_a != NULL)) {
+        if (j == 0) {
+          for (; k < K; k++) {
+
+            QUADWORD_LOAD_B(0, 0);
+            GATHER_LOAD_A(pg_true, 0, 0);
+            VECTOR_PACK_A(0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+            QUADWORD_LOAD_B(4, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
+            GATHER_LOAD_A(pg_true, 1, 0);
+            VECTOR_PACK_A(1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 0);
+          }
+        } else {
+          for (; k < K; k++) {
+
+            QUADWORD_LOAD_B(0, 0);
+            UNPACK_VECTOR_A(0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+            QUADWORD_LOAD_B(4, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
+            UNPACK_VECTOR_A(1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 0);
+            UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 0);
+          }
+        }
+      } else {
+        for (; k < K; k++) {
+
+          QUADWORD_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+          QUADWORD_LOAD_B(4, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 0, 4);
+      VECTOR_STORE(pg_true, 0, 5);
+      VECTOR_STORE(pg_true, 0, 6);
+      VECTOR_STORE(pg_true, 0, 7);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      VECTOR_STORE(pg_true, 1, 2);
+      VECTOR_STORE(pg_true, 1, 3);
+      VECTOR_STORE(pg_true, 1, 4);
+      VECTOR_STORE(pg_true, 1, 5);
+      VECTOR_STORE(pg_true, 1, 6);
+      VECTOR_STORE(pg_true, 1, 7);
+      INCR_C_POINTER(0, 8);
+      INCR_C_POINTER(1, 8);
+    }
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(1, 0);
+      DECLARE_RESULT_VECTOR(1, 1);
+      DECLARE_RESULT_VECTOR(1, 2);
+      DECLARE_RESULT_VECTOR(1, 3);
+
+      if (LIKELY(packed_a != NULL)) {
+        for (; k < K; k++) {
+
+          QUADWORD_LOAD_B(0, 0);
+          UNPACK_VECTOR_A(0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+          UNPACK_VECTOR_A(1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
+        }
+      } else {
+        for (; k < K; k++) {
+
+          QUADWORD_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
+          UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 1, 0);
+      VECTOR_STORE(pg_true, 1, 1);
+      VECTOR_STORE(pg_true, 1, 2);
+      VECTOR_STORE(pg_true, 1, 3);
+      INCR_C_POINTER(0, 4);
+      INCR_C_POINTER(1, 4);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(1, 0);
+
+      if (LIKELY(packed_a != NULL)) {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          UNPACK_VECTOR_A(0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          UNPACK_VECTOR_A(1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+        }
+      } else {
+        for (; k < K; k++) {
+
+          BROADCAST_LOAD_B(0, 0);
+          GATHER_LOAD_A(pg_true, 0, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+          GATHER_LOAD_A(pg_true, 1, 0);
+          UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
+        }
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 1, 0);
+      INCR_C_POINTER(0, 1);
+      INCR_C_POINTER(1, 1);
+    }
+
+    UPDATE_A_POINTER(v_size2);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(v_size2);
+  }
+  for (; i < v_m1; i += v_size) {
+
+    CREATE_C_POINTER(0, 0);
+    CREATE_A_POINTER(0, 0);
+
+    BLASLONG j = 0;
+    for (; j < n8; j += 8) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      CREATE_B_POINTER(4, 4);
+      CREATE_B_POINTER(5, 5);
+      CREATE_B_POINTER(6, 6);
+      CREATE_B_POINTER(7, 7);
+      UPDATE_B_POINTER(8);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(0, 4);
+      DECLARE_RESULT_VECTOR(0, 5);
+      DECLARE_RESULT_VECTOR(0, 6);
+      DECLARE_RESULT_VECTOR(0, 7);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+        QUADWORD_LOAD_B(4, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      VECTOR_STORE(pg_true, 0, 4);
+      VECTOR_STORE(pg_true, 0, 5);
+      VECTOR_STORE(pg_true, 0, 6);
+      VECTOR_STORE(pg_true, 0, 7);
+      INCR_C_POINTER(0, 8);
+    }
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      VECTOR_STORE(pg_true, 0, 1);
+      VECTOR_STORE(pg_true, 0, 2);
+      VECTOR_STORE(pg_true, 0, 3);
+      INCR_C_POINTER(0, 4);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_true, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_true, 0, 0);
+      INCR_C_POINTER(0, 1);
+    }
+
+    UPDATE_A_POINTER(v_size);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(v_size);
+  }
+  for (; i < M; i += v_size) {
+    const svbool_t pg_tail = svwhilelt_b32((uint32_t)i, (uint32_t)(M));
+    CREATE_C_POINTER(0, 0);
+    CREATE_A_POINTER(0, 0);
+
+    BLASLONG j = 0;
+    for (; j < n8; j += 8) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      CREATE_B_POINTER(4, 4);
+      CREATE_B_POINTER(5, 5);
+      CREATE_B_POINTER(6, 6);
+      CREATE_B_POINTER(7, 7);
+      UPDATE_B_POINTER(8);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+      DECLARE_RESULT_VECTOR(0, 4);
+      DECLARE_RESULT_VECTOR(0, 5);
+      DECLARE_RESULT_VECTOR(0, 6);
+      DECLARE_RESULT_VECTOR(0, 7);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+        QUADWORD_LOAD_B(4, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      VECTOR_STORE(pg_tail, 0, 2);
+      VECTOR_STORE(pg_tail, 0, 3);
+      VECTOR_STORE(pg_tail, 0, 4);
+      VECTOR_STORE(pg_tail, 0, 5);
+      VECTOR_STORE(pg_tail, 0, 6);
+      VECTOR_STORE(pg_tail, 0, 7);
+      INCR_C_POINTER(0, 8);
+    }
+    for (; j < n4; j += 4) {
+
+      CREATE_B_POINTER(0, 0);
+      CREATE_B_POINTER(1, 1);
+      CREATE_B_POINTER(2, 2);
+      CREATE_B_POINTER(3, 3);
+      UPDATE_B_POINTER(4);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+      DECLARE_RESULT_VECTOR(0, 1);
+      DECLARE_RESULT_VECTOR(0, 2);
+      DECLARE_RESULT_VECTOR(0, 3);
+
+      for (; k < K; k++) {
+
+        QUADWORD_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
+        UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      VECTOR_STORE(pg_tail, 0, 1);
+      VECTOR_STORE(pg_tail, 0, 2);
+      VECTOR_STORE(pg_tail, 0, 3);
+      INCR_C_POINTER(0, 4);
+    }
+    for (; j < N; j++) {
+
+      CREATE_B_POINTER(0, 0);
+      UPDATE_B_POINTER(1);
+
+      BLASLONG k = 0;
+      DECLARE_RESULT_VECTOR(0, 0);
+
+      for (; k < K; k++) {
+
+        BROADCAST_LOAD_B(0, 0);
+        GATHER_LOAD_A(pg_tail, 0, 0);
+        UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
+      }
+      VECTOR_STORE(pg_tail, 0, 0);
+      INCR_C_POINTER(0, 1);
+    }
+
+    UPDATE_A_POINTER(0);
+    RESET_B_POINTER();
+    UPDATE_C_POINTER(0);
+  }
+
+  if (pack_a)
+    free(packed_a);
+
+  return 0;
+}

kernel/loongarch64/scal.S

@@ -53,9 +53,86 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
    PROLOGUE
 
    li.d  TEMP, SIZE
+   ld.d XX, $sp, 0 // Load dummy2
+   slli.d XX, XX, BASE_SHIFT
    MTC  a1, $r0
    slli.d INCX, INCX, BASE_SHIFT
    bge $r0,    N, .L999
+   CMPEQ   $fcc0, ALPHA, a1
+   bceqz   $fcc0, .L50
+   beq     XX,    TEMP,  .L50  // if dummp2 == 1, do not directly copy 0
+   srai.d I, N, 3
+   bne INCX, TEMP, .L20
+   bge $r0,    I, .L15
+   .align 3
+
+.L12:
+   ST a1,  X,   0 * SIZE
+   ST a1,  X,   1 * SIZE
+   ST a1,  X,   2 * SIZE
+   ST a1,  X,   3 * SIZE
+   ST a1,  X,   4 * SIZE
+   ST a1,  X,   5 * SIZE
+   ST a1,  X,   6 * SIZE
+   ST a1,  X,   7 * SIZE
+   addi.w  I, I, -1
+   addi.d X, X, 8 * SIZE
+   blt $r0,    I, .L12
+   .align 3
+
+.L15:
+   andi    I,  N, 7
+   bge $r0,    I, .L999
+   .align  3
+.L16:
+   ST a1,  X,   0 * SIZE
+   addi.d  I, I, -1
+   addi.d X, X, SIZE
+   blt $r0,    I, .L16
+   move $r4, $r17
+   fmov.d $f0, $f22
+   jirl    $r0, $r1, 0x0
+   .align 3
+
+.L20:
+   srai.d  I, N, 3
+   bge $r0,    I, .L25
+   .align 3
+
+.L22:
+   ST a1,  X,   0 * SIZE
+   add.d   X, X, INCX
+   ST a1,  X,   0 * SIZE
+   add.d   X, X, INCX
+   ST a1,  X,   0 * SIZE
+   add.d   X, X, INCX
+   ST a1,  X,   0 * SIZE
+   add.d   X, X, INCX
+   ST a1,  X,   0 * SIZE
+   add.d   X, X, INCX
+   ST a1,  X,   0 * SIZE
+   add.d   X, X, INCX
+   ST a1,  X,   0 * SIZE
+   add.d   X, X, INCX
+   ST a1,  X,   0 * SIZE
+   addi.d  I, I, -1
+   add.d  X, X, INCX
+   blt $r0,    I, .L22
+   .align 3
+
+.L25:
+   andi    I,  N, 7
+   bge $r0,    I, .L999
+   .align  3
+.L26:
+   addi.d  I, I, -1
+   ST a1,  X,   0 * SIZE
+   add.d  X, X, INCX
+   blt $r0,    I, .L26
+   move $r4, $r17
+   fmov.d $f0, $f22
+   jirl    $r0, $r1, 0x0
+   .align 3
 
 .L50:
    srai.d I, N, 3

kernel/loongarch64/scal_lasx.S

@@ -52,17 +52,21 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     bge $r0, N, .L999
     bge $r0, INCX, .L999
     li.d TEMP, 1
+    ld.d t1, $sp,  0 // Load dummp2
     movgr2fr.d a1, $r0
     FFINT  a1, a1
     movgr2fr.d a2, TEMP
     FFINT  a2, a2
     slli.d TEMP, TEMP, BASE_SHIFT
     slli.d INCX, INCX, BASE_SHIFT
+    slli.d t1,   t1,   BASE_SHIFT
+    CMPEQ  $fcc0, ALPHA, a1
+    bcnez  $fcc0, .L20   //ALPHA==0
     CMPEQ  $fcc0, ALPHA, a2
     bcnez $fcc0, .L999  //ALPHA==1 return
-
+.L1:
     srai.d I, N, 3
-    beq INCX, TEMP, .L30    //ALPHA!=1 and INCX==1
+    beq INCX, TEMP, .L30    //ALPHA !=0|1 and INCX==1
     MTG  TEMP, ALPHA
 #ifdef DOUBLE
     xvreplgr2vr.d VALPHA, TEMP
@@ -72,7 +76,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     move XX, X
     .align 3
 
-.L10:   //ALPHA!=1 and INCX!=1
+.L10:   //ALPHA !=0|1 and INCX!=1
     bge $r0, I, .L32
     .align 3
 .L11:
@@ -165,6 +169,75 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     blt $r0, I, .L11
     b .L32
     .align 3
+
+.L20:
+    beq  t1, TEMP, .L1 // if dummp2 == 1, do not directly copy 0
+    srai.d I, N, 3
+    beq INCX, TEMP, .L24
+    bge $r0, I, .L22
+    .align 3
+
+.L21:
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    addi.d  I, I, -1
+    blt $r0, I, .L21
+    .align 3
+
+.L22:
+    andi I, N, 7
+    bge $r0, I, .L999
+    .align 3
+
+.L23:
+    ST  a1, X, 0 * SIZE
+    addi.d I, I, -1
+    add.d X, X, INCX
+    blt $r0, I, .L23
+    jirl $r0, $r1, 0
+    .align 3
+
+.L24:
+    bge $r0, I, .L26 /*N<8 INCX==1*/
+    .align 3
+.L25:
+    xvxor.v VX0, VX0, VX0
+    xvst VX0, X, 0 * SIZE
+#ifdef DOUBLE
+    xvst VX0, X, 4 * SIZE
+#endif
+    addi.d I, I, -1
+    addi.d X, X, 8 * SIZE
+    blt $r0, I, .L25
+    .align 3
+
+.L26:
+    andi I, N, 7
+    bge $r0, I, .L999
+    .align 3
+
+.L27:
+    ST  a1, X, 0 * SIZE
+    addi.d I, I, -1
+    addi.d X, X, SIZE
+    blt $r0, I, .L27
+    jirl $r0, $r1, 0
+    .align 3
+
 .L30:
     bge $r0, I, .L32/*N<8 INCX==1*/
     MTG TEMP, ALPHA

kernel/loongarch64/scal_lsx.S

@@ -51,6 +51,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
     bge $r0, N, .L999
     bge $r0, INCX, .L999
+    ld.d t1, $sp, 0 // Load dummy2
     li.d TEMP, 1
     movgr2fr.d a1, $r0
     FFINT  a1, a1
@@ -58,10 +59,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     FFINT  a2, a2
     slli.d TEMP, TEMP, BASE_SHIFT
     slli.d INCX, INCX, BASE_SHIFT
+    slli.d t1,   t1,   BASE_SHIFT
+    CMPEQ  $fcc0, ALPHA, a1
+    bcnez  $fcc0, .L20   //ALPHA==0
     CMPEQ  $fcc0, ALPHA, a2
     bcnez $fcc0, .L999  //ALPHA==1 return
+.L1:
     srai.d I, N, 3
-    beq INCX, TEMP, .L30    //ALPHA!=1 and INCX==1
+    beq INCX, TEMP, .L30    //ALPHA !=0|1 and INCX==1
     MTG  TEMP, ALPHA
 #ifdef DOUBLE
     vreplgr2vr.d VALPHA, TEMP
@@ -71,7 +76,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     move XX, X
     .align 3
 
-.L10:   //ALPHA!=1 and INCX!=1
+.L10:   //ALPHA !=0|1 and INCX!=1
     bge $r0, I, .L32
     .align 3
 
@@ -169,6 +174,79 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     b .L32
     .align 3
 
+.L20:
+    beq t1, TEMP, .L1 // if dummp2 == 1, do not directly copy 0
+    srai.d I, N, 3
+    beq INCX, TEMP, .L24
+    bge $r0, I, .L22
+    .align 3
+
+.L21:
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    ST a1, X, 0
+    add.d X, X, INCX
+    addi.d  I, I, -1
+    blt $r0, I, .L21
+    .align 3
+
+.L22:
+    andi I, N, 7
+    bge $r0, I, .L999
+    .align 3
+
+.L23:
+    ST  a1, X, 0 * SIZE
+    addi.d I, I, -1
+    add.d X, X, INCX
+    blt $r0, I, .L23
+    jirl $r0, $r1, 0
+    .align 3
+
+.L24:
+    bge $r0, I, .L26 /*N<8 INCX==1*/
+    .align 3
+
+.L25:
+    vxor.v VX0, VX0, VX0
+    vst VX0, X, 0 * SIZE
+#ifdef DOUBLE
+    vst VX0, X, 2 * SIZE
+    vst VX0, X, 4 * SIZE
+    vst VX0, X, 6 * SIZE
+#else
+    vst VX0, X, 4 * SIZE
+#endif
+    addi.d I, I, -1
+    addi.d X, X, 8 * SIZE
+    blt $r0, I, .L25
+    .align 3
+
+.L26:
+    andi I, N, 7
+    bge $r0, I, .L999
+    .align 3
+
+.L27:
+    ST  a1, X, 0 * SIZE
+    addi.d I, I, -1
+    addi.d X, X, SIZE
+    blt $r0, I, .L27
+    jirl $r0, $r1, 0
+    .align 3
+
 .L30:
     bge $r0, I, .L32/*N<8 INCX==1*/
     MTG  TEMP, ALPHA

kernel/mips/dscal_msa.c

@@ -42,7 +42,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x,
 
     if (1 == inc_x)
     {
-        if (0)     //if (0.0 == da )
+        if (0.0 == da && !dummy2)
         {
             v2f64 zero_v = {0.0, 0.0};
 
@@ -240,14 +240,12 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x,
     }
     else
     {
-        if (da == 0.0)
+        if (da == 0.0 && !dummy2)
         {
             for (i = n; i--;)
-            {         
-                if (isfinite(*x))
-                          *x = 0.0;
-                else
-                          *x = NAN;
+            {
+                *x = 0.0;
+
                 x += inc_x;
             }
         }

kernel/mips/scal.c

@@ -29,27 +29,34 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
 {
-	BLASLONG i=0,j=0;
-
-	while(j < n)
-	{
-
-		if ( da == 0.0 )
-			if (isnan(x[i])||isinf(x[i]))
-				x[i]=NAN;
-			else
-				x[i]=0.0;
-		else if (isnan(da))
-			x[i]=NAN;
-		else
-			x[i] = da * x[i] ;
-
-		i += inc_x ;
-		j++;
-
-	}
-	return 0;
-
+    BLASLONG i = 0, j = 0;
+
+    // Resolved issue 4728 when the caller is {s/d}scal
+    if (da == 0.0 && dummy2 == 1)
+    {
+        while(j < n)
+        {
+            x[i] = da * x[i] ;
+
+            i += inc_x ;
+            j++;
+        }
+    }
+    else
+    {
+        while(j < n)
+        {
+
+            if ( da == 0.0 )
+                x[i] = 0.0;
+            else
+                x[i] = da * x[i] ;
+
+            i += inc_x ;
+            j++;
+        }
+    }
+    return 0;
 }
 
 

kernel/mips/sscal_msa.c

@@ -42,7 +42,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x,
 
     if (1 == inc_x)
     {
-        if (0) // if (0.0 == da)
+        if (0.0 == da && !dummy2)
         {
             v4f32 zero_v = {0.0, 0.0, 0.0, 0.0};
 
@@ -255,14 +255,11 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x,
     }
     else
     {
-        if (0.0 == da)
+        if (0.0 == da && !dummy2)
         {
             for (i = n; i--;)
             {
-                if (isfinite(*x))
-                          *x = 0;
-                else
-                          *x = NAN;
+                *x = 0;
                 x += inc_x;
             }
         }

kernel/mips64/scal.S

@@ -48,6 +48,7 @@
 #define TEMP	$3
 
 #define XX	$5
+#define DUMMY2 	$6
 
 #define ALPHA	$f15
 
@@ -73,13 +74,13 @@
 	blez	N, .L999
 	dsll	INCX, INCX, BASE_SHIFT
 
-	CMPEQ	$fcc0, ALPHA, a1
-	NOP
+	CMPEQ	$fcc0,  ALPHA, a1
+	LDARG   DUMMY2, 8($sp)
 
 	bc1f	$fcc0, .L50
-	NOP
+	dsll    DUMMY2, DUMMY2, BASE_SHIFT
 
-	bc1t	$fcc0, .L50
+	beq     DUMMY2, TEMP, .L50 // If dummy2 == 1, do not directly copy 0
 	NOP
 
 	bne	INCX, TEMP, .L20

kernel/power/dscal.c

@@ -73,6 +73,15 @@ static void dscal_kernel_8_zero (BLASLONG n, FLOAT *x)
 
         for( i=0; i<n; i+=8 )
         {
+		x[0] = alpha;
+		x[1] = alpha;
+		x[2] = alpha;
+		x[3] = alpha;
+		x[4] = alpha;
+		x[5] = alpha;
+		x[6] = alpha;
+		x[7] = alpha;
+#if 0
                 if(isfinite(x[0]))
 			x[0] = alpha;
 		else	
@@ -106,7 +115,8 @@ static void dscal_kernel_8_zero (BLASLONG n, FLOAT *x)
 		else
 			x[7] = NAN;
                 x+=8;
-        }
+#endif 
+	}
 
 }
 
@@ -130,6 +140,11 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 			if ( n >= 16 )
 			{
 				BLASLONG align = ((32 - ((uintptr_t)x & (uintptr_t)0x1F)) >> 3) & 0x3;
+				if (dummy2 == 0)
+				for (j = 0; j < align; j++) {
+					x [j] = 0.0;
+				}
+				else
 				for (j = 0; j < align; j++) {
 					if (isfinite(x[j]))
 						x[j] = 0.0;
@@ -151,7 +166,13 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 				j=n1;
 			}
 #endif
-
+			if (dummy2 == 0)
+			while(j < n)
+			{
+				x[j]=0.0;
+				j++;
+			}
+			else
 			while(j < n)
 			{
 				if (!isfinite(x[j]))
@@ -202,7 +223,14 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 
 		if ( da == 0.0 )
 		{		
-
+		if (dummy2 == 0)
+			while(j < n)
+			{
+				x[i]=0.0;
+				i += inc_x;
+				j++;
+			}
+		else
 			while(j < n)
 			{
 				if (!isfinite(x[i]))

kernel/power/gemm_kernel_power6.S

@@ -864,15 +864,15 @@ LL(22):
 	LFD	f22, 10 * SIZE(BO)
 	LFD	f23, 11 * SIZE(BO)
 
-	FMADD	f2,  f18, f24, f2
-	FMADD	f3,  f19, f24, f3
-	FMADD	f6,  f18, f25, f6
-	FMADD	f7,  f19, f25, f7
+	FMADD	f0,  f18, f24, f0
+	FMADD	f1,  f19, f24, f1
+	FMADD	f4,  f18, f25, f4
+	FMADD	f5,  f19, f25, f5
 
-	FMADD	f10, f18, f26, f10
-	FMADD	f11, f19, f26, f11
-	FMADD	f14, f18, f27, f14
-	FMADD	f15, f19, f27, f15
+	FMADD	f8, f18, f26, f8
+	FMADD	f9, f19, f26, f9
+	FMADD	f12, f18, f27, f12
+	FMADD	f13, f19, f27, f13
 
 	LFD	f16,  4 * SIZE(AO)
 	LFD	f17,  5 * SIZE(AO)
@@ -899,15 +899,15 @@ LL(22):
 	LFD	f22, 18 * SIZE(BO)
 	LFD	f23, 19 * SIZE(BO)
 
-	FMADD	f2,  f18, f24, f2
-	FMADD	f3,  f19, f24, f3
-	FMADD	f6,  f18, f25, f6
-	FMADD	f7,  f19, f25, f7
+	FMADD	f0,  f18, f24, f0
+	FMADD	f1,  f19, f24, f1
+	FMADD	f4,  f18, f25, f4
+	FMADD	f5,  f19, f25, f5
 
-	FMADD	f10, f18, f26, f10
-	FMADD	f11, f19, f26, f11
-	FMADD	f14, f18, f27, f14
-	FMADD	f15, f19, f27, f15
+	FMADD	f8, f18, f26, f8
+	FMADD	f9, f19, f26, f9
+	FMADD	f12, f18, f27, f12
+	FMADD	f13, f19, f27, f13
 
 	LFD	f16,  8 * SIZE(AO)
 	LFD	f17,  9 * SIZE(AO)
@@ -923,14 +923,6 @@ LL(22):
 	addi	BO, BO, 16 * SIZE
 	bdnz	LL(22)
 
-	fadd	f0,  f2,  f0
-	fadd	f1,  f3,  f1
-	fadd	f4,  f6,  f4
-	fadd	f5,  f7,  f5
-	fadd	f8,  f10, f8
-	fadd	f9,  f11, f9
-	fadd	f12, f14, f12
-	fadd	f13, f15, f13
 	.align 4
 
 LL(25):
@@ -1161,10 +1153,10 @@ LL(32):
 	LFD	f22, 10 * SIZE(BO)
 	LFD	f23, 11 * SIZE(BO)
 
-	FMADD	f1,  f17, f24, f1
-	FMADD	f5,  f17, f25, f5
-	FMADD	f9,  f17, f26, f9
-	FMADD	f13, f17, f27, f13
+	FMADD	f0,  f17, f24, f0
+	FMADD	f4,  f17, f25, f4
+	FMADD	f8,  f17, f26, f8
+	FMADD	f12, f17, f27, f12
 
 	LFD	f24, 12 * SIZE(BO)
 	LFD	f25, 13 * SIZE(BO)
@@ -1181,10 +1173,10 @@ LL(32):
 	LFD	f22, 18 * SIZE(BO)
 	LFD	f23, 19 * SIZE(BO)
 
-	FMADD	f1,  f19, f24, f1
-	FMADD	f5,  f19, f25, f5
-	FMADD	f9,  f19, f26, f9
-	FMADD	f13, f19, f27, f13
+	FMADD	f0,  f19, f24, f0
+	FMADD	f4,  f19, f25, f4
+	FMADD	f8,  f19, f26, f8
+	FMADD	f12, f19, f27, f12
 
 	LFD	f16,  4 * SIZE(AO)
 	LFD	f17,  5 * SIZE(AO)
@@ -1200,10 +1192,6 @@ LL(32):
 	addi	BO, BO, 16 * SIZE
 	bdnz	LL(32)
 
-	fadd	f0,  f1,   f0
-	fadd	f4,  f5,   f4
-	fadd	f8,  f9,   f8
-	fadd	f12, f13, f12
 	.align 4
 
 LL(35):
@@ -1691,10 +1679,10 @@ LL(52):
 	FMADD	f2,  f16, f21, f2
 	FMADD	f3,  f17, f21, f3
 
-	FMADD	f4,  f18, f22, f4
-	FMADD	f5,  f19, f22, f5
-	FMADD	f6,  f18, f23, f6
-	FMADD	f7,  f19, f23, f7
+	FMADD	f0,  f18, f22, f0
+	FMADD	f1,  f19, f22, f1
+	FMADD	f2,  f18, f23, f2
+	FMADD	f3,  f19, f23, f3
 
 	LFD	f16,  4 * SIZE(AO)
 	LFD	f17,  5 * SIZE(AO)
@@ -1711,10 +1699,10 @@ LL(52):
 	FMADD	f2,  f16, f25, f2
 	FMADD	f3,  f17, f25, f3
 
-	FMADD	f4,  f18, f26, f4
-	FMADD	f5,  f19, f26, f5
-	FMADD	f6,  f18, f27, f6
-	FMADD	f7,  f19, f27, f7
+	FMADD	f0,  f18, f26, f0
+	FMADD	f1,  f19, f26, f1
+	FMADD	f2,  f18, f27, f2
+	FMADD	f3,  f19, f27, f3
 
 	LFD	f16,  8 * SIZE(AO)
 	LFD	f17,  9 * SIZE(AO)
@@ -1775,21 +1763,11 @@ LL(58):
 	LFD	f18, 0 * SIZE(CO2)
 	LFD	f19, 1 * SIZE(CO2)
 
-	FADD	f0, f4,  f0
-	FADD	f1, f5,  f1
-	FADD	f2, f6,  f2
-	FADD	f3, f7,  f3
-
 	FMADD	f0,  f0, f30, f16
 	FMADD	f1,  f1, f30, f17
 	FMADD	f2,  f2, f30, f18
 	FMADD	f3,  f3, f30, f19
 #else
-	FADD	f0, f4,  f0
-	FADD	f1, f5,  f1
-	FADD	f2, f6,  f2
-	FADD	f3, f7,  f3
-
 	FMUL	f0,  f0, f30
 	FMUL	f1,  f1, f30
 	FMUL	f2,  f2, f30
@@ -1916,8 +1894,8 @@ LL(60):
 LL(62):
 	FMADD	f0,  f16, f20, f0
 	FMADD	f1,  f16, f21, f1
-	FMADD	f2,  f17, f22, f2
-	FMADD	f3,  f17, f23, f3
+	FMADD	f0,  f17, f22, f0
+	FMADD	f1,  f17, f23, f1
 
 	LFD	f20,  8 * SIZE(BO)
 	LFD	f21,  9 * SIZE(BO)
@@ -1926,8 +1904,8 @@ LL(62):
 
 	FMADD	f0,  f18, f24, f0
 	FMADD	f1,  f18, f25, f1
-	FMADD	f2,  f19, f26, f2
-	FMADD	f3,  f19, f27, f3
+	FMADD	f0,  f19, f26, f0
+	FMADD	f1,  f19, f27, f1
 
 	LFD	f16,  4 * SIZE(AO)
 	LFD	f17,  5 * SIZE(AO)
@@ -1986,15 +1964,9 @@ LL(68):
 	LFD	f16, 0 * SIZE(CO1)
 	LFD	f18, 0 * SIZE(CO2)
 
-	FADD	f0, f2, f0
-	FADD	f1, f3, f1
-
 	FMADD	f0,  f0,  f30, f16
 	FMADD	f1,  f1,  f30, f18
 #else
-	FADD	f0, f2, f0
-	FADD	f1, f3, f1
-
 	FMUL	f0,  f0,  f30
 	FMUL	f1,  f1,  f30
 #endif
@@ -2007,7 +1979,6 @@ LL(68):
 	fmr	f4,  f0
 	fmr	f5,  f0
 
-
 #ifdef TRMMKERNEL
 #if ( defined(LEFT) &&  defined(TRANSA)) || \
     (!defined(LEFT) && !defined(TRANSA))
@@ -2332,8 +2303,8 @@ LL(80):
 LL(82):
 	FMADD	f0,  f16, f20, f0
 	FMADD	f1,  f17, f20, f1
-	FMADD	f2,  f18, f21, f2
-	FMADD	f3,  f19, f21, f3
+	FMADD	f0,  f18, f21, f0
+	FMADD	f1,  f19, f21, f1
 
 	LFD	f16,  4 * SIZE(AO)
 	LFD	f17,  5 * SIZE(AO)
@@ -2342,8 +2313,8 @@ LL(82):
 
 	FMADD	f0,  f16, f22, f0
 	FMADD	f1,  f17, f22, f1
-	FMADD	f2,  f18, f23, f2
-	FMADD	f3,  f19, f23, f3
+	FMADD	f0,  f18, f23, f0
+	FMADD	f1,  f19, f23, f1
 
 	LFD	f16,  8 * SIZE(AO)
 	LFD	f17,  9 * SIZE(AO)
@@ -2401,15 +2372,9 @@ LL(88):
 	LFD	f16, 0 * SIZE(CO1)
 	LFD	f17, 1 * SIZE(CO1)
 
-	FADD	f0, f2, f0
-	FADD	f1, f3, f1
-
 	FMADD	f0,  f0, f30, f16
 	FMADD	f1,  f1, f30, f17
 #else
-	FADD	f0, f2, f0
-	FADD	f1, f3, f1
-
 	FMUL	f0,  f0, f30
 	FMUL	f1,  f1, f30
 #endif
@@ -2418,9 +2383,6 @@ LL(88):
 	STFD	f1,  1 * SIZE(CO1)
 
 	lfs	f0,  FZERO
- 	fmr	f1,  f0
-	fmr	f2,  f0
-	fmr	f3,  f0
 
 	addi	CO1, CO1, 2 * SIZE
 
@@ -2512,9 +2474,9 @@ LL(90):
 
 LL(92):
 	FMADD	f0,  f16, f20, f0
-	FMADD	f1,  f17, f21, f1
-	FMADD	f2,  f18, f22, f2
-	FMADD	f3,  f19, f23, f3
+	FMADD	f0,  f17, f21, f0
+	FMADD	f0,  f18, f22, f0
+	FMADD	f0,  f19, f23, f0
 
 	LFD	f16,  4 * SIZE(AO)
 	LFD	f17,  5 * SIZE(AO)
@@ -2527,9 +2489,9 @@ LL(92):
 	LFD	f23,  7 * SIZE(BO)
 
 	FMADD	f0,  f16, f20, f0
-	FMADD	f1,  f17, f21, f1
-	FMADD	f2,  f18, f22, f2
-	FMADD	f3,  f19, f23, f3
+	FMADD	f0,  f17, f21, f0
+	FMADD	f0,  f18, f22, f0
+	FMADD	f0,  f19, f23, f0
 
 	LFD	f16,  8 * SIZE(AO)
 	LFD	f17,  9 * SIZE(AO)
@@ -2583,16 +2545,8 @@ LL(98):
 #ifndef TRMMKERNEL
 	LFD	f16, 0 * SIZE(CO1)
 
-	FADD	f0, f1, f0
-	FADD	f2, f3, f2
-	FADD	f0, f2, f0
-
 	FMADD	f0,  f0,  f30, f16
 #else
-	FADD	f0, f1, f0
-	FADD	f2, f3, f2
-	FADD	f0, f2, f0
-
 	FMUL	f0,  f0,  f30
 #endif
 

kernel/power/scal.S

@@ -47,9 +47,11 @@
 #ifndef __64BIT__
 #define X r6
 #define INCX r7
+#define FLAG r11
 #else
 #define X r7
 #define INCX r8
+#define FLAG r12
 #endif
 #endif
 
@@ -57,9 +59,11 @@
 #if !defined(__64BIT__) && defined(DOUBLE)
 #define X r8
 #define INCX r9
+#define FLAG r13
 #else
 #define X r7
 #define INCX r8
+#define FLAG r12
 #endif
 #endif
 
@@ -84,9 +88,12 @@
 	cmpwi	cr0, N, 0
 	blelr-	cr0
 
-//	fcmpu	cr0, FZERO, ALPHA
-//	bne-	cr0, LL(A1I1)
-	b LL(A1I1)
+	fcmpu	cr0, FZERO, ALPHA
+	bne-	cr0, LL(A1I1)
+
+	ld      FLAG,    48+64+8(SP)
+	cmpwi   cr0, FLAG, 1
+	beq-   cr0, LL(A1I1)
 
 	cmpwi	cr0, INCX, SIZE
 	bne-	cr0, LL(A0IN)

kernel/power/sscal.c

@@ -74,7 +74,24 @@ static void sscal_kernel_16_zero( BLASLONG n, FLOAT *x )
 
         for( i=0; i<n; i+=8 )
         {
-                if (isfinite(x[0]))
+		x[0] = alpha;
+		x[1] = alpha;
+		x[2] = alpha;
+		x[3] = alpha;
+		x[4] = alpha;
+		x[5] = alpha;
+		x[6] = alpha;
+		x[7] = alpha;
+		x[8] = alpha;
+		x[9] = alpha;
+		x[10] = alpha;
+		x[11] = alpha;
+		x[12] = alpha;
+		x[13] = alpha;
+		x[14] = alpha;
+		x[15] = alpha;
+#if 0
+		if (isfinite(x[0]))
 			x[0] = alpha;
 		else
 			x[0] = NAN;
@@ -107,7 +124,8 @@ static void sscal_kernel_16_zero( BLASLONG n, FLOAT *x )
 		else
 			x[7] = NAN;
                 x+=8;
-        }
+#endif
+    	}
 
 }
 
@@ -132,6 +150,11 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 			if ( n >= 32 )
 			{
 				BLASLONG align = ((32 - ((uintptr_t)x & (uintptr_t)0x1F)) >> 2) & 0x7;
+				if (dummy2 == 0)
+					for (j = 0; j < align; j++){
+						x[j] = 0.0;
+					}
+				else
 				for (j = 0; j < align; j++) {
 					if (isfinite(x[j]))
 						x[j] = 0.0;
@@ -153,9 +176,15 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 				j=n1;
 			}
 #endif
-
+			if (dummy2 == 0)
 			while(j < n)
 			{
+				x[j] = 0.0;
+				j++;
+			}
+			else
+			while(j < n)
+			{	
 				if (isfinite(x[j]))
 					x[j]=0.0;
 				else
@@ -204,7 +233,14 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 
 		if ( da == 0.0 )
 		{		
-
+		if (dummy2 == 0)
+			while(j < n)
+			{
+				x[i]=0.0;
+				i += inc_x;
+				j++;
+			}
+		else
 			while(j < n)
 			{
 				if (isfinite(x[i]))

kernel/power/zgemm_kernel_power6.S

@@ -1159,9 +1159,9 @@ LL(20):
 
 LL(22):
 	FMA1	f0,  f16, f20, f0
-	FMA4	f3,  f17, f20, f3
 	FMA2	f1,  f16, f21, f1
-	FMA3	f2,  f17, f21, f2
+	FMA4	f1,  f17, f20, f1
+	FMA3	f0,  f17, f21, f0
 
 	LFD	f28,  4 * SIZE(AO)
 	LFD	f29,  5 * SIZE(AO)
@@ -1169,9 +1169,9 @@ LL(22):
 	LFD	f31,  7 * SIZE(AO)
 
 	FMA1	f4,  f16, f22, f4
-	FMA4	f7,  f17, f22, f7
 	FMA2	f5,  f16, f23, f5
-	FMA3	f6,  f17, f23, f6
+	FMA4	f5,  f17, f22, f5
+	FMA3	f4,  f17, f23, f4
 
 	LFD	f20,  8 * SIZE(BO)
 	LFD	f21,  9 * SIZE(BO)
@@ -1179,14 +1179,14 @@ LL(22):
 	LFD	f23, 11 * SIZE(BO)
 
 	FMA1	f8,  f16, f24, f8
-	FMA4	f11, f17, f24, f11
 	FMA2	f9,  f16, f25, f9
-	FMA3	f10, f17, f25, f10
+	FMA4	f9, f17, f24, f9
+	FMA3	f8, f17, f25, f8
 
 	FMA1	f12, f16, f26, f12
-	FMA4	f15, f17, f26, f15
 	FMA2	f13, f16, f27, f13
-	FMA3	f14, f17, f27, f14
+	FMA4	f13, f17, f26, f13
+	FMA3	f12, f17, f27, f12
 
 	LFD	f24, 12 * SIZE(BO)
 	LFD	f25, 13 * SIZE(BO)
@@ -1194,14 +1194,14 @@ LL(22):
 	LFD	f27, 15 * SIZE(BO)
 
 	FMA1	f0,  f18, f20, f0
-	FMA4	f3,  f19, f20, f3
 	FMA2	f1,  f18, f21, f1
-	FMA3	f2,  f19, f21, f2
+	FMA4	f1,  f19, f20, f1
+	FMA3	f0,  f19, f21, f0
 
 	FMA1	f4,  f18, f22, f4
-	FMA4	f7,  f19, f22, f7
 	FMA2	f5,  f18, f23, f5
-	FMA3	f6,  f19, f23, f6
+	FMA4	f5,  f19, f22, f5
+	FMA3	f4,  f19, f23, f4
 
 	LFD	f20, 16 * SIZE(BO)
 	LFD	f21, 17 * SIZE(BO)
@@ -1209,14 +1209,14 @@ LL(22):
 	LFD	f23, 19 * SIZE(BO)
 
 	FMA1	f8,  f18, f24, f8
-	FMA4	f11, f19, f24, f11
 	FMA2	f9,  f18, f25, f9
-	FMA3	f10, f19, f25, f10
+	FMA4	f9, f19, f24, f9
+	FMA3	f8, f19, f25, f8
 
 	FMA1	f12, f18, f26, f12
-	FMA4	f15, f19, f26, f15
 	FMA2	f13, f18, f27, f13
-	FMA3	f14, f19, f27, f14
+	FMA4	f13, f19, f26, f13
+	FMA3	f12, f19, f27, f12
 
 	LFD	f24, 20 * SIZE(BO)
 	LFD	f25, 21 * SIZE(BO)
@@ -1224,9 +1224,9 @@ LL(22):
 	LFD	f27, 23 * SIZE(BO)
 
 	FMA1	f0,  f28, f20, f0
-	FMA4	f3,  f29, f20, f3
 	FMA2	f1,  f28, f21, f1
-	FMA3	f2,  f29, f21, f2
+	FMA4	f1,  f29, f20, f1
+	FMA3	f0,  f29, f21, f0
 
 	LFD	f16,  8 * SIZE(AO)
 	LFD	f17,  9 * SIZE(AO)
@@ -1234,9 +1234,9 @@ LL(22):
 	LFD	f19, 11 * SIZE(AO)
 
 	FMA1	f4,  f28, f22, f4
-	FMA4	f7,  f29, f22, f7
 	FMA2	f5,  f28, f23, f5
-	FMA3	f6,  f29, f23, f6
+	FMA4	f5,  f29, f22, f5
+	FMA3	f4,  f29, f23, f4
 
 	LFD	f20, 24 * SIZE(BO)
 	LFD	f21, 25 * SIZE(BO)
@@ -1244,14 +1244,14 @@ LL(22):
 	LFD	f23, 27 * SIZE(BO)
 
 	FMA1	f8,  f28, f24, f8
-	FMA4	f11, f29, f24, f11
 	FMA2	f9,  f28, f25, f9
-	FMA3	f10, f29, f25, f10
+	FMA4	f9, f29, f24, f9
+	FMA3	f8, f29, f25, f8
 
 	FMA1	f12, f28, f26, f12
-	FMA4	f15, f29, f26, f15
 	FMA2	f13, f28, f27, f13
-	FMA3	f14, f29, f27, f14
+	FMA4	f13, f29, f26, f13
+	FMA3	f12, f29, f27, f12
 
 	LFD	f24, 28 * SIZE(BO)
 	LFD	f25, 29 * SIZE(BO)
@@ -1259,14 +1259,14 @@ LL(22):
 	LFD	f27, 31 * SIZE(BO)
 
 	FMA1	f0,  f30, f20, f0
-	FMA4	f3,  f31, f20, f3
 	FMA2	f1,  f30, f21, f1
-	FMA3	f2,  f31, f21, f2
+	FMA4	f1,  f31, f20, f1
+	FMA3	f0,  f31, f21, f0
 
 	FMA1	f4,  f30, f22, f4
-	FMA4	f7,  f31, f22, f7
 	FMA2	f5,  f30, f23, f5
-	FMA3	f6,  f31, f23, f6
+	FMA4	f5,  f31, f22, f5
+	FMA3	f4,  f31, f23, f4
 
 	LFD	f20, 32 * SIZE(BO)
 	LFD	f21, 33 * SIZE(BO)
@@ -1274,14 +1274,14 @@ LL(22):
 	LFD	f23, 35 * SIZE(BO)
 
 	FMA1	f8,  f30, f24, f8
-	FMA4	f11, f31, f24, f11
 	FMA2	f9,  f30, f25, f9
-	FMA3	f10, f31, f25, f10
+	FMA4	f9, f31, f24, f9
+	FMA3	f8, f31, f25, f8
 
 	FMA1	f12, f30, f26, f12
-	FMA4	f15, f31, f26, f15
 	FMA2	f13, f30, f27, f13
-	FMA3	f14, f31, f27, f14
+	FMA4	f13, f31, f26, f13
+	FMA3	f12, f31, f27, f12
 
 	LFD	f24, 36 * SIZE(BO)
 	LFD	f25, 37 * SIZE(BO)
@@ -1318,14 +1318,14 @@ LL(25):
 
 LL(26):
 	FMA1	f0,  f16, f20, f0
-	FMA4	f3,  f17, f20, f3
 	FMA2	f1,  f16, f21, f1
-	FMA3	f2,  f17, f21, f2
+	FMA4	f1,  f17, f20, f1
+	FMA3	f0,  f17, f21, f0
 
 	FMA1	f4,  f16, f22, f4
-	FMA4	f7,  f17, f22, f7
 	FMA2	f5,  f16, f23, f5
-	FMA3	f6,  f17, f23, f6
+	FMA4	f5,  f17, f22, f5
+	FMA3	f4,  f17, f23, f4
 
 	LFD	f20,  8 * SIZE(BO)
 	LFD	f21,  9 * SIZE(BO)
@@ -1333,14 +1333,14 @@ LL(26):
 	LFD	f23, 11 * SIZE(BO)
 
 	FMA1	f8,  f16, f24, f8
-	FMA4	f11, f17, f24, f11
 	FMA2	f9,  f16, f25, f9
-	FMA3	f10, f17, f25, f10
+	FMA4	f9, f17, f24, f9
+	FMA3	f8, f17, f25, f8
 
 	FMA1	f12, f16, f26, f12
-	FMA4	f15, f17, f26, f15
 	FMA2	f13, f16, f27, f13
-	FMA3	f14, f17, f27, f14
+	FMA4	f13, f17, f26, f13
+	FMA3	f12, f17, f27, f12
 
 	LFD	f16,  2 * SIZE(AO)
 	LFD	f17,  3 * SIZE(AO)
@@ -1363,47 +1363,42 @@ LL(28):
 	LFD	f18, 0 * SIZE(CO2)
 	LFD	f19, 1 * SIZE(CO2)
 
-	FADD	f0,  f0,  f2
-	FADD	f1,  f1,  f3
-	FADD	f4,  f4,  f6
-	FADD	f5,  f5,  f7
-
 	LFD	f20, 0 * SIZE(CO3)
 	LFD	f21, 1 * SIZE(CO3)
 	LFD	f22, 0 * SIZE(CO4)
 	LFD	f23, 1 * SIZE(CO4)
 
-	FADD	f8,  f8,  f10
-	FADD	f9,  f9,  f11
-	FADD	f12, f12, f14
-	FADD	f13, f13, f15
+	fmr   f2, f0
+	fmr   f3, f1
+	fmr   f6, f4
+	fmr   f7, f5
 
-	FNMSUB	f24, f31, f1, f16
-	FMADD	f25, f31, f0, f17
-	FNMSUB	f26, f31, f5, f18
-	FMADD	f27, f31, f4, f19
+	FMADD	f24,  f30, f0, f16
+	FMADD	f25,  f30, f1, f17
+	FMADD	f26,  f30, f4,  f18
+	FMADD	f27,  f30, f5,  f19
 
-	FMADD	f0,  f30, f0, f24
-	FMADD	f1,  f30, f1, f25
-	FMADD	f4,  f30, f4,  f26
-	FMADD	f5,  f30, f5,  f27
+	FNMSUB	f0, f31, f3, f24
+	FMADD	f1, f31, f2, f25
+	FNMSUB	f4, f31, f7, f26
+	FMADD	f5, f31, f6, f27
 
-	FNMSUB	f24, f31, f9,  f20
-	FMADD	f25, f31, f8,  f21
-	FNMSUB	f26, f31, f13, f22
-	FMADD	f27, f31, f12, f23
+	fmr   f10, f8
+	fmr   f11, f9
+	fmr   f14, f12
+	fmr   f15, f13
 
-	FMADD	f8,  f30, f8,  f24
-	FMADD	f9,  f30, f9,  f25
-	FMADD	f12,  f30, f12, f26
-	FMADD	f13,  f30, f13, f27
+	FMADD	f24,  f30, f8,  f20
+	FMADD	f25,  f30, f9,  f21
+	FMADD	f26,  f30, f12, f22
+	FMADD	f27,  f30, f13, f23
 
-#else
-	FADD	f0,  f0,  f2
-	FADD	f1,  f1,  f3
-	FADD	f4,  f4,  f6
-	FADD	f5,  f5,  f7
+	FNMSUB	f8, f31, f11,  f24
+	FMADD	f9, f31, f10,  f25
+	FNMSUB	f12, f31, f15, f26
+	FMADD	f13, f31, f14, f27
 
+#else
 	FMUL	f16, f31, f1
 	FMUL	f17, f31, f0
 	FMUL	f18, f31, f5
@@ -1414,11 +1409,6 @@ LL(28):
 	FMSUB	f4,  f30, f4, f18
 	FMADD	f5,  f30, f5, f19
 
-	FADD	f8,  f8,  f10
-	FADD	f9,  f9,  f11
-	FADD	f12, f12, f14
-	FADD	f13, f13, f15
-
 	FMUL	f20, f31, f9
 	FMUL	f21, f31, f8
 	FMUL	f22, f31, f13
@@ -1616,15 +1606,15 @@ LL(32):
 	FMA2	f5,  f16, f23, f5
 	FMA2	f7,  f18, f23, f7
 
-	FMA4	f9,  f17, f20, f9
-	FMA4	f11, f19, f20, f11
-	FMA3	f8,  f17, f21, f8
-	FMA3	f10, f19, f21, f10
+	FMA4	f1,  f17, f20, f1
+	FMA4	f3, f19, f20, f3
+	FMA3	f0,  f17, f21, f0
+	FMA3	f2, f19, f21, f2
 
-	FMA4	f13, f17, f22, f13
-	FMA4	f15, f19, f22, f15
-	FMA3	f12, f17, f23, f12
-	FMA3	f14, f19, f23, f14
+	FMA4	f5, f17, f22, f5
+	FMA4	f7, f19, f22, f7
+	FMA3	f4, f17, f23, f4
+	FMA3	f6, f19, f23, f6
 
 	LFD	f20,  8 * SIZE(BO)
 	LFD	f21,  9 * SIZE(BO)
@@ -1646,15 +1636,15 @@ LL(32):
 	FMA2	f5,  f28, f27, f5
 	FMA2	f7,  f30, f27, f7
 
-	FMA4	f9,  f29, f24, f9
-	FMA4	f11, f31, f24, f11
-	FMA3	f8,  f29, f25, f8
-	FMA3	f10, f31, f25, f10
+	FMA4	f1,  f29, f24, f1
+	FMA4	f3, f31, f24, f3
+	FMA3	f0,  f29, f25, f0
+	FMA3	f2, f31, f25, f2
 
-	FMA4	f13, f29, f26, f13
-	FMA4	f15, f31, f26, f15
-	FMA3	f12, f29, f27, f12
-	FMA3	f14, f31, f27, f14
+	FMA4	f5, f29, f26, f5
+	FMA4	f7, f31, f26, f7
+	FMA3	f4, f29, f27, f4
+	FMA3	f6, f31, f27, f6
 
 	LFD	f24, 12 * SIZE(BO)
 	LFD	f25, 13 * SIZE(BO)
@@ -1676,15 +1666,15 @@ LL(32):
 	FMA2	f5,  f16, f23, f5
 	FMA2	f7,  f18, f23, f7
 
-	FMA4	f9,  f17, f20, f9
-	FMA4	f11, f19, f20, f11
-	FMA3	f8,  f17, f21, f8
-	FMA3	f10, f19, f21, f10
+	FMA4	f1,  f17, f20, f1
+	FMA4	f3, f19, f20, f3
+	FMA3	f0,  f17, f21, f0
+	FMA3	f2, f19, f21, f2
 
-	FMA4	f13, f17, f22, f13
-	FMA4	f15, f19, f22, f15
-	FMA3	f12, f17, f23, f12
-	FMA3	f14, f19, f23, f14
+	FMA4	f5, f17, f22, f5
+	FMA4	f7, f19, f22, f7
+	FMA3	f4, f17, f23, f4
+	FMA3	f6, f19, f23, f6
 
 	LFD	f20, 16 * SIZE(BO)
 	LFD	f21, 17 * SIZE(BO)
@@ -1706,15 +1696,15 @@ LL(32):
 	FMA2	f5,  f28, f27, f5
 	FMA2	f7,  f30, f27, f7
 
-	FMA4	f9,  f29, f24, f9
-	FMA4	f11, f31, f24, f11
-	FMA3	f8,  f29, f25, f8
-	FMA3	f10, f31, f25, f10
+	FMA4	f1,  f29, f24, f1
+	FMA4	f3, f31, f24, f3
+	FMA3	f0,  f29, f25, f0
+	FMA3	f2, f31, f25, f2
 
-	FMA4	f13, f29, f26, f13
-	FMA4	f15, f31, f26, f15
-	FMA3	f12, f29, f27, f12
-	FMA3	f14, f31, f27, f14
+	FMA4	f5, f29, f26, f5
+	FMA4	f7, f31, f26, f7
+	FMA3	f4, f29, f27, f4
+	FMA3	f6, f31, f27, f6
 
 	LFD	f24, 20 * SIZE(BO)
 	LFD	f25, 21 * SIZE(BO)
@@ -1736,15 +1726,15 @@ LL(32):
 	FMA2	f5,  f16, f23, f5
 	FMA2	f7,  f18, f23, f7
 
-	FMA4	f9,  f17, f20, f9
-	FMA4	f11, f19, f20, f11
-	FMA3	f8,  f17, f21, f8
-	FMA3	f10, f19, f21, f10
+	FMA4	f1,  f17, f20, f1
+	FMA4	f3, f19, f20, f3
+	FMA3	f0,  f17, f21, f0
+	FMA3	f2, f19, f21, f2
 
-	FMA4	f13, f17, f22, f13
-	FMA4	f15, f19, f22, f15
-	FMA3	f12, f17, f23, f12
-	FMA3	f14, f19, f23, f14
+	FMA4	f5, f17, f22, f5
+	FMA4	f7, f19, f22, f7
+	FMA3	f4, f17, f23, f4
+	FMA3	f6, f19, f23, f6
 
 	LFD	f20, 24 * SIZE(BO)
 	LFD	f21, 25 * SIZE(BO)
@@ -1766,15 +1756,15 @@ LL(32):
 	FMA2	f5,  f28, f27, f5
 	FMA2	f7,  f30, f27, f7
 
-	FMA4	f9,  f29, f24, f9
-	FMA4	f11, f31, f24, f11
-	FMA3	f8,  f29, f25, f8
-	FMA3	f10, f31, f25, f10
+	FMA4	f1,  f29, f24, f1
+	FMA4	f3, f31, f24, f3
+	FMA3	f0,  f29, f25, f0
+	FMA3	f2, f31, f25, f2
 
-	FMA4	f13, f29, f26, f13
-	FMA4	f15, f31, f26, f15
-	FMA3	f12, f29, f27, f12
-	FMA3	f14, f31, f27, f14
+	FMA4	f5, f29, f26, f5
+	FMA4	f7, f31, f26, f7
+	FMA3	f4, f29, f27, f4
+	FMA3	f6, f31, f27, f6
 
 	LFD	f24, 28 * SIZE(BO)
 	LFD	f25, 29 * SIZE(BO)
@@ -1796,15 +1786,15 @@ LL(32):
 	FMA2	f5,  f16, f23, f5
 	FMA2	f7,  f18, f23, f7
 
-	FMA4	f9,  f17, f20, f9
-	FMA4	f11, f19, f20, f11
-	FMA3	f8,  f17, f21, f8
-	FMA3	f10, f19, f21, f10
+	FMA4	f1,  f17, f20, f1
+	FMA4	f3, f19, f20, f3
+	FMA3	f0,  f17, f21, f0
+	FMA3	f2, f19, f21, f2
 
-	FMA4	f13, f17, f22, f13
-	FMA4	f15, f19, f22, f15
-	FMA3	f12, f17, f23, f12
-	FMA3	f14, f19, f23, f14
+	FMA4	f5, f17, f22, f5
+	FMA4	f7, f19, f22, f7
+	FMA3	f4, f17, f23, f4
+	FMA3	f6, f19, f23, f6
 
 	LFD	f20, 32 * SIZE(BO)
 	LFD	f21, 33 * SIZE(BO)
@@ -1826,15 +1816,15 @@ LL(32):
 	FMA2	f5,  f28, f27, f5
 	FMA2	f7,  f30, f27, f7
 
-	FMA4	f9,  f29, f24, f9
-	FMA4	f11, f31, f24, f11
-	FMA3	f8,  f29, f25, f8
-	FMA3	f10, f31, f25, f10
+	FMA4	f1,  f29, f24, f1
+	FMA4	f3, f31, f24, f3
+	FMA3	f0,  f29, f25, f0
+	FMA3	f2, f31, f25, f2
 
-	FMA4	f13, f29, f26, f13
-	FMA4	f15, f31, f26, f15
-	FMA3	f12, f29, f27, f12
-	FMA3	f14, f31, f27, f14
+	FMA4	f5, f29, f26, f5
+	FMA4	f7, f31, f26, f7
+	FMA3	f4, f29, f27, f4
+	FMA3	f6, f31, f27, f6
 
 	LFD	f24, 36 * SIZE(BO)
 	LFD	f25, 37 * SIZE(BO)
@@ -1883,20 +1873,20 @@ LL(36):
 	FMA2	f5,  f16, f23, f5
 	FMA2	f7,  f18, f23, f7
 
-	FMA4	f9,  f17, f20, f9
-	FMA4	f11, f19, f20, f11
-	FMA3	f8,  f17, f21, f8
-	FMA3	f10, f19, f21, f10
+	FMA4	f1,  f17, f20, f1
+	FMA4	f3, f19, f20, f3
+	FMA3	f0,  f17, f21, f0
+	FMA3	f2, f19, f21, f2
 
 	LFD	f16,  4 * SIZE(AO)
 	LFD	f18,  6 * SIZE(AO)
 	LFD	f20,  4 * SIZE(BO)
 	LFD	f21,  5 * SIZE(BO)
 
-	FMA4	f13, f17, f22, f13
-	FMA4	f15, f19, f22, f15
-	FMA3	f12, f17, f23, f12
-	FMA3	f14, f19, f23, f14
+	FMA4	f5, f17, f22, f5
+	FMA4	f7, f19, f22, f7
+	FMA3	f4, f17, f23, f4
+	FMA3	f6, f19, f23, f6
 
 	LFD	f17,  5 * SIZE(AO)
 	LFD	f19,  7 * SIZE(AO)
@@ -1916,52 +1906,42 @@ LL(38):
 	LFD	f18, 2 * SIZE(CO1)
 	LFD	f19, 3 * SIZE(CO1)
 
-	FADD	f0, f0, f8
-	FADD	f1, f1, f9
-	FADD	f2, f2, f10
-	FADD	f3, f3, f11
-
 	LFD	f20, 0 * SIZE(CO2)
 	LFD	f21, 1 * SIZE(CO2)
 	LFD	f22, 2 * SIZE(CO2)
 	LFD	f23, 3 * SIZE(CO2)
 
-	FADD	f4, f4, f12
-	FADD	f5, f5, f13
-	FADD	f6, f6, f14
-	FADD	f7, f7, f15
+	fmr f8,  f0
+	fmr f9,  f1
+	fmr f10, f2
+	fmr f11, f3
 
-	FNMSUB	f24, f31, f1, f16
-	FMADD	f25, f31, f0, f17
-	FNMSUB	f26, f31, f3, f18
-	FMADD	f27, f31, f2, f19
+	FMADD	f24,  f30, f0, f16
+	FMADD	f25,  f30, f1, f17
+	FMADD	f26,  f30, f2, f18
+	FMADD	f27,  f30, f3, f19
 
-	FMADD	f0,  f30, f0, f24
-	FMADD	f1,  f30, f1, f25
-	FMADD	f2,  f30, f2, f26
-	FMADD	f3,  f30, f3, f27
+	FNMSUB	f0, f31, f9, f24
+	FMADD	f1, f31, f8, f25
+	FNMSUB	f2, f31, f11, f26
+	FMADD	f3, f31, f10, f27
 
-	FNMSUB	f24, f31, f5, f20
-	FMADD	f25, f31, f4, f21
-	FNMSUB	f26, f31, f7, f22
-	FMADD	f27, f31, f6, f23
+	fmr f12, f4
+	fmr f13, f5
+	fmr f14, f6
+	fmr f15, f7
 
-	FMADD	f4,  f30, f4,  f24
-	FMADD	f5,  f30, f5,  f25
-	FMADD	f6,  f30, f6,  f26
-	FMADD	f7,  f30, f7,  f27
+	FMADD	f24,  f30, f4,  f20
+	FMADD	f25,  f30, f5,  f21
+	FMADD	f26,  f30, f6,  f22
+	FMADD	f27,  f30, f7,  f23
 
-#else
-	FADD	f0, f0, f8
-	FADD	f1, f1, f9
-	FADD	f2, f2, f10
-	FADD	f3, f3, f11
-
-	FADD	f4, f4, f12
-	FADD	f5, f5, f13
-	FADD	f6, f6, f14
-	FADD	f7, f7, f15
+	FNMSUB	f4, f31, f13, f24
+	FMADD	f5, f31, f12, f25
+	FNMSUB	f6, f31, f15, f26
+	FMADD	f7, f31, f14, f27
 
+#else
 	FMUL	f16, f31, f1
 	FMUL	f17, f31, f0
 	FMUL	f18, f31, f3
@@ -2101,14 +2081,14 @@ LL(40):
 
 LL(42):
 	FMA1	f0,  f16, f20, f0
-	FMA4	f3,  f17, f20, f3
 	FMA2	f1,  f16, f21, f1
-	FMA3	f2,  f17, f21, f2
+	FMA4	f1,  f17, f20, f1
+	FMA3	f0,  f17, f21, f0
 
 	FMA1	f4,  f16, f22, f4
-	FMA4	f7,  f17, f22, f7
 	FMA2	f5,  f16, f23, f5
-	FMA3	f6,  f17, f23, f6
+	FMA4	f5,  f17, f22, f5
+	FMA3	f4,  f17, f23, f4
 
 	LFD	f16,  2 * SIZE(AO)
 	LFD	f17,  3 * SIZE(AO)
@@ -2119,14 +2099,14 @@ LL(42):
 	LFD	f23,  7 * SIZE(BO)
 
 	FMA1	f0,  f16, f20, f0
-	FMA4	f3,  f17, f20, f3
 	FMA2	f1,  f16, f21, f1
-	FMA3	f2,  f17, f21, f2
+	FMA4	f1,  f17, f20, f1
+	FMA3	f0,  f17, f21, f0
 
 	FMA1	f4,  f16, f22, f4
-	FMA4	f7,  f17, f22, f7
 	FMA2	f5,  f16, f23, f5
-	FMA3	f6,  f17, f23, f6
+	FMA4	f5,  f17, f22, f5
+	FMA3	f4,  f17, f23, f4
 
 	LFD	f16,  4 * SIZE(AO)
 	LFD	f17,  5 * SIZE(AO)
@@ -2137,14 +2117,14 @@ LL(42):
 	LFD	f23, 11 * SIZE(BO)
 
 	FMA1	f0,  f16, f20, f0
-	FMA4	f3,  f17, f20, f3
 	FMA2	f1,  f16, f21, f1
-	FMA3	f2,  f17, f21, f2
+	FMA4	f1,  f17, f20, f1
+	FMA3	f0,  f17, f21, f0
 
 	FMA1	f4,  f16, f22, f4
-	FMA4	f7,  f17, f22, f7
 	FMA2	f5,  f16, f23, f5
-	FMA3	f6,  f17, f23, f6
+	FMA4	f5,  f17, f22, f5
+	FMA3	f4,  f17, f23, f4
 
 	LFD	f16,  6 * SIZE(AO)
 	LFD	f17,  7 * SIZE(AO)
@@ -2155,14 +2135,14 @@ LL(42):
 	LFD	f23, 15 * SIZE(BO)
 
 	FMA1	f0,  f16, f20, f0
-	FMA4	f3,  f17, f20, f3
 	FMA2	f1,  f16, f21, f1
-	FMA3	f2,  f17, f21, f2
+	FMA4	f1,  f17, f20, f1
+	FMA3	f0,  f17, f21, f0
 
 	FMA1	f4,  f16, f22, f4
-	FMA4	f7,  f17, f22, f7
 	FMA2	f5,  f16, f23, f5
-	FMA3	f6,  f17, f23, f6
+	FMA4	f5,  f17, f22, f5
+	FMA3	f4,  f17, f23, f4
 
 	LFD	f16,  8 * SIZE(AO)
 	LFD	f17,  9 * SIZE(AO)
@@ -2202,14 +2182,14 @@ LL(45):
 
 LL(46):
 	FMA1	f0,  f16, f20, f0
-	FMA4	f3,  f17, f20, f3
 	FMA2	f1,  f16, f21, f1
-	FMA3	f2,  f17, f21, f2
+	FMA4	f1,  f17, f20, f1
+	FMA3	f0,  f17, f21, f0
 
 	FMA1	f4,  f16, f22, f4
-	FMA4	f7,  f17, f22, f7
 	FMA2	f5,  f16, f23, f5
-	FMA3	f6,  f17, f23, f6
+	FMA4	f5,  f17, f22, f5
+	FMA3	f4,  f17, f23, f4
 
 	LFD	f16,  2 * SIZE(AO)
 	LFD	f17,  3 * SIZE(AO)
@@ -2231,27 +2211,22 @@ LL(48):
 	LFD	f20, 0 * SIZE(CO2)
 	LFD	f21, 1 * SIZE(CO2)
 
-	FADD	f0,  f0,  f2
-	FADD	f1,  f1,  f3
-	FADD	f4,  f4,  f6
-	FADD	f5,  f5,  f7
+	fmr f2, f0
+	fmr f3, f1
+	fmr f6, f4
+	fmr f7, f5
 
-	FNMSUB	f24, f31, f1, f16
-	FMADD	f25, f31, f0, f17
-	FNMSUB	f26, f31, f5, f20
-	FMADD	f27, f31, f4, f21
+	FMADD	f24,  f30, f0, f16
+	FMADD	f25,  f30, f1, f17
+	FMADD	f26,  f30, f4,  f20
+	FMADD	f27,  f30, f5,  f21
 
-	FMADD	f0,  f30, f0, f24
-	FMADD	f1,  f30, f1, f25
-	FMADD	f4,  f30, f4,  f26
-	FMADD	f5,  f30, f5,  f27
+	FNMSUB	f0, f31, f3, f24
+	FMADD	f1, f31, f2, f25
+	FNMSUB	f4, f31, f7, f26
+	FMADD	f5, f31, f6, f27
 
 #else
-	FADD	f0,  f0,  f2
-	FADD	f1,  f1,  f3
-	FADD	f4,  f4,  f6
-	FADD	f5,  f5,  f7
-
 	FMUL	f16, f31, f1
 	FMUL	f17, f31, f0
 	FMUL	f18, f31, f5
@@ -2401,10 +2376,10 @@ LL(52):
 	FMA2	f1,  f16, f21, f1
 	FMA2	f3,  f18, f21, f3
 
-	FMA4	f9,  f17, f20, f9
-	FMA4	f11, f19, f20, f11
-	FMA3	f8,  f17, f21, f8
-	FMA3	f10, f19, f21, f10
+	FMA4	f1,  f17, f20, f1
+	FMA4	f3, f19, f20, f3
+	FMA3	f0,  f17, f21, f0
+	FMA3	f2, f19, f21, f2
 
 	LFD	f16,  4 * SIZE(AO)
 	LFD	f17,  5 * SIZE(AO)
@@ -2416,10 +2391,10 @@ LL(52):
 	FMA2	f1,  f16, f23, f1
 	FMA2	f3,  f18, f23, f3
 
-	FMA4	f9,  f17, f22, f9
-	FMA4	f11, f19, f22, f11
-	FMA3	f8,  f17, f23, f8
-	FMA3	f10, f19, f23, f10
+	FMA4	f1,  f17, f22, f1
+	FMA4	f3, f19, f22, f3
+	FMA3	f0,  f17, f23, f0
+	FMA3	f2, f19, f23, f2
 
 	LFD	f16,  8 * SIZE(AO)
 	LFD	f17,  9 * SIZE(AO)
@@ -2436,10 +2411,10 @@ LL(52):
 	FMA2	f1,  f16, f21, f1
 	FMA2	f3,  f18, f21, f3
 
-	FMA4	f9,  f17, f20, f9
-	FMA4	f11, f19, f20, f11
-	FMA3	f8,  f17, f21, f8
-	FMA3	f10, f19, f21, f10
+	FMA4	f1,  f17, f20, f1
+	FMA4	f3, f19, f20, f3
+	FMA3	f0,  f17, f21, f0
+	FMA3	f2, f19, f21, f2
 
 	LFD	f16, 12 * SIZE(AO)
 	LFD	f17, 13 * SIZE(AO)
@@ -2451,10 +2426,10 @@ LL(52):
 	FMA2	f1,  f16, f23, f1
 	FMA2	f3,  f18, f23, f3
 
-	FMA4	f9,  f17, f22, f9
-	FMA4	f11, f19, f22, f11
-	FMA3	f8,  f17, f23, f8
-	FMA3	f10, f19, f23, f10
+	FMA4	f1,  f17, f22, f1
+	FMA4	f3, f19, f22, f3
+	FMA3	f0,  f17, f23, f0
+	FMA3	f2, f19, f23, f2
 
 	LFD	f16, 16 * SIZE(AO)
 	LFD	f17, 17 * SIZE(AO)
@@ -2471,10 +2446,10 @@ LL(52):
 	FMA2	f1,  f16, f21, f1
 	FMA2	f3,  f18, f21, f3
 
-	FMA4	f9,  f17, f20, f9
-	FMA4	f11, f19, f20, f11
-	FMA3	f8,  f17, f21, f8
-	FMA3	f10, f19, f21, f10
+	FMA4	f1,  f17, f20, f1
+	FMA4	f3, f19, f20, f3
+	FMA3	f0,  f17, f21, f0
+	FMA3	f2, f19, f21, f2
 
 	LFD	f16, 20 * SIZE(AO)
 	LFD	f17, 21 * SIZE(AO)
@@ -2486,10 +2461,10 @@ LL(52):
 	FMA2	f1,  f16, f23, f1
 	FMA2	f3,  f18, f23, f3
 
-	FMA4	f9,  f17, f22, f9
-	FMA4	f11, f19, f22, f11
-	FMA3	f8,  f17, f23, f8
-	FMA3	f10, f19, f23, f10
+	FMA4	f1,  f17, f22, f1
+	FMA4	f3, f19, f22, f3
+	FMA3	f0,  f17, f23, f0
+	FMA3	f2, f19, f23, f2
 
 	LFD	f16, 24 * SIZE(AO)
 	LFD	f17, 25 * SIZE(AO)
@@ -2506,10 +2481,10 @@ LL(52):
 	FMA2	f1,  f16, f21, f1
 	FMA2	f3,  f18, f21, f3
 
-	FMA4	f9,  f17, f20, f9
-	FMA4	f11, f19, f20, f11
-	FMA3	f8,  f17, f21, f8
-	FMA3	f10, f19, f21, f10
+	FMA4	f1,  f17, f20, f1
+	FMA4	f3, f19, f20, f3
+	FMA3	f0,  f17, f21, f0
+	FMA3	f2, f19, f21, f2
 
 	LFD	f16, 28 * SIZE(AO)
 	LFD	f17, 29 * SIZE(AO)
@@ -2521,10 +2496,10 @@ LL(52):
 	FMA2	f1,  f16, f23, f1
 	FMA2	f3,  f18, f23, f3
 
-	FMA4	f9,  f17, f22, f9
-	FMA4	f11, f19, f22, f11
-	FMA3	f8,  f17, f23, f8
-	FMA3	f10, f19, f23, f10
+	FMA4	f1,  f17, f22, f1
+	FMA4	f3, f19, f22, f3
+	FMA3	f0,  f17, f23, f0
+	FMA3	f2, f19, f23, f2
 
 	LFD	f16, 32 * SIZE(AO)
 	LFD	f17, 33 * SIZE(AO)
@@ -2573,10 +2548,10 @@ LL(56):
 	LFD	f16,  4 * SIZE(AO)
 	LFD	f18,  6 * SIZE(AO)
 
-	FMA4	f9,  f17, f20, f9
-	FMA4	f11, f19, f20, f11
-	FMA3	f8,  f17, f21, f8
-	FMA3	f10, f19, f21, f10
+	FMA4	f1,  f17, f20, f1
+	FMA4	f3, f19, f20, f3
+	FMA3	f0,  f17, f21, f0
+	FMA3	f2, f19, f21, f2
 
 	LFD	f17,  5 * SIZE(AO)
 	LFD	f19,  7 * SIZE(AO)
@@ -2595,27 +2570,22 @@ LL(58):
 	LFD	f18, 2 * SIZE(CO1)
 	LFD	f19, 3 * SIZE(CO1)
 
-	FADD	f0, f0, f8
-	FADD	f1, f1, f9
-	FADD	f2, f2, f10
-	FADD	f3, f3, f11
+	fmr f8,  f0
+	fmr f9,  f1
+	fmr f10, f2
+	fmr f11, f3
 
-	FNMSUB	f24, f31, f1, f16
-	FMADD	f25, f31, f0, f17
-	FNMSUB	f26, f31, f3, f18
-	FMADD	f27, f31, f2, f19
+	FMADD	f24,  f30, f0, f16
+	FMADD	f25,  f30, f1, f17
+	FMADD	f26,  f30, f2, f18
+	FMADD	f27,  f30, f3, f19
 
-	FMADD	f0,  f30, f0, f24
-	FMADD	f1,  f30, f1, f25
-	FMADD	f2,  f30, f2, f26
-	FMADD	f3,  f30, f3, f27
+	FNMSUB	f0, f31, f9, f24
+	FMADD	f1, f31, f8, f25
+	FNMSUB	f2, f31, f11, f26
+	FMADD	f3, f31, f10, f27
 
 #else
-	FADD	f0, f0, f8
-	FADD	f1, f1, f9
-	FADD	f2, f2, f10
-	FADD	f3, f3, f11
-
 	FMUL	f16, f31, f1
 	FMUL	f17, f31, f0
 	FMUL	f18, f31, f3
@@ -2735,9 +2705,9 @@ LL(60):
 
 LL(62):
 	FMA1	f0,  f16, f20, f0
-	FMA4	f3,  f17, f20, f3
 	FMA2	f1,  f16, f21, f1
-	FMA3	f2,  f17, f21, f2
+	FMA4	f1,  f17, f20, f1
+	FMA3	f0,  f17, f21, f0
 
 	LFD	f16,  4 * SIZE(AO)
 	LFD	f17,  5 * SIZE(AO)
@@ -2745,9 +2715,9 @@ LL(62):
 	LFD	f21,  5 * SIZE(BO)
 
 	FMA1	f0,  f18, f22, f0
-	FMA4	f3,  f19, f22, f3
 	FMA2	f1,  f18, f23, f1
-	FMA3	f2,  f19, f23, f2
+	FMA4	f1,  f19, f22, f1
+	FMA3	f0,  f19, f23, f0
 
 	LFD	f18,  6 * SIZE(AO)
 	LFD	f19,  7 * SIZE(AO)
@@ -2755,9 +2725,9 @@ LL(62):
 	LFD	f23,  7 * SIZE(BO)
 
 	FMA1	f0,  f16, f20, f0
-	FMA4	f3,  f17, f20, f3
 	FMA2	f1,  f16, f21, f1
-	FMA3	f2,  f17, f21, f2
+	FMA4	f1,  f17, f20, f1
+	FMA3	f0,  f17, f21, f0
 
 	LFD	f16,  8 * SIZE(AO)
 	LFD	f17,  9 * SIZE(AO)
@@ -2765,9 +2735,9 @@ LL(62):
 	LFD	f21,  9 * SIZE(BO)
 
 	FMA1	f0,  f18, f22, f0
-	FMA4	f3,  f19, f22, f3
 	FMA2	f1,  f18, f23, f1
-	FMA3	f2,  f19, f23, f2
+	FMA4	f1,  f19, f22, f1
+	FMA3	f0,  f19, f23, f0
 
 	LFD	f18, 10 * SIZE(AO)
 	LFD	f19, 11 * SIZE(AO)
@@ -2803,11 +2773,11 @@ LL(65):
 
 LL(66):
 	FMA1	f0,  f16, f20, f0
-	FMA4	f3,  f17, f20, f3
-	LFD	f20,  2 * SIZE(BO)
 	FMA2	f1,  f16, f21, f1
 	LFD	f16,  2 * SIZE(AO)
-	FMA3	f2,  f17, f21, f2
+	FMA4	f1,  f17, f20, f1
+	LFD	f20,  2 * SIZE(BO)
+	FMA3	f0,  f17, f21, f0
 	LFD	f17,  3 * SIZE(AO)
 
 	LFD	f21,  3 * SIZE(BO)
@@ -2821,20 +2791,17 @@ LL(68):
 	LFD	f16, 0 * SIZE(CO1)
 	LFD	f17, 1 * SIZE(CO1)
 
-	FADD	f0,  f0,  f2
-	FADD	f1,  f1,  f3
+	fmr f2, f0
+	fmr f3, f1
 
-	FNMSUB	f24, f31, f1, f16
-	FMADD	f25, f31, f0, f17
+	FMADD	f24,  f30, f0, f16
+	FMADD	f25,  f30, f1, f17
 
-	FMADD	f0,  f30, f0, f24
-	FMADD	f1,  f30, f1, f25
+	FNMSUB	f0, f31, f3, f24
+	FMADD	f1, f31, f2, f25
 
 #else
 
-	FADD	f0,  f0,  f2
-	FADD	f1,  f1,  f3
-
 	FMUL	f16, f31, f1
 	FMUL	f17, f31, f0
 

kernel/riscv64/KERNEL.RISCV64_GENERIC

@@ -99,26 +99,26 @@ ZTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
 SGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
 SGEMMONCOPY    =  ../generic/gemm_ncopy_2.c
 SGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c
-SGEMMONCOPYOBJ =  sgemm_oncopy.o
-SGEMMOTCOPYOBJ =  sgemm_otcopy.o
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
 DGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
 DGEMMONCOPY    = ../generic/gemm_ncopy_2.c
 DGEMMOTCOPY    = ../generic/gemm_tcopy_2.c
-DGEMMONCOPYOBJ = dgemm_oncopy.o
-DGEMMOTCOPYOBJ = dgemm_otcopy.o
+DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
 CGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
 CGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
 CGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
-CGEMMONCOPYOBJ =  cgemm_oncopy.o
-CGEMMOTCOPYOBJ =  cgemm_otcopy.o
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
 ZGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
 ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
 ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
-ZGEMMONCOPYOBJ =  zgemm_oncopy.o
-ZGEMMOTCOPYOBJ =  zgemm_otcopy.o
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
 STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
 STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c

kernel/riscv64/scal.c

@@ -43,9 +43,9 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 	if ( (n <= 0) || (inc_x <= 0))
 		return(0);
 	
-
-	while(j < n)
-	{
+	if (dummy2 == 1) {
+		while(j < n)
+		{
 
 		if ( da == 0.0 )
 			if (isfinite(x[i]))
@@ -57,7 +57,19 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 
 		i += inc_x ;
 		j++;
+		}
+	} else {
+		while(j < n)
+		{
 
+		if ( da == 0.0 )
+			x[i]=0.0;
+		else
+			x[i] = da * x[i] ;
+
+		i += inc_x ;
+		j++;
+		}
 	}
 	return 0;
 

kernel/riscv64/scal_rvv.c

@@ -56,7 +56,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
     FLOAT_V_T v0;
  
     if(inc_x == 1) {
-        if(da == 0.0) {
+        if(dummy2 == 0 && da == 0.0) {
             int gvl = VSETVL_MAX;
             v0 = VFMVVF_FLOAT(0.0, gvl);
             for (size_t vl; n > 0; n -= vl, x += vl) {
@@ -75,7 +75,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
     } else {
         BLASLONG stride_x = inc_x * sizeof(FLOAT);
 
-        if(da == 0.0) {
+        if(dummy2 == 0 && da == 0.0) {
             int gvl = VSETVL_MAX;
             v0 = VFMVVF_FLOAT(0.0, gvl);
             for (size_t vl; n > 0; n -= vl, x += vl*inc_x) {

kernel/riscv64/scal_vector.c

@@ -71,7 +71,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
         FLOAT_V_T v0, v1;
         unsigned int gvl = 0;
         if(inc_x == 1){
-                if (0){ //if(da == 0.0){
+                if(dummy2 == 0 && da == 0.0){
                         memset(&x[0], 0, n * sizeof(FLOAT));
                 }else{
                         gvl = VSETVL(n);
@@ -96,7 +96,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
                         }
                 }
         }else{
-                if (0) { //if(da == 0.0){
+                if(dummy2 == 0 && da == 0.0){
                         BLASLONG stride_x = inc_x * sizeof(FLOAT);
                         BLASLONG ix = 0;
                         gvl = VSETVL(n);

kernel/setparam-ref.c

@@ -1244,6 +1244,36 @@ static void init_parameter(void) {
 }
 #else //ZARCH
 
+#if (ARCH_RISCV64)
+static void init_parameter(void) {
+
+#ifdef BUILD_BFLOAT16
+  TABLE_NAME.sbgemm_p = SBGEMM_DEFAULT_P;
+#endif
+  TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
+  TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
+  TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
+  TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
+
+#ifdef BUILD_BFLOAT16
+  TABLE_NAME.sbgemm_r = SBGEMM_DEFAULT_R;
+#endif
+  TABLE_NAME.sgemm_r = SGEMM_DEFAULT_R;
+  TABLE_NAME.dgemm_r = DGEMM_DEFAULT_R;
+  TABLE_NAME.cgemm_r = CGEMM_DEFAULT_R;
+  TABLE_NAME.zgemm_r = ZGEMM_DEFAULT_R;
+
+
+#ifdef BUILD_BFLOAT16
+  TABLE_NAME.sbgemm_q = SBGEMM_DEFAULT_Q;
+#endif
+  TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
+  TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
+  TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
+  TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
+}
+#else //RISCV64
+
 #ifdef ARCH_X86
 static int get_l2_size_old(void){
   int i, eax, ebx, ecx, edx, cpuid_level;
@@ -2046,6 +2076,7 @@ static void init_parameter(void) {
 
 
 }
+#endif //RISCV64
 #endif //POWER
 #endif //ZARCH
 #endif //(ARCH_LOONGARCH64)

kernel/x86/scal.S

@@ -57,19 +57,24 @@
 #ifdef XDOUBLE
 	movl	44(%esp),%edi
 	movl	48(%esp),%esi
+	movl    64(%esp),%ecx
 #elif defined(DOUBLE)
 	movl	36(%esp),%edi
 	movl	40(%esp),%esi
+	movl	56(%esp),%ecx
 #else
 	movl	32(%esp),%edi
 	movl	36(%esp),%esi
+	movl	52(%esp),%ecx
 #endif
 
 	ftst
 	fnstsw	%ax
 	andb	$68, %ah
-//	je	.L300		# Alpha != ZERO
-	jmp	.L300
+	je	.L300		# Alpha != ZERO
+
+	cmpl	$1,%ecx		# dummy2 flag
+	je	.L300
 
 /* Alpha == ZERO */
 	cmpl	$1,%esi

kernel/x86_64/dscal.c

@@ -43,21 +43,21 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 static void dscal_kernel_8( BLASLONG n, FLOAT *da , FLOAT *x )
 {
 
-	BLASLONG i;
-	FLOAT alpha = *da;
-
-	for( i=0; i<n; i+=8 )
-	{
-		x[0] *= alpha;	
-		x[1] *= alpha;	
-		x[2] *= alpha;	
-		x[3] *= alpha;	
-		x[4] *= alpha;	
-		x[5] *= alpha;	
-		x[6] *= alpha;	
-		x[7] *= alpha;	
-		x+=8;
-	}
+    BLASLONG i;
+    FLOAT alpha = *da;
+
+    for( i=0; i<n; i+=8 )
+    {
+        x[0] *= alpha;
+        x[1] *= alpha;
+        x[2] *= alpha;
+        x[3] *= alpha;
+        x[4] *= alpha;
+        x[5] *= alpha;
+        x[6] *= alpha;
+        x[7] *= alpha;
+        x+=8;
+    }
 
 }
 
@@ -65,19 +65,19 @@ static void dscal_kernel_8( BLASLONG n, FLOAT *da , FLOAT *x )
 static void dscal_kernel_8_zero( BLASLONG n, FLOAT *alpha , FLOAT *x )
 {
 
-	BLASLONG i;
-	for( i=0; i<n; i+=8 )
-	{
-		x[0] = 0.0;	
-		x[1] = 0.0;	
-		x[2] = 0.0;	
-		x[3] = 0.0;	
-		x[4] = 0.0;	
-		x[5] = 0.0;	
-		x[6] = 0.0;	
-		x[7] = 0.0;	
-		x+=8;
-	}
+    BLASLONG i;
+    for( i=0; i<n; i+=8 )
+    {
+        x[0] = 0.0;
+        x[1] = 0.0;
+        x[2] = 0.0;
+        x[3] = 0.0;
+        x[4] = 0.0;
+        x[5] = 0.0;
+        x[6] = 0.0;
+        x[7] = 0.0;
+        x+=8;
+    }
 
 }
 
@@ -89,51 +89,51 @@ static void dscal_kernel_inc_8(BLASLONG n, FLOAT *alpha, FLOAT *x, BLASLONG inc_
 static void dscal_kernel_inc_8(BLASLONG n, FLOAT *alpha, FLOAT *x, BLASLONG inc_x)
 {
 
-	FLOAT *x1=NULL;
-	BLASLONG inc_x3;
+    FLOAT *x1=NULL;
+    BLASLONG inc_x3;
 
-	inc_x <<= 3;
-	inc_x3 = (inc_x << 1) + inc_x;
+    inc_x <<= 3;
+    inc_x3 = (inc_x << 1) + inc_x;
 
         __asm__  __volatile__
         (
-        "movddup               (%3), %%xmm0                 \n\t"  // alpha     
+        "movddup (%3),      %%xmm0              \n\t"  // alpha
 
-	"leaq		(%1,%4,4), %2		            \n\t"
+        "leaq    (%1,%4,4), %2                  \n\t"
 
-        ".p2align 4                                          \n\t"
+        ".p2align 4                             \n\t"
 
-        "1:                                                 \n\t"
-	"movsd	(%1)     , %%xmm4			    \n\t"
-	"movhpd (%1,%4,1), %%xmm4			    \n\t"
-	"movsd	(%1,%4,2), %%xmm5			    \n\t"
-	"movhpd (%1,%5,1), %%xmm5			    \n\t"
+        "1:                                     \n\t"
+        "movsd  (%1)     , %%xmm4               \n\t"
+        "movhpd (%1,%4,1), %%xmm4               \n\t"
+        "movsd  (%1,%4,2), %%xmm5               \n\t"
+        "movhpd (%1,%5,1), %%xmm5               \n\t"
 
-	"movsd	(%2)     , %%xmm6			    \n\t"
-	"movhpd (%2,%4,1), %%xmm6			    \n\t"
-	"movsd	(%2,%4,2), %%xmm7			    \n\t"
-	"movhpd (%2,%5,1), %%xmm7			    \n\t"
+        "movsd  (%2)     , %%xmm6               \n\t"
+        "movhpd (%2,%4,1), %%xmm6               \n\t"
+        "movsd  (%2,%4,2), %%xmm7               \n\t"
+        "movhpd (%2,%5,1), %%xmm7               \n\t"
 
-	"mulpd  %%xmm0, %%xmm4				    \n\t"
-	"mulpd  %%xmm0, %%xmm5				    \n\t"
-	"mulpd  %%xmm0, %%xmm6				    \n\t"
-	"mulpd  %%xmm0, %%xmm7				    \n\t"
+        "mulpd  %%xmm0, %%xmm4                  \n\t"
+        "mulpd  %%xmm0, %%xmm5                  \n\t"
+        "mulpd  %%xmm0, %%xmm6                  \n\t"
+        "mulpd  %%xmm0, %%xmm7                  \n\t"
 
-	"movsd  %%xmm4 , (%1)				    \n\t"
-	"movhpd %%xmm4 , (%1,%4,1)			    \n\t"
-	"movsd  %%xmm5 , (%1,%4,2)			    \n\t"
-	"movhpd %%xmm5 , (%1,%5,1)			    \n\t"
+        "movsd  %%xmm4 , (%1)                   \n\t"
+        "movhpd %%xmm4 , (%1,%4,1)              \n\t"
+        "movsd  %%xmm5 , (%1,%4,2)              \n\t"
+        "movhpd %%xmm5 , (%1,%5,1)              \n\t"
 
-	"movsd  %%xmm6 , (%2)				    \n\t"
-	"movhpd %%xmm6 , (%2,%4,1)			    \n\t"
-	"movsd  %%xmm7 , (%2,%4,2)			    \n\t"
-	"movhpd %%xmm7 , (%2,%5,1)			    \n\t"
+        "movsd  %%xmm6 , (%2)                   \n\t"
+        "movhpd %%xmm6 , (%2,%4,1)              \n\t"
+        "movsd  %%xmm7 , (%2,%4,2)              \n\t"
+        "movhpd %%xmm7 , (%2,%5,1)              \n\t"
 
-	"leaq   (%1,%4,8), %1				    \n\t"
-	"leaq   (%2,%4,8), %2				    \n\t"
+        "leaq   (%1,%4,8), %1                   \n\t"
+        "leaq   (%2,%4,8), %2                   \n\t"
 
-	"subq	$8, %0					    \n\t"
-	"jnz    1b					    \n\t"
+        "subq   $8, %0                          \n\t"
+        "jnz    1b                              \n\t"
 
         :
           "+r" (n),     // 0
@@ -150,106 +150,96 @@ static void dscal_kernel_inc_8(BLASLONG n, FLOAT *alpha, FLOAT *x, BLASLONG inc_
           "%xmm12", "%xmm13", "%xmm14", "%xmm15",
           "memory"
         );
-
-
 }
 
 int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
 {
-	BLASLONG i=0,j=0;
-
-	if ( inc_x != 1 )
-	{
-
-		if ( da == 0.0 )
-		{
-
-			BLASLONG n1 = n & -2;
-
-			while(j < n1)
-			{
-			
-				if (isinf(x[i])||isnan(x[i]))
-					x[i]=NAN;
-				else x[i]=0.0;
-				if (isinf(x[i+inc_x])||isnan(x[i+inc_x]))
-					x[i+inc_x]=NAN;
-				else x[i+inc_x]=0.0;
-				i += 2*inc_x ;
-				j+=2;
-
-			}
-
-			while(j < n)
-			{
-			
-				if (isinf(x[i])||isnan(x[i]))
-					x[i]=NAN;
-				else x[i]=0.0;
-				i += inc_x ;
-				j++;
-
-			}
-		}
-		else
-		{
-
-			BLASLONG n1 = n & -8;
-			if ( n1 > 0 )
-			{
-				dscal_kernel_inc_8(n1, &da, x, inc_x);
-				i = n1 * inc_x;
-				j = n1;
-		        }			
-
-			while(j < n)
-			{
-			
-				x[i] *= da;
-				i += inc_x ;
-				j++;
-
-			}
-
-		}
-
-		return(0);
-	}
-
-	BLASLONG n1 = n & -8;
-	if ( n1 > 0 )
-	{
-//		if ( da == 0.0 )
-//			dscal_kernel_8_zero(n1 , &da , x);
-//		else
-			dscal_kernel_8(n1 , &da , x);
-	}
-
-	if ( da == 0.0 )
-	{
-		for ( i=n1 ; i<n; i++ )
-		{
-			if(isinf(x[i])||isnan(x[i]))
-				x[i]=NAN;
-			else x[i] = 0.0;
-		}
-	}
-	else if (isinf(da)){
-		for ( i=n1 ; i<n; i++)
-			if (x[i]==0.) x[i]=NAN;
-			else x[i] *=da;
-	}	
-	else
-	{
-
-		for ( i=n1 ; i<n; i++ )
-		{
-			if(isinf(x[i]))
-				x[i]=NAN;
-			else x[i] *= da;
-		}
-	}
-	return(0);
+    BLASLONG i = 0, j = 0;
+
+    // Resolved issue 4728 when the caller is dscal
+    if (dummy2 == 1 && da == 0.0)
+    {
+        if ( inc_x != 1 )
+        {
+            BLASLONG n1 = n & -8;
+            if ( n1 > 0 )
+            {
+                dscal_kernel_inc_8(n1, &da, x, inc_x);
+                i = n1 * inc_x;
+                j = n1;
+            }
+            while(j < n)
+            {
+                x[i] *= da;
+                i += inc_x ;
+                j++;
+            }
+        }
+        else
+        {
+            BLASLONG n1 = n & -8;
+            if ( n1 > 0)
+                dscal_kernel_8(n1 , &da , x);
+            for ( i = n1 ; i < n; i++ )
+                x[i] *= da;
+        }
+    }
+    else
+    {
+        if ( inc_x != 1 )
+        {
+            if( da == 0.0)
+            {
+                BLASLONG n1 = n & -2;
+                while(j < n1)
+                {
+                    x[i] = 0.0;
+                    x[i+inc_x] = 0.0;
+                    i += 2 * inc_x ;
+                    j += 2;
+                }
+                while(j < n)
+                {
+                    x[i] = 0.0;
+                    i += inc_x ;
+                    j++;
+                }
+            }
+            else
+            {
+                BLASLONG n1 = n & -8;
+                if ( n1 > 0 )
+                {
+                    dscal_kernel_inc_8(n1, &da, x, inc_x);
+                    i = n1 * inc_x;
+                    j = n1;
+                }
+                while(j < n)
+                {
+                    x[i] *= da;
+                    i += inc_x ;
+                    j++;
+                }
+            }
+        }
+        else
+        {
+            if ( da == 0.0 )
+            {
+                BLASLONG n1 = n & -8;
+                if ( n1 > 0)
+                   dscal_kernel_8_zero(n1, &da, x);
+                for ( i = n1 ; i < n; i++ )
+                    x[i] = 0.0;
+            }
+            else
+            {
+                BLASLONG n1 = n & -8;
+                if ( n1 > 0)
+                    dscal_kernel_8(n1 , &da , x);
+                for ( i = n1 ; i < n; i++ )
+                    x[i] *= da;
+            }
+        }
+    }
 }
-
-

kernel/x86_64/scal_atom.S

@@ -60,8 +60,10 @@
 #ifdef WINDOWS_ABI
 	movq	40(%rsp), X
 	movq	48(%rsp), INCX
-
+	movq    64(%rsp), %r9
 	movaps	%xmm3, %xmm0
+#else
+	movq	24(%rsp), %r9
 #endif
 
 	SAVEREGISTERS
@@ -73,6 +75,10 @@
 	lea	(, INCX, SIZE), INCX
 	comisd	%xmm0, %xmm1
 	jne	.L100
+	jp	.L100
+
+	cmpq	$1, %r9
+	je	.L100
 
 /* Alpha == ZERO */
 	cmpq	$SIZE, INCX

kernel/x86_64/scal_sse.S

@@ -60,8 +60,10 @@
 #ifdef WINDOWS_ABI
 	movq	40(%rsp), X
 	movq	48(%rsp), INCX
-
+	movq	64(%rsp), %r9
 	movaps	%xmm3, %xmm0
+#else
+	movq    24(%rsp), %r9
 #endif
 
 	SAVEREGISTERS
@@ -76,6 +78,8 @@
 	shufps	$0, %xmm0, %xmm0
 
 	jne	.L100		# Alpha != ZERO
+
+	cmpq   $1, %r9
 	je	.L100
 /* Alpha == ZERO */
 	cmpq	$SIZE, INCX

kernel/x86_64/scal_sse2.S

@@ -48,6 +48,7 @@
 #define X	ARG2
 #define INCX	ARG3
 #endif
+#define FLAG    %r9
 
 #define XX	%r10
 #define I	%rax
@@ -60,8 +61,10 @@
 #ifdef WINDOWS_ABI
 	movq	40(%rsp), X
 	movq	48(%rsp), INCX
-
+	movq	64(%rsp), FLAG
 	movaps	%xmm3, %xmm0
+#else
+	movq    24(%rsp), FLAG
 #endif
 
 	SAVEREGISTERS
@@ -75,6 +78,8 @@
 	comisd	%xmm0, %xmm1
 	jne	.L100		# Alpha != ZERO
 	jp	.L100		# For Alpha = NaN
+
+	cmpq $1, FLAG
 	je 	.L100		# disable the Alpha=zero path as it does not handle x=inf or nan
 /* Alpha == ZERO */
 	cmpq	$SIZE, INCX

kernel/x86_64/sscal.c

@@ -39,21 +39,21 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 static void sscal_kernel_16( BLASLONG n, FLOAT *da , FLOAT *x )
 {
 
-	BLASLONG i;
-	FLOAT alpha = *da;
-
-	for( i=0; i<n; i+=8 )
-	{
-		x[0] *= alpha;
-		x[1] *= alpha;
-		x[2] *= alpha;
-		x[3] *= alpha;
-		x[4] *= alpha;
-		x[5] *= alpha;
-		x[6] *= alpha;
-		x[7] *= alpha;
-		x+=8;
-	}
+    BLASLONG i;
+    FLOAT alpha = *da;
+
+    for( i=0; i<n; i+=8 )
+    {
+        x[0] *= alpha;
+        x[1] *= alpha;
+        x[2] *= alpha;
+        x[3] *= alpha;
+        x[4] *= alpha;
+        x[5] *= alpha;
+        x[6] *= alpha;
+        x[7] *= alpha;
+        x+=8;
+    }
 
 }
 
@@ -61,19 +61,19 @@ static void sscal_kernel_16( BLASLONG n, FLOAT *da , FLOAT *x )
 static void sscal_kernel_16_zero( BLASLONG n, FLOAT *alpha , FLOAT *x )
 {
 
-	BLASLONG i;
-	for( i=0; i<n; i+=8 )
-	{
-		x[0] = 0.0;
-		x[1] = 0.0;
-		x[2] = 0.0;
-		x[3] = 0.0;
-		x[4] = 0.0;
-		x[5] = 0.0;
-		x[6] = 0.0;
-		x[7] = 0.0;
-		x+=8;
-	}
+    BLASLONG i;
+    for( i=0; i<n; i+=8 )
+    {
+        x[0] = 0.0;
+        x[1] = 0.0;
+        x[2] = 0.0;
+        x[3] = 0.0;
+        x[4] = 0.0;
+        x[5] = 0.0;
+        x[6] = 0.0;
+        x[7] = 0.0;
+        x+=8;
+    }
 
 }
 
@@ -85,126 +85,119 @@ static void sscal_kernel_inc_8(BLASLONG n, FLOAT *alpha, FLOAT *x, BLASLONG inc_
 static void sscal_kernel_inc_8(BLASLONG n, FLOAT *alpha, FLOAT *x, BLASLONG inc_x)
 {
 
-	BLASLONG i;
-	BLASLONG inc_x2 = 2 * inc_x;
-	BLASLONG inc_x3 = inc_x2 + inc_x;
-	FLOAT t0,t1,t2,t3;
-	FLOAT da = alpha[0];
+    BLASLONG i;
+    BLASLONG inc_x2 = 2 * inc_x;
+    BLASLONG inc_x3 = inc_x2 + inc_x;
+    FLOAT t0,t1,t2,t3;
+    FLOAT da = alpha[0];
 
-	for ( i=0; i<n; i+=4 )
-	{
-		t0 = da * x[0];
-		t1 = da * x[inc_x];
-		t2 = da * x[inc_x2];
-		t3 = da * x[inc_x3];
+    for ( i=0; i<n; i+=4 )
+    {
+        t0 = da * x[0];
+        t1 = da * x[inc_x];
+        t2 = da * x[inc_x2];
+        t3 = da * x[inc_x3];
 
-		x[0]        = t0;
-		x[inc_x]    = t1;
-		x[inc_x2]   = t2;
-		x[inc_x3]   = t3;
+        x[0]        = t0;
+        x[inc_x]    = t1;
+        x[inc_x2]   = t2;
+        x[inc_x3]   = t3;
 
-		x+=4*inc_x;
+        x+=4*inc_x;
 
-	}
+    }
 
 
 }
 
 int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
 {
-	BLASLONG i=0,j=0;
-
-	if ( inc_x != 1 )
-	{
-
-		if ( da == 0.0 )
-		{
-			BLASLONG n1 = n & -2;
-
-			while(j < n1)
-			{
-                                if (isinf(x[i])||isnan(x[i]))
-                                        x[i]=NAN;
-                                else x[i]=0.0;
-                                if (isinf(x[i+inc_x])||isnan(x[i+inc_x]))
-                                        x[i+inc_x]=NAN;
-				else	x[i+inc_x]=0.0;
-				i += 2*inc_x ;
-				j+=2;
-
-			}
-
-			while(j < n)
-			{
-                                if (isinf(x[i])||isnan(x[i]))
-                                        x[i]=NAN;
-				else x[i]=0.0;
-				i += inc_x ;
-				j++;
-
-			}
-		}
-		else
-		{
-#if 1
-			BLASLONG n1 = n & -8;
-			if ( n1 > 0 )
-			{
-				sscal_kernel_inc_8(n1, &da, x, inc_x);
-				i = n1 * inc_x;
-				j = n1;
-		        }
-#endif
-			while(j < n)
-			{
-				x[i] *= da;
-				i += inc_x ;
-				j++;
-
-			}
-
-		}
-		return(0);
-	}
-
-	BLASLONG n1 = n & -16;
-	if ( n1 > 0 )
-	{
-		//if ( da == 0.0 )
-		//	sscal_kernel_16_zero(n1 , &da , x);
-		//else
-			sscal_kernel_16(n1 , &da , x);
-	}
-
-	if ( da == 0.0 )
-	{
-		for ( i=n1 ; i<n; i++ )
-		{
-                        if (isinf(x[i])||isnan(x[i]))
-                                x[i]=NAN;
-			else x[i]=0.0;
-		}
-	}
-	else if ( isinf(da) )
-	{
-		for ( i=n1 ; i<n; i++ )
-		{
-                        if (x[i] == 0.0)
-                                x[i]=NAN;
-			else x[i] *= da;
-		}
-	}
-	else
-	{
-
-		for ( i=n1 ; i<n; i++ )
-		{
-                        if (isinf(x[i]))
-                                x[i]=NAN;
-			else x[i] *= da;
-		}
-	}
-	return(0);
+    BLASLONG i = 0, j = 0;
+
+    // Resolved issue 4728 when the caller is sscal
+    if (dummy2 == 1 && da == 0.0)
+    {
+        if ( inc_x != 1 )
+        {
+            BLASLONG n1 = n & -8;
+            if ( n1 > 0 )
+            {
+                sscal_kernel_inc_8(n1, &da, x, inc_x);
+                i = n1 * inc_x;
+                j = n1;
+            }
+            while(j < n)
+            {
+                x[i] *= da;
+                i += inc_x ;
+                j++;
+            }
+        }
+        else
+        {
+            BLASLONG n1 = n & -16;
+            if ( n1 > 0)
+                sscal_kernel_16(n1 , &da , x);
+            for ( i = n1 ; i < n; i++ )
+                x[i] *= da;
+        }
+    }
+    else
+    {
+        if ( inc_x != 1 )
+        {
+            if( da == 0.0)
+            {
+                BLASLONG n1 = n & -2;
+                while(j < n1)
+                {
+                    x[i] = 0.0;
+                    x[i+inc_x] = 0.0;
+                    i += 2 * inc_x ;
+                    j += 2;
+                }
+                while(j < n)
+                {
+                    x[i] = 0.0;
+                    i += inc_x ;
+                    j++;
+                }
+            }
+            else
+            {
+                BLASLONG n1 = n & -8;
+                if ( n1 > 0 )
+                {
+                    sscal_kernel_inc_8(n1, &da, x, inc_x);
+                    i = n1 * inc_x;
+                    j = n1;
+                }
+                while(j < n)
+                {
+                    x[i] *= da;
+                    i += inc_x ;
+                    j++;
+                }
+            }
+        }
+        else
+        {
+            if ( da == 0.0 )
+            {
+                BLASLONG n1 = n & -16;
+                if ( n1 > 0)
+                   sscal_kernel_16_zero(n1, &da, x);
+                for ( i = n1 ; i < n; i++ )
+                    x[i] = 0.0;
+            }
+            else
+            {
+                BLASLONG n1 = n & -16;
+                if ( n1 > 0)
+                    sscal_kernel_16(n1 , &da , x);
+                for ( i = n1 ; i < n; i++ )
+                    x[i] *= da;
+            }
+        }
+    }
 }
-
-

kernel/x86_64/zscal_atom.S

@@ -74,7 +74,7 @@
 	pxor	%xmm15, %xmm15
 	comisd	%xmm0, %xmm15
 	jne	.L30		# Alpha_r != ZERO
-
+	jp	.L30
 	comisd	%xmm1, %xmm15
 	jne	.L30		# Alpha_i != ZERO
 

kernel/x86_64/zscal_sse.S

@@ -76,7 +76,7 @@
 	pxor	%xmm15, %xmm15
 	comiss	%xmm0, %xmm15
 	jne	.L100		# Alpha_r != ZERO
-
+	jp	.L100		# Alpha_r == NAN	
 	comiss	%xmm1, %xmm15
 	jne	.L100		# Alpha_i != ZERO
 

kernel/zarch/cscal.c

@@ -234,12 +234,23 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i,
       } else {
 
         while (j < n1) {
-
-          temp0 = -da_i * x[i + 1];
-          x[i + 1] = da_i * x[i];
+          if (isnan(x[i]) || isinf(x[i]))
+            temp0   = NAN;
+	  else
+	    temp0 = -da_i * x[i + 1];
+	  if (!isinf(x[i + 1]))
+            x[i + 1] = da_i * x[i];
+	  else
+	    x[i + 1] = NAN;
           x[i] = temp0;
-          temp1 = -da_i * x[i + 1 + inc_x];
-          x[i + 1 + inc_x] = da_i * x[i + inc_x];
+	  if (isnan(x[i+inc_x]) || isinf(x[i+inc_x]))
+	    temp1 = NAN;
+	  else
+	    temp1 = -da_i * x[i + 1 + inc_x];
+	  if (!isinf(x[i + 1 + inc_x]))
+            x[i + 1 + inc_x] = da_i * x[i + inc_x];
+	  else
+	    x[i + 1 + inc_x] = NAN;
           x[i + inc_x] = temp1;
           i += 2 * inc_x;
           j += 2;
@@ -247,9 +258,14 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i,
         }
 
         while (j < n) {
-
-          temp0 = -da_i * x[i + 1];
-          x[i + 1] = da_i * x[i];
+	  if (isnan(x[i]) || isinf(x[i]))
+	    temp0 = NAN;
+	  else
+            temp0 = -da_i * x[i + 1];
+	  if (isinf(x[i + 1]))
+	    x[i + 1] = NAN;
+	  else
+            x[i + 1] = da_i * x[i];
           x[i] = temp0;
           i += inc_x;
           j++;
@@ -332,26 +348,42 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i,
     j = n1;
   }
 
-  if (da_r == 0.0) {
+  if (da_r == 0.0 || isnan(da_r)) {
 
     if (da_i == 0.0) {
-
+      float res = 0.0;
+      if (isnan(da_r)) res = da_r;
       while (j < n) {
 
-        x[i] = 0.0;
-        x[i + 1] = 0.0;
+        x[i] = res;
+        x[i + 1] = res;
         i += 2;
         j++;
 
       }
+    } else if (isinf(da_r)) {
+        while(j < n)
+        {
+
+	  x[i]= NAN;
+          x[i+1] = da_r;
+          i += 2 ;
+          j++;
+
+        }
 
     } else {
 
       while (j < n) {
 
         temp0 = -da_i * x[i + 1];
-        x[i + 1] = da_i * x[i];
-        x[i] = temp0;
+	if (isinf(x[i])) temp0 = NAN;
+	if (!isinf(x[i + 1]))
+          x[i + 1] = da_i * x[i];
+	else
+	  x[i + 1] = NAN;
+	if (x[i] == x[i])
+          x[i] = temp0;
         i += 2;
         j++;
 

kernel/zarch/dscal.c

@@ -96,20 +96,28 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x,
   if (inc_x == 1) {
 
     if (da == 0.0) {
-
-      BLASLONG n1 = n & -16;
-      if (n1 > 0) {
-
-        dscal_kernel_16_zero(n1, x);
-        j = n1;
+    
+      if (dummy2 == 0) {
+        BLASLONG n1 = n & -16;
+        if (n1 > 0) {
+          dscal_kernel_16_zero(n1, x);
+          j = n1;
+        }
+
+        while (j < n) {
+          x[j] = 0.0;
+          j++;
+        }
+      } else {
+        while (j < n) {
+          if (isfinite(x[j]))
+            x[j] = 0.0;
+          else
+            x[j] = NAN;
+          j++;
+        }
       }
-
-      while (j < n) {
-
-        x[j] = 0.0;
-        j++;
-      }
-
+      
     } else {
 
       BLASLONG n1 = n & -16;
@@ -127,11 +135,9 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x,
   } else {
 
     if (da == 0.0) {
-
+     if (dummy2 == 0) {
       BLASLONG n1 = n & -4;
-
       while (j < n1) {
-
         x[i] = 0.0;
         x[i + inc_x] = 0.0;
         x[i + 2 * inc_x] = 0.0;
@@ -139,11 +145,13 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x,
 
         i += inc_x * 4;
         j += 4;
-
       }
+     }
       while (j < n) {
-
-        x[i] = 0.0;
+        if (dummy2==0 || isfinite(x[i]))
+          x[i] = 0.0;
+        else
+          x[i] = NAN;
         i += inc_x;
         j++;
       }

kernel/zarch/sscal.c

@@ -95,21 +95,31 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x,
 
   if (inc_x == 1) {
 
-    if (da == 0.0) {
-
-      BLASLONG n1 = n & -32;
-      if (n1 > 0) {
-
-        sscal_kernel_32_zero(n1, x);
-        j = n1;
-      }
-
-      while (j < n) {
-
-        x[j] = 0.0;
-        j++;
+    if (da == 0.0 || !isfinite(da)) {
+      if (dummy2 == 0) {
+        BLASLONG n1 = n & -32;
+        if (n1 > 0) {
+
+          sscal_kernel_32_zero(n1, x);
+          j = n1;
+        }
+
+        while (j < n) {
+
+          x[j] = 0.0;
+          j++;
+        }
+      } else {
+        float res = 0.0;
+        if (!isfinite(da)) res = NAN;
+        while (j < n) {
+          if (isfinite(x[i]))
+            x[j] = res;
+          else
+            x[j] = NAN;
+          j++;
+        }
       }
-
     } else {
 
       BLASLONG n1 = n & -32;
@@ -126,26 +136,37 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x,
 
   } else {
 
-    if (da == 0.0) {
-
-      BLASLONG n1 = n & -2;
-
-      while (j < n1) {
-
-        x[i] = 0.0;
-        x[i + inc_x] = 0.0;
-
-        i += inc_x * 2;
-        j += 2;
-
-      }
-      while (j < n) {
-
-        x[i] = 0.0;
-        i += inc_x;
-        j++;
-      }
-
+    if (da == 0.0 || !isfinite(da)) {
+      if (dummy2 == 0) {
+        BLASLONG n1 = n & -2;
+
+        while (j < n1) {
+
+          x[i] = 0.0;
+          x[i + inc_x] = 0.0;
+
+          i += inc_x * 2;
+          j += 2;
+
+        }
+        while (j < n) {
+
+          x[i] = 0.0;
+          i += inc_x;
+          j++;
+        }
+      } else {
+        while (j < n) {
+          float res = 0.0;
+          if (!isfinite(da)) res = NAN;
+          if (isfinite(x[i]))
+            x[i] = res;
+          else
+            x[i] = NAN;
+          i += inc_x;
+          j++;
+        }
+      }     
     } else {
       BLASLONG n1 = n & -2;
 

kernel/zarch/zscal.c

@@ -237,13 +237,19 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i,
 		temp0 = NAN;
 	  else
 		temp0 = -da_i * x[i + 1];
-          x[i + 1] = da_i * x[i];
+	  if (!isinf(x[i + 1]))
+          	x[i + 1] = da_i * x[i];
+	  else
+		x[i + 1] = NAN;
           x[i] = temp0;
 	  if (isnan(x[i + inc_x]) || isinf(x[i + inc_x]))
 		temp1 = NAN;
 	  else
           temp1 = -da_i * x[i + 1 + inc_x];
-          x[i + 1 + inc_x] = da_i * x[i + inc_x];
+	  if (!isinf(x[i + 1 + inc_x]))
+            x[i + 1 + inc_x] = da_i * x[i + inc_x];
+	  else
+	    x[i + 1 + inc_x] = NAN;
           x[i + inc_x] = temp1;
           i += 2 * inc_x;
           j += 2;
@@ -256,7 +262,10 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i,
 		temp0 = NAN;
 	  else
           	temp0 = -da_i * x[i + 1];
-          x[i + 1] = da_i * x[i];
+	  if (!isinf(x[i +1]))
+          	x[i + 1] = da_i * x[i];
+	  else
+		x[i + 1] = NAN;
           x[i] = temp0;
           i += inc_x;
           j++;
@@ -330,7 +339,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i,
         zscal_kernel_8_zero(n1, x);
       else
         zscal_kernel_8(n1, da_r, da_i, x);
-    else if (da_i == 0)
+    else if (da_i == 0 && da_r == da_r)
       zscal_kernel_8_zero_i(n1, alpha, x);
     else
       zscal_kernel_8(n1, da_r, da_i, x);
@@ -339,29 +348,41 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i,
     j = n1;
   }
 
-  if (da_r == 0.0) {
+  if (da_r == 0.0 || isnan(da_r)) {
 
     if (da_i == 0.0) {
-
+      double res= 0.0;
+      if (isnan(da_r)) res = da_r;
       while (j < n) {
 
-        x[i] = 0.0;
-        x[i + 1] = 0.0;
+        x[i] = res;
+        x[i + 1] = res;
         i += 2;
         j++;
 
       }
 
+    } else if (isinf(da_r)) {
+      while (j < n) {
+	x[i] = NAN;
+	x[i + 1] = da_r;
+	i += 2;
+	j++;
+      }
     } else {
 
       while (j < n) {
 
-	if (isnan(x[i]) || isinf(x[i]))
+	if (isinf(x[i]))
 		temp0 = NAN;
 	  else
         	temp0 = -da_i * x[i + 1];
-        x[i + 1] = da_i * x[i];
-        x[i] = temp0;
+	if (!isinf(x[i + 1]))
+        	x[i + 1] = da_i * x[i];
+	else
+		x[i + 1] = NAN;
+	if (x[i]==x[i])
+          x[i] = temp0;
         i += 2;
         j++;
 

lapack-netlib/SRC/cpstf2.c

@@ -256,7 +256,7 @@ static char junk[] = "\n@(#)LIBF77 VERSION 19990503\n";
 #define myceiling_(w) {ceil(w)}
 #define myhuge_(w) {HUGE_VAL}
 //#define mymaxloc_(w,s,e,n) {if (sizeof(*(w)) == sizeof(double)) dmaxloc_((w),*(s),*(e),n); else dmaxloc_((w),*(s),*(e),n);}
-#define mymaxloc_(w,s,e,n) dmaxloc_(w,*(s),*(e),n)
+#define mymaxloc_(w,s,e,n) smaxloc_(w,*(s),*(e),n)
 
 /* procedure parameter types for -A and -C++ */
 

lapack-netlib/SRC/cpstrf.c

@@ -256,7 +256,7 @@ static char junk[] = "\n@(#)LIBF77 VERSION 19990503\n";
 #define myceiling_(w) {ceil(w)}
 #define myhuge_(w) {HUGE_VAL}
 //#define mymaxloc_(w,s,e,n) {if (sizeof(*(w)) == sizeof(double)) dmaxloc_((w),*(s),*(e),n); else dmaxloc_((w),*(s),*(e),n);}
-#define mymaxloc_(w,s,e,n) dmaxloc_(w,*(s),*(e),n)
+#define mymaxloc_(w,s,e,n) smaxloc_(w,*(s),*(e),n)
 
 /* procedure parameter types for -A and -C++ */
 

lapack-netlib/SRC/spstf2.c

@@ -256,7 +256,7 @@ static char junk[] = "\n@(#)LIBF77 VERSION 19990503\n";
 #define myceiling_(w) {ceil(w)}
 #define myhuge_(w) {HUGE_VAL}
 //#define mymaxloc_(w,s,e,n) {if (sizeof(*(w)) == sizeof(double)) dmaxloc_((w),*(s),*(e),n); else dmaxloc_((w),*(s),*(e),n);}
-#define mymaxloc_(w,s,e,n) dmaxloc_(w,*(s),*(e),n)
+#define mymaxloc_(w,s,e,n) smaxloc_(w,*(s),*(e),n)
 
 /* procedure parameter types for -A and -C++ */
 

lapack-netlib/SRC/spstrf.c

@@ -256,7 +256,7 @@ static char junk[] = "\n@(#)LIBF77 VERSION 19990503\n";
 #define myceiling_(w) {ceil(w)}
 #define myhuge_(w) {HUGE_VAL}
 //#define mymaxloc_(w,s,e,n) {if (sizeof(*(w)) == sizeof(double)) dmaxloc_((w),*(s),*(e),n); else dmaxloc_((w),*(s),*(e),n);}
-#define mymaxloc_(w,s,e,n) dmaxloc_(w,*(s),*(e),n)
+#define mymaxloc_(w,s,e,n) smaxloc_(w,*(s),*(e),n)
 
 /* procedure parameter types for -A and -C++ */
 

lapack/laswp/riscv64/Makefile

@@ -1,6 +1,11 @@
 TOPDIR	= ../../..
 include ../../../Makefile.system
 
+ifeq ($(DYNAMIC_ARCH), 1)
+LASWP	= ../generic/laswp_k_4.c
+ZLASWP	= ../generic/zlaswp_k_4.c
+endif
+
 ifndef LASWP
 LASWP	= ../generic/laswp_k.c
 endif

test/Makefile

@@ -2,7 +2,7 @@ TOPDIR	= ..
 include ../Makefile.system
 ifeq ($(F_COMPILER),GFORTRAN)
 ifneq (, $(filter $(CORE),LOONGSON3R3 LOONGSON3R4))
-	override FFLAGS = $(filter_out(-O2 -O3,$(FFLAGS)) -O0
+	override FFLAGS = $(filter_out(-O2 -O3,$(FFLAGS))) -O0
 endif
         override FFLAGS += -fno-tree-vectorize
 endif

utest/CMakeLists.txt

@@ -18,6 +18,7 @@ else ()
     test_zscal.c
     test_amin.c
     test_axpby.c
+    test_gemv.c
   )
 endif ()
 

utest/Makefile

@@ -14,7 +14,7 @@ UTESTEXTBIN=openblas_utest_ext
 include $(TOPDIR)/Makefile.system
 
 OBJS=utest_main.o test_min.o test_amax.o test_ismin.o test_rotmg.o test_axpy.o test_dotu.o test_dsdot.o test_swap.o test_rot.o test_dnrm2.o test_zscal.o \
-     test_amin.o test_axpby.o
+     test_amin.o test_axpby.o test_gemv.o
 #test_rot.o test_swap.o test_axpy.o test_dotu.o test_dsdot.o test_fork.o
 OBJS_EXT=utest_main.o $(DIR_EXT)/xerbla.o $(DIR_EXT)/common.o 
 OBJS_EXT+=$(DIR_EXT)/test_isamin.o $(DIR_EXT)/test_idamin.o $(DIR_EXT)/test_icamin.o $(DIR_EXT)/test_izamin.o 

utest/test_gemv.c

@@ -0,0 +1,130 @@
+#include "openblas_utest.h"
+#include <cblas.h>
+
+#ifndef NAN
+#define NAN 0.0/0.0
+#endif
+#ifndef INFINITY
+#define INFINITY 1.0/0.0
+#endif
+
+#ifdef BUILD_SINGLE
+
+CTEST(sgemv, 0_nan_inf)
+{
+    int i;
+    blasint N = 17;
+    blasint incX = 1;
+    blasint incY = 1;
+    float alpha = 0.0;
+    float beta = 0.0;
+    char  trans = 'N';
+    float A[17 * 17];
+    float X[17];
+    float Y[17];
+
+    memset(A, 0, sizeof(A));
+    memset(X, 0, sizeof(X));
+    for (i = 0; i < (N - 1); i += 2)
+    {
+        Y[i]     = NAN;
+        Y[i + 1] = INFINITY;
+    }
+    Y[N - 1] = NAN;
+    BLASFUNC(sgemv)(&trans, &N, &N, &alpha, A, &N, X, &incX, &beta, Y, &incY);
+    for (i = 0; i < N; i ++)
+        ASSERT_TRUE(Y[i] == 0.0);
+}
+
+CTEST(sgemv, 0_nan_inf_incy_2)
+{
+    int i;
+    blasint N = 17;
+    blasint Ny = 33;
+    blasint incX = 1;
+    blasint incY = 2;
+    float alpha = 0.0;
+    float beta = 0.0;
+    char  trans = 'N';
+    float A[17 * 17];
+    float X[17];
+    float Y[33];
+    float *ay = Y;
+
+    memset(A, 0, sizeof(A));
+    memset(X, 0, sizeof(X));
+    memset(Y, 0, sizeof(Y));
+    for (i = 0; i < (N - 1); i += 2)
+    {
+        ay[0]   = NAN;
+        ay     += 2;
+        ay[0]   = INFINITY;
+        ay     += 2;
+    }
+    Y[Ny - 1] = NAN;
+    BLASFUNC(sgemv)(&trans, &N, &N, &alpha, A, &N, X, &incX, &beta, Y, &incY);
+    for (i = 0; i < Ny; i ++)
+        ASSERT_TRUE(Y[i] == 0.0);
+}
+
+#endif
+
+#ifdef BUILD_DOUBLE
+CTEST(dgemv, 0_nan_inf)
+{
+    int i;
+    blasint N = 17;
+    blasint incX = 1;
+    blasint incY = 1;
+    double alpha = 0.0;
+    double beta = 0.0;
+    char  trans = 'N';
+    double A[17 * 17];
+    double X[17];
+    double Y[17];
+
+    memset(A, 0, sizeof(A));
+    memset(X, 0, sizeof(X));
+    for (i = 0; i < (N - 1); i += 2)
+    {
+        Y[i]     = NAN;
+        Y[i + 1] = INFINITY;
+    }
+    Y[N - 1] = NAN;
+    BLASFUNC(dgemv)(&trans, &N, &N, &alpha, A, &N, X, &incX, &beta, Y, &incY);
+    for (i = 0; i < N; i ++)
+        ASSERT_TRUE(Y[i] == 0.0);
+}
+
+CTEST(dgemv, 0_nan_inf_incy_2)
+{
+    int i;
+    blasint N = 17;
+    blasint Ny = 33;
+    blasint incX = 1;
+    blasint incY = 2;
+    double alpha = 0.0;
+    double beta = 0.0;
+    char  trans = 'N';
+    double A[17 * 17];
+    double X[17];
+    double Y[33];
+    double *ay = Y;
+
+    memset(A, 0, sizeof(A));
+    memset(X, 0, sizeof(X));
+    memset(Y, 0, sizeof(Y));
+    for (i = 0; i < (N - 1); i += 2)
+    {
+        ay[0]   = NAN;
+        ay     += 2;
+        ay[0]   = INFINITY;
+        ay     += 2;
+    }
+    Y[Ny - 1] = NAN;
+    BLASFUNC(dgemv)(&trans, &N, &N, &alpha, A, &N, X, &incX, &beta, Y, &incY);
+    for (i = 0; i < Ny; i ++)
+        ASSERT_TRUE(Y[i] == 0.0);
+}
+
+#endif

utest/test_potrs.c

@@ -32,7 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 **********************************************************************************/
 
 #include "openblas_utest.h"
-
+#pragma GCC optimize("no-gcse")
 /*
 void BLASFUNC(cpotrf)(char*, BLASINT*, complex float*, BLASINT*, BLASINT*);
 void BLASFUNC(zpotrs_(char*, BLASINT*, BLASINT*, complex double*,