Added bli_mem_init() to level 3 facade functions. Use Perl script to generate wrapper...

[dense-linear-algebra-libraries/linalg.git] / blasblisacc / src / ti_cblas_cblas_scopy.c

diff --git a/blasblisacc/src/ti_cblas_cblas_scopy.c b/blasblisacc/src/ti_cblas_cblas_scopy.c
index f1bdaedb41c464f58c3ed9fed988f6918d8c5c42..099f5ecf17f60fc1ac60cfff4a94a6d5f3881e36 100644 (file)
--- a/blasblisacc/src/ti_cblas_cblas_scopy.c
+++ b/blasblisacc/src/ti_cblas_cblas_scopy.c
@@ -44,20 +44,19 @@ void cblas_scopy(const int N, const float *X, const int incX, float *Y, const in
        if (!ti_cblas_init_done) ti_cblas_init();
        TI_CBLAS_DEBUG_PRINT("Intercepted call to %s\n", "cblas_scopy");
 
-        TI_CBLAS_PROFILE_START();
-
+    TI_CBLAS_PROFILE_START();
        /* Dynamic condtional offload to ARM */
-        if ((TI_CBLAS_L1_OFFLOAD == TI_CBLAS_OFFLOAD_NONE)) { 
-               TI_CBLAS_DEBUG_PRINT("Executing ARM %s\n", "cblas_scopy"); 
-               __real_cblas_scopy(N,X,incX,Y,incY); 
-               TI_CBLAS_PROFILE_REPORT("  Entire %s call (ARM) took %8.2f us\n","cblas_scopy", (float) clock_diff);
-               return ;
-        }
-        /* End ARM offload */
+    if ((TI_CBLAS_L1_OFFLOAD == TI_CBLAS_OFFLOAD_NONE)) { 
+       TI_CBLAS_DEBUG_PRINT("Executing ARM %s\n", "cblas_scopy"); 
+       __real_cblas_scopy(N,X,incX,Y,incY); 
+    TI_CBLAS_PROFILE_REPORT("  Entire %s call (ARM) took %8.2f us\n","cblas_scopy", (float) clock_diff);
+       return ;
+    }
+    /* End ARM offload */
 
        /******************************************************************/
        /* DSP offload WILL be done if control reaches here */
-               TI_CBLAS_DEBUG_PRINT("Offloading to DSP %s\n", "cblas_scopy"); 
+       TI_CBLAS_DEBUG_PRINT("Offloading to DSP %s\n", "cblas_scopy"); 
 
        /* Lookup kernel pointer from global table */
 #ifdef __cplusplus
@@ -74,7 +73,6 @@ void cblas_scopy(const int N, const float *X, const int incX, float *Y, const in
 #endif
        {
 
-
 #ifdef __cplusplus
                __K->setArg(0, N);
 #else
@@ -135,6 +133,7 @@ void cblas_scopy(const int N, const float *X, const int incX, float *Y, const in
                TI_CBLAS_OCL_CHKERROR("clReleaseEvent",err);
 
 #endif
+
                ti_cblas_delete_kernel(__K);
 
                TI_CBLAS_DEBUG_PRINT("Finished executing %s\n", "cblas_scopy");