/*
* This file is part of the GROMACS molecular simulation package.
*
- * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
- * Copyright (c) 2001-2012, The GROMACS Development Team
- * Copyright (c) 2012, by the GROMACS development team, led by
- * David van der Spoel, Berk Hess, Erik Lindahl, and including many
- * others, as listed in the AUTHORS file in the top-level source
- * directory and at http://www.gromacs.org.
+ * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
*
* GROMACS is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public License
* To help us fund GROMACS development, we humbly ask that you cite
* the research papers on the package. Check out http://www.gromacs.org.
*/
-#ifndef _nbnxn_kernel_sse_utils_h_
-#define _nbnxn_kernel_sse_utils_h_
+#ifndef _nbnxn_kernel_simd_utils_h_
+#define _nbnxn_kernel_simd_utils_h_
-/* This files contains all functions/macros for the SIMD kernels
- * which have explicit dependencies on the j-cluster size and/or SIMD-width.
+#include "config.h"
+
+#include "gromacs/legacyheaders/types/simple.h"
+
+/*! \brief Provides hardware-specific utility routines for the SIMD kernels.
+ *
+ * Defines all functions, typedefs, constants and macros that have
+ * explicit dependencies on the j-cluster size, precision, or SIMD
+ * width. This includes handling diagonal, Newton and topology
+ * exclusions.
+ *
* The functionality which depends on the j-cluster size is:
* LJ-parameter lookup
* force table lookup
* energy group pair energy storage
*/
-#ifdef GMX_X86_SSE2
-
-/* Transpose 2 double precision registers */
-#define GMX_MM_TRANSPOSE2_OP_PD(in0, in1, out0, out1) \
- { \
- out0 = _mm_unpacklo_pd(in0, in1); \
- out1 = _mm_unpackhi_pd(in0, in1); \
- }
-
-#if defined GMX_MM128_HERE || !defined GMX_DOUBLE
-/* Collect element 0 and 1 of the 4 inputs to out0 and out1, respectively */
-#define GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(in0, in1, in2, in3, out0, out1) \
- { \
- __m128 _c01, _c23; \
- _c01 = _mm_movelh_ps(in0, in1); \
- _c23 = _mm_movelh_ps(in2, in3); \
- out0 = _mm_shuffle_ps(_c01, _c23, _MM_SHUFFLE(2, 0, 2, 0)); \
- out1 = _mm_shuffle_ps(_c01, _c23, _MM_SHUFFLE(3, 1, 3, 1)); \
- }
-#else
-/* Collect element 0 and 1 of the 4 inputs to out0 and out1, respectively */
-#define GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(in0, in1, in2, in3, out0, out1) \
- { \
- __m256d _c01, _c23; \
- _c01 = _mm256_shuffle_pd(in0, in1, _MM_SHUFFLE(1, 0, 1, 0)); \
- _c23 = _mm256_shuffle_pd(in2, in3, _MM_SHUFFLE(1, 0, 1, 0)); \
- out0 = _mm256_shuffle_pd(_c01, _c23, _MM_SHUFFLE(2, 0, 2, 0)); \
- out1 = _mm256_shuffle_pd(_c01, _c23, _MM_SHUFFLE(3, 1, 3, 1)); \
- }
-#endif
-
-/* Collect element 2 of the 4 inputs to out */
-#define GMX_MM_SHUFFLE_4_PS_FIL2_TO_1_PS(in0, in1, in2, in3, out) \
- { \
- __m128 _c01, _c23; \
- _c01 = _mm_shuffle_ps(in0, in1, _MM_SHUFFLE(3, 2, 3, 2)); \
- _c23 = _mm_shuffle_ps(in2, in3, _MM_SHUFFLE(3, 2, 3, 2)); \
- out = _mm_shuffle_ps(_c01, _c23, _MM_SHUFFLE(2, 0, 2, 0)); \
- }
-
-#ifndef GMX_MM256_HERE
-#ifndef GMX_DOUBLE
-/* Sum the elements within each input register and store the sums in out */
-#define GMX_MM_TRANSPOSE_SUM4_PR(in0, in1, in2, in3, out) \
- { \
- _MM_TRANSPOSE4_PS(in0, in1, in2, in3); \
- in0 = _mm_add_ps(in0, in1); \
- in2 = _mm_add_ps(in2, in3); \
- out = _mm_add_ps(in0, in2); \
- }
-#else
-/* Sum the elements within each input register and store the sums in out */
-#define GMX_MM_TRANSPOSE_SUM2_PD(in0, in1, out) \
- { \
- GMX_MM_TRANSPOSE2_PD(in0, in1); \
- out = _mm_add_pd(in0, in1); \
- }
-#endif
-#else
-#ifndef GMX_DOUBLE
-/* Sum the elements within each input register and store the sums in out */
-#define GMX_MM_TRANSPOSE_SUM4_PR(in0, in1, in2, in3, out) \
- { \
- in0 = _mm256_hadd_ps(in0, in1); \
- in2 = _mm256_hadd_ps(in2, in3); \
- in1 = _mm256_hadd_ps(in0, in2); \
- out = _mm_add_ps(_mm256_castps256_ps128(in1), _mm256_extractf128_ps(in1, 1)); \
- }
-/* Sum the elements of halfs of each input register and store sums in out */
-#define GMX_MM_TRANSPOSE_SUM4H_PR(in0, in2, out) \
- { \
- in0 = _mm256_hadd_ps(in0, _mm256_setzero_ps()); \
- in2 = _mm256_hadd_ps(in2, _mm256_setzero_ps()); \
- in0 = _mm256_hadd_ps(in0, in2); \
- in2 = _mm256_permute_ps(in0, _MM_SHUFFLE(2, 3, 0, 1)); \
- out = _mm_add_ps(_mm256_castps256_ps128(in0), _mm256_extractf128_ps(in2, 1)); \
- }
-#else
-/* Sum the elements within each input register and store the sums in out */
-#define GMX_MM_TRANSPOSE_SUM4_PR(in0, in1, in2, in3, out) \
- { \
- in0 = _mm256_hadd_pd(in0, in1); \
- in2 = _mm256_hadd_pd(in2, in3); \
- out = _mm256_add_pd(_mm256_permute2f128_pd(in0, in2, 0x20), _mm256_permute2f128_pd(in0, in2, 0x31)); \
- }
-#endif
+#if !defined GMX_NBNXN_SIMD_2XNN && !defined GMX_NBNXN_SIMD_4XN
+#error "Must define an NBNxN kernel flavour before including NBNxN kernel utility functions"
#endif
-#ifdef GMX_MM128_HERE
+#ifdef GMX_SIMD_REFERENCE
-static inline __m128
-gmx_mm128_invsqrt_ps_single(__m128 x)
+/* Align a stack-based thread-local working array. */
+static gmx_inline int *
+prepare_table_load_buffer(const int gmx_unused *array)
{
- const __m128 half = _mm_set_ps(0.5, 0.5, 0.5, 0.5);
- const __m128 three = _mm_set_ps(3.0, 3.0, 3.0, 3.0);
-
- __m128 lu = _mm_rsqrt_ps(x);
-
- return _mm_mul_ps(half, _mm_mul_ps(_mm_sub_ps(three, _mm_mul_ps(_mm_mul_ps(lu, lu), x)), lu));
+ return NULL;
}
-/* Do 2 double precision invsqrt operations.
- * Doing the SIMD rsqrt and the first Newton Raphson iteration
- * in single precision gives full double precision accuracy.
- * The speed is more than double that of two gmx_mm_invsqrt_pd calls.
- */
-#define GMX_MM128_INVSQRT2_PD(in0, in1, out0, out1) \
- { \
- const __m128d half = _mm_set1_pd(0.5); \
- const __m128d three = _mm_set1_pd(3.0); \
- __m128 s, ir; \
- __m128d lu0, lu1; \
- \
- s = _mm_movelh_ps(_mm_cvtpd_ps(in0), _mm_cvtpd_ps(in1)); \
- ir = gmx_mm128_invsqrt_ps_single(s); \
- lu0 = _mm_cvtps_pd(ir); \
- lu1 = _mm_cvtps_pd(_mm_movehl_ps(ir, ir)); \
- out0 = _mm_mul_pd(half, _mm_mul_pd(_mm_sub_pd(three, _mm_mul_pd(_mm_mul_pd(lu0, lu0), in0)), lu0)); \
- out1 = _mm_mul_pd(half, _mm_mul_pd(_mm_sub_pd(three, _mm_mul_pd(_mm_mul_pd(lu1, lu1), in1)), lu1)); \
- }
-
-#define GMX_MM_INVSQRT2_PD GMX_MM128_INVSQRT2_PD
+#include "gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_simd_utils_ref.h"
-#endif
-
-#ifdef GMX_MM256_HERE
-
-static inline __m256
-gmx_mm256_invsqrt_ps_single(__m256 x)
-{
- const __m256 half = _mm256_set_ps(0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5);
- const __m256 three = _mm256_set_ps(3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0);
-
- __m256 lu = _mm256_rsqrt_ps(x);
+#else /* GMX_SIMD_REFERENCE */
- return _mm256_mul_ps(half, _mm256_mul_ps(_mm256_sub_ps(three, _mm256_mul_ps(_mm256_mul_ps(lu, lu), x)), lu));
-}
+#if defined GMX_TARGET_X86 && !defined __MIC__
+/* Include x86 SSE2 compatible SIMD functions */
-/* Do 4 double precision invsqrt operations.
- * Doing the SIMD rsqrt and the first Newton Raphson iteration
- * in single precision gives full double precision accuracy.
+/* Set the stride for the lookup of the two LJ parameters from their
+ * (padded) array. We use the minimum supported SIMD memory alignment.
*/
-#define GMX_MM256_INVSQRT2_PD(in0, in1, out0, out1) \
- { \
- const __m256d half = _mm256_set1_pd(0.5); \
- const __m256d three = _mm256_set1_pd(3.0); \
- __m256 s, ir; \
- __m256d lu0, lu1; \
- \
- s = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm256_cvtpd_ps(in0)), _mm256_cvtpd_ps(in1), 1); \
- ir = gmx_mm256_invsqrt_ps_single(s); \
- lu0 = _mm256_cvtps_pd(_mm256_castps256_ps128(ir)); \
- lu1 = _mm256_cvtps_pd(_mm256_extractf128_ps(ir, 1)); \
- out0 = _mm256_mul_pd(half, _mm256_mul_pd(_mm256_sub_pd(three, _mm256_mul_pd(_mm256_mul_pd(lu0, lu0), in0)), lu0)); \
- out1 = _mm256_mul_pd(half, _mm256_mul_pd(_mm256_sub_pd(three, _mm256_mul_pd(_mm256_mul_pd(lu1, lu1), in1)), lu1)); \
- }
-
-#define GMX_MM_INVSQRT2_PD GMX_MM256_INVSQRT2_PD
-
-#endif
-
-/* Force and energy table load and interpolation routines */
-
-#if defined GMX_MM128_HERE && !defined GMX_DOUBLE
-
-#define load_lj_pair_params(nbfp, type, aj, c6_SSE, c12_SSE) \
- { \
- gmx_mm_pr clj_SSE[UNROLLJ]; \
- int p; \
- \
- for (p = 0; p < UNROLLJ; p++) \
- { \
- /* Here we load 4 aligned floats, but we need just 2 */ \
- clj_SSE[p] = gmx_load_pr(nbfp+type[aj+p]*NBFP_STRIDE); \
- } \
- GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(clj_SSE[0], clj_SSE[1], clj_SSE[2], clj_SSE[3], c6_SSE, c12_SSE); \
- }
-
+#if defined GMX_DOUBLE
+static const int nbfp_stride = 2;
+#else
+static const int nbfp_stride = 4;
#endif
-#if defined GMX_MM256_HERE && !defined GMX_DOUBLE
-
-/* Put two 128-bit 4-float registers into one 256-bit 8-float register */
-#define GMX_2_MM_TO_M256(in0, in1, out) \
- { \
- out = _mm256_insertf128_ps(_mm256_castps128_ps256(in0), in1, 1); \
- }
-
-#define load_lj_pair_params(nbfp, type, aj, c6_SSE, c12_SSE) \
- { \
- __m128 clj_SSE[UNROLLJ], c6t_SSE[2], c12t_SSE[2]; \
- int p; \
- \
- for (p = 0; p < UNROLLJ; p++) \
- { \
- /* Here we load 4 aligned floats, but we need just 2 */ \
- clj_SSE[p] = _mm_load_ps(nbfp+type[aj+p]*NBFP_STRIDE); \
- } \
- GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(clj_SSE[0], clj_SSE[1], clj_SSE[2], clj_SSE[3], c6t_SSE[0], c12t_SSE[0]); \
- GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(clj_SSE[4], clj_SSE[5], clj_SSE[6], clj_SSE[7], c6t_SSE[1], c12t_SSE[1]); \
- \
- GMX_2_MM_TO_M256(c6t_SSE[0], c6t_SSE[1], c6_SSE); \
- GMX_2_MM_TO_M256(c12t_SSE[0], c12t_SSE[1], c12_SSE); \
- }
-
-#define load_lj_pair_params2(nbfp0, nbfp1, type, aj, c6_SSE, c12_SSE) \
- { \
- __m128 clj_SSE0[UNROLLJ], clj_SSE1[UNROLLJ], c6t_SSE[2], c12t_SSE[2]; \
- int p; \
- \
- for (p = 0; p < UNROLLJ; p++) \
- { \
- /* Here we load 4 aligned floats, but we need just 2 */ \
- clj_SSE0[p] = _mm_load_ps(nbfp0+type[aj+p]*NBFP_STRIDE); \
- } \
- for (p = 0; p < UNROLLJ; p++) \
- { \
- /* Here we load 4 aligned floats, but we need just 2 */ \
- clj_SSE1[p] = _mm_load_ps(nbfp1+type[aj+p]*NBFP_STRIDE); \
- } \
- GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(clj_SSE0[0], clj_SSE0[1], clj_SSE0[2], clj_SSE0[3], c6t_SSE[0], c12t_SSE[0]); \
- GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(clj_SSE1[0], clj_SSE1[1], clj_SSE1[2], clj_SSE1[3], c6t_SSE[1], c12t_SSE[1]); \
- \
- GMX_2_MM_TO_M256(c6t_SSE[0], c6t_SSE[1], c6_SSE); \
- GMX_2_MM_TO_M256(c12t_SSE[0], c12t_SSE[1], c12_SSE); \
- }
-
+/* Align a stack-based thread-local working array. Table loads on
+ * 256-bit AVX use the array, but other implementations do not.
+ */
+static gmx_inline int *
+prepare_table_load_buffer(int gmx_unused *array)
+{
+#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_MM128_HERE && defined GMX_DOUBLE
-
-#define load_lj_pair_params(nbfp, type, aj, c6_SSE, c12_SSE) \
- { \
- gmx_mm_pr clj_SSE[UNROLLJ]; \
- int p; \
- \
- for (p = 0; p < UNROLLJ; p++) \
- { \
- clj_SSE[p] = gmx_load_pr(nbfp+type[aj+p]*NBFP_STRIDE); \
- } \
- GMX_MM_TRANSPOSE2_OP_PD(clj_SSE[0], clj_SSE[1], c6_SSE, c12_SSE); \
- }
-
+#ifdef GMX_DOUBLE
+#if GMX_SIMD_REAL_WIDTH == 2
+#include "gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_simd_utils_x86_128d.h"
+#else
+#include "gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_simd_utils_x86_256d.h"
#endif
-
-#if defined GMX_MM256_HERE && defined GMX_DOUBLE
-
-#define load_lj_pair_params(nbfp, type, aj, c6_SSE, c12_SSE) \
- { \
- __m128d clj_SSE[UNROLLJ], c6t_SSE[2], c12t_SSE[2]; \
- int p; \
- \
- for (p = 0; p < UNROLLJ; p++) \
- { \
- clj_SSE[p] = _mm_load_pd(nbfp+type[aj+p]*NBFP_STRIDE); \
- } \
- GMX_MM_TRANSPOSE2_OP_PD(clj_SSE[0], clj_SSE[1], c6t_SSE[0], c12t_SSE[0]); \
- GMX_MM_TRANSPOSE2_OP_PD(clj_SSE[2], clj_SSE[3], c6t_SSE[1], c12t_SSE[1]); \
- GMX_2_M128D_TO_M256D(c6t_SSE[0], c6t_SSE[1], c6_SSE); \
- GMX_2_M128D_TO_M256D(c12t_SSE[0], c12t_SSE[1], c12_SSE); \
- }
-
+#else /* GMX_DOUBLE */
+/* In single precision aligned FDV0 table loads are optimal */
+#define TAB_FDV0
+#if GMX_SIMD_REAL_WIDTH == 4
+#include "gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_simd_utils_x86_128s.h"
+#else
+#include "gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_simd_utils_x86_256s.h"
#endif
+#endif /* GMX_DOUBLE */
+#else /* GMX_TARGET_X86 && !__MIC__ */
-/* The load_table functions below are performance critical.
- * The routines issue UNROLLI*UNROLLJ _mm_load_ps calls.
- * As these all have latencies, scheduling is crucial.
- * The Intel compilers and CPUs seem to do a good job at this.
- * But AMD CPUs perform significantly worse with gcc than with icc.
- * Performance is improved a bit by using the extract function UNROLLJ times,
- * instead of doing an _mm_store_si128 for every i-particle.
- * This is only faster when we use FDV0 formatted tables, where we also need
- * to multiple the index by 4, which can be done by a SIMD bit shift.
- * With single precision AVX, 8 extracts are much slower than 1 store.
- * Because of this, the load_table_f macro always takes the ti parameter,
- * but it is only used with AVX.
- */
-
-#if defined GMX_MM128_HERE && !defined GMX_DOUBLE
-
-#define load_table_f(tab_coul_FDV0, ti_SSE, ti, ctab0_SSE, ctab1_SSE) \
- { \
- int idx[4]; \
- __m128 ctab_SSE[4]; \
- \
- /* Table has 4 entries, left-shift index by 2 */ \
- ti_SSE = _mm_slli_epi32(ti_SSE, 2); \
- /* Without SSE4.1 the extract macro needs an immediate: unroll */ \
- idx[0] = gmx_mm_extract_epi32(ti_SSE, 0); \
- ctab_SSE[0] = _mm_load_ps(tab_coul_FDV0+idx[0]); \
- idx[1] = gmx_mm_extract_epi32(ti_SSE, 1); \
- ctab_SSE[1] = _mm_load_ps(tab_coul_FDV0+idx[1]); \
- idx[2] = gmx_mm_extract_epi32(ti_SSE, 2); \
- ctab_SSE[2] = _mm_load_ps(tab_coul_FDV0+idx[2]); \
- idx[3] = gmx_mm_extract_epi32(ti_SSE, 3); \
- ctab_SSE[3] = _mm_load_ps(tab_coul_FDV0+idx[3]); \
- \
- /* Shuffle the force table entries to a convenient order */ \
- GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(ctab_SSE[0], ctab_SSE[1], ctab_SSE[2], ctab_SSE[3], ctab0_SSE, ctab1_SSE); \
- }
-
-#define load_table_f_v(tab_coul_FDV0, ti_SSE, ti, ctab0_SSE, ctab1_SSE, ctabv_SSE) \
- { \
- int idx[4]; \
- __m128 ctab_SSE[4]; \
- \
- /* Table has 4 entries, left-shift index by 2 */ \
- ti_SSE = _mm_slli_epi32(ti_SSE, 2); \
- /* Without SSE4.1 the extract macro needs an immediate: unroll */ \
- idx[0] = gmx_mm_extract_epi32(ti_SSE, 0); \
- ctab_SSE[0] = _mm_load_ps(tab_coul_FDV0+idx[0]); \
- idx[1] = gmx_mm_extract_epi32(ti_SSE, 1); \
- ctab_SSE[1] = _mm_load_ps(tab_coul_FDV0+idx[1]); \
- idx[2] = gmx_mm_extract_epi32(ti_SSE, 2); \
- ctab_SSE[2] = _mm_load_ps(tab_coul_FDV0+idx[2]); \
- idx[3] = gmx_mm_extract_epi32(ti_SSE, 3); \
- ctab_SSE[3] = _mm_load_ps(tab_coul_FDV0+idx[3]); \
- \
- /* Shuffle the force table entries to a convenient order */ \
- GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(ctab_SSE[0], ctab_SSE[1], ctab_SSE[2], ctab_SSE[3], ctab0_SSE, ctab1_SSE); \
- /* Shuffle the energy table entries to a convenient order */ \
- GMX_MM_SHUFFLE_4_PS_FIL2_TO_1_PS(ctab_SSE[0], ctab_SSE[1], ctab_SSE[2], ctab_SSE[3], ctabv_SSE); \
- }
-
+#if GMX_SIMD_REAL_WIDTH > 4
+/* For width>4 we use unaligned loads. And thus we can use the minimal stride */
+static const int nbfp_stride = 2;
+#else
+static const int nbfp_stride = GMX_SIMD_REAL_WIDTH;
#endif
-#if defined GMX_MM256_HERE && !defined GMX_DOUBLE
-
-#define load_table_f(tab_coul_FDV0, ti_SSE, ti, ctab0_SSE, ctab1_SSE) \
- { \
- __m128 ctab_SSE[8], ctabt_SSE[4]; \
- int j; \
- \
- /* Bit shifting would be faster, but AVX doesn't support that */ \
- _mm256_store_si256((__m256i *)ti, ti_SSE); \
- for (j = 0; j < 8; j++) \
- { \
- ctab_SSE[j] = _mm_load_ps(tab_coul_FDV0+ti[j]*4); \
- } \
- GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(ctab_SSE[0], ctab_SSE[1], ctab_SSE[2], ctab_SSE[3], ctabt_SSE[0], ctabt_SSE[2]); \
- GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(ctab_SSE[4], ctab_SSE[5], ctab_SSE[6], ctab_SSE[7], ctabt_SSE[1], ctabt_SSE[3]); \
- \
- GMX_2_MM_TO_M256(ctabt_SSE[0], ctabt_SSE[1], ctab0_SSE); \
- GMX_2_MM_TO_M256(ctabt_SSE[2], ctabt_SSE[3], ctab1_SSE); \
- }
-
-#define load_table_f_v(tab_coul_FDV0, ti_SSE, ti, ctab0_SSE, ctab1_SSE, ctabv_SSE) \
- { \
- __m128 ctab_SSE[8], ctabt_SSE[4], ctabvt_SSE[2]; \
- int j; \
- \
- /* Bit shifting would be faster, but AVX doesn't support that */ \
- _mm256_store_si256((__m256i *)ti, ti_SSE); \
- for (j = 0; j < 8; j++) \
- { \
- ctab_SSE[j] = _mm_load_ps(tab_coul_FDV0+ti[j]*4); \
- } \
- GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(ctab_SSE[0], ctab_SSE[1], ctab_SSE[2], ctab_SSE[3], ctabt_SSE[0], ctabt_SSE[2]); \
- GMX_MM_SHUFFLE_4_PS_FIL01_TO_2_PS(ctab_SSE[4], ctab_SSE[5], ctab_SSE[6], ctab_SSE[7], ctabt_SSE[1], ctabt_SSE[3]); \
- \
- GMX_2_MM_TO_M256(ctabt_SSE[0], ctabt_SSE[1], ctab0_SSE); \
- GMX_2_MM_TO_M256(ctabt_SSE[2], ctabt_SSE[3], ctab1_SSE); \
- \
- GMX_MM_SHUFFLE_4_PS_FIL2_TO_1_PS(ctab_SSE[0], ctab_SSE[1], ctab_SSE[2], ctab_SSE[3], ctabvt_SSE[0]); \
- GMX_MM_SHUFFLE_4_PS_FIL2_TO_1_PS(ctab_SSE[4], ctab_SSE[5], ctab_SSE[6], ctab_SSE[7], ctabvt_SSE[1]); \
- \
- GMX_2_MM_TO_M256(ctabvt_SSE[0], ctabvt_SSE[1], ctabv_SSE); \
- }
-
+/* We use the FDV0 table layout when we can use aligned table loads */
+#if GMX_SIMD_REAL_WIDTH == 4
+#define TAB_FDV0
#endif
-#if defined GMX_MM128_HERE && defined GMX_DOUBLE
-
-#define load_table_f(tab_coul_F, ti_SSE, ti, ctab0_SSE, ctab1_SSE) \
- { \
- int idx[2]; \
- __m128d ctab_SSE[2]; \
- \
- /* Without SSE4.1 the extract macro needs an immediate: unroll */ \
- idx[0] = gmx_mm_extract_epi32(ti_SSE, 0); \
- ctab_SSE[0] = _mm_loadu_pd(tab_coul_F+idx[0]); \
- idx[1] = gmx_mm_extract_epi32(ti_SSE, 1); \
- ctab_SSE[1] = _mm_loadu_pd(tab_coul_F+idx[1]); \
- \
- /* Shuffle the force table entries to a convenient order */ \
- GMX_MM_TRANSPOSE2_OP_PD(ctab_SSE[0], ctab_SSE[1], ctab0_SSE, ctab1_SSE); \
- /* The second force table entry should contain the difference */ \
- ctab1_SSE = _mm_sub_pd(ctab1_SSE, ctab0_SSE); \
- }
-
-#define load_table_f_v(tab_coul_F, tab_coul_V, ti_SSE, ti, ctab0_SSE, ctab1_SSE, ctabv_SSE) \
- { \
- int idx[2]; \
- __m128d ctab_SSE[4]; \
- \
- /* Without SSE4.1 the extract macro needs an immediate: unroll */ \
- idx[0] = gmx_mm_extract_epi32(ti_SSE, 0); \
- ctab_SSE[0] = _mm_loadu_pd(tab_coul_F+idx[0]); \
- idx[1] = gmx_mm_extract_epi32(ti_SSE, 1); \
- ctab_SSE[1] = _mm_loadu_pd(tab_coul_F+idx[1]); \
- \
- /* Shuffle the force table entries to a convenient order */ \
- GMX_MM_TRANSPOSE2_OP_PD(ctab_SSE[0], ctab_SSE[1], ctab0_SSE, ctab1_SSE); \
- /* The second force table entry should contain the difference */ \
- ctab1_SSE = _mm_sub_pd(ctab1_SSE, ctab0_SSE); \
- \
- ctab_SSE[2] = _mm_loadu_pd(tab_coul_V+idx[0]); \
- ctab_SSE[3] = _mm_loadu_pd(tab_coul_V+idx[1]); \
- \
- /* Shuffle the energy table entries to a single register */ \
- ctabv_SSE = _mm_shuffle_pd(ctab_SSE[2], ctab_SSE[3], _MM_SHUFFLE2(0, 0)); \
- }
+#ifdef GMX_SIMD_IBM_QPX
+#include "gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_simd_utils_ibm_qpx.h"
+#endif /* GMX_SIMD_IBM_QPX */
+#ifdef __MIC__
+#include "gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_simd_utils_x86_mic.h"
#endif
-#if defined GMX_MM256_HERE && defined GMX_DOUBLE
-
-/* Put two 128-bit 2-double registers into one 256-bit 4-ouble register */
-#define GMX_2_M128D_TO_M256D(in0, in1, out) \
- { \
- out = _mm256_insertf128_pd(_mm256_castpd128_pd256(in0), in1, 1); \
- }
-
-#define load_table_f(tab_coul_F, ti_SSE, ti, ctab0_SSE, ctab1_SSE) \
- { \
- __m128d ctab_SSE[4], tr_SSE[4]; \
- int j; \
- \
- _mm_store_si128((__m128i *)ti, ti_SSE); \
- for (j = 0; j < 4; j++) \
- { \
- ctab_SSE[j] = _mm_loadu_pd(tab_coul_F+ti[j]); \
- } \
- /* Shuffle the force table entries to a convenient order */ \
- GMX_MM_TRANSPOSE2_OP_PD(ctab_SSE[0], ctab_SSE[1], tr_SSE[0], tr_SSE[1]); \
- GMX_MM_TRANSPOSE2_OP_PD(ctab_SSE[2], ctab_SSE[3], tr_SSE[2], tr_SSE[3]); \
- GMX_2_M128D_TO_M256D(tr_SSE[0], tr_SSE[2], ctab0_SSE); \
- GMX_2_M128D_TO_M256D(tr_SSE[1], tr_SSE[3], ctab1_SSE); \
- /* The second force table entry should contain the difference */ \
- ctab1_SSE = _mm256_sub_pd(ctab1_SSE, ctab0_SSE); \
- }
+#endif /* GMX_TARGET_X86 && !__MIC__ */
-#define load_table_f_v(tab_coul_F, tab_coul_V, ti_SSE, ti, ctab0_SSE, ctab1_SSE, ctabv_SSE) \
- { \
- __m128d ctab_SSE[8], tr_SSE[4]; \
- int j; \
- \
- _mm_store_si128((__m128i *)ti, ti_SSE); \
- for (j = 0; j < 4; j++) \
- { \
- ctab_SSE[j] = _mm_loadu_pd(tab_coul_F+ti[j]); \
- } \
- /* Shuffle the force table entries to a convenient order */ \
- GMX_MM_TRANSPOSE2_OP_PD(ctab_SSE[0], ctab_SSE[1], tr_SSE[0], tr_SSE[1]); \
- GMX_MM_TRANSPOSE2_OP_PD(ctab_SSE[2], ctab_SSE[3], tr_SSE[2], tr_SSE[3]); \
- GMX_2_M128D_TO_M256D(tr_SSE[0], tr_SSE[2], ctab0_SSE); \
- GMX_2_M128D_TO_M256D(tr_SSE[1], tr_SSE[3], ctab1_SSE); \
- /* The second force table entry should contain the difference */ \
- ctab1_SSE = _mm256_sub_pd(ctab1_SSE, ctab0_SSE); \
- \
- for (j = 0; j < 4; j++) \
- { \
- ctab_SSE[4+j] = _mm_loadu_pd(tab_coul_V+ti[j]); \
- } \
- /* Shuffle the energy table entries to a single register */ \
- GMX_2_M128D_TO_M256D(_mm_shuffle_pd(ctab_SSE[4], ctab_SSE[5], _MM_SHUFFLE2(0, 0)), _mm_shuffle_pd(ctab_SSE[6], ctab_SSE[7], _MM_SHUFFLE2(0, 0)), ctabv_SSE); \
- }
+#endif /* GMX_SIMD_REFERENCE */
+/* If the simd width is 4, but simd4 instructions are not defined,
+ * reuse the simd real type and the four instructions we need.
+ */
+#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
-
-/* Add energy register to possibly multiple terms in the energy array.
- * This function is the same for SSE/AVX single/double.
- */
-static inline void add_ener_grp(gmx_mm_pr e_SSE, real *v, const int *offset_jj)
+#ifdef UNROLLJ
+/* Add energy register to possibly multiple terms in the energy array */
+static gmx_inline void add_ener_grp(gmx_simd_real_t e_S, real *v, const int *offset_jj)
{
int jj;
*/
for (jj = 0; jj < (UNROLLJ/2); jj++)
{
- gmx_mm_pr v_SSE;
+ gmx_simd_real_t v_S;
- v_SSE = gmx_load_pr(v+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE);
- gmx_store_pr(v+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE, gmx_add_pr(v_SSE, e_SSE));
+ v_S = gmx_simd_load_r(v+offset_jj[jj]+jj*GMX_SIMD_REAL_WIDTH);
+ gmx_simd_store_r(v+offset_jj[jj]+jj*GMX_SIMD_REAL_WIDTH, gmx_simd_add_r(v_S, e_S));
}
}
+#endif
-#if defined GMX_X86_AVX_256 && GMX_SIMD_WIDTH_HERE == 8
-/* As add_ener_grp above, but for two groups of UNROLLJ/2 stored in
+#if defined GMX_NBNXN_SIMD_2XNN && defined UNROLLJ
+/* As add_ener_grp, but for two groups of UNROLLJ/2 stored in
* a single SIMD register.
*/
-static inline void add_ener_grp_halves(gmx_mm_pr e_SSE,
- real *v0, real *v1, const int *offset_jj)
+static gmx_inline void
+add_ener_grp_halves(gmx_simd_real_t e_S, real *v0, real *v1, const int *offset_jj)
{
- gmx_mm_hpr e_SSE0, e_SSE1;
+ gmx_mm_hpr e_S0, e_S1;
int jj;
- e_SSE0 = _mm256_extractf128_ps(e_SSE, 0);
- e_SSE1 = _mm256_extractf128_ps(e_SSE, 1);
+ gmx_pr_to_2hpr(e_S, &e_S0, &e_S1);
for (jj = 0; jj < (UNROLLJ/2); jj++)
{
- gmx_mm_hpr v_SSE;
+ gmx_mm_hpr v_S;
- v_SSE = gmx_load_hpr(v0+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE/2);
- gmx_store_hpr(v0+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE/2, gmx_add_hpr(v_SSE, e_SSE0));
+ gmx_load_hpr(&v_S, v0+offset_jj[jj]+jj*GMX_SIMD_REAL_WIDTH/2);
+ gmx_store_hpr(v0+offset_jj[jj]+jj*GMX_SIMD_REAL_WIDTH/2, gmx_add_hpr(v_S, e_S0));
}
for (jj = 0; jj < (UNROLLJ/2); jj++)
{
- gmx_mm_hpr v_SSE;
+ gmx_mm_hpr v_S;
- v_SSE = gmx_load_hpr(v1+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE/2);
- gmx_store_hpr(v1+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE/2, gmx_add_hpr(v_SSE, e_SSE1));
+ gmx_load_hpr(&v_S, v1+offset_jj[jj]+jj*GMX_SIMD_REAL_WIDTH/2);
+ gmx_store_hpr(v1+offset_jj[jj]+jj*GMX_SIMD_REAL_WIDTH/2, gmx_add_hpr(v_S, e_S1));
}
}
#endif
-#endif /* GMX_X86_SSE2 */
-
-#endif /* _nbnxn_kernel_sse_utils_h_ */
+#endif /* _nbnxn_kernel_simd_utils_h_ */
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff