Merge branch 'OpenMathLib:develop' into fix_dyn_armv9sme

Author: martin-frbg

.github/workflows/c910v.yml

@@ -31,7 +31,7 @@ jobs:
 
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: install build deps
         run: |
@@ -40,18 +40,18 @@ jobs:
           gcc-${{ matrix.apt_triple }} gfortran-${{ matrix.apt_triple }} libgomp1-riscv64-cross libglib2.0-dev
 
       - name: checkout qemu
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
-          repository: T-head-Semi/qemu
+          repository: XUANTIE-RV/qemu
           path: qemu
-          ref: 1e692ebb43d396c52352406323fc782c1ac99a42
+          ref: e0ace167effcd36d1f82c7ccb4522b3126011479 # xuantie-qemu-9.0
 
       - name: build qemu
         run: |
           # Force use c910v qemu-user
-          wget https://github.com/revyos/qemu/commit/5164bca5a4bcde4534dc1a9aa3a7f619719874cf.patch
+          wget https://github.com/revyos/qemu/commit/222729c7455784dd855216d7a2bec4bd8f2a6800.patch
           cd qemu
-          patch -p1 < ../5164bca5a4bcde4534dc1a9aa3a7f619719874cf.patch
+          patch -p1 < ../222729c7455784dd855216d7a2bec4bd8f2a6800.patch
           export CXXFLAGS="-Wno-error"; export CFLAGS="-Wno-error"
           ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=riscv64-linux-user --disable-system
           make -j$(nproc)

CMakeLists.txt

@@ -299,23 +299,44 @@ if (USE_OPENMP)
   endif()
 endif()
 
-# Seems that this hack doesn't required since macOS 11 Big Sur
-if (APPLE AND BUILD_SHARED_LIBS AND CMAKE_HOST_SYSTEM_VERSION VERSION_LESS 20)
-  set (CMAKE_C_USE_RESPONSE_FILE_FOR_OBJECTS 1)
-  if (NOT NOFORTRAN)
-  set (CMAKE_Fortran_USE_RESPONSE_FILE_FOR_OBJECTS 1)
-  set (CMAKE_Fortran_CREATE_SHARED_LIBRARY
- "sh -c 'cat ${CMAKE_BINARY_DIR}/CMakeFiles/openblas_shared.dir/objects*.rsp | xargs -n 1024 ${CMAKE_AR} -ru libopenblas.a && exit 0' "
- "sh -c '${CMAKE_AR} -rs libopenblas.a ${CMAKE_BINARY_DIR}/driver/others/CMakeFiles/driver_others.dir/xerbla.c.o && exit 0' "
- "sh -c 'echo \"\" | ${CMAKE_Fortran_COMPILER} -o dummy.o -c -x f95-cpp-input - '"
- "sh -c '${CMAKE_Fortran_COMPILER} -fpic -shared -Wl,-all_load -Wl,-force_load,libopenblas.a -Wl,-noall_load dummy.o -o ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/libopenblas.${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.dylib'"
- "sh -c 'ls -l ${CMAKE_BINARY_DIR}/lib'")
-  else ()
-  set (CMAKE_C_CREATE_SHARED_LIBRARY
-   "sh -c 'cat ${CMAKE_BINARY_DIR}/CMakeFiles/openblas_shared.dir/objects*.rsp | xargs -n 1024 ${CMAKE_AR} -ru libopenblas.a && exit 0' "
-   "sh -c '${CMAKE_AR} -rs libopenblas.a ${CMAKE_BINARY_DIR}/driver/others/CMakeFiles/driver_others.dir/xerbla.c.o && exit 0' "
-   "sh -c '${CMAKE_C_COMPILER} -fpic -shared -Wl,-all_load -Wl,-force_load,libopenblas.a -Wl,-noall_load -o ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/libopenblas.${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.dylib'")
-  endif ()
+# Fix "Argument list too long" for macOS with Intel CPUs and DYNAMIC_ARCH turned on
+if(APPLE AND DYNAMIC_ARCH AND (NOT CMAKE_HOST_SYSTEM_PROCESSOR STREQUAL "arm64"))
+  # Use response files
+  set(CMAKE_C_USE_RESPONSE_FILE_FOR_OBJECTS 1)
+  # Always build static library first
+  if(BUILD_STATIC_LIBS)
+    set(STATIC_PATH "${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/lib${OpenBLAS_LIBNAME}.a")
+  else()
+    add_library(${OpenBLAS_LIBNAME}_static STATIC ${TARGET_OBJS} ${OpenBLAS_DEF_FILE})
+    set(STATIC_PATH "lib${OpenBLAS_LIBNAME}.a")
+  endif()
+  set(CREATE_STATIC_LIBRARY_COMMAND
+    "sh -c 'cat ${CMAKE_BINARY_DIR}/CMakeFiles/${OpenBLAS_LIBNAME}_static.dir/objects*.rsp | xargs -n 1024 ${CMAKE_AR} -ru ${STATIC_PATH} && exit 0' "
+    "sh -c '${CMAKE_AR} -rs ${STATIC_PATH} ${CMAKE_BINARY_DIR}/driver/others/CMakeFiles/driver_others.dir/xerbla.c.o && exit 0' ")
+  if(BUILD_SHARED_LIBS)
+    add_dependencies(${OpenBLAS_LIBNAME}_shared ${OpenBLAS_LIBNAME}_static)
+    set(SHARED_PATH "${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/libopenblas.${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.dylib")
+  endif()
+  if(USE_OPENMP)
+    get_target_property(OMP_LIB OpenMP::OpenMP_C INTERFACE_LINK_LIBRARIES)
+  else()
+    set(OMP_LIB "")
+  endif()
+  if(NOT NOFORTRAN)
+    set(CMAKE_Fortran_USE_RESPONSE_FILE_FOR_OBJECTS 1)
+    set(CMAKE_Fortran_CREATE_STATIC_LIBRARY ${CREATE_STATIC_LIBRARY_COMMAND})
+    if(BUILD_SHARED_LIBS)
+      set(CMAKE_Fortran_CREATE_SHARED_LIBRARY
+	"sh -c 'echo \"\" | ${CMAKE_Fortran_COMPILER} -o dummy.o -c -x f95-cpp-input - '"
+	"sh -c '${CMAKE_Fortran_COMPILER} -fpic -shared -Wl,-all_load -Wl,-force_load,${STATIC_PATH} dummy.o -o ${SHARED_PATH} ${OMP_LIB}'")
+    endif()
+  else()
+    set(CMAKE_C_CREATE_STATIC_LIBRARY ${CREATE_STATIC_LIBRARY_COMMAND})
+    if(BUILD_SHARED_LIBS)
+      set(CMAKE_C_CREATE_SHARED_LIBRARY
+	"sh -c '${CMAKE_C_COMPILER} -fpic -shared -Wl,-all_load -Wl,-force_load,${STATIC_PATH} -o ${SHARED_PATH} ${OMP_LIB}'")
+    endif()
+  endif()
 endif()
 
 # Handle MSVC exports

CONTRIBUTORS.md

@@ -250,4 +250,7 @@ In chronological order:
 
 * Ye Tao <ye.tao@arm.com>
   * [2025-02-03] Optimize SBGEMM kernel on NEOVERSEV1
-  * [2025-02-27] Add sbgemv_n_neon kernel
+  * [2025-02-27] Add sbgemv_n_neon kernel
+
+* Abhishek Kumar <https://github.com/abhishek-iitmadras>
+  * [2025-04-22] Optimise dot kernel for NEOVERSE V1

cmake/prebuild.cmake

@@ -1006,15 +1006,15 @@ endif ()
       "#define HAVE_SVE\n"
       "#define ARMV8\n")
     set(SGEMM_UNROLL_M 16)
-    set(SGEMM_UNROLL_N 4)
-    set(DGEMM_UNROLL_M 8)
-    set(DGEMM_UNROLL_N 4)
-    set(CGEMM_UNROLL_M 8)
+    set(SGEMM_UNROLL_N 8)
+    set(DGEMM_UNROLL_M 4)
+    set(DGEMM_UNROLL_N 8)
+    set(CGEMM_UNROLL_M 2)
     set(CGEMM_UNROLL_N 4)
-    set(ZGEMM_UNROLL_M 4)
+    set(ZGEMM_UNROLL_M 2)
     set(ZGEMM_UNROLL_N 4)
     set(SYMV_P 16)
-  elseif ("${TCORE}" STREQUAL "NEOVERSEN2" or "${TCORE}" STREQUAL "ARMV9SME")
+  elseif ("${TCORE}" STREQUAL "NEOVERSEN2" OR "${TCORE}" STREQUAL "ARMV9SME")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_CODE_SIZE\t65536\n"
       "#define L1_CODE_LINESIZE\t64\n"
@@ -1249,6 +1249,25 @@ endif ()
     set(ZGEMM_UNROLL_M 2)
     set(ZGEMM_UNROLL_N 4)
     set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "ARMV8SVE" OR "${TCORE}" STREQUAL "CORTEXA510" OR "${TCORE}" STREQUAL "CORTEXX2" OR "${TCORE}" STREQUAL "ARMV9")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_DATA_SIZE\t32768\n"
+      "#define L1_DATA_LINESIZE\t64\n"
+      "#define L2_SIZE\t262144\n"
+      "#define L2_LINESIZE\t64\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define L2_ASSOCIATIVE\t32\n"
+      "#define ARMV8\n")
+    set(SGEMM_UNROLL_M 4)
+    set(SGEMM_UNROLL_N 8)
+    set(DGEMM_UNROLL_M 4)
+    set(DGEMM_UNROLL_N 8)
+    set(CGEMM_UNROLL_M 2)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 2)
+    set(ZGEMM_UNROLL_N 4)
+    set(SYMV_P 16)
   elseif ("${TCORE}" STREQUAL "P5600")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L2_SIZE 1048576\n"
@@ -1409,9 +1428,11 @@ endif ()
   # GetArch_2nd
   foreach(float_char S;D;Q;C;Z;X)
     if (NOT DEFINED ${float_char}GEMM_UNROLL_M)
+	    message(STATUS "setting unrollm=2")
       set(${float_char}GEMM_UNROLL_M 2)
     endif()
     if (NOT DEFINED ${float_char}GEMM_UNROLL_N)
+	    message(STATUS "setting unrolln=2")
       set(${float_char}GEMM_UNROLL_N 2)
     endif()
   endforeach()

cpuid_arm64.c

@@ -374,15 +374,20 @@ int detect(void)
 	}
 #else
 #ifdef __APPLE__
+	length64 = sizeof(value64);
 	sysctlbyname("hw.ncpu",&value64,&length64,NULL,0);
 	cpulowperf=value64;
+	length64 = sizeof(value64);
 	sysctlbyname("hw.nperflevels",&value64,&length64,NULL,0);
 	if (value64 > 1) {
-	sysctlbyname("hw.perflevel0.cpusperl",&value64,&length64,NULL,0);
+	length64 = sizeof(value64);
+	sysctlbyname("hw.perflevel0.cpusperl2",&value64,&length64,NULL,0);
 	cpuhiperf=value64;
-	sysctlbyname("hw.perflevel1.cpusperl",&value64,&length64,NULL,0);
+	length64 = sizeof(value64);
+	sysctlbyname("hw.perflevel1.cpusperl2",&value64,&length64,NULL,0);
 	cpulowperf=value64;
 	}
+	length64 = sizeof(value64);
 	sysctlbyname("hw.cpufamily",&value64,&length64,NULL,0);
 	if (value64 ==131287967|| value64 == 458787763 ) return CPU_VORTEX; //A12/M1
 	if (value64 == 3660830781) return CPU_VORTEX; //A15/M2
@@ -467,6 +472,7 @@ int n=0;
 	printf("#define NUM_CORES_HP %d\n",cpuhiperf);
 #endif
 #ifdef __APPLE__
+	length64 = sizeof(value64);
 	sysctlbyname("hw.physicalcpu_max",&value,&length,NULL,0);
 	printf("#define NUM_CORES %d\n",value);
 	if (cpulowperf >0)
@@ -698,12 +704,17 @@ void get_cpuconfig(void)
 	    case CPU_VORTEX:
 		printf("#define VORTEX			      \n");
 #ifdef __APPLE__
+		length64 = sizeof(value64);
 		sysctlbyname("hw.l1icachesize",&value64,&length64,NULL,0);
 		printf("#define L1_CODE_SIZE	     %lld       \n",value64);
+		length64 = sizeof(value64);
 		sysctlbyname("hw.cachelinesize",&value64,&length64,NULL,0);
 		printf("#define L1_CODE_LINESIZE     %lld       \n",value64);
+		printf("#define L1_DATA_LINESIZE     %lld       \n",value64);
+		length64 = sizeof(value64);
 		sysctlbyname("hw.l1dcachesize",&value64,&length64,NULL,0);
 		printf("#define L1_DATA_SIZE	     %lld       \n",value64);
+		length64 = sizeof(value64);
 		sysctlbyname("hw.l2cachesize",&value64,&length64,NULL,0);
 		printf("#define L2_SIZE	     %lld       \n",value64);
 #endif	

cpuid_x86.c

@@ -1578,6 +1578,7 @@ int get_cpuname(void){
       case 12: //family 6 exmodel 12
 	switch (model) {
 	  case 15:
+	  case 6: // Arrow Lake
 	    if(support_avx512())
               return CPUTYPE_SAPPHIRERAPIDS;
             if(support_avx2())
@@ -2421,6 +2422,22 @@ int get_coretype(void){
 	  else
 	  return CORE_NEHALEM;
 	}
+      case 12:
+	switch (model) {
+	  case 6: // Arrow Lake
+	  if(support_amx_bf16())
+	    return CORE_SAPPHIRERAPIDS;
+	  if(support_avx512_bf16())
+            return CORE_COOPERLAKE;
+          if(support_avx512())
+            return CORE_SKYLAKEX;
+          if(support_avx2())
+            return CORE_HASWELL;
+          if(support_avx())
+	    return CORE_SANDYBRIDGE;
+	  else
+	    return CORE_NEHALEM;
+	}
       }
     case 15:
       if (model <= 0x2) return CORE_NORTHWOOD;

kernel/arm64/KERNEL.A64FX

@@ -1,6 +1,6 @@
 include $(KERNELDIR)/KERNEL.ARMV8SVE
 
-SGEMVNKERNEL = gemv_n_sve.c
-DGEMVNKERNEL = gemv_n_sve.c
+SGEMVNKERNEL = gemv_n_sve_v4x3.c
+DGEMVNKERNEL = gemv_n_sve_v4x3.c
 SGEMVTKERNEL = gemv_t_sve_v4x3.c
 DGEMVTKERNEL = gemv_t_sve_v4x3.c

kernel/arm64/KERNEL.ARMV8SVE

@@ -74,8 +74,8 @@ DSCALKERNEL  = scal.S
 CSCALKERNEL  = zscal.S
 ZSCALKERNEL  = zscal.S
 
-SGEMVNKERNEL = gemv_n_sve.c
-DGEMVNKERNEL = gemv_n.S
+SGEMVNKERNEL = gemv_n_sve_v1x3.c
+DGEMVNKERNEL = gemv_n_sve_v1x3.c
 CGEMVNKERNEL = zgemv_n.S
 ZGEMVNKERNEL = zgemv_n.S
 
@@ -84,6 +84,11 @@ DGEMVTKERNEL = gemv_t_sve_v1x3.c
 CGEMVTKERNEL = zgemv_t.S
 ZGEMVTKERNEL = zgemv_t.S
 
+SSYMV_L_KERNEL = symv_L_sve_v1x4.c
+SSYMV_U_KERNEL = symv_U_sve_v1x4.c
+DSYMV_L_KERNEL = symv_L_sve_v1x4.c
+DSYMV_U_KERNEL = symv_U_sve_v1x4.c
+
 SASUMKERNEL    = sasum_thunderx2t99.c
 DASUMKERNEL    = dasum_thunderx2t99.c
 CASUMKERNEL    = casum_thunderx2t99.c

kernel/arm64/KERNEL.NEOVERSEN1

@@ -60,7 +60,7 @@ DSCALKERNEL  = scal.S
 CSCALKERNEL  = zscal.S
 ZSCALKERNEL  = zscal.S
 
-SGEMVNKERNEL = gemv_n.S
+SGEMVNKERNEL = sgemv_n_neon.c
 DGEMVNKERNEL = gemv_n.S
 CGEMVNKERNEL = zgemv_n.S
 ZGEMVNKERNEL = zgemv_n.S
@@ -70,6 +70,10 @@ DGEMVTKERNEL = gemv_t.S
 CGEMVTKERNEL = zgemv_t.S
 ZGEMVTKERNEL = zgemv_t.S
 
+SSYMV_L_KERNEL = symv_L_asimd_4x4.c
+SSYMV_U_KERNEL = symv_U_asimd_4x4.c
+DSYMV_L_KERNEL = symv_L_asimd_4x4.c
+DSYMV_U_KERNEL = symv_U_asimd_4x4.c
 
 SASUMKERNEL    = sasum_thunderx2t99.c
 DASUMKERNEL    = dasum_thunderx2t99.c

kernel/arm64/KERNEL.NEOVERSEN2

@@ -60,8 +60,8 @@ DSCALKERNEL  = scal.S
 CSCALKERNEL  = zscal.S
 ZSCALKERNEL  = zscal.S
 
-SGEMVNKERNEL = gemv_n.S
-DGEMVNKERNEL = gemv_n.S
+SGEMVNKERNEL = gemv_n_sve_v1x3.c
+DGEMVNKERNEL = gemv_n_sve_v1x3.c
 CGEMVNKERNEL = zgemv_n.S
 ZGEMVNKERNEL = zgemv_n.S