Change "HALF" and "sh" to "BFLOAT16" and "sb"

diff --git a/driver/others/blas_server.c b/driver/others/blas_server.c
index acfaed7..30e0cc6 100644 (file)
--- a/driver/others/blas_server.c
+++ b/driver/others/blas_server.c
@@ -227,7 +227,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
                   args -> a, args -> lda,
                   args -> b, args -> ldb,
                   args -> c, args -> ldc, sb);
-#ifdef BUILD_HALF
+#ifdef BUILD_BFLOAT16
          } else if ((mode & BLAS_PREC) == BLAS_BFLOAT16){
             /* REAL / BFLOAT16 */
             void (*afunc)(BLASLONG, BLASLONG, BLASLONG, bfloat16,

diff --git a/driver/others/blas_server_omp.c b/driver/others/blas_server_omp.c
index bfbe3a6..d546553 100644 (file)
--- a/driver/others/blas_server_omp.c
+++ b/driver/others/blas_server_omp.c
@@ -192,7 +192,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
                  args -> a, args -> lda,
                  args -> b, args -> ldb,
                  args -> c, args -> ldc, sb);
-#ifdef BUILD_HALF
+#ifdef BUILD_BFLOAT16
           } else if ((mode & BLAS_PREC) == BLAS_BFLOAT16){
             /* REAL / BFLOAT16 */
             void (*afunc)(BLASLONG, BLASLONG, BLASLONG, bfloat16,

diff --git a/driver/others/blas_server_win32.c b/driver/others/blas_server_win32.c
index d2cc917..4624085 100644 (file)
--- a/driver/others/blas_server_win32.c
+++ b/driver/others/blas_server_win32.c
@@ -112,7 +112,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
                  args -> a, args -> lda,
                  args -> b, args -> ldb,
                  args -> c, args -> ldc, sb);
-#ifdef BUILD_HALF
+#ifdef BUILD_BFLOAT16
           } else if ((mode & BLAS_PREC) == BLAS_BFLOAT16){
             /* REAL / BFLOAT16 */
             void (*afunc)(BLASLONG, BLASLONG, BLASLONG, bfloat16,

diff --git a/driver/others/parameter.c b/driver/others/parameter.c
index 5d312fa..35fc0a2 100644 (file)
--- a/driver/others/parameter.c
+++ b/driver/others/parameter.c
@@ -62,10 +62,10 @@ BLASLONG gemm_offset_b = DEFAULT_GEMM_OFFSET_B;
 BLASLONG gemm_offset_b = GEMM_OFFSET_B;
 #endif
 
-#if SHGEMM_P == shgemm_p
-BLASLONG shgemm_p = DEFAULT_GEMM_P;
+#if SBGEMM_P == sbgemm_p
+BLASLONG sbgemm_p = DEFAULT_GEMM_P;
 #else
-BLASLONG shgemm_p = SHGEMM_P;
+BLASLONG sbgemm_p = SBGEMM_P;
 #endif
 #if SGEMM_P == sgemm_p
 BLASLONG sgemm_p = DEFAULT_GEMM_P;
@@ -88,10 +88,10 @@ BLASLONG zgemm_p = DEFAULT_GEMM_P;
 BLASLONG zgemm_p = ZGEMM_P;
 #endif
 
-#if SHGEMM_Q == shgemm_q
-BLASLONG shgemm_q = DEFAULT_GEMM_Q;
+#if SBGEMM_Q == sbgemm_q
+BLASLONG sbgemm_q = DEFAULT_GEMM_Q;
 #else
-BLASLONG shgemm_q = SHGEMM_Q;
+BLASLONG sbgemm_q = SBGEMM_Q;
 #endif
 #if SGEMM_Q == sgemm_q
 BLASLONG sgemm_q = DEFAULT_GEMM_Q;
@@ -114,10 +114,10 @@ BLASLONG zgemm_q = DEFAULT_GEMM_Q;
 BLASLONG zgemm_q = ZGEMM_Q;
 #endif
 
-#if SHGEMM_R == shgemm_r
-BLASLONG shgemm_r = DEFAULT_GEMM_R;
+#if SBGEMM_R == sbgemm_r
+BLASLONG sbgemm_r = DEFAULT_GEMM_R;
 #else
-BLASLONG shgemm_r = SHGEMM_R;
+BLASLONG sbgemm_r = SBGEMM_R;
 #endif
 #if SGEMM_R == sgemm_r
 BLASLONG sgemm_r = DEFAULT_GEMM_R;
@@ -615,7 +615,7 @@ void blas_set_parameter(void){
 
   size = BITMASK(cpuid3, 16, 0xff);
 
-  shgemm_p = 192 * (size + 1);
+  sbgemm_p = 192 * (size + 1);
   sgemm_p = 192 * (size + 1);
   dgemm_p =  96 * (size + 1);
   cgemm_p =  96 * (size + 1);
@@ -629,7 +629,7 @@ void blas_set_parameter(void){
   xgemm_p =  16 * (size + 1);
 #endif
 
-  shgemm_r = (((BUFFER_SIZE - ((SHGEMM_P * SHGEMM_Q *  4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SHGEMM_Q *  4)) - 15) & ~15;
+  sbgemm_r = (((BUFFER_SIZE - ((SBGEMM_P * SBGEMM_Q *  4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SBGEMM_Q *  4)) - 15) & ~15;
   sgemm_r = (((BUFFER_SIZE - ((SGEMM_P * SGEMM_Q *  4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SGEMM_Q *  4)) - 15) & ~15;
   dgemm_r = (((BUFFER_SIZE - ((DGEMM_P * DGEMM_Q *  8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (DGEMM_Q *  8)) - 15) & ~15;
   cgemm_r = (((BUFFER_SIZE - ((CGEMM_P * CGEMM_Q *  8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (CGEMM_Q *  8)) - 15) & ~15;