Fixed bug with precision in cublasXgemm

SGo-Go · Oct 11, 2013 · 38be3f5 · 38be3f5
1 parent 99dee84
commit 38be3f5
Show file tree

Hide file tree

Showing 4 changed files with 11 additions and 8 deletions.
diff --git a/bench/bench.py b/bench/bench.py
@@ -30,7 +30,7 @@
     # os.environ['MKL_NUM_THREADS'] = str(nthreads)
 
     for n in n_list:
-        ntests = 4*(1024/n) - 3
+        ntests = 44*(1024/n) - 3
         print     cmd_all.format(implement = implementation, ofolder = ofolder_local,
                                  nthreads=nthreads, ntests=ntests, blknum=L, blksize=n)
         os.system(cmd_all.format(implement = implementation, ofolder = ofolder_local,

diff --git a/examples/ctest/ctest.c b/examples/ctest/ctest.c
@@ -52,8 +52,8 @@ void Init_DQMC_matrix(int n, int nb, scalar_t A[])
       }
 }
 
-#define n    1728 //1024  //1024//1728 ////512
-#define nb   10 //45   //45  //16//26 ////90
+#define n   1728  /* 1024 1024 1728 512 */
+#define nb  10    /* 45 45 16 26 90 */
 
 int main(void)
 {

diff --git a/include/bsofi_macro.h b/include/bsofi_macro.h
@@ -15,7 +15,8 @@
 #include "third_party/cublas.h"
 
 #define GPU_ALIGN 64 /* 32 */
-#define GET_LD_DEV(__len) ((((__len) + (GPU_ALIGN)-1)/(GPU_ALIGN))*(GPU_ALIGN))
+#define GET_LD_DEV(__len) \
+  ((((__len) + (GPU_ALIGN)-1)/(GPU_ALIGN))*(GPU_ALIGN))
 
 #define IDX_OFFSET(__ld,__i, __j) ((__j)*(__ld) + __i)
 #define BLK_OFFSET(__ld,__i, __j) (n*((__j)*(__ld) + __i))
@@ -54,7 +55,8 @@ typedef struct _bsofi_profile_t {
     tim1 = getwalltime();				\
     {__code_do;}					\
     __total_time += elapsed(getwalltime(), tim1); }
-#  define RESET_BSOFI_PROFILE(__counter) __counter = (const struct _bsofi_profile_t){0}
+#  define RESET_BSOFI_PROFILE(__counter) \
+  __counter = (const struct _bsofi_profile_t){0}
 
 #else
 #  define BENCH_CUMMULATIVE(__total_time, __code_do) {__code_do;}
@@ -67,7 +69,8 @@ typedef struct _bsofi_profile_t {
 #  define CHECK_CUMALLOC(__code)					\
   if (cudaSuccess != (__code)) {					\
     DBGERROR("CUDA: GPU device memory allocation failed");		\
-    cudaFree(dwork); cublasShutdown();/* cublasDestroy(handle); */	\
+    /* cudaFree(dwork); */						\
+    cublasShutdown();/* cublasDestroy(handle); */			\
     return -1;								\
   }
 

diff --git a/include/third_party/cublas.h b/include/third_party/cublas.h
@@ -42,9 +42,9 @@
 			__A, __lda, __B, __ldb), "cublasGetMatrix")
 
 #  ifdef __SINGLE_PREC__
-#    define cublasXgemm  cublasDgemm
-#  else
 #    define cublasXgemm  cublasSgemm
+#  else
+#    define cublasXgemm  cublasDgemm
 #  endif
 
 #else