Merge branch 'develop' into betterPowerGEMVTail

cmake/system.cmake

@@ -263,7 +263,7 @@ if (DEFINED TARGET)
   endif()
 
   if (${TARGET} STREQUAL POWER10)
-    if (CMAKE_C_COMPILER VERSION VERSION_GREATER 10.2 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 10.2)
+    if (CMAKE_C_COMPILER_VERSION VERSION_GREATER 10.2 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 10.2)
       set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -mcpu=power10 -mtune=power10 -mvsx -fno-fast-math")
     else ()
 	    message(FATAL_ERROR "Compiler GCC ${CMAKE_C_COMPILER_VERSION} does not support Power10.")

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

kernel/x86_64/sbgemv_t_microk_cooperlake_template.c

@@ -2680,7 +2680,7 @@ static int sbgemv_kernel_1x128_lda_direct(BLASLONG m, BLASLONG n, float alpha, b
     BLASLONG tag_n_32x  = n & (~31);
     BLASLONG tag_n_128x = n & (~127);
 
-    __m512 accum512_bridge[8];
+    __m512 accum512_bridge[16];
     __m512 accum512_t_0, accum512_t_1, accum512_t_2, accum512_t_3;
     __m256 accum256_0;
     __m128 accum128;

lapack-netlib/TESTING/stest_rfp.in

@@ -5,5 +5,5 @@ Data file for testing REAL LAPACK linear equation routines RFP format
 1 2 15                         Values of NRHS (number of right hand sides)
 9                              Number of matrix types (list types on next line if 0 < NTYPES <  9)
 1 2 3 4 5 6 7 8 9              Matrix Types
-42.0                           Threshold value of test ratio
+45.0                           Threshold value of test ratio
 T                              Put T to test the error exits

lapack/potrf/potrf_L_parallel.c

@@ -121,7 +121,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
       HERK_THREAD_LN(&newarg, NULL, NULL, sa, sb, 0);
 #else
       syrk_thread(mode | BLAS_TRANSA_N | BLAS_TRANSB_T | BLAS_UPLO,
-		  &newarg, NULL, NULL, (int (*)(void))HERK_LN, sa, sb, args -> nthreads);
+		  &newarg, NULL, NULL, (int (*)(blas_arg_t *, BLASLONG *, BLASLONG *, FLOAT *, FLOAT *, BLASLONG))HERK_LN, sa, sb, args -> nthreads);
 #endif
     }
   }

lapack/potrf/potrf_U_parallel.c

@@ -121,7 +121,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
       HERK_THREAD_UC(&newarg, NULL, NULL, sa, sb, 0);
 #else
       syrk_thread(mode | BLAS_TRANSA_N | BLAS_TRANSB_T,
-		  &newarg, NULL, NULL, (int (*)(void))HERK_UC, sa, sb, args -> nthreads);
+		  &newarg, NULL, NULL, (int (*)(blas_arg_t *, BLASLONG *, BLASLONG *, FLOAT *, FLOAT *, BLASLONG))HERK_UC, sa, sb, args -> nthreads);
 #endif
     }
   }

param.h

@@ -2637,8 +2637,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #undef SBGEMM_DEFAULT_Q
 #define SBGEMM_DEFAULT_UNROLL_M 16
 #define SBGEMM_DEFAULT_UNROLL_N 8
-#define SBGEMM_DEFAULT_P 832
-#define SBGEMM_DEFAULT_Q 1026
+#define SBGEMM_DEFAULT_P 512
+#define SBGEMM_DEFAULT_Q 1024
 #define SBGEMM_DEFAULT_R 4096
 #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

test/compare_sgemm_sbgemm.c

@@ -81,6 +81,8 @@ float16to32 (bfloat16_bits f16)
   return f32.v;
 }
 
+#define SBGEMM_LARGEST  256
+
 int
 main (int argc, char *argv[])
 {
@@ -88,12 +90,45 @@ main (int argc, char *argv[])
   int i, j, l;
   blasint x, y;
   int ret = 0;
-  int loop = 100;
+  int loop = SBGEMM_LARGEST;
   char transA = 'N', transB = 'N';
   float alpha = 1.0, beta = 0.0;
 
   for (x = 0; x <= loop; x++)
   {
+    if ((x > 100) && (x != SBGEMM_LARGEST)) continue;
+    m = k = n = x;
+    float *A = (float *)malloc(m * k * sizeof(FLOAT));
+    float *B = (float *)malloc(k * n * sizeof(FLOAT));
+    float *C = (float *)malloc(m * n * sizeof(FLOAT));
+    bfloat16_bits *AA = (bfloat16_bits *)malloc(m * k * sizeof(bfloat16_bits));
+    bfloat16_bits *BB = (bfloat16_bits *)malloc(k * n * sizeof(bfloat16_bits));
+    float *DD = (float *)malloc(m * n * sizeof(FLOAT));
+    float *CC = (float *)malloc(m * n * sizeof(FLOAT));
+    if ((A == NULL) || (B == NULL) || (C == NULL) || (AA == NULL) || (BB == NULL) ||
+        (DD == NULL) || (CC == NULL))
+      return 1;
+    bfloat16 atmp,btmp;
+    blasint one=1;
+
+    for (j = 0; j < m; j++)
+    {
+      for (i = 0; i < k; i++)
+      {
+        A[j * k + i] = ((FLOAT) rand () / (FLOAT) RAND_MAX) + 0.5;
+        sbstobf16_(&one, &A[j*k+i], &one, &atmp, &one);
+        AA[j * k + i].v = atmp;
+      }
+    }
+    for (j = 0; j < n; j++)
+    {
+      for (i = 0; i < k; i++)
+      {
+        B[j * k + i] = ((FLOAT) rand () / (FLOAT) RAND_MAX) + 0.5;
+        sbstobf16_(&one, &B[j*k+i], &one, &btmp, &one);
+        BB[j * k + i].v = btmp;
+      }
+    }
     for (y = 0; y < 4; y++)
     {
       if ((y == 0) || (y == 2)) {
@@ -106,40 +141,19 @@ main (int argc, char *argv[])
       } else {
         transB = 'T';
       }
-      m = k = n = x;
-      float A[m * k];
-      float B[k * n];
-      float C[m * n];
-      bfloat16_bits AA[m * k], BB[k * n];
-      float DD[m * n], CC[m * n];
-      bfloat16 atmp,btmp;
-      blasint one=1;
 
-      for (j = 0; j < m; j++)
-      {
-        for (i = 0; i < m; i++)
-        {
-          A[j * k + i] = ((FLOAT) rand () / (FLOAT) RAND_MAX) + 0.5;
-          B[j * k + i] = ((FLOAT) rand () / (FLOAT) RAND_MAX) + 0.5;
-          C[j * k + i] = 0;
-          sbstobf16_(&one, &A[j*k+i], &one, &atmp, &one);
-          sbstobf16_(&one, &B[j*k+i], &one, &btmp, &one);
-          AA[j * k + i].v = atmp;
-          BB[j * k + i].v = btmp;
-          CC[j * k + i] = 0;
-          DD[j * k + i] = 0;
-        }
-      }
+      memset(CC, 0, m * n * sizeof(FLOAT));
+      memset(DD, 0, m * n * sizeof(FLOAT));
+      memset(C, 0, m * n * sizeof(FLOAT));
+
       SGEMM (&transA, &transB, &m, &n, &k, &alpha, A,
         &m, B, &k, &beta, C, &m);
       SBGEMM (&transA, &transB, &m, &n, &k, &alpha, (bfloat16*) AA,
         &m, (bfloat16*)BB, &k, &beta, CC, &m);
+
       for (i = 0; i < n; i++)
         for (j = 0; j < m; j++)
-          if (fabs (CC[i * m + j] - C[i * m + j]) > 1.0)
-            ret++;
-      for (i = 0; i < n; i++)
-        for (j = 0; j < m; j++)
+        {
           for (l = 0; l < k; l++)
             if (transA == 'N' && transB == 'N')
             {
@@ -158,11 +172,19 @@ main (int argc, char *argv[])
               DD[i * m + j] +=
                 float16to32 (AA[k * j + l]) * float16to32 (BB[i + l * n]);
             }
-      for (i = 0; i < n; i++)
-        for (j = 0; j < m; j++)
-          if (CC[i * m + j] != DD[i * m + j])
+          if (fabs (CC[i * m + j] - C[i * m + j]) > 1.0)
+            ret++;
+          if (fabs (CC[i * m + j] - DD[i * m + j]) > 1.0)
             ret++;
+        }
     }
+    free(A);
+    free(B);
+    free(C);
+    free(AA);
+    free(BB);
+    free(DD);
+    free(CC);
   }
 
   if (ret != 0)