Fixed clang build error

 error: embedding a #pragma directive within
      macro arguments is not supported

Author: notorca

src/library/blas/trtri/diag_dtrtri_lower_128_16.cpp

@@ -62,7 +62,7 @@ __local double Bs[BLOCK_SIZE*BLOCK_SIZE]; \n
 __local double workspace[BLOCK_SIZE]; \n   // workspace used to store the current working column
 
 // load A
-#pragma unroll\n
+_Pragma("unroll")\n
 for (i = 0; i < BLOCK_SIZE; i++)\n
 { \n
     if (tx >= i && gx < na)\n
@@ -138,7 +138,7 @@ for (i = BLOCK_SIZE - 2; i >= 0; i--) {\n
 
 	txw = (tx - i - 1); \n
 
-#pragma unroll\n
+_Pragma("unroll")\n
 	for (j = 0; j < BLOCK_SIZE - i - 1; j++)\n
 		Ystx += switcher*(*(Bw + j*BLOCK_SIZE + txw)*x[j]); \n
 
@@ -163,7 +163,7 @@ for (i = BLOCK_SIZE - 2; i >= 0; i--) {\n
 }\n
 
 // write back A
-#pragma unroll\n
+_Pragma("unroll")\n
 for (i = 0; i < BLOCK_SIZE; i++)\n
 	*(d_dinvA + i*NB + tx) = Bs[i*BLOCK_SIZE + tx]; \n
 }\n

src/library/blas/trtri/diag_dtrtri_upper_128_16.cpp

@@ -59,7 +59,7 @@ uint na)\n
   __local double workspace[BLOCK_SIZE]; \n   // workspace used to store the current working column
 
   // load A
-  #pragma unroll \n
+  _Pragma("unroll")\n
   for( i=0; i < BLOCK_SIZE; i++ )\n
     {\n
       if(tx <= i && i+bx*BLOCK_SIZE < na )\n
@@ -111,7 +111,7 @@ uint na)\n
     workspace[tx] = *(Bs+i*BLOCK_SIZE+tx);\n
     y = Bs+i*BLOCK_SIZE;\n
 
-    #pragma unroll\n
+    _Pragma("unroll")\n
     //for( j=tx; j < i; j++ )
     for( j=0; j < i; j++ )\n
 	{\n
@@ -139,7 +139,7 @@ uint na)\n
   }\n
 
     // write back A
-#pragma unroll\n
+  _Pragma("unroll")\n
   for( i=0; i < BLOCK_SIZE; i++ )\n
   {\n
     *(d_dinvA+i*NB+tx) = Bs[i*BLOCK_SIZE+tx];\n

src/library/blas/trtri/diag_dtrtri_upper_192_12.cpp

@@ -59,7 +59,7 @@ __local double Bs[BLOCK_SIZE*BLOCK_SIZE]; \n
 __local double workspace[BLOCK_SIZE];\n    // workspace used to store the current working column 
 
 // load A \n
-#pragma unroll\n
+_Pragma("unroll")\n
 for (i = 0; i < BLOCK_SIZE; i++)\n
 { \n
     if (tx <= i && i + bx*BLOCK_SIZE < na)\n
@@ -112,7 +112,7 @@ for (i = 0; i < BLOCK_SIZE; i++) {\n
 	workspace[tx] = *(Bs + i*BLOCK_SIZE + tx); \n
 	y = Bs + i*BLOCK_SIZE; \n
 
-#pragma unroll\n
+_Pragma("unroll")\n
 	//for( j=tx; j < i; j++ )
 	for (j = 0; j < i; j++)\n
 		Ystx += switcher * (*(Bs + j*BLOCK_SIZE + tx)*workspace[j]); \n
@@ -138,7 +138,7 @@ for (i = 0; i < BLOCK_SIZE; i++) {\n
 }\n
 
 // write back A
-#pragma unroll\n
+_Pragma("unroll")\n
 for (i = 0; i < BLOCK_SIZE; i++)\n
 	*(d_dinvA + i*NB + tx) = Bs[i*BLOCK_SIZE + tx]; \n