grid: Fix unroll pragmas ignored by icc

cp2k · Apr 22, 2021 · 50d0dad · 50d0dad
1 parent 0cd1d52
commit 50d0dad
Show file tree

Hide file tree

Showing 2 changed files with 21 additions and 17 deletions.
diff --git a/src/grid/common/grid_common.h b/src/grid/common/grid_common.h
@@ -9,26 +9,30 @@
 
 #define GRID_STRINGIFY(SYMBOL) #SYMBOL
 
-// GCC introduced the unroll pragma with version 8 using a custom syntax.
-#if defined(__INTEL_COMPILER)
-#define GRID_PRAGMA_UNROLL(N) _Pragma(GRID_STRINGIFY(unroll(N)))
-#if __INTEL_COMPILER >= 1900
-#define GRID_PRAGMA_SIMD(OBJS, N)                                              \
-  _Pragma(GRID_STRINGIFY(omp simd linear OBJS simdlen(N)))
-#else
+// GCC added the simd pragma with version 6.
+#if defined(__GNUC__) && !defined(__INTEL_COMPILER) && __GNUC__ < 6
 #define GRID_PRAGMA_SIMD(OBJS, N)
-#endif
-#elif defined(__GNUC__) && __GNUC__ >= 8
-#define GRID_PRAGMA_SIMD(OBJS, N)                                              \
-  _Pragma(GRID_STRINGIFY(omp simd linear OBJS simdlen(N)))
-#define GRID_PRAGMA_UNROLL(N) _Pragma(GRID_STRINGIFY(GCC unroll N))
-#elif defined(__GNUC__) && __GNUC__ < 8
+// Intel added the simd pragma with version 19.00.
+#elif defined(__INTEL_COMPILER) && __INTEL_COMPILER < 1900
 #define GRID_PRAGMA_SIMD(OBJS, N)
-#define GRID_PRAGMA_UNROLL(N)
-#else // Most other compilers support a common syntax.
+// All compilers support the same syntax defined by the OpenMP standard.
+#else
 #define GRID_PRAGMA_SIMD(OBJS, N)                                              \
   _Pragma(GRID_STRINGIFY(omp simd linear OBJS simdlen(N)))
+#endif
+
+// GCC added the unroll pragma with version 8 and...
+#if defined(__GNUC__) && !defined(__INTEL_COMPILER) && __GNUC__ < 8
+#define GRID_PRAGMA_UNROLL(N)
+#define GRID_PRAGMA_UNROLL_UP_TO(N)
+// ...chose a custom syntax.
+#elif defined(__GNUC__) && !defined(__INTEL_COMPILER) && __GNUC__ >= 8
+#define GRID_PRAGMA_UNROLL(N) _Pragma(GRID_STRINGIFY(GCC unroll N))
+#define GRID_PRAGMA_UNROLL_UP_TO(N) _Pragma(GRID_STRINGIFY(GCC unroll N))
+// Most other compilers support a common syntax.
+#else
 #define GRID_PRAGMA_UNROLL(N) _Pragma(GRID_STRINGIFY(unroll(N)))
+#define GRID_PRAGMA_UNROLL_UP_TO(N) _Pragma("unroll")
 #endif
 
 #if defined(__CUDACC__)

diff --git a/src/grid/ref/grid_ref_collint.h b/src/grid/ref/grid_ref_collint.h
@@ -98,7 +98,7 @@ ortho_cx_to_grid_avx2(const int lp, const int cmax, const int i,
   __m256d r_vec_3 = _mm256_mul_pd(p_vec, _mm256_set1_pd(cx[3]));
 
   // Remaining iterations for lxp > 0 use fused multiply adds.
-  GRID_PRAGMA_UNROLL(GRID_MAX_LP_OPTIMIZED)
+  GRID_PRAGMA_UNROLL_UP_TO(GRID_MAX_LP_OPTIMIZED)
   for (int lxp = 1; lxp <= lp; lxp++) {
     const double *cx_base = &cx[lxp * 4];
     p_vec = _mm256_loadu_pd(&pol[0][lxp][icmax]);
@@ -121,7 +121,7 @@ ortho_cx_to_grid_avx2(const int lp, const int cmax, const int i,
   __m256d grid_vec_2 = _mm256_loadu_pd(grid_2);
   __m256d grid_vec_3 = _mm256_loadu_pd(grid_3);
 
-  GRID_PRAGMA_UNROLL(GRID_MAX_LP_OPTIMIZED + 1)
+  GRID_PRAGMA_UNROLL_UP_TO(GRID_MAX_LP_OPTIMIZED + 1)
   for (int lxp = 0; lxp <= lp; lxp++) {
     __m256d p_vec = _mm256_loadu_pd(&pol[0][lxp][icmax]);