Inhibit vectorization of short loops

In the code that mirrors independent harmonics up to
Nyquist frequency, the short loop containing computation
of location of the harmonics and its dependent counterpart
was being vectorized by ICC with detrimental effects on
performance for FFT-dimensions smaller than vector register
width.

Use pragma novector for now. This improves performance of
mkl_fft.fft2 on (8K,8K) array of doubles making ICC compiled
code 18% faster than GCC, rather than 2.5X slower.

Author: oleksandr-pavlyk

mkl_fft/src/mklfft.c.src

@@ -560,6 +560,9 @@ int @name@_mkl_@mode@_in(PyArrayObject* x_inout, npy_intp n, int axis)
         while(!MultiIter_Done(mit)) {
             char *tmp;
 
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+            #pragma novector
+#endif
             for(tmp = (char *) x_data, i = 0; i < x_rank; i++)
                 tmp += x_strides[i] * MultiIter_IndexElem(mit, i);
 
@@ -747,6 +750,9 @@ int @REALIN@_@COMPLEXOUT@_mkl_@mode@_out(
         while(!MultiIter_Done(mit)) {
             char *tmp1, *tmp2;
 
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+            #pragma novector
+#endif
             for(tmp1 = (char *) xin_data,
                 tmp2 = (char *) xout_data,
                 i = 0; i < xin_rank; i++) {
@@ -808,6 +814,9 @@ int @REALIN@_@COMPLEXOUT@_mkl_@mode@_out(
                 @MKL_OUT_TYPE@ *dest, *src;
                 /* npy_intp k_last = MultiIter_IndexElem(mit, axis); */
 
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+                #pragma novector
+#endif
                 for(tmp1 = (char *) xout_data, tmp2 = (char *) xout_data,
                     i = 0; i < xout_rank; i++) {
                     npy_intp si = xout_strides[i],
@@ -997,6 +1006,9 @@ int @COMPLEXIN@_@COMPLEXOUT@_mkl_@mode@_out(
         while(!MultiIter_Done(mit)) {
             char *tmp1, *tmp2;
 
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+            #pragma novector
+#endif
             for(tmp1 = (char *) xin_data,
                 tmp2 = (char *) xout_data,
                 i = 0; i < xin_rank; i++) {
@@ -1135,6 +1147,9 @@ int @name@_mkl_@mode@_in(PyArrayObject* x_inout, npy_intp n, int axis)
         while(!MultiIter_Done(mit)) {
             char *tmp;
 
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+            #pragma novector
+#endif
             for(tmp = (char *) x_data, i = 0; i < x_rank; i++)
                 tmp += x_strides[i] * MultiIter_IndexElem(mit, i);
 
@@ -1297,6 +1312,9 @@ int
         while(!MultiIter_Done(mit)) {
             char *tmp1, *tmp2;
 
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+            #pragma novector
+#endif
             for(tmp1 = (char *) xin_data,
                 tmp2 = (char *) xout_data,
                 i = 0; i < xin_rank; i++) {
@@ -1475,6 +1493,9 @@ int @name@_@name@_mkl_@mode@_out(
         while(!MultiIter_Done(mit)) {
             char *tmp1, *tmp2;
 
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+            #pragma novector
+#endif
             for(tmp1 = (char *) xin_data,
                 tmp2 = (char *) xout_data,
                 i = 0; i < xin_rank; i++) {
@@ -1872,6 +1893,9 @@ int
             char *tmp1, *tmp2;
             @MKL_OUT_TYPE@ *dest, *src;
 
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+            #pragma novector
+#endif
             for(tmp1 = (char *) xout_data, tmp2 = (char *) xout_data,
                 i = 0; i < xout_rank; i++) {
                 npy_intp si = xout_strides[i],
@@ -1890,12 +1914,10 @@ int
                 tmp1 += si * dest_ki;
                 tmp2 += si * src_ki;
             }
+
             dest = (@MKL_OUT_TYPE@*) tmp1;
             src  = (@MKL_OUT_TYPE@*) tmp2;
-
-            /* is there a nicer way to complex conjugate ? */
-            dest->real = src->real;
-            dest->imag = -src->imag;
+            SET_CONJ(dest, src);
 
             if (multi_iter_next(&mit))
                 break;