DGEMV BugFixes and code cleanup (#134)

- Modified gemv (matrix-vector multiply) reference for better handling of transpose flags.
- Modified Zen4 kernel implementations for better handling of transpose flags and vector stride (incy).
- The changes refine kernel selection logic and move variable definition in macro guards.

kernels/zen/2/bli_gemv_zen_ref.c

@@ -175,7 +175,7 @@ void bli_dgemv_zen_ref
     // If alpha == 0, return.
     if ( bli_deq0( *alpha ) ) return;
 
-    if ( bli_is_notrans( transa ) )     // BLIS_NO_TRANSPOSE
+    if ( bli_does_notrans( transa ) ) // transa = N or C
     {
         if ( incy == 1 )
         {
@@ -292,7 +292,7 @@ void bli_sgemv_zen_ref
     // If alpha == 0, return.
     if ( bli_seq0( *alpha ) ) return;
 
-    if ( bli_is_notrans( transa ) )     // BLIS_NO_TRANSPOSE
+    if ( bli_does_notrans( transa ) ) // transa = N or C
     {
         if ( incy == 1 )
         {

kernels/zen4/2/bli_gemv_n_zen4_int.c

@@ -1465,9 +1465,10 @@ void bli_dgemv_n_zen4_int (
                       double*, 
                       inc_t, cntx_t* ) = NULL;
 
-    dim_t size = m * n;
-
 // If AOCL_DYNAMIC is enabled, call ST kernels for small sizes.
+#if (defined(AOCL_DYNAMIC) || (defined(BLIS_ENABLE_OPENMP)))
+    dim_t size = m * n;
+#endif
 #ifdef AOCL_DYNAMIC
     if ( size < 95000 )
     {
@@ -1514,6 +1515,12 @@ void bli_dgemv_n_zen4_int (
 #endif
     }
 
+    // Use 32x8 kernel when transa = "C" or "H"
+    // and if incy != 1, which uses packing to handle non unit stride y
+    if ( incy != 1 || transa != BLIS_NO_TRANSPOSE)
+    {
+        ker_ft = bli_dgemv_n_zen4_32x8_int_st;
+    }
     ker_ft
     (
         transa,