Merge branch 'OpenMathLib:develop' into gemmt_tests

Author: martin-frbg

.github/workflows/c910v.yml

@@ -83,9 +83,39 @@ jobs:
 
       - name: test
         run: |
-          export PATH=$GITHUB_WORKSPACE/qemu-install/bin/:$PATH
-          qemu-riscv64 ./utest/openblas_utest
-          qemu-riscv64 ./utest/openblas_utest_ext
+          run_with_retry() {
+              local cmd="$1"
+              local time_out=10
+              local retries=10
+              local attempt=0
+
+              for ((i=1; i<=retries; i++)); do
+                  attempt=$((i))
+                  if timeout -s 12  --preserve-status $time_out $cmd; then
+                      echo "Command succeeded on attempt $i."
+                      return 0
+                  else
+                      local exit_code=$?
+                      if [ $exit_code -eq 140 ]; then
+                          echo "Attempt $i timed out (retrying...)"
+                          time_out=$((time_out + 5))
+                      else
+                          echo "Attempt $i failed with exit code $exit_code. Aborting workflow."
+                          exit $exit_code
+                      fi
+                  fi
+              done
+              echo "All $retries attempts failed, giving up."
+              echo "Final failure was due to timeout."
+              echo "Aborting workflow."
+              exit $exit_code
+          }
+          export PATH=$GITHUB_WORKSPACE/qemu-install/bin:$PATH
+          which qemu-riscv64
+          export QEMU_BIN=$(which qemu-riscv64)
+          run_with_retry "$QEMU_BIN ./utest/openblas_utest"
+          run_with_retry "$QEMU_BIN ./utest/openblas_utest_ext"
+
           OPENBLAS_NUM_THREADS=2 qemu-riscv64 ./ctest/xscblat1
           OPENBLAS_NUM_THREADS=2 qemu-riscv64 ./ctest/xdcblat1
           OPENBLAS_NUM_THREADS=2 qemu-riscv64 ./ctest/xccblat1

CMakeLists.txt

@@ -62,6 +62,7 @@ option(CPP_THREAD_SAFETY_TEST "Run a massively parallel DGEMM test to confirm th
 
 option(CPP_THREAD_SAFETY_GEMV "Run a massively parallel DGEMV test to confirm thread safety of the library (requires OpenMP)" OFF)
 option(BUILD_STATIC_LIBS "Build static library" OFF)
+option(BUILD_SHARED_LIBS "Build shared library" OFF)
 if(NOT BUILD_STATIC_LIBS AND NOT BUILD_SHARED_LIBS)
   set(BUILD_STATIC_LIBS ON CACHE BOOL "Build static library" FORCE)
 endif()
@@ -123,7 +124,12 @@ message(WARNING "CMake support is experimental. It does not yet support all buil
 include("${PROJECT_SOURCE_DIR}/cmake/utils.cmake")
 include("${PROJECT_SOURCE_DIR}/cmake/system.cmake")
 
-set(OpenBLAS_LIBNAME ${LIBNAMEPREFIX}openblas${LIBNAMESUFFIX}${SUFFIX64_UNDERSCORE})
+string(FIND "${LIBNAMESUFFIX}" "${SUFFIX64_UNDERSCORE}" HAVE64)
+if (${HAVE64} GREATER -1)
+	set(OpenBLAS_LIBNAME ${LIBNAMEPREFIX}openblas${LIBNAMESUFFIX})
+else ()
+	set(OpenBLAS_LIBNAME ${LIBNAMEPREFIX}openblas${LIBNAMESUFFIX}${SUFFIX64_UNDERSCORE})
+endif ()
 
 set(BLASDIRS interface driver/level2 driver/level3 driver/others)
 
@@ -716,4 +722,5 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${PN}ConfigVersion.cmake
         DESTINATION ${CMAKECONFIG_INSTALL_DIR})
 install(EXPORT "${PN}${SUFFIX64}Targets"
         NAMESPACE "${PN}${SUFFIX64}::"
-        DESTINATION ${CMAKECONFIG_INSTALL_DIR})
+        DESTINATION ${CMAKECONFIG_INSTALL_DIR})
+

Makefile.system

@@ -276,6 +276,7 @@ SMALL_MATRIX_OPT = 1
 endif
 ifeq ($(ARCH), arm64)
 GEMM_GEMV_FORWARD = 1
+GEMM_GEMV_FORWARD_BF16 = 1
 endif
 ifeq ($(ARCH), riscv)
 GEMM_GEMV_FORWARD = 1

cmake/cc.cmake

@@ -229,9 +229,9 @@ if (${CORE} STREQUAL NEOVERSEN1)
     if (${CMAKE_C_COMPILER_ID} STREQUAL "NVC" AND NOT NO_SVE)
 	set (CCOMMON_OPT  "${CCOMMON_OPT} -tp=neoverse-n1")
     elseif (${GCC_VERSION} VERSION_GREATER 9.4 OR ${GCC_VERSION} VERSION_EQUAL 9.4)
-      set (CCOMMON_OPT  "${CCOMMON_OPT} -march=armv8.2-a+sve -mtune=neoverse-n1")
+      set (CCOMMON_OPT  "${CCOMMON_OPT} -march=armv8.2-a -mtune=neoverse-n1")
     else ()
-      set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.2-a+sve")
+      set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.2-a")
     endif()
   endif ()
 endif ()
@@ -260,13 +260,13 @@ endif ()
 
 if (${CORE} STREQUAL CORTEXA510)
   if (NOT DYNAMIC_ARCH)
-    set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8-a+sve")
+    set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.4-a+sve")
   endif ()
 endif ()
 
 if (${CORE} STREQUAL CORTEXA710)
   if (NOT DYNAMIC_ARCH)
-    set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8-a+sve")
+    set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.4-a+sve")
   endif ()
 endif ()
 
@@ -278,7 +278,7 @@ endif ()
 
 if (${CORE} STREQUAL CORTEXX2)
   if (NOT DYNAMIC_ARCH)
-    set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8-a+sve")
+    set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.4-a+sve")
   endif ()
 endif ()
 

ctest/CMakeLists.txt

@@ -6,7 +6,7 @@ enable_language(Fortran)
 endif()
 
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DADD${BU} -DCBLAS")
-if (BINARY32 AND CMAKE_C_PLATFORM_ID MATCHES "MinGW" AND CMAKE_Fortran_COMPILER_VERSION VERSION_EQUAL 14.2)
+if (BINARY32 AND CMAKE_C_PLATFORM_ID MATCHES "MinGW" AND CMAKE_Fortran_COMPILER_VERSION VERSION_GREATER 14.1)
        list(REMOVE_ITEM ${CMAKE_Fortran_FLAGS} -O3 -O2 -O1 -Os)
        set (CMAKE_Fortran_FLAGS_RELEASE "" CACHE STRING "" FORCE)
 endif()

driver/level3/level3_thread.c

@@ -851,9 +851,20 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, IFLOAT *sa, IF
     /* Objective function come from sum of partitions in m and n.             */
     /* (n / nthreads_n) + (m / nthreads_m)                                    */
     /* = (n * nthreads_m + m * nthreads_n) / (nthreads_n * nthreads_m)        */
-    while (nthreads_m % 2 == 0 && n * nthreads_m + m * nthreads_n > n * (nthreads_m / 2) + m * (nthreads_n * 2)) {
-      nthreads_m /= 2;
-      nthreads_n *= 2;
+    BLASLONG cost = 0, div = 0;
+    BLASLONG i;
+    for (i = 1; i <= sqrt(nthreads_m); i++) {
+      if (nthreads_m % i) continue;
+      BLASLONG j = nthreads_m / i;
+      BLASLONG cost_i = n * j + m * nthreads_n * i;
+      BLASLONG cost_j = n * i + m * nthreads_n * j;
+      if (cost == 0 ||
+          cost_i < cost) {cost = cost_i; div = i;}
+      if (cost_j < cost) {cost = cost_j; div = j;}
+    }
+    if (div > 1) {
+      nthreads_m /= div;
+      nthreads_n *= div;
     }
   }
 

interface/gemm.c

@@ -417,21 +417,24 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
 
   PRINT_DEBUG_CNAME;
 
-#if !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16) && defined(USE_SGEMM_KERNEL_DIRECT)
-#if defined(DYNAMIC_ARCH) && defined(ARCH_x86)
- if (support_avx512() )
+#if !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16) 
+#if defined(ARCH_x86) && (defined(USE_SGEMM_KERNEL_DIRECT)||defined(DYNAMIC_ARCH))
+#if defined(DYNAMIC_ARCH)
+  if (support_avx512() )
+#endif
   if (beta == 0 && alpha == 1.0 && order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans && SGEMM_DIRECT_PERFORMANT(m,n,k)) {
 	SGEMM_DIRECT(m, n, k, a, lda, b, ldb, c, ldc);
 	return;
   }
 #endif
-#if defined(DYNAMIC_ARCH) && defined(ARCH_ARM64)
- if (support_sme1()){
+#if defined(ARCH_ARM64) && (defined(USE_SGEMM_KERNEL_DIRECT)||defined(DYNAMIC_ARCH))
+#if defined(DYNAMIC_ARCH)
+ if (support_sme1())
+#endif
   if (beta == 0 && alpha == 1.0 && order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans) {
 	SGEMM_DIRECT(m, n, k, a, lda, b, ldb, c, ldc);
 	return;
   }
- }
 #endif
 #endif
 

interface/zscal.c

@@ -98,7 +98,7 @@ void CNAME(blasint n, FLOAT alpha_r, void *vx, blasint incx){
   if (nthreads == 1) {
 #endif
 
-  SCAL_K(n, 0, 0, alpha[0], alpha[1], x, incx, NULL, 0, NULL, 0);
+  SCAL_K(n, 0, 0, alpha[0], alpha[1], x, incx, NULL, 0, NULL, 1);
 
 #ifdef SMP
   } else {
@@ -108,7 +108,7 @@ void CNAME(blasint n, FLOAT alpha_r, void *vx, blasint incx){
     mode  =  BLAS_SINGLE | BLAS_COMPLEX;
 #endif
 
-    blas_level1_thread(mode, n, 0, 0,  alpha, x, incx, NULL, 0, NULL, 0, (int (*)(void))SCAL_K, nthreads);
+    blas_level1_thread(mode, n, 0, 0,  alpha, x, incx, NULL, 0, NULL, 1, (int (*)(void))SCAL_K, nthreads);
 
   }
 #endif

interface/zsyr.c

@@ -116,12 +116,12 @@ void NAME(char *UPLO, blasint *N, FLOAT  *ALPHA,
 
 #else
 
-void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, int n, FLOAT alpha, FLOAT *x, int incx, FLOAT *a, int lda) {
+void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, int n, void* valpha, FLOAT *x, int incx, FLOAT *a, int lda) {
 
   FLOAT *buffer;
   int uplo;
   blasint info;
-  FLOAT * ALPHA = α
+  FLOAT * ALPHA = (FLOAT*)valpha;
   FLOAT alpha_r	= ALPHA[0];
   FLOAT alpha_i	= ALPHA[1];
 #ifdef SMP

kernel/arm/zscal.c

@@ -27,65 +27,56 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 /**************************************************************************************
 * 2013/09/14 Saar
-*	 BLASTEST float		: OK
-* 	 BLASTEST double	: OK
-* 	 CTEST			: OK
-* 	 TEST			: OK
+*    BLASTEST float     : OK
+*    BLASTEST double    : OK
+*    CTEST          : OK
+*    TEST           : OK
 *
 **************************************************************************************/
 
 #include "common.h"
 
+// The c/zscal_k function is called not only by cblas_c/zscal but also by other upper-level interfaces.
+// In certain cases, the expected return values for cblas_s/zscal differ from those of other upper-level interfaces.
+// To handle this, we use the dummy2 parameter to differentiate between them.
 int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r,FLOAT da_i, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
 {
-	BLASLONG i=0;
-	BLASLONG inc_x2;
-	BLASLONG ip = 0;
-	FLOAT temp;
+    BLASLONG i = 0;
+    BLASLONG inc_x2;
+    BLASLONG ip = 0;
+    FLOAT temp;
 
-        if ( (n <= 0) || (inc_x <= 0))
-                return(0);
+    if ((n <= 0) || (inc_x <= 0))
+        return(0);
 
+    inc_x2 = 2 * inc_x;
+    if (dummy2 == 0) {
+        for (i = 0; i < n; i++)
+        {
+            if (da_r == 0.0 && da_i == 0.0)
+            {
+                x[ip] = 0.0;
+                x[ip+1] = 0.0;
+            }
+            else
+            {
+                temp    = da_r * x[ip]   - da_i * x[ip+1];
+                x[ip+1] = da_r * x[ip+1] + da_i * x[ip]  ;
+                x[ip] = temp;
+            }
 
-	inc_x2 = 2 * inc_x;
-	for ( i=0; i<n; i++ )
-	{
-		if ( da_r == 0.0 )
-		{
-			if ( da_i == 0.0 )
-			{
-				temp = 0.0;
-				x[ip+1] = 0.0 ;
-			}
-			else
-			{
-				temp = - da_i * x[ip+1] ;
-				if (isnan(x[ip]) || isinf(x[ip])) temp = NAN;
-				if (!isinf(x[ip+1])) 
-					x[ip+1] = da_i * x[ip]  ;
-				else 	x[ip+1] = NAN;
-			}
-		}
-		else
-		{
-			if ( da_i == 0.0 )
-			{
-				temp    = da_r * x[ip]  ;
-				x[ip+1] = da_r * x[ip+1];
-			}
-			else
-			{
-				temp    = da_r * x[ip]   - da_i * x[ip+1] ;
-				x[ip+1] = da_r * x[ip+1] + da_i * x[ip]   ;
-			}
-		}
-		x[ip]   = temp;
+            ip += inc_x2;
+        }
+        return(0);
+    }
+    for (i = 0; i < n; i++)
+    {
+        temp    = da_r * x[ip]   - da_i * x[ip+1];
+        x[ip+1] = da_r * x[ip+1] + da_i * x[ip]  ;
 
-		ip += inc_x2;
-	}
-
-	return(0);
+        x[ip]   = temp;
+        ip += inc_x2;
+    }
 
+    return(0);
 }
-
-