Merge pull request #4524 from martin-frbg/test3m

Enable GEMM3M tests when the architecture supports GEMM3M

.github/workflows/dynamic_arch.yml

@@ -42,6 +42,7 @@ jobs:
       - name: Install Dependencies
         run: |
           if [ "$RUNNER_OS" == "Linux" ]; then
+            sudo apt-get update
             sudo apt-get install -y gfortran cmake ccache libtinfo5
           elif [ "$RUNNER_OS" == "macOS" ]; then
             # It looks like "gfortran" isn't working correctly unless "gcc" is re-installed.

ctest/CMakeLists.txt

@@ -88,6 +88,17 @@ if (NOT NOFORTRAN)
     auxiliary.c
     c_xerbla.c
     constant.c)
+  if (USE_GEMM3M)
+  if ((${float_char} STREQUAL "c") OR (${float_char} STREQUAL "z"))
+  add_executable(x${float_char}cblat3_3m
+    c_${float_char}blat3_3m.f
+    c_${float_char}blas3_3m.c
+    c_${float_char}3chke_3m.c
+    auxiliary.c
+    c_xerbla.c
+    constant.c)
+  endif()
+  endif()
 else()
   add_executable(x${float_char}cblat3
     c_${float_char}blat3c.c
@@ -96,6 +107,17 @@ else()
     auxiliary.c
     c_xerbla.c
     constant.c)
+  if (USE_GEMM3M)
+  if ((${float_char} STREQUAL "c") OR (${float_char} STREQUAL "z"))
+  add_executable(x${float_char}cblat3_3m
+    c_${float_char}blat3c_3m.c
+    c_${float_char}blas3_3m.c
+    c_${float_char}3chke_3m.c
+    auxiliary.c
+    c_xerbla.c
+    constant.c)
+  endif()
+  endif()
 endif()
   target_link_libraries(x${float_char}cblat3 ${OpenBLAS_LIBNAME})
   if (USE_OPENMP AND (${CMAKE_Fortran_COMPILER_ID} STREQUAL GNU) AND (${CMAKE_C_COMPILER_ID} STREQUAL Clang))
@@ -105,7 +127,24 @@ endif()
   if(${CMAKE_SYSTEM_NAME} MATCHES "Linux" OR ${CMAKE_SYSTEM_NAME} MATCHES "FreeBSD" OR ${CMAKE_SYSTEM_NAME} MATCHES "QNX")
     target_link_libraries(x${float_char}cblat3 m)
   endif()
+  if (USE_GEMM3M)
+  if ((${float_char} STREQUAL "c") OR (${float_char} STREQUAL "z"))
+  target_link_libraries(x${float_char}cblat3_3m ${OpenBLAS_LIBNAME})
+  if (USE_OPENMP AND (${CMAKE_Fortran_COMPILER_ID} STREQUAL GNU) AND (${CMAKE_C_COMPILER_ID} STREQUAL Clang))
+        string(REGEX REPLACE "-fopenmp" "" CMAKE_Fortran_FLAGS "${CMAKE_Fortran_FLAGS}")
+        target_link_libraries(x${float_char}cblat3 omp pthread)
+  endif()
+  if(${CMAKE_SYSTEM_NAME} MATCHES "Linux" OR ${CMAKE_SYSTEM_NAME} MATCHES "FreeBSD" OR ${CMAKE_SYSTEM_NAME} MATCHES "QNX")
+    target_link_libraries(x${float_char}cblat3_3m m)
+  endif()
+  endif()
+  endif()
   add_test(NAME "x${float_char}cblat3"
     COMMAND ${test_helper} $<TARGET_FILE:x${float_char}cblat3> "${PROJECT_SOURCE_DIR}/ctest/${float_char}in3")
-
+    if (USE_GEMM3M)
+    if ((${float_char} STREQUAL "c") OR (${float_char} STREQUAL "z"))
+  add_test(NAME "x${float_char}cblat3_3m"
+    COMMAND ${test_helper} $<TARGET_FILE:x${float_char}cblat3_3m> "${PROJECT_SOURCE_DIR}/ctest/${float_char}in3_3m")
+    endif()
+  endif()
 endforeach()

ctest/Makefile

@@ -5,6 +5,24 @@
 TOPDIR = ..
 include $(TOPDIR)/Makefile.system
 
+SUPPORT_GEMM3M = 0
+
+ifeq ($(ARCH), x86)
+SUPPORT_GEMM3M = 1
+endif
+
+ifeq ($(ARCH), x86_64)
+SUPPORT_GEMM3M = 1
+endif
+
+ifeq ($(ARCH), ia64)
+SUPPORT_GEMM3M = 1
+endif
+
+ifeq ($(ARCH), MIPS)
+SUPPORT_GEMM3M = 1
+endif
+
 override CFLAGS += -DADD$(BU) -DCBLAS
 ifeq ($(F_COMPILER),GFORTRAN)
 	override FFLAGS += -fno-tree-vectorize
@@ -144,9 +162,15 @@ all3targets += xdcblat3
 endif
 ifeq ($(BUILD_COMPLEX),1)
 all3targets += xccblat3
+ifeq ($(SUPPORT_GEMM3M),1)
+all3targets += xccblat3_3m
+endif
 endif
 ifeq ($(BUILD_COMPLEX16),1)
 all3targets += xzcblat3
+ifeq ($(SUPPORT_GEMM3M),1)
+all3targets += xzcblat3_3m
+endif
 endif
 
 all3: $(all3targets)
@@ -181,9 +205,9 @@ endif
 endif
 endif
 
-all3_3m: xzcblat3_3m xccblat3_3m
+ifeq ($(SUPPORT_GEMM3M),1)
 ifeq ($(USE_OPENMP), 1)
-ifeq ($(BUILD_SINGLE),1)
+ifeq ($(BUILD_COMPLEX),1)
 	OMP_NUM_THREADS=2 ./xccblat3_3m < cin3_3m
 endif
 ifeq ($(BUILD_COMPLEX16),1)
@@ -197,6 +221,7 @@ ifeq ($(BUILD_COMPLEX16),1)
 	OPENBLAS_NUM_THREADS=2 ./xzcblat3_3m < zin3_3m
 endif
 endif
+endif
 
 
 
@@ -271,8 +296,10 @@ xccblat2: $(ctestl2o) c_cblat2.o $(TOPDIR)/$(LIBNAME)
 	$(FC) $(FLDFLAGS) -o xccblat2 c_cblat2.o $(ctestl2o) $(LIB) $(EXTRALIB) $(CEXTRALIB)
 xccblat3: $(ctestl3o) c_cblat3.o $(TOPDIR)/$(LIBNAME)
 	$(FC) $(FLDFLAGS) -o xccblat3 c_cblat3.o $(ctestl3o) $(LIB) $(EXTRALIB) $(CEXTRALIB)
+ifeq ($(SUPPORT_GEMM3M),1)
 xccblat3_3m: $(ctestl3o_3m) c_cblat3_3m.o $(TOPDIR)/$(LIBNAME)
 	$(FC) $(FLDFLAGS) -o xccblat3_3m c_cblat3_3m.o $(ctestl3o_3m) $(LIB) $(EXTRALIB) $(CEXTRALIB)
+endif
 else
 xccblat1: $(ctestl1o) c_cblat1c.o $(TOPDIR)/$(LIBNAME)
 	$(CC) $(CFLAGS) -o xccblat1 c_cblat1c.o $(ctestl1o) $(LIB) $(CEXTRALIB) $(filter-out -lgfortran,$(EXTRALIB)) 
@@ -280,6 +307,10 @@ xccblat2: $(ctestl2o) c_cblat2c.o $(TOPDIR)/$(LIBNAME)
 	$(CC) $(CFLAGS) -o xccblat2 c_cblat2c.o $(ctestl2o) $(LIB) $(CEXTRALIB) $(filter-out -lgfortran,$(EXTRALIB))
 xccblat3: $(ctestl3o) c_cblat3c.o $(TOPDIR)/$(LIBNAME)
 	$(CC) $(CFLAGS) -o xccblat3 c_cblat3c.o $(ctestl3o) $(LIB) $(CEXTRALIB) $(filter-out -lgfortran,$(EXTRALIB))
+ifeq ($(SUPPORT_GEMM3M),1)
+xccblat3_3m: $(ctestl3o_3m) c_cblat3c_3m.o $(TOPDIR)/$(LIBNAME)
+	$(CC) $(CFLAGS) -o xccblat3_3m c_cblat3c_3m.o $(ctestl3o_3m) $(LIB) $(EXTRALIB) $(CEXTRALIB)
+endif
 endif
 endif
 
@@ -293,8 +324,10 @@ xzcblat2: $(ztestl2o) c_zblat2.o $(TOPDIR)/$(LIBNAME)
 	$(FC) $(FLDFLAGS) -o xzcblat2 c_zblat2.o $(ztestl2o) $(LIB) $(EXTRALIB) $(CEXTRALIB)
 xzcblat3: $(ztestl3o) c_zblat3.o $(TOPDIR)/$(LIBNAME)
 	$(FC) $(FLDFLAGS) -o xzcblat3 c_zblat3.o $(ztestl3o) $(LIB) $(EXTRALIB) $(CEXTRALIB)
+ifeq ($(SUPPORT_GEMM3M),1)
 xzcblat3_3m: $(ztestl3o_3m) c_zblat3_3m.o $(TOPDIR)/$(LIBNAME)
 	$(FC) $(FLDFLAGS) -o xzcblat3_3m c_zblat3_3m.o $(ztestl3o_3m) $(LIB) $(EXTRALIB) $(CEXTRALIB)
+endif
 else
 xzcblat1: $(ztestl1o) c_zblat1c.o $(TOPDIR)/$(LIBNAME)
 	$(CC) $(CFLAGS) -o xzcblat1 c_zblat1c.o $(ztestl1o) $(LIB) $(CEXTRALIB) $(filter-out -lgfortran,$(EXTRALIB)) 
@@ -302,6 +335,10 @@ xzcblat2: $(ztestl2o) c_zblat2c.o $(TOPDIR)/$(LIBNAME)
 	$(CC) $(CFLAGS) -o xzcblat2 c_zblat2c.o $(ztestl2o) $(LIB) $(CEXTRALIB) $(filter-out -lgfortran,$(EXTRALIB))
 xzcblat3: $(ztestl3o) c_zblat3c.o $(TOPDIR)/$(LIBNAME)
 	$(CC) $(CFLAGS) -o xzcblat3 c_zblat3c.o $(ztestl3o) $(LIB) $(CEXTRALIB) $(filter-out -lgfortran,$(EXTRALIB))
+ifeq ($(SUPPORT_GEMM3M),1)
+xzcblat3_3m: $(ztestl3o_3m) c_zblat3c_3m.o $(TOPDIR)/$(LIBNAME)
+	$(CC) $(CFLAGS) -o xzcblat3_3m c_zblat3c_3m.o $(ztestl3o_3m) $(LIB) $(EXTRALIB) $(CEXTRALIB)
+endif
 endif
 endif
 

test/CMakeLists.txt

@@ -21,7 +21,16 @@ endif()
 if (BUILD_COMPLEX16)
 	list (APPEND OpenBLAS_Tests zblat1 zblat2 zblat3)
 endif()
-message (STATUS CCOMP ${CMAKE_C_COMPILER_ID} FCOMP ${CMAKE_Fortran_COMPILER_ID})
+
+if (USE_GEMM3M)
+	if (BUILD_COMPLEX)
+		list (APPEND OpenBLAS_Tests cblat3_3m)
+	endif ()
+	if (BUILD_COMPLEX16)
+		list (APPEND OpenBLAS_Tests zblat3_3m)
+	endif ()
+endif ()
+
 foreach(test_bin ${OpenBLAS_Tests})
 	add_executable(${test_bin} ${test_bin}.f)
 target_link_libraries(${test_bin} ${OpenBLAS_LIBNAME})
@@ -82,4 +91,10 @@ add_test(NAME "${float_type}blas2"
   COMMAND ${helper_prefix} $<TARGET_FILE:${float_type}blat2> "${PROJECT_SOURCE_DIR}/test/${float_type}blat2.dat" ${float_type_upper}BLAT2.SUMM)
 add_test(NAME "${float_type}blas3" 
   COMMAND ${helper_prefix} $<TARGET_FILE:${float_type}blat3> "${PROJECT_SOURCE_DIR}/test/${float_type}blat3.dat" ${float_type_upper}BLAT3.SUMM)
+if (USE_GEMM3M)
+if ((${float_type} STREQUAL "c") OR (${float_type} STREQUAL "z"))
+add_test(NAME "${float_type}blas3_3m" 
+  COMMAND ${helper_prefix} $<TARGET_FILE:${float_type}blat3_3m> "${PROJECT_SOURCE_DIR}/test/${float_type}blat3_3m.dat" ${float_type_upper}BLAT3_3M.SUMM)
+endif()
+endif()
 endforeach()

test/Makefile

@@ -4,6 +4,24 @@ ifeq ($(F_COMPILER),GFORTRAN)
         override FFLAGS += -fno-tree-vectorize
 endif
 
+SUPPORT_GEMM3M = 0
+
+ifeq ($(ARCH), x86)
+SUPPORT_GEMM3M = 1
+endif
+
+ifeq ($(ARCH), x86_64)
+SUPPORT_GEMM3M = 1
+endif
+
+ifeq ($(ARCH), ia64)
+SUPPORT_GEMM3M = 1
+endif
+
+ifeq ($(ARCH), MIPS)
+SUPPORT_GEMM3M = 1
+endif
+
 ifeq ($(NOFORTRAN),1)
 all ::
 else
@@ -153,11 +171,20 @@ ifeq ($(BUILD_DOUBLE),1)
 D3=dblat3
 endif
 ifeq ($(BUILD_COMPLEX),1)
+ifeq ($(SUPPORT_GEMM3M),1)
+C3=cblat3 cblat3_3m
+else
 C3=cblat3
 endif
+endif
 ifeq ($(BUILD_COMPLEX16),1)
+ifeq ($(SUPPORT_GEMM3M),1)
+Z3=zblat3 zblat3_3m
+else
 Z3=zblat3
 endif
+endif
+
 
 level3: $(B3) $(S3) $(D3) $(C3) $(Z3)
 

utest/test_extensions/test_cgemv_t.c

@@ -126,7 +126,7 @@ static float check_cgemv(char api, char order, char trans, blasint m, blasint n,
     srand_generate(data_cgemv_t.y_test, m * inc_y * 2);
 
     // Copy vector y for reference funcs
-    for (int i = 0; i < m * inc_y * 2; i++) {
+    for (i = 0; i < m * inc_y * 2; i++) {
         data_cgemv_t.y_verify[i] = data_cgemv_t.y_test[i];
     }
 
@@ -1129,4 +1129,4 @@ CTEST(cgemv, c_api_xerbla_invalid_order_col_major)
     int passed = c_api_check_badargs(corder, ctrans, m, n, lda, inc_x, inc_y, expected_info);
     ASSERT_EQUAL(TRUE, passed);
 }
-#endif
+#endif

utest/test_extensions/test_csbmv.c

@@ -188,7 +188,7 @@ static float check_csbmv(char uplo, blasint n, blasint k, float *alpha, blasint
     char trans = 'N';
 
     // Symmetric band packed matrix for sbmv
-    float a[lda * n * 2];
+    float *a = (float*) malloc(lda * n * 2 * sizeof(float));
 
     // Fill symmetric packed matrix sp_matrix, vector b_test, vector c_test 
     srand_generate(data_csbmv.sp_matrix, n * (n + 1));
@@ -216,7 +216,7 @@ static float check_csbmv(char uplo, blasint n, blasint k, float *alpha, blasint
     // Find the differences between output vector caculated by csbmv and cgemv
     for (i = 0; i < n * inc_c * 2; i++)
         data_csbmv.c_test[i] -= data_csbmv.c_verify[i];
-
+    free(a);
     // Find the norm of differences
     return BLASFUNC(scnrm2)(&n, data_csbmv.c_test, &inc_c);
 }
@@ -603,4 +603,4 @@ CTEST(csbmv, xerbla_lda_invalid)
     int passed = check_badargs(uplo, n, k, lda, inc_b, inc_c, expected_info);
     ASSERT_EQUAL(TRUE, passed);
 }
-#endif
+#endif

utest/test_extensions/test_idamin.c

@@ -402,13 +402,14 @@ CTEST(idamin, min_idx_in_vec_tail){
 CTEST(idamin, min_idx_in_vec_tail_inc_1){
    blasint i;
    blasint N = ELEMENTS, inc = 1;
-   double x[ELEMENTS * inc];
+   double *x = (double*)malloc(ELEMENTS * inc * sizeof(double));
    for (i = 0; i < N * inc; i ++) {
       x[i] = i + 1000;
    }
 
    x[(N - 1) * inc] = 0.0f;
    blasint index = BLASFUNC(idamin)(&N, x, &inc);
+   free(x);
    ASSERT_EQUAL(N, index);
 }
 
@@ -775,13 +776,14 @@ CTEST(idamin, c_api_min_idx_in_vec_tail){
 CTEST(idamin, c_api_min_idx_in_vec_tail_inc_1){
    blasint i;
    blasint N = ELEMENTS, inc = 1;
-   double x[ELEMENTS * inc];
+   double *x = (double*) malloc(ELEMENTS * inc * sizeof(double));
    for (i = 0; i < N * inc; i ++) {
       x[i] = i + 1000;
    }
 
    x[(N - 1) * inc] = 0.0;
    blasint index = cblas_idamin(N, x, inc);
+   free(x);
    ASSERT_EQUAL(N - 1, index);
 }
-#endif
+#endif

utest/test_extensions/test_isamin.c

@@ -402,13 +402,14 @@ CTEST(isamin, min_idx_in_vec_tail){
 CTEST(isamin, min_idx_in_vec_tail_inc_1){
    blasint i;
    blasint N = ELEMENTS, inc = 1;
-   float x[ELEMENTS * inc];
+   float *x = (float*) malloc(ELEMENTS * inc * sizeof(float));
    for (i = 0; i < N * inc; i ++) {
       x[i] = i + 1000;
    }
 
    x[(N - 1) * inc] = 0.0f;
    blasint index = BLASFUNC(isamin)(&N, x, &inc);
+   free(x);
    ASSERT_EQUAL(N, index);
 }
 
@@ -775,13 +776,14 @@ CTEST(isamin, c_api_min_idx_in_vec_tail){
 CTEST(isamin, c_api_min_idx_in_vec_tail_inc_1){
    blasint i;
    blasint N = ELEMENTS, inc = 1;
-   float x[ELEMENTS * inc];
+   float *x = (float*)malloc(ELEMENTS * inc * sizeof(float));
    for (i = 0; i < N * inc; i ++) {
       x[i] = i + 1000;
    }
 
    x[(N - 1) * inc] = 0.0f;
    blasint index = cblas_isamin(N, x, inc);
+   free(x);
    ASSERT_EQUAL(N - 1, index);
 }
-#endif
+#endif

utest/test_extensions/test_zgemv_t.c

@@ -126,7 +126,7 @@ static double check_zgemv(char api, char order, char trans, blasint m, blasint n
     drand_generate(data_zgemv_t.y_test, m * inc_y * 2);
 
     // Copy vector y for reference funcs
-    for (int i = 0; i < m * inc_y * 2; i++)
+    for (i = 0; i < m * inc_y * 2; i++)
     {
         data_zgemv_t.y_verify[i] = data_zgemv_t.y_test[i];
     }
@@ -1133,4 +1133,4 @@ CTEST(zgemv, c_api_xerbla_invalid_order_col_major)
     int passed = c_api_check_badargs(corder, ctrans, m, n, lda, inc_x, inc_y, expected_info);
     ASSERT_EQUAL(TRUE, passed);
 }
-#endif
+#endif

utest/test_extensions/test_zsbmv.c

@@ -188,7 +188,7 @@ static double check_zsbmv(char uplo, blasint n, blasint k, double *alpha, blasin
     char trans = 'N';
 
     // Symmetric band packed matrix for sbmv
-    double a[lda * n * 2];
+    double *a = (double*) malloc(lda * n * 2 * sizeof(double));
 
     // Fill symmetric packed matrix sp_matrix, vector b_test, vector c_test 
     drand_generate(data_zsbmv.sp_matrix, n * (n + 1));
@@ -213,6 +213,7 @@ static double check_zsbmv(char uplo, blasint n, blasint k, double *alpha, blasin
     BLASFUNC(zsbmv)(&uplo, &n, &k, alpha, a, &lda,
                     data_zsbmv.b_test, &inc_b, beta, data_zsbmv.c_test, &inc_c);
 
+    free(a);
     // Find the differences between output vector caculated by zsbmv and zgemv
     for (i = 0; i < n * inc_c * 2; i++)
         data_zsbmv.c_test[i] -= data_zsbmv.c_verify[i];
@@ -603,4 +604,4 @@ CTEST(zsbmv, xerbla_lda_invalid)
     int passed = check_badargs(uplo, n, k, lda, inc_b, inc_c, expected_info);
     ASSERT_EQUAL(TRUE, passed);
 }
-#endif
+#endif