Redesigned SIMD module and unit tests.

[alexxy/gromacs.git] / src / gromacs / mdlib / nbnxn_kernels / nbnxn_kernel_simd_utils.h

diff --git a/src/gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_simd_utils.h b/src/gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_simd_utils.h
index bb84a499c48fcdb2364f634f31179029a3f35473..adb9613af39c751a70b7abe9a036794c22755495 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_simd_utils.h
+++ b/src/gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_simd_utils.h
@@ -67,7 +67,7 @@ prepare_table_load_buffer(const int gmx_unused *array)
 
 #else /* GMX_SIMD_REFERENCE */
 
-#if defined  GMX_SIMD_X86_SSE2_OR_HIGHER && !defined __MIC__
+#if defined  GMX_TARGET_X86 && !defined __MIC__
 /* Include x86 SSE2 compatible SIMD functions */
 
 /* Set the stride for the lookup of the two LJ parameters from their
@@ -80,46 +80,35 @@ static const int nbfp_stride = 4;
 #endif
 
 /* Align a stack-based thread-local working array. Table loads on
- * full-width AVX_256 use the array, but other implementations do
- * not. */
+ * 256-bit AVX use the array, but other implementations do not.
+ */
 static gmx_inline int *
-prepare_table_load_buffer(const int gmx_unused *array)
+prepare_table_load_buffer(int gmx_unused *array)
 {
-#if defined GMX_SIMD_X86_AVX_256_OR_HIGHER && !defined GMX_USE_HALF_WIDTH_SIMD_HERE
+#if GMX_SIMD_REAL_WIDTH >= 8 || (defined GMX_DOUBLE && GMX_SIMD_REAL_WIDTH >= 4)
     return gmx_simd_align_i(array);
 #else
     return NULL;
 #endif
 }
 
-#if defined GMX_SIMD_X86_AVX_256_OR_HIGHER && !defined GMX_USE_HALF_WIDTH_SIMD_HERE
-
-/* With full AVX-256 SIMD, half SIMD-width table loads are optimal */
-#if GMX_SIMD_REAL_WIDTH == 8
-#define TAB_FDV0
-#endif
 #ifdef GMX_DOUBLE
+#if GMX_SIMD_REAL_WIDTH == 2
+#include "nbnxn_kernel_simd_utils_x86_128d.h"
+#else
 #include "nbnxn_kernel_simd_utils_x86_256d.h"
-#else  /* GMX_DOUBLE */
-#include "nbnxn_kernel_simd_utils_x86_256s.h"
-#endif /* GMX_DOUBLE */
-
-#else  /* defined GMX_SIMD_X86_AVX_256_OR_HIGHER && !defined GMX_USE_HALF_WIDTH_SIMD_HERE */
-
-/* We use the FDV0 table layout when we can use aligned table loads */
-#if GMX_SIMD_REAL_WIDTH == 4
-#define TAB_FDV0
 #endif
-
-#ifdef GMX_DOUBLE
-#include "nbnxn_kernel_simd_utils_x86_128d.h"
-#else  /* GMX_DOUBLE */
+#else /* GMX_DOUBLE */
+/* In single precision aligned FDV0 table loads are optimal */
+#define TAB_FDV0
+#if GMX_SIMD_REAL_WIDTH == 4
 #include "nbnxn_kernel_simd_utils_x86_128s.h"
+#else
+#include "nbnxn_kernel_simd_utils_x86_256s.h"
+#endif
 #endif /* GMX_DOUBLE */
 
-#endif /* defined GMX_SIMD_X86_AVX_256_OR_HIGHER && !defined GMX_USE_HALF_WIDTH_SIMD_HERE */
-
-#else  /* GMX_SIMD_X86_SSE2_OR_HIGHER */
+#else  /* GMX_TARGET_X86 && !__MIC__ */
 
 #if GMX_SIMD_REAL_WIDTH > 4
 /* For width>4 we use unaligned loads. And thus we can use the minimal stride */
@@ -141,64 +130,22 @@ static const int nbfp_stride = GMX_SIMD_REAL_WIDTH;
 #include "nbnxn_kernel_simd_utils_x86_mic.h"
 #endif
 
-#endif /* GMX_SIMD_X86_SSE2_OR_HIGHER */
-#endif /* GMX_SIMD_REFERENCE */
-
-#if GMX_SIMD_REAL_WIDTH == 4
-#define gmx_mm_pr4    gmx_simd_real_t
-#define gmx_load_pr4  gmx_simd_load_r
-#define gmx_store_pr4 gmx_simd_store_r
-#define gmx_add_pr4   gmx_simd_add_r
-#endif
+#endif /* GMX_TARGET_X86 && !__MIC__ */
 
-#ifndef HAVE_GMX_SUM_SIMD /* should be defined for arch with hardware reduce */
-static gmx_inline real
-gmx_sum_simd2(gmx_simd_real_t x, real* b)
-{
-    gmx_simd_store_r(b, x);
-    return b[0]+b[1];
-}
-
-#if GMX_SIMD_REAL_WIDTH >= 4
-static gmx_inline real
-gmx_sum_simd4(gmx_mm_pr4 x, real* b)
-{
-    gmx_store_pr4(b, x);
-    return b[0]+b[1]+b[2]+b[3];
-}
-#endif
+#endif /* GMX_SIMD_REFERENCE */
 
-#if GMX_SIMD_REAL_WIDTH == 2
-static gmx_inline real gmx_sum_simd(gmx_simd_real_t x, real* b)
-{
-    gmx_simd_store_r(b, x);
-    return b[0]+b[1];
-}
-#elif GMX_SIMD_REAL_WIDTH == 4
-static gmx_inline real gmx_sum_simd(gmx_simd_real_t x, real* b)
-{
-    gmx_simd_store_r(b, x);
-    return b[0]+b[1]+b[2]+b[3];
-}
-#elif GMX_SIMD_REAL_WIDTH == 8
-static gmx_inline real gmx_sum_simd(gmx_simd_real_t x, real* b)
-{
-    gmx_simd_store_r(b, x);
-    return b[0]+b[1]+b[2]+b[3]+b[4]+b[5]+b[6]+b[7];
-}
-#elif GMX_SIMD_REAL_WIDTH == 16
-/* This is getting ridiculous, SIMD horizontal adds would help,
- * but this is not performance critical (only used to reduce energies)
+/* If the simd width is 4, but simd4 instructions are not defined,
+ * reuse the simd real type and the four instructions we need.
  */
-static gmx_inline real gmx_sum_simd(gmx_simd_real_t x, real* b)
-{
-    gmx_simd_store_r(b, x);
-    return b[0]+b[1]+b[2]+b[3]+b[4]+b[5]+b[6]+b[7]+b[8]+b[9]+b[10]+b[11]+b[12]+b[13]+b[14]+b[15];
-}
-#else
-#error "unsupported kernel configuration"
+#if GMX_SIMD_REAL_WIDTH == 4 && \
+    !((!defined GMX_DOUBLE && defined GMX_SIMD4_HAVE_FLOAT) || \
+    (defined GMX_DOUBLE && defined GMX_SIMD4_HAVE_DOUBLE))
+#define gmx_simd4_real_t    gmx_simd_real_t
+#define gmx_simd4_load_r    gmx_simd_load_r
+#define gmx_simd4_store_r   gmx_simd_store_r
+#define gmx_simd4_add_r     gmx_simd_add_r
+#define gmx_simd4_reduce_r  gmx_simd_reduce_r
 #endif
-#endif //HAVE_GMX_SUM_SIMD
 
 #ifdef UNROLLJ
 /* Add energy register to possibly multiple terms in the energy array */