Skip to content

GitLab

Commit 79fd69c3 authored by peastman's avatar peastman
Browse files

Merge pull request #182 from peastman/master 

Lots of optimizations to CPU platform
parents b5da7380 062a2db0
Hide whitespace changes
Inline Side-by-side
Showing with 1029 additions and 350 deletions
CMakeLists.txt
View file @ 79fd69c3
......@@ -252,6 +252,7 @@ FOREACH(subdir ${OPENMM_SOURCE_SUBDIRS})
## OpenMM was previously installed there.
INCLUDE_DIRECTORIES(BEFORE ${CMAKE_CURRENT_SOURCE_DIR}/${subdir}/include)
ENDFOREACH(subdir)
SET_SOURCE_FILES_PROPERTIES(${CMAKE_SOURCE_DIR}/libraries/sfmt/src/SFMT.cpp PROPERTIES COMPILE_FLAGS "-msse2 -DHAVE_SSE2=1")
# If API wrappers are being generated, and add them to the build.
FIND_PROGRAM(GCCXML_PATH gccxml PATH
......
libraries/sfmt/include/sfmt/SFMT-sse2.h
View file @ 79fd69c3
......@@ -49,22 +49,23 @@ PRE_ALWAYS static __m128i mm_recursion(__m128i *a, __m128i *b,
* This function fills the internal state array with pseudorandom
* integers.
*/
inline static void gen_rand_all(void) {
inline static void gen_rand_all(SFMT& sfmt) {
int i;
__m128i r, r1, r2, mask;
mask = _mm_set_epi32(MSK4, MSK3, MSK2, MSK1);
r1 = _mm_load_si128(&sfmt[N - 2].si);
r2 = _mm_load_si128(&sfmt[N - 1].si);
SFMTData& data = *sfmt.data;
r1 = _mm_load_si128(&data.sfmt[N - 2].si);
r2 = _mm_load_si128(&data.sfmt[N - 1].si);
for (i = 0; i < N - POS1; i++) {
r = mm_recursion(&sfmt[i].si, &sfmt[i + POS1].si, r1, r2, mask);
_mm_store_si128(&sfmt[i].si, r);
r = mm_recursion(&data.sfmt[i].si, &data.sfmt[i + POS1].si, r1, r2, mask);
_mm_store_si128(&data.sfmt[i].si, r);
r1 = r2;
r2 = r;
}
for (; i < N; i++) {
r = mm_recursion(&sfmt[i].si, &sfmt[i + POS1 - N].si, r1, r2, mask);
_mm_store_si128(&sfmt[i].si, r);
r = mm_recursion(&data.sfmt[i].si, &data.sfmt[i + POS1 - N].si, r1, r2, mask);
_mm_store_si128(&data.sfmt[i].si, r);
r1 = r2;
r2 = r;
}
......@@ -77,21 +78,22 @@ inline static void gen_rand_all(void) {
* @param array an 128-bit array to be filled by pseudorandom numbers.
* @param size number of 128-bit pesudorandom numbers to be generated.
*/
inline static void gen_rand_array(w128_t *array, int size) {
inline static void gen_rand_array(w128_t *array, int size, SFMT& sfmt) {
int i, j;
__m128i r, r1, r2, mask;
mask = _mm_set_epi32(MSK4, MSK3, MSK2, MSK1);
r1 = _mm_load_si128(&sfmt[N - 2].si);
r2 = _mm_load_si128(&sfmt[N - 1].si);
SFMTData& data = *sfmt.data;
r1 = _mm_load_si128(&data.sfmt[N - 2].si);
r2 = _mm_load_si128(&data.sfmt[N - 1].si);
for (i = 0; i < N - POS1; i++) {
r = mm_recursion(&sfmt[i].si, &sfmt[i + POS1].si, r1, r2, mask);
r = mm_recursion(&data.sfmt[i].si, &data.sfmt[i + POS1].si, r1, r2, mask);
_mm_store_si128(&array[i].si, r);
r1 = r2;
r2 = r;
}
for (; i < N; i++) {
r = mm_recursion(&sfmt[i].si, &array[i + POS1 - N].si, r1, r2, mask);
r = mm_recursion(&data.sfmt[i].si, &array[i + POS1 - N].si, r1, r2, mask);
_mm_store_si128(&array[i].si, r);
r1 = r2;
r2 = r;
......@@ -106,13 +108,13 @@ inline static void gen_rand_array(w128_t *array, int size) {
}
for (j = 0; j < 2 * N - size; j++) {
r = _mm_load_si128(&array[j + size - N].si);
_mm_store_si128(&sfmt[j].si, r);
_mm_store_si128(&data.sfmt[j].si, r);
}
for (; i < size; i++) {
r = mm_recursion(&array[i - N].si, &array[i + POS1 - N].si, r1, r2,
mask);
_mm_store_si128(&array[i].si, r);
_mm_store_si128(&sfmt[j++].si, r);
_mm_store_si128(&data.sfmt[j++].si, r);
r1 = r2;
r2 = r;
}
......
libraries/sfmt/src/SFMT.cpp
View file @ 79fd69c3
......@@ -144,9 +144,9 @@ inline static void swap(w128_t *array, int size);
#endif
#if defined(HAVE_ALTIVEC)
#include "SFMT-alti.h"
#include "sfmt/SFMT-alti.h"
#elif defined(HAVE_SSE2)
#include "SFMT-sse2.h"
#include "sfmt/SFMT-sse2.h"
#endif
/**
......
openmmapi/include/openmm/internal/vectorize.h 0 → 100644
View file @ 79fd69c3
#ifndef OPENMM_VECTORIZE_H_
#define OPENMM_VECTORIZE_H_
/* -------------------------------------------------------------------------- *
* OpenMM *
* -------------------------------------------------------------------------- *
* This is part of the OpenMM molecular simulation toolkit originating from *
* Simbios, the NIH National Center for Physics-Based Simulation of *
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2013 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
* Permission is hereby granted, free of charge, to any person obtaining a *
* copy of this software and associated documentation files (the "Software"), *
* to deal in the Software without restriction, including without limitation *
* the rights to use, copy, modify, merge, publish, distribute, sublicense, *
* and/or sell copies of the Software, and to permit persons to whom the *
* Software is furnished to do so, subject to the following conditions: *
* *
* The above copyright notice and this permission notice shall be included in *
* all copies or substantial portions of the Software. *
* *
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR *
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, *
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL *
* THE AUTHORS, CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, *
* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR *
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE *
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
#include <smmintrin.h>
// This file defines classes and functions to simplify vectorizing code with SSE.
class ivec4;
/**
* A four element vector of floats.
*/
class fvec4 {
public:
__m128 val;
fvec4() {}
fvec4(float v) : val(_mm_set1_ps(v)) {}
fvec4(float v1, float v2, float v3, float v4) : val(_mm_set_ps(v4, v3, v2, v1)) {}
fvec4(__m128 v) : val(v) {}
fvec4(const float* v) : val(_mm_loadu_ps(v)) {}
operator __m128() const {
return val;
}
float operator[](int i) const {
int resultBits = _mm_extract_ps(val, i);
return *((float*) &resultBits);
}
void store(float* v) const {
_mm_storeu_ps(v, val);
}
fvec4 operator+(fvec4 other) const {
return _mm_add_ps(val, other);
}
fvec4 operator-(fvec4 other) const {
return _mm_sub_ps(val, other);
}
fvec4 operator*(fvec4 other) const {
return _mm_mul_ps(val, other);
}
fvec4 operator/(fvec4 other) const {
return _mm_div_ps(val, other);
}
void operator+=(fvec4 other) {
val = _mm_add_ps(val, other);
}
void operator-=(fvec4 other) {
val = _mm_sub_ps(val, other);
}
void operator*=(fvec4 other) {
val = _mm_mul_ps(val, other);
}
void operator/=(fvec4 other) {
val = _mm_div_ps(val, other);
}
fvec4 operator-() const {
return _mm_sub_ps(_mm_set1_ps(0.0f), val);
}
fvec4 operator&(fvec4 other) const {
return _mm_and_ps(val, other);
}
fvec4 operator==(fvec4 other) const {
return _mm_cmpeq_ps(val, other);
}
fvec4 operator!=(fvec4 other) const {
return _mm_cmpneq_ps(val, other);
}
fvec4 operator>(fvec4 other) const {
return _mm_cmpgt_ps(val, other);
}
fvec4 operator<(fvec4 other) const {
return _mm_cmplt_ps(val, other);
}
fvec4 operator>=(fvec4 other) const {
return _mm_cmpge_ps(val, other);
}
fvec4 operator<=(fvec4 other) const {
return _mm_cmple_ps(val, other);
}
operator ivec4() const;
};
/**
* A four element vector of ints.
*/
class ivec4 {
public:
__m128i val;
ivec4() {}
ivec4(int v) : val(_mm_set1_epi32(v)) {}
ivec4(int v1, int v2, int v3, int v4) : val(_mm_set_epi32(v4, v3, v2, v1)) {}
ivec4(__m128i v) : val(v) {}
ivec4(const int* v) : val(_mm_loadu_si128((const __m128i*) v)) {}
operator __m128i() const {
return val;
}
int operator[](int i) const {
return _mm_extract_epi32(val, i);
}
void store(int* v) const {
_mm_storeu_si128((__m128i*) v, val);
}
ivec4 operator+(ivec4 other) const {
return _mm_add_epi32(val, other);
}
ivec4 operator-(ivec4 other) const {
return _mm_sub_epi32(val, other);
}
ivec4 operator*(ivec4 other) const {
return _mm_mul_epi32(val, other);
}
void operator+=(ivec4 other) {
val = _mm_add_epi32(val, other);
}
void operator-=(ivec4 other) {
val = _mm_sub_epi32(val, other);
}
void operator*=(ivec4 other) {
val = _mm_mul_epi32(val, other);
}
ivec4 operator-() const {
return _mm_sub_epi32(_mm_set1_epi32(0), val);
}
ivec4 operator&(ivec4 other) const {
return _mm_and_si128(val, other);
}
ivec4 operator==(ivec4 other) const {
return _mm_cmpeq_epi32(val, other);
}
operator fvec4() const;
};
// Conversion operators.
inline fvec4::operator ivec4() const {
return _mm_cvttps_epi32(val);
}
inline ivec4::operator fvec4() const {
return _mm_cvtepi32_ps(val);
}
// Functions that operate on fvec4s.
static inline fvec4 floor(fvec4 v) {
return fvec4(_mm_floor_ps(v.val));
}
static inline fvec4 ceil(fvec4 v) {
return fvec4(_mm_ceil_ps(v.val));
}
static inline fvec4 round(fvec4 v) {
return fvec4(_mm_round_ps(v.val, _MM_FROUND_TO_NEAREST_INT));
}
static inline fvec4 min(fvec4 v1, fvec4 v2) {
return fvec4(_mm_min_ps(v1.val, v2.val));
}
static inline fvec4 max(fvec4 v1, fvec4 v2) {
return fvec4(_mm_max_ps(v1.val, v2.val));
}
static inline fvec4 abs(fvec4 v) {
static const __m128 mask = _mm_castsi128_ps(_mm_set1_epi32(0x7FFFFFFF));
return fvec4(_mm_and_ps(v.val, mask));
}
static inline fvec4 sqrt(fvec4 v) {
return fvec4(_mm_sqrt_ps(v.val));
}
static inline float dot3(fvec4 v1, fvec4 v2) {
return _mm_cvtss_f32(_mm_dp_ps(v1, v2, 0x71));
}
static inline float dot4(fvec4 v1, fvec4 v2) {
return _mm_cvtss_f32(_mm_dp_ps(v1, v2, 0xF1));
}
// Functions that operate on ivec4s.
static inline ivec4 min(ivec4 v1, ivec4 v2) {
return ivec4(_mm_min_epi32(v1.val, v2.val));
}
static inline ivec4 max(ivec4 v1, ivec4 v2) {
return ivec4(_mm_max_epi32(v1.val, v2.val));
}
static inline ivec4 abs(ivec4 v) {
return ivec4(_mm_abs_epi32(v.val));
}
// Mathematical operators involving a scalar and a vector.
static inline fvec4 operator+(float v1, fvec4 v2) {
return fvec4(v1)+v2;
}
static inline fvec4 operator-(float v1, fvec4 v2) {
return fvec4(v1)-v2;
}
static inline fvec4 operator*(float v1, fvec4 v2) {
return fvec4(v1)*v2;
}
static inline fvec4 operator/(float v1, fvec4 v2) {
return fvec4(v1)/v2;
}
#endif /*OPENMM_VECTORIZE_H_*/
platforms/cpu/include/CpuNeighborList.h
View file @ 79fd69c3
......@@ -12,26 +12,42 @@ namespace OpenMM {
class OPENMM_EXPORT_CPU CpuNeighborList {
public:
class ThreadData;
class VoxelHash;
class Voxels;
static const int BlockSize;
CpuNeighborList();
~CpuNeighborList();
void computeNeighborList(int numAtoms, const std::vector<float>& atomLocations, const std::vector<std::set<int> >& exclusions,
const float* periodicBoxSize, bool usePeriodic, float maxDistance);
const std::vector<std::pair<int, int> >& getNeighbors();
int getNumBlocks() const;
const std::vector<int>& getSortedAtoms() const;
const std::vector<int>& getBlockNeighbors(int blockIndex) const;
const std::vector<char>& getBlockExclusions(int blockIndex) const;
/**
* This routine contains the code executed by each thread.
*/
void runThread(int index, std::vector<std::pair<int, int> >& threadNeighbors);
void runThread(int index);
private:
/**
* This is called by the worker threads to wait until the master thread instructs them to advance.
*/
void threadWait();
/**
* This is called by the master thread to instruct all the worker threads to advance.
*/
void advanceThreads();
bool isDeleted;
int numThreads, waitCount;
std::vector<std::pair<int, int> > neighbors;
std::vector<int> sortedAtoms;
std::vector<std::vector<int> > blockNeighbors;
std::vector<std::vector<char> > blockExclusions;
std::vector<pthread_t> thread;
std::vector<ThreadData*> threadData;
pthread_cond_t startCondition, endCondition;
pthread_mutex_t lock;
// The following variables are used to make information accessible to the individual threads.
VoxelHash* voxelHash;
float minx, maxx, miny, maxy, minz, maxz;
std::vector<std::pair<int, int> > atomBins;
Voxels* voxels;
const std::vector<std::set<int> >* exclusions;
const float* atomLocations;
const float* periodicBoxSize;
......
platforms/cpu/include/CpuNonbondedForce.h
View file @ 79fd69c3
......@@ -25,14 +25,17 @@
#ifndef OPENMM_CPU_NONBONDED_FORCE_H__
#define OPENMM_CPU_NONBONDED_FORCE_H__
#include "CpuNeighborList.h"
#include "ReferencePairIxn.h"
#include "openmm/internal/vectorize.h"
#include <pthread.h>
#include <set>
#include <utility>
#include <vector>
#include <smmintrin.h>
// ---------------------------------------------------------------------------------------
namespace OpenMM {
class CpuNonbondedForce {
public:
class ThreadData;
......@@ -63,7 +66,7 @@ class CpuNonbondedForce {
--------------------------------------------------------------------------------------- */
void setUseCutoff(float distance, const std::vector<std::pair<int, int> >& neighbors, float solventDielectric);
void setUseCutoff(float distance, const CpuNeighborList& neighbors, float solventDielectric);
/**---------------------------------------------------------------------------------------
......@@ -127,9 +130,9 @@ class CpuNonbondedForce {
--------------------------------------------------------------------------------------- */
void calculateReciprocalIxn(int numberOfAtoms, float* posq, std::vector<OpenMM::RealVec>& atomCoordinates,
void calculateReciprocalIxn(int numberOfAtoms, float* posq, std::vector<RealVec>& atomCoordinates,
const std::vector<std::pair<float, float> >& atomParameters, const std::vector<std::set<int> >& exclusions,
std::vector<OpenMM::RealVec>& forces, float* totalEnergy) const;
std::vector<RealVec>& forces, float* totalEnergy) const;
/**---------------------------------------------------------------------------------------
......@@ -159,14 +162,14 @@ private:
bool periodic;
bool ewald;
bool pme;
const std::vector<std::pair<int, int> >* neighborList;
const CpuNeighborList* neighborList;
float periodicBoxSize[3];
float cutoffDistance, switchingDistance;
float krf, crf;
float alphaEwald;
int numRx, numRy, numRz;
int meshDim[3];
std::vector<float> ewaldScaleX, ewaldScaleY, ewaldScaleDeriv;
std::vector<float> ewaldScaleTable;
float ewaldDX, ewaldDXInv;
bool isDeleted;
int numThreads, waitCount;
......@@ -195,31 +198,42 @@ private:
--------------------------------------------------------------------------------------- */
void calculateOneIxn(int atom1, int atom2, float* forces, double* totalEnergy, const __m128& boxSize, const __m128& invBoxSize);
void calculateOneIxn(int atom1, int atom2, float* forces, double* totalEnergy, const fvec4& boxSize, const fvec4& invBoxSize);
/**---------------------------------------------------------------------------------------
Calculate LJ Coulomb pair ixn between two atoms
Calculate all the interactions for one atom block.
@param atom1 the index of the first atom
@param atom2 the index of the second atom
@param blockIndex the index of the atom block
@param forces force array (forces added)
@param totalEnergy total energy
--------------------------------------------------------------------------------------- */
void calculateBlockIxn(int blockIndex, float* forces, double* totalEnergy, const fvec4& boxSize, const fvec4& invBoxSize);
/**---------------------------------------------------------------------------------------
Calculate all the interactions for one atom block.
@param blockIndex the index of the atom block
@param forces force array (forces added)
@param totalEnergy total energy
--------------------------------------------------------------------------------------- */
void calculateOneEwaldIxn(int atom1, int atom2, float* forces, double* totalEnergy, const __m128& boxSize, const __m128& invBoxSize);
void calculateBlockEwaldIxn(int blockIndex, float* forces, double* totalEnergy, const fvec4& boxSize, const fvec4& invBoxSize);
/**
* Compute the displacement and squared distance between two points, optionally using
* periodic boundary conditions.
*/
void getDeltaR(const __m128& posI, const __m128& posJ, __m128& deltaR, float& r2, bool periodic, const __m128& boxSize, const __m128& invBoxSize) const;
void getDeltaR(const fvec4& posI, const fvec4& posJ, fvec4& deltaR, float& r2, bool periodic, const fvec4& boxSize, const fvec4& invBoxSize) const;
/**
* Compute a fast approximation to erfc(x).
*/
static float erfcApprox(float x);
static fvec4 erfcApprox(fvec4 x);
/**
* Create a lookup table for the scale factor used with Ewald and PME.
......@@ -229,9 +243,11 @@ private:
/**
* Evaluate the scale factor used with Ewald and PME: erfc(alpha*r) + 2*alpha*r*exp(-alpha*alpha*r*r)/sqrt(PI)
*/
float ewaldScaleFunction(float x);
fvec4 ewaldScaleFunction(fvec4 x);
};
} // namespace OpenMM
// ---------------------------------------------------------------------------------------
#endif // OPENMM_CPU_NONBONDED_FORCE_H__
platforms/cpu/src/CpuKernels.cpp
View file @ 79fd69c3
......@@ -221,20 +221,48 @@ double CpuCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeFo
if (nonbondedMethod != NoCutoff) {
// Determine whether we need to recompute the neighbor list.
double padding = 0.1*nonbondedCutoff;
double padding = 0.15*nonbondedCutoff;
bool needRecompute = false;
double closeCutoff2 = 0.25*padding*padding;
double farCutoff2 = 0.5*padding*padding;
int maxNumMoved = numParticles/10;
vector<int> moved;
for (int i = 0; i < numParticles; i++) {
RealVec delta = posData[i]-lastPositions[i];
if (delta.dot(delta) > 0.25*padding*padding) {
needRecompute = true;
break;
double dist2 = delta.dot(delta);
if (dist2 > closeCutoff2) {
moved.push_back(i);
if (dist2 > farCutoff2 || moved.size() > maxNumMoved) {
needRecompute = true;
break;
}
}
}
if (!needRecompute && moved.size() > 0) {
// Some particles have moved further than half the padding distance. Look for pairs
// that are missing from the neighbor list.
int numMoved = moved.size();
double cutoff2 = nonbondedCutoff*nonbondedCutoff;
for (int i = 1; i < numMoved && !needRecompute; i++)
for (int j = 0; j < i; j++) {
RealVec delta = posData[moved[i]]-posData[moved[j]];
if (delta.dot(delta) < cutoff2) {
// These particles should interact. See if they are in the neighbor list.
RealVec oldDelta = lastPositions[moved[i]]-lastPositions[moved[j]];
if (oldDelta.dot(oldDelta) > cutoff2) {
needRecompute = true;
break;
}
}
}
}
if (needRecompute) {
neighborList.computeNeighborList(numParticles, posq, exclusions, floatBoxSize, periodic || ewald || pme, nonbondedCutoff+padding);
lastPositions = posData;
}
nonbonded.setUseCutoff(nonbondedCutoff, neighborList.getNeighbors(), rfDielectric);
nonbonded.setUseCutoff(nonbondedCutoff, neighborList, rfDielectric);
}
if (periodic || ewald || pme) {
double minAllowedSize = 1.999999*nonbondedCutoff;
......
platforms/cpu/src/CpuNeighborList.cpp
View file @ 79fd69c3
#include "CpuNeighborList.h"
#include "openmm/internal/hardware.h"
#include "openmm/internal/vectorize.h"
#include "hilbert.h"
#include <algorithm>
#include <set>
#include <map>
#include <cmath>
......@@ -9,137 +12,236 @@ using namespace std;
namespace OpenMM {
const int CpuNeighborList::BlockSize = 4;
class VoxelIndex
{
public:
VoxelIndex(int xx, int yy, int zz) : x(xx), y(yy), z(zz) {
}
// operator<() needed for map
bool operator<(const VoxelIndex& other) const {
if (x < other.x) return true;
else if (x > other.x) return false;
else if (y < other.y) return true;
else if (y > other.y) return false;
else if (z < other.z) return true;
else return false;
VoxelIndex(int x, int y) : x(x), y(y) {
}
int x;
int y;
int z;
};
typedef pair<const float*, int> VoxelItem;
typedef vector< VoxelItem > Voxel;
class CpuNeighborList::VoxelHash {
/**
* This data structure organizes the particles spatially. It divides them into bins along the x and y axes,
* then sorts each bin along the z axis so ranges can be identified quickly with a binary search.
*/
class CpuNeighborList::Voxels {
public:
VoxelHash(float vsx, float vsy, float vsz, const float* periodicBoxSize, bool usePeriodic) :
voxelSizeX(vsx), voxelSizeY(vsy), voxelSizeZ(vsz), periodicBoxSize(periodicBoxSize), usePeriodic(usePeriodic) {
Voxels(float vsx, float vsy, float minx, float maxx, float miny, float maxy, const float* periodicBoxSize, bool usePeriodic) :
voxelSizeX(vsx), voxelSizeY(vsy), minx(minx), maxx(maxx), miny(miny), maxy(maxy), periodicBoxSize(periodicBoxSize), usePeriodic(usePeriodic) {
if (usePeriodic) {
nx = (int) floorf(periodicBoxSize[0]/voxelSizeX+0.5f);
ny = (int) floorf(periodicBoxSize[1]/voxelSizeY+0.5f);
nz = (int) floorf(periodicBoxSize[2]/voxelSizeZ+0.5f);
voxelSizeX = periodicBoxSize[0]/nx;
voxelSizeY = periodicBoxSize[1]/ny;
voxelSizeZ = periodicBoxSize[2]/nz;
}
else {
nx = max(1, (int) floorf((maxx-minx)/voxelSizeX+0.5f));
ny = max(1, (int) floorf((maxy-miny)/voxelSizeY+0.5f));
if (maxx > minx)
voxelSizeX = (maxx-minx)/nx;
if (maxy > miny)
voxelSizeY = (maxy-miny)/ny;
}
bins.resize(nx);
for (int i = 0; i < nx; i++) {
bins[i].resize(ny);
for (int j = 0; j < ny; j++)
bins[i][j].resize(0);
}
}
void insert(const int& item, const float* location) {
/**
* Insert a particle into the voxel data structure.
*/
void insert(const int& atom, const float* location) {
VoxelIndex voxelIndex = getVoxelIndex(location);
if (voxelMap.find(voxelIndex) == voxelMap.end())
voxelMap[voxelIndex] = Voxel();
Voxel& voxel = voxelMap.find(voxelIndex)->second;
voxel.push_back(VoxelItem(location, item));
bins[voxelIndex.x][voxelIndex.y].push_back(make_pair(location[2], VoxelItem(location, atom)));
}
/**
* Sort the particles in each voxel by z coordinate.
*/
void sortItems() {
for (int i = 0; i < nx; i++)
for (int j = 0; j < ny; j++)
sort(bins[i][j].begin(), bins[i][j].end());
}
/**
* Find the index of the first particle in voxel (x,y) whose z coordinate in >= the specified value.
*/
int findLowerBound(int x, int y, double z) const {
const vector<pair<float, VoxelItem> >& bin = bins[x][y];
int lower = 0;
int upper = bin.size();
while (lower < upper) {
int middle = (lower+upper)/2;
if (bin[middle].first < z)
lower = middle+1;
else
upper = middle;
}
return lower;
}
/**
* Find the index of the first particle in voxel (x,y) whose z coordinate in greater than the specified value.
*/
int findUpperBound(int x, int y, double z) const {
const vector<pair<float, VoxelItem> >& bin = bins[x][y];
int lower = 0;
int upper = bin.size();
while (lower < upper) {
int middle = (lower+upper)/2;
if (bin[middle].first > z)
upper = middle;
else
lower = middle+1;
}
return upper;
}
/**
* Get the voxel index containing a particular location.
*/
VoxelIndex getVoxelIndex(const float* location) const {
float xperiodic, yperiodic, zperiodic;
float xperiodic, yperiodic;
if (!usePeriodic) {
xperiodic = location[0];
yperiodic = location[1];
zperiodic = location[2];
xperiodic = location[0]-minx;
yperiodic = location[1]-miny;
}
else {
xperiodic = location[0]-periodicBoxSize[0]*floorf(location[0]/periodicBoxSize[0]);
yperiodic = location[1]-periodicBoxSize[1]*floorf(location[1]/periodicBoxSize[1]);
zperiodic = location[2]-periodicBoxSize[2]*floorf(location[2]/periodicBoxSize[2]);
}
int x = int(floorf(xperiodic / voxelSizeX));
int y = int(floorf(yperiodic / voxelSizeY));
int z = int(floorf(zperiodic / voxelSizeZ));
int x = min(nx-1, int(floorf(xperiodic / voxelSizeX)));
int y = min(ny-1, int(floorf(yperiodic / voxelSizeY)));
return VoxelIndex(x, y, z);
return VoxelIndex(x, y);
}
void getNeighbors(vector<pair<int, int> >& neighbors, const VoxelItem& referencePoint, const vector<set<int> >& exclusions, float maxDistance) const {
// Loop over neighboring voxels
// TODO use more clever selection of neighboring voxels
const int atomI = referencePoint.second;
const float* locationI = referencePoint.first;
__m128 posI = _mm_loadu_ps(locationI);
__m128 boxSize = _mm_set_ps(0, periodicBoxSize[2], periodicBoxSize[1], periodicBoxSize[0]);
__m128 invBoxSize = _mm_set_ps(0, (1/periodicBoxSize[2]), (1/periodicBoxSize[1]), (1/periodicBoxSize[0]));
__m128 half = _mm_set1_ps(0.5);
void getNeighbors(vector<int>& neighbors, int blockIndex, fvec4 blockCenter, fvec4 blockWidth, const vector<int>& sortedAtoms, vector<char>& exclusions, float maxDistance, const vector<int> blockAtoms, const float* atomLocations) const {
neighbors.resize(0);
exclusions.resize(0);
fvec4 boxSize(periodicBoxSize[0], periodicBoxSize[1], periodicBoxSize[2], 0);
fvec4 invBoxSize(1/periodicBoxSize[0], 1/periodicBoxSize[1], 1/periodicBoxSize[2], 0);
float maxDistanceSquared = maxDistance * maxDistance;
float refineCutoff = maxDistance-max(max(blockWidth[0], blockWidth[1]), blockWidth[2]);
float refineCutoffSquared = refineCutoff*refineCutoff;
int dIndexX = int(maxDistance / voxelSizeX) + 1; // How may voxels away do we have to look?
int dIndexY = int(maxDistance / voxelSizeY) + 1;
int dIndexZ = int(maxDistance / voxelSizeZ) + 1;
VoxelIndex centerVoxelIndex = getVoxelIndex(locationI);
int lastx = centerVoxelIndex.x+dIndexX;
int lasty = centerVoxelIndex.y+dIndexY;
int lastz = centerVoxelIndex.z+dIndexZ;
int dIndexX = min(nx/2, int((maxDistance+blockWidth[0])/voxelSizeX)+1); // How may voxels away do we have to look?
int dIndexY = min(ny/2, int((maxDistance+blockWidth[1])/voxelSizeY)+1);
float centerPos[4];
blockCenter.store(centerPos);
VoxelIndex centerVoxelIndex = getVoxelIndex(centerPos);
int startx = centerVoxelIndex.x-dIndexX;
int starty = centerVoxelIndex.y-dIndexY;
int endx = centerVoxelIndex.x+dIndexX;
int endy = centerVoxelIndex.y+dIndexY;
int numRanges;
if (usePeriodic) {
lastx = min(lastx, centerVoxelIndex.x-dIndexX+nx-1);
lasty = min(lasty, centerVoxelIndex.y-dIndexY+ny-1);
lastz = min(lastz, centerVoxelIndex.z-dIndexZ+nz-1);
endx = min(endx, centerVoxelIndex.x-dIndexX+nx-1);
endy = min(endy, centerVoxelIndex.y-dIndexY+ny-1);
numRanges = 2;
}
VoxelIndex voxelIndex(0, 0, 0);
for (int x = centerVoxelIndex.x - dIndexX; x <= lastx; ++x) {
else {
startx = max(startx, 0);
starty = max(starty, 0);
endx = min(endx, nx-1);
endy = min(endy, ny-1);
numRanges = 1;
}
int lastSortedIndex = BlockSize*(blockIndex+1);
VoxelIndex voxelIndex(0, 0);
for (int x = startx; x <= endx; ++x) {
voxelIndex.x = x;
if (usePeriodic)
voxelIndex.x = (x < 0 ? x+nx : (x >= nx ? x-nx : x));
for (int y = centerVoxelIndex.y - dIndexY; y <= lasty; ++y) {
float dx = max(0.0f, voxelSizeX*max(0, abs(centerVoxelIndex.x-x)-1)-blockWidth[0]);
for (int y = starty; y <= endy; ++y) {
voxelIndex.y = y;
if (usePeriodic)
voxelIndex.y = (y < 0 ? y+ny : (y >= ny ? y-ny : y));
for (int z = centerVoxelIndex.z - dIndexZ; z <= lastz; ++z) {
voxelIndex.z = z;
if (usePeriodic)
voxelIndex.z = (z < 0 ? z+nz : (z >= nz ? z-nz : z));
const map<VoxelIndex, Voxel>::const_iterator voxelEntry = voxelMap.find(voxelIndex);
if (voxelEntry == voxelMap.end())
continue; // no such voxel; skip
const Voxel& voxel = voxelEntry->second;
for (Voxel::const_iterator itemIter = voxel.begin(); itemIter != voxel.end(); ++itemIter) {
const int atomJ = itemIter->second;
float dy = max(0.0f, voxelSizeY*max(0, abs(centerVoxelIndex.y-y)-1)-blockWidth[1]);
// Identify the range of atoms within this bin we need to search. When using periodic boundary
// conditions, there may be two separate ranges.
float dz = maxDistance+blockWidth[2];
dz = sqrtf(max(0.0f, dz*dz-dx*dx-dy*dy));
int rangeStart[2];
int rangeEnd[2];
rangeStart[0] = findLowerBound(voxelIndex.x, voxelIndex.y, centerPos[2]-dz);
if (usePeriodic) {
rangeEnd[0] = findUpperBound(voxelIndex.x, voxelIndex.y, centerPos[2]+dz);
if (rangeStart[0] > 0) {
rangeStart[1] = 0;
rangeEnd[1] = min(findUpperBound(voxelIndex.x, voxelIndex.y, centerPos[2]+dz-periodicBoxSize[2]), rangeStart[0]);
}
else {
rangeStart[1] = max(findLowerBound(voxelIndex.x, voxelIndex.y, centerPos[2]-dz+periodicBoxSize[2]), rangeEnd[0]);
rangeEnd[1] = bins[voxelIndex.x][voxelIndex.y].size();
}
}
else
rangeEnd[0] = findUpperBound(voxelIndex.x, voxelIndex.y, centerPos[2]+dz);
// Loop over atoms and check to see if they are neighbors of this block.
for (int range = 0; range < numRanges; range++) {
for (int item = rangeStart[range]; item < rangeEnd[range]; item++) {
const int sortedIndex = bins[voxelIndex.x][voxelIndex.y][item].second.second;
// Avoid duplicate entries.
if (atomJ >= atomI)
break;
if (sortedIndex >= lastSortedIndex)
continue;
__m128 posJ = _mm_loadu_ps(itemIter->first);
__m128 delta = _mm_sub_ps(posJ, posI);
fvec4 atomPos(bins[voxelIndex.x][voxelIndex.y][item].second.first);
fvec4 delta = atomPos-centerPos;
if (usePeriodic) {
__m128 base = _mm_mul_ps(_mm_floor_ps(_mm_add_ps(_mm_mul_ps(delta, invBoxSize), half)), boxSize);
delta = _mm_sub_ps(delta, base);
fvec4 base = round(delta*invBoxSize)*boxSize;
delta = delta-base;
}
float dSquared = _mm_cvtss_f32(_mm_dp_ps(delta, delta, 0x71));
delta = max(0.0f, abs(delta)-blockWidth);
float dSquared = dot3(delta, delta);
if (dSquared > maxDistanceSquared)
continue;
// Ignore exclusions.
if (exclusions[atomI].find(atomJ) != exclusions[atomI].end())
continue;
neighbors.push_back(make_pair(atomI, atomJ));
if (dSquared > refineCutoffSquared) {
// The distance is large enough that there might not be any actual interactions.
// Check individual atom pairs to be sure.
bool any = false;
for (int k = 0; k < (int) blockAtoms.size(); k++) {
fvec4 pos1(&atomLocations[4*blockAtoms[k]]);
delta = atomPos-pos1;
if (usePeriodic) {
fvec4 base = round(delta*invBoxSize)*boxSize;
delta = delta-base;
}
float r2 = dot3(delta, delta);
if (r2 < maxDistanceSquared) {
any = true;
break;
}
}
if (!any)
continue;
}
// Add this atom to the list of neighbors.
neighbors.push_back(sortedAtoms[sortedIndex]);
if (sortedIndex < BlockSize*blockIndex)
exclusions.push_back(0);
else
exclusions.push_back(0xF & (0xF<<(sortedIndex-BlockSize*blockIndex)));
}
}
}
......@@ -147,11 +249,12 @@ public:
}
private:
float voxelSizeX, voxelSizeY, voxelSizeZ;
int nx, ny, nz;
float voxelSizeX, voxelSizeY;
float minx, maxx, miny, maxy;
int nx, ny;
const float* periodicBoxSize;
const bool usePeriodic;
map<VoxelIndex, Voxel> voxelMap;
vector<vector<vector<pair<float, VoxelItem> > > > bins;
};
class CpuNeighborList::ThreadData {
......@@ -160,12 +263,11 @@ public:
}
int index;
CpuNeighborList& owner;
vector<pair<int, int> > threadNeighbors;
};
static void* threadBody(void* args) {
CpuNeighborList::ThreadData& data = *reinterpret_cast<CpuNeighborList::ThreadData*>(args);
data.owner.runThread(data.index, data.threadNeighbors);
data.owner.runThread(data.index);
delete &data;
return 0;
}
......@@ -203,23 +305,13 @@ CpuNeighborList::~CpuNeighborList() {
void CpuNeighborList::computeNeighborList(int numAtoms, const vector<float>& atomLocations, const vector<set<int> >& exclusions,
const float* periodicBoxSize, bool usePeriodic, float maxDistance) {
// Build the voxel hash.
float edgeSizeX, edgeSizeY, edgeSizeZ;
if (!usePeriodic)
edgeSizeX = edgeSizeY = edgeSizeZ = maxDistance; // TODO - adjust this as needed
else {
edgeSizeX = 0.5f*periodicBoxSize[0]/floorf(periodicBoxSize[0]/maxDistance);
edgeSizeY = 0.5f*periodicBoxSize[1]/floorf(periodicBoxSize[1]/maxDistance);
edgeSizeZ = 0.5f*periodicBoxSize[2]/floorf(periodicBoxSize[2]/maxDistance);
}
VoxelHash voxelHash(edgeSizeX, edgeSizeY, edgeSizeZ, periodicBoxSize, usePeriodic);
for (int i = 0; i < numAtoms; i++)
voxelHash.insert(i, &atomLocations[4*i]);
int numBlocks = (numAtoms+BlockSize-1)/BlockSize;
blockNeighbors.resize(numBlocks);
blockExclusions.resize(numBlocks);
sortedAtoms.resize(numAtoms);
// Record the parameters for the threads.
this->voxelHash = &voxelHash;
this->exclusions = &exclusions;
this->atomLocations = &atomLocations[0];
this->periodicBoxSize = periodicBoxSize;
......@@ -227,43 +319,158 @@ void CpuNeighborList::computeNeighborList(int numAtoms, const vector<float>& ato
this->usePeriodic = usePeriodic;
this->maxDistance = maxDistance;
// Signal the threads to start running and wait for them to finish.
// Identify the range of atom positions along each axis.
fvec4 minPos(&atomLocations[0]);
fvec4 maxPos = minPos;
for (int i = 0; i < numAtoms; i++) {
fvec4 pos(&atomLocations[4*i]);
minPos = min(minPos, pos);
maxPos = max(maxPos, pos);
}
minx = minPos[0];
maxx = maxPos[0];
miny = minPos[1];
maxy = maxPos[1];
minz = minPos[2];
maxz = maxPos[2];
// Sort the atoms based on a Hilbert curve.
atomBins.resize(numAtoms);
pthread_mutex_lock(&lock);
waitCount = 0;
pthread_cond_broadcast(&startCondition);
while (waitCount < numThreads)
pthread_cond_wait(&endCondition, &lock);
advanceThreads();
sort(atomBins.begin(), atomBins.end());
// Build the voxel hash.
float edgeSizeX, edgeSizeY;
if (!usePeriodic)
edgeSizeX = edgeSizeY = maxDistance; // TODO - adjust this as needed
else {
edgeSizeX = 0.5f*periodicBoxSize[0]/floorf(periodicBoxSize[0]/maxDistance);
edgeSizeY = 0.5f*periodicBoxSize[1]/floorf(periodicBoxSize[1]/maxDistance);
}
Voxels voxels(edgeSizeX, edgeSizeY, minx, maxx, miny, maxy, periodicBoxSize, usePeriodic);
for (int i = 0; i < numAtoms; i++) {
int atomIndex = atomBins[i].second;
sortedAtoms[i] = atomIndex;
voxels.insert(i, &atomLocations[4*atomIndex]);
}
voxels.sortItems();
this->voxels = &voxels;
// Signal the threads to start running and wait for them to finish.
advanceThreads();
pthread_mutex_unlock(&lock);
// Combine the results from all the threads.
// Add padding atoms to fill up the last block.
neighbors.clear();
for (int i = 0; i < numThreads; i++)
neighbors.insert(neighbors.end(), threadData[i]->threadNeighbors.begin(), threadData[i]->threadNeighbors.end());
int numPadding = numBlocks*BlockSize-numAtoms;
if (numPadding > 0) {
char mask = (0xF0 >> numPadding) & 0xF;
for (int i = 0; i < numPadding; i++)
sortedAtoms.push_back(0);
vector<char>& exc = blockExclusions[blockExclusions.size()-1];
for (int i = 0; i < (int) exc.size(); i++)
exc[i] |= mask;
}
}
int CpuNeighborList::getNumBlocks() const {
return sortedAtoms.size()/BlockSize;
}
const vector<pair<int, int> >& CpuNeighborList::getNeighbors() {
return neighbors;
const std::vector<int>& CpuNeighborList::getSortedAtoms() const {
return sortedAtoms;
}
const std::vector<int>& CpuNeighborList::getBlockNeighbors(int blockIndex) const {
return blockNeighbors[blockIndex];
}
const std::vector<char>& CpuNeighborList::getBlockExclusions(int blockIndex) const {
return blockExclusions[blockIndex];
}
void CpuNeighborList::runThread(int index, vector<pair<int, int> >& threadNeighbors) {
void CpuNeighborList::runThread(int index) {
while (true) {
// Wait for the signal to start running.
pthread_mutex_lock(&lock);
waitCount++;
pthread_cond_signal(&endCondition);
pthread_cond_wait(&startCondition, &lock);
pthread_mutex_unlock(&lock);
threadWait();
if (isDeleted)
break;
// Compute the positions of atoms along the Hilbert curve.
float binWidth = max(max(maxx-minx, maxy-miny), maxz-minz)/255.0f;
float invBinWidth = 1.0f/binWidth;
bitmask_t coords[3];
for (int i = index; i < numAtoms; i += numThreads) {
const float* pos = &atomLocations[4*i];
coords[0] = (bitmask_t) ((pos[0]-minx)*invBinWidth);
coords[1] = (bitmask_t) ((pos[1]-miny)*invBinWidth);
coords[2] = (bitmask_t) ((pos[2]-minz)*invBinWidth);
int bin = (int) hilbert_c2i(3, 8, coords);
atomBins[i] = pair<int, int>(bin, i);
}
threadWait();
// Compute this thread's subset of neighbors.
threadNeighbors.clear();
for (int i = index; i < numAtoms; i += numThreads)
voxelHash->getNeighbors(threadNeighbors, VoxelItem(&atomLocations[4*i], i), *exclusions, maxDistance);
int numBlocks = blockNeighbors.size();
vector<int> blockAtoms;
for (int i = index; i < numBlocks; i += numThreads) {
{
int firstIndex = BlockSize*i;
int atomsInBlock = min(BlockSize, numAtoms-firstIndex);
blockAtoms.resize(atomsInBlock);
for (int j = 0; j < atomsInBlock; j++)
blockAtoms[j] = sortedAtoms[firstIndex+j];
}
int firstIndex = BlockSize*i;
fvec4 minPos(&atomLocations[4*sortedAtoms[firstIndex]]);
fvec4 maxPos = minPos;
int atomsInBlock = min(BlockSize, numAtoms-firstIndex);
for (int j = 1; j < atomsInBlock; j++) {
fvec4 pos(&atomLocations[4*sortedAtoms[firstIndex+j]]);
minPos = min(minPos, pos);
maxPos = max(maxPos, pos);
}
voxels->getNeighbors(blockNeighbors[i], i, (maxPos+minPos)*0.5f, (maxPos-minPos)*0.5f, sortedAtoms, blockExclusions[i], maxDistance, blockAtoms, atomLocations);
// Record the exclusions for this block.
for (int j = 0; j < atomsInBlock; j++) {
const set<int>& atomExclusions = (*exclusions)[sortedAtoms[firstIndex+j]];
char mask = 1<<j;
for (int k = 0; k < (int) blockNeighbors[i].size(); k++) {
int atomIndex = blockNeighbors[i][k];
if (atomExclusions.find(atomIndex) != atomExclusions.end())
blockExclusions[i][k] |= mask;
}
}
}
}
}
void CpuNeighborList::threadWait() {
pthread_mutex_lock(&lock);
waitCount++;
pthread_cond_signal(&endCondition);
pthread_cond_wait(&startCondition, &lock);
pthread_mutex_unlock(&lock);
}
void CpuNeighborList::advanceThreads() {
waitCount = 0;
pthread_cond_broadcast(&startCondition);
while (waitCount < numThreads) {
pthread_cond_wait(&endCondition, &lock);
}
}
......
platforms/cpu/src/CpuNonbondedForce.cpp
View file @ 79fd69c3
......@@ -32,6 +32,7 @@
#include "ReferencePME.h"
#include "openmm/internal/hardware.h"
#include "openmm/internal/SplineFitter.h"
#include "openmm/internal/vectorize.h"
// In case we're using some primitive version of Visual Studio this will
// make sure that erf() and erfc() are defined.
......@@ -113,7 +114,7 @@ CpuNonbondedForce::~CpuNonbondedForce(){
--------------------------------------------------------------------------------------- */
void CpuNonbondedForce::setUseCutoff(float distance, const vector<pair<int, int> >& neighbors, float solventDielectric) {
void CpuNonbondedForce::setUseCutoff(float distance, const CpuNeighborList& neighbors, float solventDielectric) {
cutoff = true;
cutoffDistance = distance;
......@@ -199,23 +200,22 @@ void CpuNonbondedForce::setUseSwitchingFunction(float distance) {
void CpuNonbondedForce::tabulateEwaldScaleFactor() {
ewaldDX = cutoffDistance/(NUM_TABLE_POINTS-2);
ewaldDXInv = 1.0f/ewaldDX;
vector<double> x(NUM_TABLE_POINTS);
vector<double> y(NUM_TABLE_POINTS);
vector<double> x(NUM_TABLE_POINTS+1);
vector<double> y(NUM_TABLE_POINTS+1);
vector<double> deriv;
for (int i = 0; i < NUM_TABLE_POINTS; i++) {
for (int i = 0; i < NUM_TABLE_POINTS+1; i++) {
double r = i*cutoffDistance/(NUM_TABLE_POINTS-2);
double alphaR = alphaEwald*r;
x[i] = r;
y[i] = erfc(alphaR) + TWO_OVER_SQRT_PI*alphaR*exp(-alphaR*alphaR);
}
SplineFitter::createNaturalSpline(x, y, deriv);
ewaldScaleX.resize(NUM_TABLE_POINTS);
ewaldScaleY.resize(NUM_TABLE_POINTS);
ewaldScaleDeriv.resize(NUM_TABLE_POINTS);
ewaldScaleTable.resize(4*NUM_TABLE_POINTS);
for (int i = 0; i < NUM_TABLE_POINTS; i++) {
ewaldScaleX[i] = (float) x[i];
ewaldScaleY[i] = (float) y[i];
ewaldScaleDeriv[i] = (float) (deriv[i]*ewaldDX*ewaldDX/6);
ewaldScaleTable[4*i] = (float) y[i];
ewaldScaleTable[4*i+1] = (float) y[i+1];
ewaldScaleTable[4*i+2] = (float) (deriv[i]*ewaldDX*ewaldDX/6);
ewaldScaleTable[4*i+3] = (float) (deriv[i+1]*ewaldDX*ewaldDX/6);
}
}
......@@ -358,42 +358,12 @@ void CpuNonbondedForce::calculateDirectIxn(int numberOfAtoms, float* posq, const
for (int i = 0; i < numThreads; i++)
directEnergy += threadData[i]->threadEnergy;
for (int i = 0; i < numberOfAtoms; i++) {
__m128 f = _mm_loadu_ps(forces+4*i);
fvec4 f(forces+4*i);
for (int j = 0; j < numThreads; j++)
f = _mm_add_ps(f, _mm_loadu_ps(&threadData[j]->threadForce[4*i]));
_mm_storeu_ps(forces+4*i, f);
f += fvec4(&threadData[j]->threadForce[4*i]);
f.store(forces+4*i);
}
if (ewald || pme) {
// Now subtract off the exclusions, since they were implicitly included in the reciprocal space sum.
__m128 boxSize = _mm_set_ps(0, periodicBoxSize[2], periodicBoxSize[1], periodicBoxSize[0]);
__m128 invBoxSize = _mm_set_ps(0, (1/periodicBoxSize[2]), (1/periodicBoxSize[1]), (1/periodicBoxSize[0]));
for (int i = 0; i < numberOfAtoms; i++)
for (set<int>::const_iterator iter = exclusions[i].begin(); iter != exclusions[i].end(); ++iter) {
if (*iter > i) {
int ii = i;
int jj = *iter;
__m128 deltaR;
__m128 posI = _mm_loadu_ps(posq+4*ii);
__m128 posJ = _mm_loadu_ps(posq+4*jj);
float r2;
getDeltaR(posJ, posI, deltaR, r2, false, boxSize, invBoxSize);
float r = sqrtf(r2);
float inverseR = 1/r;
float chargeProd = ONE_4PI_EPS0*posq[4*ii+3]*posq[4*jj+3];
float alphaR = alphaEwald*r;
float erfcAlphaR = erfcApprox(alphaR);
float dEdR = (float) (chargeProd * inverseR * inverseR * inverseR);
dEdR = (float) (dEdR * (1.0f-erfcAlphaR-TWO_OVER_SQRT_PI*alphaR*exp(-alphaR*alphaR)));
__m128 result = _mm_mul_ps(deltaR, _mm_set1_ps(dEdR));
_mm_storeu_ps(forces+4*ii, _mm_sub_ps(_mm_loadu_ps(forces+4*ii), result));
_mm_storeu_ps(forces+4*jj, _mm_add_ps(_mm_loadu_ps(forces+4*jj), result));
if (includeEnergy)
directEnergy -= chargeProd*inverseR*(1.0f-erfcAlphaR);
}
}
}
if (totalEnergy != NULL)
*totalEnergy += (float) directEnergy;
}
......@@ -418,23 +388,47 @@ void CpuNonbondedForce::runThread(int index, vector<float>& threadForce, double&
threadForce.resize(4*numberOfAtoms, 0.0f);
for (int i = 0; i < 4*numberOfAtoms; i++)
threadForce[i] = 0.0f;
__m128 boxSize = _mm_set_ps(0, periodicBoxSize[2], periodicBoxSize[1], periodicBoxSize[0]);
__m128 invBoxSize = _mm_set_ps(0, (1/periodicBoxSize[2]), (1/periodicBoxSize[1]), (1/periodicBoxSize[0]));
fvec4 boxSize(periodicBoxSize[0], periodicBoxSize[1], periodicBoxSize[2], 0);
fvec4 invBoxSize((1/periodicBoxSize[0]), (1/periodicBoxSize[1]), (1/periodicBoxSize[2]), 0);
if (ewald || pme) {
// Compute the interactions from the neighbor list.
for (int i = index; i < (int) neighborList->size(); i += numThreads) {
pair<int, int> pair = (*neighborList)[i];
calculateOneEwaldIxn(pair.first, pair.second, &threadForce[0], energyPtr, boxSize, invBoxSize);
}
for (int i = index; i < neighborList->getNumBlocks(); i += numThreads)
calculateBlockEwaldIxn(i, &threadForce[0], energyPtr, boxSize, invBoxSize);
// Now subtract off the exclusions, since they were implicitly included in the reciprocal space sum.
fvec4 boxSize(periodicBoxSize[0], periodicBoxSize[1], periodicBoxSize[2], 0);
fvec4 invBoxSize((1/periodicBoxSize[0]), (1/periodicBoxSize[1]), (1/periodicBoxSize[2]), 0);
for (int i = index; i < numberOfAtoms; i += numThreads)
for (set<int>::const_iterator iter = exclusions[i].begin(); iter != exclusions[i].end(); ++iter) {
if (*iter > i) {
int j = *iter;
fvec4 deltaR;
fvec4 posI(posq+4*i);
fvec4 posJ(posq+4*j);
float r2;
getDeltaR(posJ, posI, deltaR, r2, false, boxSize, invBoxSize);
float r = sqrtf(r2);
float inverseR = 1/r;
float chargeProd = ONE_4PI_EPS0*posq[4*i+3]*posq[4*j+3];
float alphaR = alphaEwald*r;
float erfcAlphaR = erfcApprox(alphaR)[0];
float dEdR = (float) (chargeProd * inverseR * inverseR * inverseR);
dEdR = (float) (dEdR * (1.0f-erfcAlphaR-TWO_OVER_SQRT_PI*alphaR*exp(-alphaR*alphaR)));
fvec4 result = deltaR*dEdR;
(fvec4(&threadForce[4*i])-result).store(&threadForce[4*i]);
(fvec4(&threadForce[4*j])+result).store(&threadForce[4*j]);
if (includeEnergy)
threadEnergy -= chargeProd*inverseR*(1.0f-erfcAlphaR);
}
}
}
else if (cutoff) {
// Compute the interactions from the neighbor list.
for (int i = index; i < (int) neighborList->size(); i += numThreads) {
pair<int, int> pair = (*neighborList)[i];
calculateOneIxn(pair.first, pair.second, &threadForce[0], energyPtr, boxSize, invBoxSize);
}
for (int i = index; i < neighborList->getNumBlocks(); i += numThreads)
calculateBlockIxn(i, &threadForce[0], energyPtr, boxSize, invBoxSize);
}
else {
// Loop over all atom pairs
......@@ -448,12 +442,12 @@ void CpuNonbondedForce::runThread(int index, vector<float>& threadForce, double&
}
}
void CpuNonbondedForce::calculateOneIxn(int ii, int jj, float* forces, double* totalEnergy, const __m128& boxSize, const __m128& invBoxSize) {
void CpuNonbondedForce::calculateOneIxn(int ii, int jj, float* forces, double* totalEnergy, const fvec4& boxSize, const fvec4& invBoxSize) {
// get deltaR, R2, and R between 2 atoms
__m128 deltaR;
__m128 posI = _mm_loadu_ps(posq+4*ii);
__m128 posJ = _mm_loadu_ps(posq+4*jj);
fvec4 deltaR;
fvec4 posI(posq+4*ii);
fvec4 posJ(posq+4*jj);
float r2;
getDeltaR(posJ, posI, deltaR, r2, periodic, boxSize, invBoxSize);
if (cutoff && r2 >= cutoffDistance*cutoffDistance)
......@@ -497,83 +491,239 @@ void CpuNonbondedForce::calculateOneIxn(int ii, int jj, float* forces, double* t
// accumulate forces
__m128 result = _mm_mul_ps(deltaR, _mm_set1_ps(dEdR));
_mm_storeu_ps(forces+4*ii, _mm_add_ps(_mm_loadu_ps(forces+4*ii), result));
_mm_storeu_ps(forces+4*jj, _mm_sub_ps(_mm_loadu_ps(forces+4*jj), result));
fvec4 result = deltaR*dEdR;
(fvec4(forces+4*ii)+result).store(forces+4*ii);
(fvec4(forces+4*jj)-result).store(forces+4*jj);
}
void CpuNonbondedForce::calculateOneEwaldIxn(int ii, int jj, float* forces, double* totalEnergy, const __m128& boxSize, const __m128& invBoxSize) {
__m128 deltaR;
__m128 posI = _mm_loadu_ps(posq+4*ii);
__m128 posJ = _mm_loadu_ps(posq+4*jj);
float r2;
getDeltaR(posJ, posI, deltaR, r2, true, boxSize, invBoxSize);
if (r2 < cutoffDistance*cutoffDistance) {
float r = sqrtf(r2);
float inverseR = 1/r;
float switchValue = 1, switchDeriv = 0;
if (useSwitch && r > switchingDistance) {
float t = (r-switchingDistance)/(cutoffDistance-switchingDistance);
switchValue = 1+t*t*t*(-10+t*(15-t*6));
switchDeriv = t*t*(-30+t*(60-t*30))/(cutoffDistance-switchingDistance);
void CpuNonbondedForce::calculateBlockIxn(int blockIndex, float* forces, double* totalEnergy, const fvec4& boxSize, const fvec4& invBoxSize) {
// Load the positions and parameters of the atoms in the block.
int blockAtom[4];
fvec4 blockAtomPosq[4];
fvec4 blockAtomForce[4];
for (int i = 0; i < 4; i++) {
blockAtom[i] = neighborList->getSortedAtoms()[4*blockIndex+i];
blockAtomPosq[i] = fvec4(posq+4*blockAtom[i]);
blockAtomForce[i] = fvec4(0.0f);
}
fvec4 blockAtomCharge = fvec4(ONE_4PI_EPS0)*fvec4(blockAtomPosq[0][3], blockAtomPosq[1][3], blockAtomPosq[2][3], blockAtomPosq[3][3]);
fvec4 blockAtomSigma(atomParameters[blockAtom[0]].first, atomParameters[blockAtom[1]].first, atomParameters[blockAtom[2]].first, atomParameters[blockAtom[3]].first);
fvec4 blockAtomEpsilon(atomParameters[blockAtom[0]].second, atomParameters[blockAtom[1]].second, atomParameters[blockAtom[2]].second, atomParameters[blockAtom[3]].second);
// Loop over neighbors for this block.
const vector<int>& neighbors = neighborList->getBlockNeighbors(blockIndex);
const vector<char>& exclusions = neighborList->getBlockExclusions(blockIndex);
float blockAtomR2[4];
bool include[4];
fvec4 blockAtomDelta[4];
for (int i = 0; i < (int) neighbors.size(); i++) {
// Load the next neighbor.
int atom = neighbors[i];
fvec4 atomPosq(posq+4*atom);
// Compute the distances to the block atoms.
bool any = false;
for (int j = 0; j < 4; j++) {
getDeltaR(atomPosq, blockAtomPosq[j], blockAtomDelta[j], blockAtomR2[j], periodic, boxSize, invBoxSize);
include[j] = (((exclusions[i]>>j)&1) == 0 && (!cutoff || blockAtomR2[j] < cutoffDistance*cutoffDistance));
any |= include[j];
}
if (!any)
continue; // No interactions to compute.
// Compute the interactions.
fvec4 r2(blockAtomR2);
fvec4 r = sqrt(r2);
fvec4 inverseR = fvec4(1.0f)/r;
fvec4 switchValue(1.0f), switchDeriv(0.0f);
if (useSwitch) {
fvec4 t = (r>switchingDistance) & ((r-switchingDistance)/(cutoffDistance-switchingDistance));
switchValue = 1+t*t*t*(-10.0f+t*(15.0f-t*6.0f));
switchDeriv = t*t*(-30.0f+t*(60.0f-t*30.0f))/(cutoffDistance-switchingDistance);
}
float chargeProd = ONE_4PI_EPS0*posq[4*ii+3]*posq[4*jj+3];
float dEdR = chargeProd*inverseR*ewaldScaleFunction(r);
float sig = atomParameters[ii].first + atomParameters[jj].first;
float sig2 = inverseR*sig;
sig2 *= sig2;
float sig6 = sig2*sig2*sig2;
float eps = atomParameters[ii].second*atomParameters[jj].second;
dEdR += switchValue*eps*(12.0f*sig6 - 6.0f)*sig6;
fvec4 sig = blockAtomSigma+atomParameters[atom].first;
fvec4 sig2 = inverseR*sig;
sig2 *= sig2;
fvec4 sig6 = sig2*sig2*sig2;
fvec4 eps = blockAtomEpsilon*atomParameters[atom].second;
fvec4 dEdR = switchValue*eps*(12.0f*sig6 - 6.0f)*sig6;
fvec4 chargeProd = blockAtomCharge*posq[4*atom+3];
if (cutoff)
dEdR += chargeProd*(inverseR-2.0f*krf*r2);
else
dEdR += chargeProd*inverseR;
dEdR *= inverseR*inverseR;
float energy = eps*(sig6-1.0f)*sig6;
fvec4 energy = eps*(sig6-1.0f)*sig6;
if (useSwitch) {
dEdR -= energy*switchDeriv*inverseR;
energy *= switchValue;
}
// accumulate forces
// Accumulate energies.
__m128 result = _mm_mul_ps(deltaR, _mm_set1_ps(dEdR));
_mm_storeu_ps(forces+4*ii, _mm_add_ps(_mm_loadu_ps(forces+4*ii), result));
_mm_storeu_ps(forces+4*jj, _mm_sub_ps(_mm_loadu_ps(forces+4*jj), result));
if (totalEnergy) {
if (cutoff)
energy += chargeProd*(inverseR+krf*r2-crf);
else
energy += chargeProd*inverseR;
for (int j = 0; j < 4; j++)
if (include[j])
*totalEnergy += energy[j];
}
// accumulate energies
// Accumulate forces.
fvec4 atomForce(forces+4*atom);
for (int j = 0; j < 4; j++) {
if (include[j]) {
fvec4 result = blockAtomDelta[j]*dEdR[j];
blockAtomForce[j] += result;
atomForce -= result;
}
}
atomForce.store(forces+4*atom);
}
// Record the forces on the block atoms.
for (int j = 0; j < 4; j++)
(fvec4(forces+4*blockAtom[j])+blockAtomForce[j]).store(forces+4*blockAtom[j]);
}
void CpuNonbondedForce::calculateBlockEwaldIxn(int blockIndex, float* forces, double* totalEnergy, const fvec4& boxSize, const fvec4& invBoxSize) {
// Load the positions and parameters of the atoms in the block.
int blockAtom[4];
fvec4 blockAtomPosq[4];
fvec4 blockAtomForce[4];
for (int i = 0; i < 4; i++) {
blockAtom[i] = neighborList->getSortedAtoms()[4*blockIndex+i];
blockAtomPosq[i] = fvec4(posq+4*blockAtom[i]);
blockAtomForce[i] = fvec4(0.0f);
}
fvec4 blockAtomCharge = fvec4(ONE_4PI_EPS0)*fvec4(blockAtomPosq[0][3], blockAtomPosq[1][3], blockAtomPosq[2][3], blockAtomPosq[3][3]);
fvec4 blockAtomSigma(atomParameters[blockAtom[0]].first, atomParameters[blockAtom[1]].first, atomParameters[blockAtom[2]].first, atomParameters[blockAtom[3]].first);
fvec4 blockAtomEpsilon(atomParameters[blockAtom[0]].second, atomParameters[blockAtom[1]].second, atomParameters[blockAtom[2]].second, atomParameters[blockAtom[3]].second);
// Loop over neighbors for this block.
const vector<int>& neighbors = neighborList->getBlockNeighbors(blockIndex);
const vector<char>& exclusions = neighborList->getBlockExclusions(blockIndex);
float blockAtomR2[4];
bool include[4];
fvec4 blockAtomDelta[4];
for (int i = 0; i < (int) neighbors.size(); i++) {
// Load the next neighbor.
int atom = neighbors[i];
fvec4 atomPosq(posq+4*atom);
// Compute the distances to the block atoms.
bool any = false;
for (int j = 0; j < 4; j++) {
getDeltaR(atomPosq, blockAtomPosq[j], blockAtomDelta[j], blockAtomR2[j], periodic, boxSize, invBoxSize);
include[j] = (((exclusions[i]>>j)&1) == 0 && blockAtomR2[j] < cutoffDistance*cutoffDistance);
any |= include[j];
}
if (!any)
continue; // No interactions to compute.
// Compute the interactions.
fvec4 r2(blockAtomR2);
fvec4 r = sqrt(r2);
fvec4 inverseR = fvec4(1.0f)/r;
fvec4 switchValue(1.0f), switchDeriv(0.0f);
if (useSwitch) {
fvec4 t = (r>switchingDistance) & ((r-switchingDistance)/(cutoffDistance-switchingDistance));
switchValue = 1+t*t*t*(-10.0f+t*(15.0f-t*6.0f));
switchDeriv = t*t*(-30.0f+t*(60.0f-t*30.0f))/(cutoffDistance-switchingDistance);
}
fvec4 chargeProd = blockAtomCharge*posq[4*atom+3];
fvec4 dEdR = chargeProd*inverseR*ewaldScaleFunction(r);
fvec4 sig = blockAtomSigma+atomParameters[atom].first;
fvec4 sig2 = inverseR*sig;
sig2 *= sig2;
fvec4 sig6 = sig2*sig2*sig2;
fvec4 eps = blockAtomEpsilon*atomParameters[atom].second;
dEdR += switchValue*eps*(12.0f*sig6 - 6.0f)*sig6;
dEdR *= inverseR*inverseR;
fvec4 energy = eps*(sig6-1.0f)*sig6;
if (useSwitch) {
dEdR -= energy*switchDeriv*inverseR;
energy *= switchValue;
}
// Accumulate energies.
if (totalEnergy) {
energy += (float) (chargeProd*inverseR*erfcApprox(alphaEwald*r));
*totalEnergy += energy;
energy += chargeProd*inverseR*erfcApprox(alphaEwald*r);
for (int j = 0; j < 4; j++)
if (include[j])
*totalEnergy += energy[j];
}
// Accumulate forces.
fvec4 atomForce(forces+4*atom);
for (int j = 0; j < 4; j++) {
if (include[j]) {
fvec4 result = blockAtomDelta[j]*dEdR[j];
blockAtomForce[j] += result;
atomForce -= result;
}
}
atomForce.store(forces+4*atom);
}
// Record the forces on the block atoms.
for (int j = 0; j < 4; j++)
(fvec4(forces+4*blockAtom[j])+blockAtomForce[j]).store(forces+4*blockAtom[j]);
}
void CpuNonbondedForce::getDeltaR(const __m128& posI, const __m128& posJ, __m128& deltaR, float& r2, bool periodic, const __m128& boxSize, const __m128& invBoxSize) const {
deltaR = _mm_sub_ps(posJ, posI);
void CpuNonbondedForce::getDeltaR(const fvec4& posI, const fvec4& posJ, fvec4& deltaR, float& r2, bool periodic, const fvec4& boxSize, const fvec4& invBoxSize) const {
deltaR = posJ-posI;
if (periodic) {
__m128 base = _mm_mul_ps(_mm_floor_ps(_mm_add_ps(_mm_mul_ps(deltaR, invBoxSize), _mm_set1_ps(0.5))), boxSize);
deltaR = _mm_sub_ps(deltaR, base);
fvec4 base = round(deltaR*invBoxSize)*boxSize;
deltaR = deltaR-base;
}
r2 = _mm_cvtss_f32(_mm_dp_ps(deltaR, deltaR, 0x71));
r2 = dot3(deltaR, deltaR);
}
float CpuNonbondedForce::erfcApprox(float x) {
fvec4 CpuNonbondedForce::erfcApprox(fvec4 x) {
// This approximation for erfc is from Abramowitz and Stegun (1964) p. 299. They cite the following as
// the original source: C. Hastings, Jr., Approximations for Digital Computers (1955). It has a maximum
// error of 3e-7.
float t = 1.0f+(0.0705230784f+(0.0422820123f+(0.0092705272f+(0.0001520143f+(0.0002765672f+0.0000430638f*x)*x)*x)*x)*x)*x;
fvec4 t = 1.0f+(0.0705230784f+(0.0422820123f+(0.0092705272f+(0.0001520143f+(0.0002765672f+0.0000430638f*x)*x)*x)*x)*x)*x;
t *= t;
t *= t;
t *= t;
return 1.0f/(t*t);
}
float CpuNonbondedForce::ewaldScaleFunction(float x) {
fvec4 CpuNonbondedForce::ewaldScaleFunction(fvec4 x) {
// Compute the tabulated Ewald scale factor: erfc(alpha*r) + 2*alpha*r*exp(-alpha*alpha*r*r)/sqrt(PI)
int lower = (int) (x*ewaldDXInv);
int upper = lower+1;
float a = (ewaldScaleX[upper]-x)*ewaldDXInv;
float b = 1.0f-a;
return a*ewaldScaleY[lower]+b*ewaldScaleY[upper]+((a*a*a-a)*ewaldScaleDeriv[lower] + (b*b*b-b)*ewaldScaleDeriv[upper]);
float y[4];
fvec4 x1 = x*ewaldDXInv;
ivec4 index = floor(x1);
fvec4 coeff[4];
coeff[1] = x1-index;
coeff[0] = 1.0f-coeff[1];
coeff[2] = coeff[0]*coeff[0]*coeff[0]-coeff[0];
coeff[3] = coeff[1]*coeff[1]*coeff[1]-coeff[1];
_MM_TRANSPOSE4_PS(coeff[0], coeff[1], coeff[2], coeff[3]);
for (int i = 0; i < 4; i++) {
if (index[i] < NUM_TABLE_POINTS)
y[i] = dot4(coeff[i], fvec4(&ewaldScaleTable[4*index[i]]));
}
return fvec4(y);
}
platforms/cpu/tests/TestCpuNeighborList.cpp
View file @ 79fd69c3
......@@ -39,6 +39,7 @@
#include "sfmt/SFMT.h"
#include <iostream>
#include <set>
#include <utility>
#include <vector>
using namespace OpenMM;
......@@ -68,10 +69,19 @@ void testNeighborList(bool periodic) {
// Convert the neighbor list to a set for faster lookup.
set<pair<int, int> > neighbors;
for (int i = 0; i < (int) neighborList.getNeighbors().size(); i++) {
pair<int, int> entry = neighborList.getNeighbors()[i];
ASSERT(neighbors.find(entry) == neighbors.end() && neighbors.find(make_pair(entry.second, entry.first)) == neighbors.end()); // No duplicates
neighbors.insert(entry);
for (int i = 0; i < (int) neighborList.getSortedAtoms().size(); i++) {
int blockIndex = i/CpuNeighborList::BlockSize;
int indexInBlock = i-blockIndex*CpuNeighborList::BlockSize;
char mask = 1<<indexInBlock;
for (int j = 0; j < (int) neighborList.getBlockExclusions(blockIndex).size(); j++) {
if ((neighborList.getBlockExclusions(blockIndex)[j] & mask) == 0) {
int atom1 = neighborList.getSortedAtoms()[i];
int atom2 = neighborList.getBlockNeighbors(blockIndex)[j];
pair<int, int> entry = make_pair(min(atom1, atom2), max(atom1, atom2));
ASSERT(neighbors.find(entry) == neighbors.end() && neighbors.find(make_pair(entry.second, entry.first)) == neighbors.end()); // No duplicates
neighbors.insert(entry);
}
}
}
// Check each particle pair and figure out whether they should be in the neighbor list.
......@@ -90,7 +100,8 @@ void testNeighborList(bool periodic) {
if (dx*dx + dy*dy + dz*dz > cutoff*cutoff)
shouldInclude = false;
bool isIncluded = (neighbors.find(make_pair(i, j)) != neighbors.end() || neighbors.find(make_pair(j, i)) != neighbors.end());
ASSERT_EQUAL(shouldInclude, isIncluded);
if (shouldInclude)
ASSERT(isIncluded);
}
}
......
plugins/cpupme/src/CpuPmeKernels.cpp
View file @ 79fd69c3
......@@ -35,9 +35,9 @@
#include "CpuPmeKernels.h"
#include "SimTKOpenMMRealType.h"
#include "openmm/internal/hardware.h"
#include "openmm/internal/vectorize.h"
#include <cmath>
#include <cstring>
#include <smmintrin.h>
using namespace OpenMM;
using namespace std;
......@@ -47,52 +47,50 @@ static const int PME_ORDER = 5;
bool CpuCalcPmeReciprocalForceKernel::hasInitializedThreads = false;
int CpuCalcPmeReciprocalForceKernel::numThreads = 0;
#define EXTRACT_FLOAT(v, element) _mm_cvtss_f32(_mm_shuffle_ps(v, v, _MM_SHUFFLE(0, 0, 0, element)))
static void spreadCharge(int start, int end, float* posq, float* grid, int gridx, int gridy, int gridz, int numParticles, Vec3 periodicBoxSize) {
float temp[4];
__m128 boxSize = _mm_set_ps(0, (float) periodicBoxSize[2], (float) periodicBoxSize[1], (float) periodicBoxSize[0]);
__m128 invBoxSize = _mm_set_ps(0, (float) (1/periodicBoxSize[2]), (float) (1/periodicBoxSize[1]), (float) (1/periodicBoxSize[0]));
__m128 gridSize = _mm_set_ps(0, gridz, gridy, gridx);
__m128i gridSizeInt = _mm_set_epi32(0, gridz, gridy, gridx);
__m128 one = _mm_set1_ps(1);
__m128 scale = _mm_set1_ps(1.0f/(PME_ORDER-1));
fvec4 boxSize((float) periodicBoxSize[0], (float) periodicBoxSize[1], (float) periodicBoxSize[2], 0);
fvec4 invBoxSize((float) (1/periodicBoxSize[0]), (float) (1/periodicBoxSize[1]), (float) (1/periodicBoxSize[2]), 0);
fvec4 gridSize(gridx, gridy, gridz, 0);
ivec4 gridSizeInt(gridx, gridy, gridz, 0);
fvec4 one(1);
fvec4 scale(1.0f/(PME_ORDER-1));
const float epsilonFactor = sqrt(ONE_4PI_EPS0);
memset(grid, 0, sizeof(float)*gridx*gridy*gridz);
for (int i = start; i < end; i++) {
// Find the position relative to the nearest grid point.
__m128 pos = _mm_loadu_ps(&posq[4*i]);
__m128 posFloor = _mm_floor_ps(_mm_mul_ps(pos, invBoxSize));
__m128 posInBox = _mm_sub_ps(pos, _mm_mul_ps(boxSize, posFloor));
__m128 t = _mm_mul_ps(_mm_mul_ps(posInBox, invBoxSize), gridSize);
__m128i ti = _mm_cvttps_epi32(t);
__m128 dr = _mm_sub_ps(t, _mm_cvtepi32_ps(ti));
__m128i gridIndex = _mm_sub_epi32(ti, _mm_and_si128(gridSizeInt, _mm_cmpeq_epi32(ti, gridSizeInt)));
fvec4 pos(&posq[4*i]);
fvec4 posFloor = floor(pos*invBoxSize);
fvec4 posInBox = pos-boxSize*posFloor;
fvec4 t = posInBox*invBoxSize*gridSize;
ivec4 ti = t;
fvec4 dr = t-ti;
ivec4 gridIndex = ti-(gridSizeInt&ti==gridSizeInt);
// Compute the B-spline coefficients.
__m128 data[PME_ORDER];
data[PME_ORDER-1] = _mm_setzero_ps();
fvec4 data[PME_ORDER];
data[PME_ORDER-1] = 0.0f;
data[1] = dr;
data[0] = _mm_sub_ps(one, dr);
data[0] = one-dr;
for (int j = 3; j < PME_ORDER; j++) {
__m128 div = _mm_set1_ps(1.0f/(j-1));
data[j-1] = _mm_mul_ps(_mm_mul_ps(div, dr), data[j-2]);
fvec4 div(1.0f/(j-1));
data[j-1] = div*dr*data[j-2];
for (int k = 1; k < j-1; k++)
data[j-k-1] = _mm_mul_ps(div, _mm_add_ps(_mm_mul_ps(_mm_add_ps(dr, _mm_set1_ps(k)), data[j-k-2]), _mm_mul_ps(_mm_sub_ps(_mm_set1_ps(j-k), dr), data[j-k-1])));
data[0] = _mm_mul_ps(_mm_mul_ps(div, _mm_sub_ps(one, dr)), data[0]);
data[j-k-1] = div*((dr+k)*data[j-k-2]+(fvec4(j-k)-dr)*data[j-k-1]);
data[0] = div*(one-dr)*data[0];
}
data[PME_ORDER-1] = _mm_mul_ps(_mm_mul_ps(scale, dr), data[PME_ORDER-2]);
data[PME_ORDER-1] = scale*dr*data[PME_ORDER-2];
for (int j = 1; j < (PME_ORDER-1); j++)
data[PME_ORDER-j-1] = _mm_mul_ps(scale, _mm_add_ps(_mm_mul_ps(_mm_add_ps(dr, _mm_set1_ps(j)), data[PME_ORDER-j-2]), _mm_mul_ps(_mm_sub_ps(_mm_set1_ps(PME_ORDER-j), dr), data[PME_ORDER-j-1])));
data[0] = _mm_mul_ps(_mm_mul_ps(scale, _mm_sub_ps(one, dr)), data[0]);
data[PME_ORDER-j-1] = scale*((dr+j)*data[PME_ORDER-j-2]+(fvec4(PME_ORDER-j)-dr)*data[PME_ORDER-j-1]);
data[0] = scale*(one-dr)*data[0];
// Spread the charges.
int gridIndexX = _mm_extract_epi32(gridIndex, 0);
int gridIndexY = _mm_extract_epi32(gridIndex, 1);
int gridIndexZ = _mm_extract_epi32(gridIndex, 2);
int gridIndexX = gridIndex[0];
int gridIndexY = gridIndex[1];
int gridIndexZ = gridIndex[2];
if (gridIndexX < 0)
return; // This happens when a simulation blows up and coordinates become NaN.
int zindex[PME_ORDER];
......@@ -101,21 +99,21 @@ static void spreadCharge(int start, int end, float* posq, float* grid, int gridx
zindex[j] -= (zindex[j] >= gridz ? gridz : 0);
}
float charge = epsilonFactor*posq[4*i+3];
__m128 zdata0to3 = _mm_set_ps(EXTRACT_FLOAT(data[3], 2), EXTRACT_FLOAT(data[2], 2), EXTRACT_FLOAT(data[1], 2), EXTRACT_FLOAT(data[0], 2));
float zdata4 = EXTRACT_FLOAT(data[4], 2);
fvec4 zdata0to3(data[0][2], data[1][2], data[2][2], data[3][2]);
float zdata4 = data[4][2];
if (gridIndexZ+4 < gridz) {
for (int ix = 0; ix < PME_ORDER; ix++) {
int xbase = gridIndexX+ix;
xbase -= (xbase >= gridx ? gridx : 0);
xbase = xbase*gridy*gridz;
float xdata = charge*EXTRACT_FLOAT(data[ix], 0);
float xdata = charge*data[ix][0];
for (int iy = 0; iy < PME_ORDER; iy++) {
int ybase = gridIndexY+iy;
ybase -= (ybase >= gridy ? gridy : 0);
ybase = xbase + ybase*gridz;
float multiplier = xdata*EXTRACT_FLOAT(data[iy], 1);
__m128 add0to3 = _mm_mul_ps(zdata0to3, _mm_set1_ps(multiplier));
_mm_storeu_ps(&grid[ybase+gridIndexZ], _mm_add_ps(_mm_loadu_ps(&grid[ybase+gridIndexZ]), add0to3));
float multiplier = xdata*data[iy][1];
fvec4 add0to3 = zdata0to3*multiplier;
(fvec4(&grid[ybase+gridIndexZ])+add0to3).store(&grid[ybase+gridIndexZ]);
grid[ybase+zindex[4]] += multiplier*zdata4;
}
}
......@@ -125,14 +123,14 @@ static void spreadCharge(int start, int end, float* posq, float* grid, int gridx
int xbase = gridIndexX+ix;
xbase -= (xbase >= gridx ? gridx : 0);
xbase = xbase*gridy*gridz;
float xdata = charge*EXTRACT_FLOAT(data[ix], 0);
float xdata = charge*data[ix][0];
for (int iy = 0; iy < PME_ORDER; iy++) {
int ybase = gridIndexY+iy;
ybase -= (ybase >= gridy ? gridy : 0);
ybase = xbase + ybase*gridz;
float multiplier = xdata*EXTRACT_FLOAT(data[iy], 1);
__m128 add0to3 = _mm_mul_ps(zdata0to3, _mm_set1_ps(multiplier));
_mm_store_ps(temp, add0to3);
float multiplier = xdata*data[iy][1];
fvec4 add0to3 = zdata0to3*multiplier;
add0to3.store(temp);
grid[ybase+zindex[0]] += temp[0];
grid[ybase+zindex[1]] += temp[1];
grid[ybase+zindex[2]] += temp[2];
......@@ -245,51 +243,51 @@ static void reciprocalConvolution(int start, int end, fftwf_complex* grid, int g
}
static void interpolateForces(int start, int end, float* posq, float* force, float* grid, int gridx, int gridy, int gridz, int numParticles, Vec3 periodicBoxSize) {
__m128 boxSize = _mm_set_ps(0, (float) periodicBoxSize[2], (float) periodicBoxSize[1], (float) periodicBoxSize[0]);
__m128 invBoxSize = _mm_set_ps(0, (float) (1/periodicBoxSize[2]), (float) (1/periodicBoxSize[1]), (float) (1/periodicBoxSize[0]));
__m128 gridSize = _mm_set_ps(0, gridz, gridy, gridx);
__m128i gridSizeInt = _mm_set_epi32(0, gridz, gridy, gridx);
__m128 one = _mm_set1_ps(1);
__m128 scale = _mm_set1_ps(1.0f/(PME_ORDER-1));
fvec4 boxSize((float) periodicBoxSize[0], (float) periodicBoxSize[1], (float) periodicBoxSize[2], 0);
fvec4 invBoxSize((float) (1/periodicBoxSize[0]), (float) (1/periodicBoxSize[1]), (float) (1/periodicBoxSize[2]), 0);
fvec4 gridSize(gridx, gridy, gridz, 0);
ivec4 gridSizeInt(gridx, gridy, gridz, 0);
fvec4 one(1);
fvec4 scale(1.0f/(PME_ORDER-1));
const float epsilonFactor = sqrt(ONE_4PI_EPS0);
for (int i = start; i < end; i++) {
// Find the position relative to the nearest grid point.
__m128 pos = _mm_loadu_ps(&posq[4*i]);
__m128 posFloor = _mm_floor_ps(_mm_mul_ps(pos, invBoxSize));
__m128 posInBox = _mm_sub_ps(pos, _mm_mul_ps(boxSize, posFloor));
__m128 t = _mm_mul_ps(_mm_mul_ps(posInBox, invBoxSize), gridSize);
__m128i ti = _mm_cvttps_epi32(t);
__m128 dr = _mm_sub_ps(t, _mm_cvtepi32_ps(ti));
__m128i gridIndex = _mm_sub_epi32(ti, _mm_and_si128(gridSizeInt, _mm_cmpeq_epi32(ti, gridSizeInt)));
fvec4 pos(&posq[4*i]);
fvec4 posFloor = floor(pos*invBoxSize);
fvec4 posInBox = pos-boxSize*posFloor;
fvec4 t = posInBox*invBoxSize*gridSize;
ivec4 ti = t;
fvec4 dr = t-ti;
ivec4 gridIndex = ti-(gridSizeInt&ti==gridSizeInt);
// Compute the B-spline coefficients.
__m128 data[PME_ORDER];
__m128 ddata[PME_ORDER];
data[PME_ORDER-1] = _mm_setzero_ps();
fvec4 data[PME_ORDER];
fvec4 ddata[PME_ORDER];
data[PME_ORDER-1] = 0.0f;
data[1] = dr;
data[0] = _mm_sub_ps(one, dr);
data[0] = one-dr;
for (int j = 3; j < PME_ORDER; j++) {
__m128 div = _mm_set1_ps(1.0f/(j-1));
data[j-1] = _mm_mul_ps(_mm_mul_ps(div, dr), data[j-2]);
fvec4 div(1.0f/(j-1));
data[j-1] = div*dr*data[j-2];
for (int k = 1; k < j-1; k++)
data[j-k-1] = _mm_mul_ps(div, _mm_add_ps(_mm_mul_ps(_mm_add_ps(dr, _mm_set1_ps(k)), data[j-k-2]), _mm_mul_ps(_mm_sub_ps(_mm_set1_ps(j-k), dr), data[j-k-1])));
data[0] = _mm_mul_ps(_mm_mul_ps(div, _mm_sub_ps(one, dr)), data[0]);
data[j-k-1] = div*((dr+k)*data[j-k-2]+(fvec4(j-k)-dr)*data[j-k-1]);
data[0] = div*(one-dr)*data[0];
}
ddata[0] = _mm_sub_ps(_mm_set1_ps(0), data[0]);
ddata[0] = -data[0];
for (int j = 1; j < PME_ORDER; j++)
ddata[j] = _mm_sub_ps(data[j-1], data[j]);
data[PME_ORDER-1] = _mm_mul_ps(_mm_mul_ps(scale, dr), data[PME_ORDER-2]);
ddata[j] = data[j-1]-data[j];
data[PME_ORDER-1] = scale*dr*data[PME_ORDER-2];
for (int j = 1; j < (PME_ORDER-1); j++)
data[PME_ORDER-j-1] = _mm_mul_ps(scale, _mm_add_ps(_mm_mul_ps(_mm_add_ps(dr, _mm_set1_ps(j)), data[PME_ORDER-j-2]), _mm_mul_ps(_mm_sub_ps(_mm_set1_ps(PME_ORDER-j), dr), data[PME_ORDER-j-1])));
data[0] = _mm_mul_ps(_mm_mul_ps(scale, _mm_sub_ps(one, dr)), data[0]);
data[PME_ORDER-j-1] = scale*((dr+j)*data[PME_ORDER-j-2]+(fvec4(PME_ORDER-j)-dr)*data[PME_ORDER-j-1]);
data[0] = scale*(one-dr)*data[0];
// Compute the force on this atom.
int gridIndexX = _mm_extract_epi32(gridIndex, 0);
int gridIndexY = _mm_extract_epi32(gridIndex, 1);
int gridIndexZ = _mm_extract_epi32(gridIndex, 2);
int gridIndexX = gridIndex[0];
int gridIndexY = gridIndex[1];
int gridIndexZ = gridIndex[2];
if (gridIndexX < 0)
return; // This happens when a simulation blows up and coordinates become NaN.
int zindex[PME_ORDER];
......@@ -297,34 +295,34 @@ static void interpolateForces(int start, int end, float* posq, float* force, flo
zindex[j] = gridIndexZ+j;
zindex[j] -= (zindex[j] >= gridz ? gridz : 0);
}
__m128 zdata[PME_ORDER];
fvec4 zdata[PME_ORDER];
for (int j = 0; j < PME_ORDER; j++)
zdata[j] = _mm_set_ps(0, EXTRACT_FLOAT(ddata[j], 2), EXTRACT_FLOAT(data[j], 2), EXTRACT_FLOAT(data[j], 2));
__m128 f = _mm_set1_ps(0);
zdata[j] = fvec4(data[j][2], data[j][2], ddata[j][2], 0);
fvec4 f = 0.0f;
for (int ix = 0; ix < PME_ORDER; ix++) {
int xbase = gridIndexX+ix;
xbase -= (xbase >= gridx ? gridx : 0);
xbase = xbase*gridy*gridz;
float dx = EXTRACT_FLOAT(data[ix], 0);
float ddx = EXTRACT_FLOAT(ddata[ix], 0);
__m128 xdata = _mm_set_ps(0, dx, dx, ddx);
float dx = data[ix][0];
float ddx = ddata[ix][0];
fvec4 xdata(ddx, dx, dx, 0);
for (int iy = 0; iy < PME_ORDER; iy++) {
int ybase = gridIndexY+iy;
ybase -= (ybase >= gridy ? gridy : 0);
ybase = xbase + ybase*gridz;
float dy = EXTRACT_FLOAT(data[iy], 1);
float ddy = EXTRACT_FLOAT(ddata[iy], 1);
__m128 xydata = _mm_mul_ps(xdata, _mm_set_ps(0, dy, ddy, dy));
float dy = data[iy][1];
float ddy = ddata[iy][1];
fvec4 xydata = xdata*fvec4(dy, ddy, dy, 0);
for (int iz = 0; iz < PME_ORDER; iz++) {
__m128 gridValue = _mm_set1_ps(grid[ybase+zindex[iz]]);
f = _mm_add_ps(f, _mm_mul_ps(xydata, _mm_mul_ps(zdata[iz], gridValue)));
fvec4 gridValue(grid[ybase+zindex[iz]]);
f = f+xydata*zdata[iz]*gridValue;
}
}
}
f = _mm_mul_ps(invBoxSize, _mm_mul_ps(gridSize, _mm_mul_ps(f, _mm_set1_ps(-epsilonFactor*posq[4*i+3]))));
_mm_store_ps(&force[4*i], f);
f = invBoxSize*gridSize*f*(-epsilonFactor*posq[4*i+3]);
f.store(&force[4*i]);
}
}
......@@ -509,10 +507,10 @@ void CpuCalcPmeReciprocalForceKernel::runThread(int index) {
threadWait();
int numGrids = threadData.size();
for (int i = gridStart; i < gridEnd; i += 4) {
__m128 sum = _mm_load_ps(&realGrid[i]);
fvec4 sum(&realGrid[i]);
for (int j = 1; j < numGrids; j++)
sum = _mm_add_ps(sum, _mm_load_ps(&threadData[j]->tempGrid[i]));
_mm_store_ps(&realGrid[i], sum);
sum += fvec4(&threadData[j]->tempGrid[i]);
sum.store(&realGrid[i]);
}
threadWait();
if (lastBoxSize != periodicBoxSize) {
......
wrappers/python/src/swig_doxygen/swigInputConfig.py
View file @ 79fd69c3
......@@ -145,6 +145,8 @@ SKIP_METHODS = [('State',),
('IntegrateDrudeSCFStepKernel',),
('XmlSerializer', 'serialize'),
('XmlSerializer', 'deserialize'),
('fvec4',),
('ivec4',),
]
# The build script assumes method args that are non-const references are
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment