Skip to content

GitLab

Commit ec39f6ff authored by Yutong Zhao's avatar Yutong Zhao
Browse files

nonbonded.cu will now use shuffles on sm_30 or higher

parent 44665537
Hide whitespace changes
Inline Side-by-side
Showing with 412 additions and 340 deletions
platforms/cuda/src/CudaContext.cpp
View file @ ec39f6ff
......@@ -394,7 +394,7 @@ CUmodule CudaContext::createModule(const string source, const map<string, string
// Write out the source to a temporary file.
stringstream tempFileName;
tempFileName << "openmmTempKernel" << /*rand() <<*/ this; // Include a pointer to this context as part of the filename to avoid collisions.
tempFileName << "openmmTempKernel" << this; // Include a pointer to this context as part of the filename to avoid collisions.
string inputFile = (tempDir+tempFileName.str()+".cu");
string outputFile = (tempDir+tempFileName.str()+".ptx");
string logFile = (tempDir+tempFileName.str()+".log");
......@@ -438,15 +438,15 @@ CUmodule CudaContext::createModule(const string source, const map<string, string
m<<"Error loading CUDA module: "<<getErrorString(result)<<" ("<<result<<")";
throw OpenMMException(m.str());
}
//remove(inputFile.c_str());
//remove(outputFile.c_str());
//remove(logFile.c_str());
remove(inputFile.c_str());
remove(outputFile.c_str());
remove(logFile.c_str());
return module;
}
catch (...) {
//remove(inputFile.c_str());
//remove(outputFile.c_str());
//remove(logFile.c_str());
remove(inputFile.c_str());
remove(outputFile.c_str());
remove(logFile.c_str());
throw;
}
}
......
platforms/cuda/src/CudaNonbondedUtilities.cpp
View file @ ec39f6ff
......@@ -416,6 +416,11 @@ void CudaNonbondedUtilities::setAtomBlockRange(double startFraction, double endF
}
CUfunction CudaNonbondedUtilities::createInteractionKernel(const string& source, vector<ParameterInfo>& params, vector<ParameterInfo>& arguments, bool useExclusions, bool isSymmetric) {
map<string, string> defines;
if (context.getComputeCapability() >= 3.0 && !context.getUseDoublePrecision())
defines["ENABLE_SHUFFLE"] = "1";
map<string, string> replacements;
replacements["COMPUTE_INTERACTION"] = source;
const string suffixes[] = {"x", "y", "z", "w"};
......@@ -446,163 +451,145 @@ CUfunction CudaNonbondedUtilities::createInteractionKernel(const string& source,
}
replacements["PARAMETER_ARGUMENTS"] = args.str();
/*
stringstream loadLocal1;
stringstream load1;
for (int i = 0; i < (int) params.size(); i++) {
if (params[i].getNumComponents() == 1) {
loadLocal1<<"localData[localAtomIndex]."<<params[i].getName()<<" = "<<params[i].getName()<<"1;\n";
}
else {
for (int j = 0; j < params[i].getNumComponents(); ++j)
loadLocal1<<"localData[localAtomIndex]."<<params[i].getName()<<"_"<<suffixes[j]<<" = "<<params[i].getName()<<"1."<<suffixes[j]<<";\n";
}
load1 << params[i].getType();
load1 << " ";
load1 << params[i].getName();
load1 << "1 = global_";
load1 << params[i].getName();
load1 << "[atom1];\n";
}
replacements["LOAD_LOCAL_PARAMETERS_FROM_1"] = loadLocal1.str();
*/
replacements["LOAD_ATOM1_PARAMETERS"] = load1.str();
bool useShuffle = (defines["ENABLE_SHUFFLE"]=="1");
// Part 1. Defines for on diagonal exclusion tiles
stringstream loadLocal1;
loadLocal1 << "tempSigmaEpsilon = sigmaEpsilon1;" << endl;
//for (int i = 0; i < (int) params.size(); i++) {
// loadLocal1<<params[i].getType()<<" temp"<<params[i].getName()<<"="<<params[i].getName()<<"1;\n";
//}
//cout << loadLocal1.str() << endl;
if(useShuffle) {
// not needed if using shuffles as we can directly fetch from
// LOAD_ATOM1_PARAMETERS
} else {
for (int i = 0; i < (int) params.size(); i++) {
if (params[i].getNumComponents() == 1) {
loadLocal1<<"localData[threadIdx.x]."<<params[i].getName()<<" = "<<params[i].getName()<<"1;\n";
}
else {
for (int j = 0; j < params[i].getNumComponents(); ++j)
loadLocal1<<"localData[threadIdx.x]."<<params[i].getName()<<"_"<<suffixes[j]<<" = "<<params[i].getName()<<"1."<<suffixes[j]<<";\n";
}
}
}
replacements["LOAD_LOCAL_PARAMETERS_FROM_1"] = loadLocal1.str();
/*
stringstream loadLocal2;
for (int i = 0; i < (int) params.size(); i++) {
if (params[i].getNumComponents() == 1) {
loadLocal2<<"localData[localAtomIndex]."<<params[i].getName()<<" = global_"<<params[i].getName()<<"[j];\n";
}
else {
loadLocal2<<params[i].getType()<<" temp_"<<params[i].getName()<<" = global_"<<params[i].getName()<<"[j];\n";
for (int j = 0; j < params[i].getNumComponents(); ++j)
loadLocal2<<"localData[localAtomIndex]."<<params[i].getName()<<"_"<<suffixes[j]<<" = temp_"<<params[i].getName()<<"."<<suffixes[j]<<";\n";
stringstream broadcastWarpData;
if(useShuffle) {
broadcastWarpData << "posq2.x = real_shfl(shflPosq.x, j);\n";
broadcastWarpData << "posq2.y = real_shfl(shflPosq.y, j);\n";
broadcastWarpData << "posq2.z = real_shfl(shflPosq.z, j);\n";
broadcastWarpData << "posq2.w = real_shfl(shflPosq.w, j);\n";
for(int i=0; i< (int) params.size();i++) {
broadcastWarpData << params[i].getType() << " shfl" << params[i].getName() << ";\n";
for(int j=0; j < params[i].getNumComponents(); j++) {
string name;
if (params[i].getNumComponents() == 1) {
broadcastWarpData << "shfl" << params[i].getName() << "=real_shfl(" << params[i].getName() <<"1,j);\n";
} else {
broadcastWarpData << "shfl" << params[i].getName()+"."+suffixes[j] << "=real_shfl(" << params[i].getName()+"1."+suffixes[j] <<",j);\n";
}
}
}
} else {
// not used if not shuffling
}
replacements["LOAD_LOCAL_PARAMETERS_FROM_GLOBAL"] = loadLocal2.str();
*/
replacements["BROADCAST_WARP_DATA"] = broadcastWarpData.str();
// Part 2. Defines for off-diagonal exclusions, and neighborlist tiles.
stringstream declareLocal2;
for(int i=0; i< (int) params.size(); i++) {
if (params[i].getNumComponents() == 1) {
// loadLocal2<<params[i].getType()<<" "<<params[i].getName()<<" = global_"<<params[i].getName()<<"[j];\n";
} else {
declareLocal2<<params[i].getType()<<" temp"<<params[i].getName()<<";\n";
if(useShuffle) {
for(int i=0; i< (int) params.size(); i++) {
declareLocal2<<params[i].getType()<<" shfl"<<params[i].getName()<<";\n";
//if (params[i].getNumComponents() == 1) {
//declareLocal2<<params[i].getType()<<" "<<params[i].getName()<<" = global_"<<params[i].getName()<<"[j];\n";
//} else {
// declareLocal2<<params[i].getType()<<" temp"<<params[i].getName()<<";\n";
//}
}
} else {
// not used if using shared memory
}
replacements["DECLARE_LOCAL_PARAMETERS"] = declareLocal2.str();
stringstream loadLocal2;
for(int i=0; i< (int) params.size(); i++) {
if (params[i].getNumComponents() == 1) {
// loadLocal2<<params[i].getType()<<" "<<params[i].getName()<<" = global_"<<params[i].getName()<<"[j];\n";
} else {
loadLocal2<<"temp"<<params[i].getName()<<" = global_"<<params[i].getName()<<"[j];\n";
if(useShuffle) {
for(int i=0; i< (int) params.size(); i++) {
loadLocal2<<"shfl"<<params[i].getName()<<" = global_"<<params[i].getName()<<"[j];\n";
}
}
/*
for (int i = 0; i < (int) params.size(); i++) {
if (params[i].getNumComponents() == 1) {
loadLocal2<<params[i].getType()<<" "<<params[i].getName()<<" = global_"<<params[i].getName()<<"[j];\n";
}
else {
loadLocal2<<params[i].getType()<<" temp_"<<params[i].getName()<<" = global_"<<params[i].getName()<<"[j];\n";
for (int j = 0; j < params[i].getNumComponents(); ++j)
loadLocal2<<params[i].getType()<<" "<<params[i].getName()<<"_"<<suffixes[j]<<" = temp_"<<params[i].getName()<<"."<<suffixes[j]<<";\n";
} else {
for (int i = 0; i < (int) params.size(); i++) {
if (params[i].getNumComponents() == 1) {
loadLocal2<<"localData[threadIdx.x]."<<params[i].getName()<<" = global_"<<params[i].getName()<<"[j];\n";
}
else {
loadLocal2<<params[i].getType()<<" temp_"<<params[i].getName()<<" = global_"<<params[i].getName()<<"[j];\n";
for (int j = 0; j < params[i].getNumComponents(); ++j)
loadLocal2<<"localData[threadIdx.x]."<<params[i].getName()<<"_"<<suffixes[j]<<" = temp_"<<params[i].getName()<<"."<<suffixes[j]<<";\n";
}
}
}
*/
replacements["LOAD_LOCAL_PARAMETERS_FROM_GLOBAL"] = loadLocal2.str();
stringstream load1;
for (int i = 0; i < (int) params.size(); i++) {
load1 << params[i].getType();
load1 << " ";
load1 << params[i].getName();
load1 << "1 = global_";
load1 << params[i].getName();
load1 << "[atom1];\n";
}
replacements["LOAD_ATOM1_PARAMETERS"] = load1.str();
/*
stringstream load2j;
for (int i = 0; i < (int) params.size(); i++) {
if (params[i].getNumComponents() == 1) {
load2j<<params[i].getType()<<" "<<params[i].getName()<<"2 = localData[atom2]."<<params[i].getName()<<";\n";
}
else {
load2j<<params[i].getType()<<" "<<params[i].getName()<<"2 = make_"<<params[i].getType()<<"(";
for (int j = 0; j < params[i].getNumComponents(); ++j) {
if (j > 0)
load2j<<", ";
load2j<<"localData[atom2]."<<params[i].getName()<<"_"<<suffixes[j];
if(useShuffle) {
for(int i = 0; i < (int) params.size(); i++)
load2j<<params[i].getType()<<" "<<params[i].getName()<<"2 = shfl"<<params[i].getName()<<";\n";
} else {
for (int i = 0; i < (int) params.size(); i++) {
if (params[i].getNumComponents() == 1) {
load2j<<params[i].getType()<<" "<<params[i].getName()<<"2 = localData[atom2]."<<params[i].getName()<<";\n";
}
load2j<<");\n";
}
}
replacements["LOAD_ATOM2_PARAMETERS"] = load2j.str();
*/
stringstream load2j;
for (int i = 0; i < (int) params.size(); i++) {
/*
if (params[i].getNumComponents() == 1) {
load2j<<params[i].getType()<<" "<<params[i].getName()<<"2 = "<<params[i].getName()<<";\n";
}
else {
load2j<<params[i].getType()<<" "<<params[i].getName()<<"2 = make_"<<params[i].getType()<<"(";
for (int j = 0; j < params[i].getNumComponents(); ++j) {
if (j > 0)
load2j<<", ";
load2j<<params[i].getName()<<"_"<<suffixes[j];
else {
load2j<<params[i].getType()<<" "<<params[i].getName()<<"2 = make_"<<params[i].getType()<<"(";
for (int j = 0; j < params[i].getNumComponents(); ++j) {
if (j > 0)
load2j<<", ";
load2j<<"localData[atom2]."<<params[i].getName()<<"_"<<suffixes[j];
}
load2j<<");\n";
}
load2j<<");\n";
}*/
load2j<<params[i].getType()<<" "<<params[i].getName()<<"2 = temp"<<params[i].getName()<<";\n";
}
}
replacements["LOAD_ATOM2_PARAMETERS"] = load2j.str();
stringstream broadcastWarpData;
broadcastWarpData << "posq2.x = __shfl(tempPosq.x, j);\n";
broadcastWarpData << "posq2.y = __shfl(tempPosq.y, j);\n";
broadcastWarpData << "posq2.z = __shfl(tempPosq.z, j);\n";
broadcastWarpData << "posq2.w = __shfl(tempPosq.w, j);\n";
for(int i=0; i< (int) params.size();i++) {
broadcastWarpData << params[i].getType() << " temp" << params[i].getName() << ";\n";
for(int j=0; j < params[i].getNumComponents(); j++) {
string name;
if (params[i].getNumComponents() == 1) {
broadcastWarpData << "temp" << params[i].getName() << "=__shfl(" << params[i].getName() <<"1,j);\n";
stringstream shuffleWarpData;
if(useShuffle) {
shuffleWarpData << "shflPosq.x = real_shfl(shflPosq.x, tgx+1);\n";
shuffleWarpData << "shflPosq.y = real_shfl(shflPosq.y, tgx+1);\n";
shuffleWarpData << "shflPosq.z = real_shfl(shflPosq.z, tgx+1);\n";
shuffleWarpData << "shflPosq.w = real_shfl(shflPosq.w, tgx+1);\n";
shuffleWarpData << "shflForce.x = real_shfl(shflForce.x, tgx+1);\n";
shuffleWarpData << "shflForce.y = real_shfl(shflForce.y, tgx+1);\n";
shuffleWarpData << "shflForce.z = real_shfl(shflForce.z, tgx+1);\n";
for(int i=0; i < (int) params.size(); i++) {
if(params[i].getNumComponents() == 1) {
shuffleWarpData<<"shfl"<<params[i].getName()<<"=real_shfl(shfl"<<params[i].getName()<<", tgx+1);\n";
} else {
broadcastWarpData << "temp" << params[i].getName()+"."+suffixes[j] << "=__shfl(" << params[i].getName()+"1."+suffixes[j] <<",j);\n";
for(int j=0;j<params[i].getNumComponents();j++) {
// looks something like
// shflsigmaEpsilon.x = real_shfl(shflsigmaEpsilon.x,tgx+1);
shuffleWarpData<<"shfl"<<params[i].getName()
<<"."<<suffixes[j]<<"=real_shfl(shfl"
<<params[i].getName()<<"."<<suffixes[j]
<<", tgx+1);\n";
}
}
}
} else {
// not used otherwise
}
replacements["BROADCAST_WARP_DATA"] = broadcastWarpData.str();
stringstream shuffleWarpData;
shuffleWarpData << "tempPosq.x = __shfl(tempPosq.x, tgx+1);\n";
shuffleWarpData << "tempPosq.y = __shfl(tempPosq.y, tgx+1);\n";
shuffleWarpData << "tempPosq.z = __shfl(tempPosq.z, tgx+1);\n";
shuffleWarpData << "tempPosq.w = __shfl(tempPosq.w, tgx+1);\n";
shuffleWarpData << "tempForces.x = __shfl(tempForces.x, tgx+1);\n";
shuffleWarpData << "tempForces.y = __shfl(tempForces.y, tgx+1);\n";
shuffleWarpData << "tempForces.z = __shfl(tempForces.z, tgx+1);\n";
shuffleWarpData << "tempsigmaEpsilon.x = __shfl(tempsigmaEpsilon.x, tgx+1);\n";
shuffleWarpData << "tempsigmaEpsilon.y = __shfl(tempsigmaEpsilon.y, tgx+1);\n";
/*
for(int i=0; i< (int) params.size(); i++) {
shuffleWarpData << params[i].getName() << "=__shfl(" << params[i].getName() << ", tgx+1);\n";
}
*/
replacements["SHUFFLE_WARP_DATA"] = shuffleWarpData.str();
map<string, string> defines;
if (useCutoff)
defines["USE_CUTOFF"] = "1";
if (usePeriodic)
......
platforms/cuda/src/kernels/nonbonded.cu
View file @ ec39f6ff
#define WARPS_PER_GROUP (THREAD_BLOCK_SIZE/TILE_SIZE)
// structs are aligned to host compiler rules by default.
// large structures can spill into cache if using registers.
// this would defeat the purpose of using shuffles!
#ifndef ENABLE_SHUFFLE
typedef struct {
real x, y, z;
real q;
......@@ -12,6 +10,20 @@ typedef struct {
real padding;
#endif
} AtomData;
#endif
//support for 64 bit shuffles
static __inline__ __device__ float real_shfl(float var, int srcLane) {
return __shfl(var, srcLane);
}
static __inline__ __device__ double real_shfl(double var, int srcLane) {
int hi, lo;
asm volatile("mov.b64 { %0, %1 }, %2;" : "=r"(lo), "=r"(hi) : "d"(var));
hi = __shfl(hi, srcLane);
lo = __shfl(lo, srcLane);
return __hiloint2double( hi, lo );
}
/**
* Compute nonbonded interactions. The kernel is separated into two parts,
......@@ -19,10 +31,7 @@ typedef struct {
* implicit warp-level synchronization. A tile is defined by two atom blocks
* each of warpsize. Each warp computes a range of tiles.
*
* On-diagonal tiles processes interaction using a naive all-against-one interaction
* accumulation scheme.
*
* Off-diagonal tiles with exclusions compute the entire set of interactions across
* Tiles with exclusions compute the entire set of interactions across
* atom blocks, equal to warpsize*warpsize. In order to avoid access conflicts
* the forces are computed and accumulated diagonally in the manner shown below
* where, suppose
......@@ -46,12 +55,15 @@ typedef struct {
* t o 3 4 5 6 7 8 1 2
* a p 2 3 4 5 6 7 8 1
*
* TODO: Implement shuffle as opposed to using nonbonded.
*
* Tiles without exclusions read off directly from the neighbourlist interactingAtoms
* and follows the same force accumulation method above. If more there are more interactingTiles
* and follows the same force accumulation method. If more there are more interactingTiles
* than the size of the neighbourlist initially allocated, the neighbourlist is rebuilt
* and the full tileset.
* and the full tileset is computed. This should happen on the first step, and very rarely
* afterwards.
*
* On CUDA devices that support the shuffle intrinsic, on diagonal exclusion tiles use
* __shfl to broadcast. For all other types of tiles __shfl is used to pass around the
* forces, positions, and parameters when computing the forces.
*
* [out]forceBuffers - forces on each atom to eventually be accumulated
* [out]energyBuffer - energyBuffer to eventually be accumulated
......@@ -89,7 +101,11 @@ extern "C" __global__ void computeNonbonded(
const unsigned int tgx = threadIdx.x & (TILE_SIZE-1); // index within the warp
const unsigned int tbx = threadIdx.x - tgx; // block warpIndex
real energy = 0.0f;
// used shared memory if the device cannot shuffle
#ifndef ENABLE_SHUFFLE
__shared__ AtomData localData[THREAD_BLOCK_SIZE];
#endif
// First loop: process tiles that contain exclusions.
const unsigned int firstExclusionTile = FIRST_EXCLUSION_TILE+warp*(LAST_EXCLUSION_TILE-FIRST_EXCLUSION_TILE)/totalWarps;
const unsigned int lastExclusionTile = FIRST_EXCLUSION_TILE+(warp+1)*(LAST_EXCLUSION_TILE-FIRST_EXCLUSION_TILE)/totalWarps;
for (int pos = firstExclusionTile; pos < lastExclusionTile; pos++) {
......@@ -99,29 +115,33 @@ extern "C" __global__ void computeNonbonded(
real3 force = make_real3(0);
unsigned int atom1 = x*TILE_SIZE + tgx;
real4 posq1 = posq[atom1];
LOAD_ATOM1_PARAMETERS
#ifdef USE_EXCLUSIONS
tileflags excl = exclusions[pos*TILE_SIZE+tgx];
#endif
const bool hasExclusions = true;
if (x == y) {
// This tile is on the diagonal.
#ifdef ENABLE_SHUFFLE
real4 shflPosq = posq1;
#elif
localData[threadIdx.x].x = posq1.x;
localData[threadIdx.x].y = posq1.y;
localData[threadIdx.x].z = posq1.z;
localData[threadIdx.x].q = posq1.w;
LOAD_LOCAL_PARAMETERS_FROM_1
#endif
const unsigned int localAtomIndex = threadIdx.x;
real4 tempPosq = posq1;
// we do not need to fetch parameters from global since this is a symmetric tile
// instead we can broadcast the values using shuffle
// LOAD_LOCAL_PARAMETERS_FROM_1
for (unsigned int j = 0; j < TILE_SIZE; j++) {
int atom2 = tbx+j;
real4 posq2;
// load in the data from other registers
#ifdef ENABLE_SHUFFLE
BROADCAST_WARP_DATA
#elif
posq2 = make_real4(localData[atom2].x, localData[atom2].y, localData[atom2].z, localData[atom2].q);
#endif
real3 delta = make_real3(posq2.x-posq1.x, posq2.y-posq1.y, posq2.z-posq1.z);
#ifdef USE_PERIODIC
delta.x -= floor(delta.x*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
......@@ -159,16 +179,24 @@ extern "C" __global__ void computeNonbonded(
#endif
}
}
else { // This is an off-diagonal tile.
const unsigned int localAtomIndex = threadIdx.x;
else {
// This is an off-diagonal tile.
unsigned int j = y*TILE_SIZE + tgx;
real4 tempPosq = posq[j];
real3 tempForces;
tempForces.x = 0.0f;
tempForces.y = 0.0f;
tempForces.z = 0.0f;
real4 shflPosq = posq[j];
#ifdef ENABLE_SHUFFLE
real3 shflForce;
shflForce.x = 0.0f;
shflForce.y = 0.0f;
shflForce.z = 0.0f;
#elif
localData[threadIdx.x].x = shflPosq.x;
localData[threadIdx.x].y = shflPosq.y;
localData[threadIdx.x].z = shflPosq.z;
localData[threadIdx.x].q = shflPosq.w;
localData[threadIdx.x].fx = 0.0f;
localData[threadIdx.x].fy = 0.0f;
localData[threadIdx.x].fz = 0.0f;
#endif
DECLARE_LOCAL_PARAMETERS
LOAD_LOCAL_PARAMETERS_FROM_GLOBAL
#ifdef USE_EXCLUSIONS
......@@ -177,7 +205,11 @@ extern "C" __global__ void computeNonbonded(
unsigned int tj = tgx;
for (j = 0; j < TILE_SIZE; j++) {
int atom2 = tbx+tj;
real4 posq2 = tempPosq;
#ifdef ENABLE_SHUFFLE
real4 posq2 = shflPosq;
#elif
real4 posq2 = make_real4(localData[atom2].x, localData[atom2].y, localData[atom2].z, localData[atom2].q);
#endif
real3 delta = make_real3(posq2.x-posq1.x, posq2.y-posq1.y, posq2.z-posq1.z);
#ifdef USE_PERIODIC
delta.x -= floor(delta.x*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
......@@ -185,77 +217,88 @@ extern "C" __global__ void computeNonbonded(
delta.z -= floor(delta.z*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
#endif
real r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
#ifdef USE_CUTOFF
if (r2 < CUTOFF_SQUARED) {
#endif
real invR = RSQRT(r2);
real r = RECIP(invR);
LOAD_ATOM2_PARAMETERS
atom2 = y*TILE_SIZE+tj;
#ifdef USE_SYMMETRIC
real dEdR = 0.0f;
#else
real3 dEdR1 = make_real3(0);
real3 dEdR2 = make_real3(0);
#endif
#ifdef USE_EXCLUSIONS
bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS || !(excl & 0x1));
#endif
real tempEnergy = 0.0f;
COMPUTE_INTERACTION
energy += tempEnergy;
#ifdef USE_SYMMETRIC
delta *= dEdR;
force.x -= delta.x;
force.y -= delta.y;
force.z -= delta.z;
#ifdef ENABLE_SHUFFLE
shflForce.x += delta.x;
shflForce.y += delta.y;
shflForce.z += delta.z;
#ifdef USE_CUTOFF
if (r2 < CUTOFF_SQUARED) {
#endif
real invR = RSQRT(r2);
real r = RECIP(invR);
LOAD_ATOM2_PARAMETERS
atom2 = y*TILE_SIZE+tj;
#ifdef USE_SYMMETRIC
real dEdR = 0.0f;
#else
real3 dEdR1 = make_real3(0);
real3 dEdR2 = make_real3(0);
#endif
#ifdef USE_EXCLUSIONS
bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS || !(excl & 0x1));
#endif
real tempEnergy = 0.0f;
COMPUTE_INTERACTION
energy += tempEnergy;
#ifdef USE_SYMMETRIC
delta *= dEdR;
force.x -= delta.x;
force.y -= delta.y;
force.z -= delta.z;
tempForces.x += delta.x;
tempForces.y += delta.y;
tempForces.z += delta.z;
#else
force.x -= dEdR1.x;
force.y -= dEdR1.y;
force.z -= dEdR1.z;
tempForces.x += dEdR2.x;
tempForces.y += dEdR2.y;
tempForces.z += dEdR2.z;
#endif
#ifdef USE_CUTOFF
}
#endif
#elif
localData[tbx+tj].fx += delta.x;
localData[tbx+tj].fy += delta.y;
localData[tbx+tj].fz += delta.z;
#endif
#else // !USE_SYMMETRIC
force.x -= dEdR1.x;
force.y -= dEdR1.y;
force.z -= dEdR1.z;
#ifdef ENABLE_SHUFFLE
shflForce.x += dEdR2.x;
shflForce.y += dEdR2.y;
shflForce.z += dEdR2.z;
#elif
localData[tbx+tj].fx += dEdR2.x;
localData[tbx+tj].fy += dEdR2.y;
localData[tbx+tj].fz += dEdR2.z;
#endif
#endif // end USE_SYMMETRIC
#ifdef USE_CUTOFF
}
#endif
#ifdef USE_EXCLUSIONS
excl >>= 1;
#endif
#ifdef ENABLE_SHUFFLE
SHUFFLE_WARP_DATA
#endif
// cycles the indices
// 0 1 2 3 4 5 6 7 -> 1 2 3 4 5 6 7 0
SHUFFLE_WARP_DATA
tj = (tj + 1) & (TILE_SIZE - 1);
}
unsigned int offset = y*TILE_SIZE + tgx;
atomicAdd(&forceBuffers[offset], static_cast<unsigned long long>((long long) (tempForces.x*0x100000000)));
atomicAdd(&forceBuffers[offset+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (tempForces.y*0x100000000)));
atomicAdd(&forceBuffers[offset+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (tempForces.z*0x100000000)));
const unsigned int offset = y*TILE_SIZE + tgx;
// write results for off diagonal tiles
#ifdef ENABLE_SHUFFLE
atomicAdd(&forceBuffers[offset], static_cast<unsigned long long>((long long) (shflForce.x*0x100000000)));
atomicAdd(&forceBuffers[offset+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (shflForce.y*0x100000000)));
atomicAdd(&forceBuffers[offset+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (shflForce.z*0x100000000)));
#elif
atomicAdd(&forceBuffers[offset], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fx*0x100000000)));
atomicAdd(&forceBuffers[offset+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fy*0x100000000)));
atomicAdd(&forceBuffers[offset+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fz*0x100000000)));
#endif
}
unsigned int offset = x*TILE_SIZE + tgx;
// Write results for on and off diagonal tiles
const unsigned int offset = x*TILE_SIZE + tgx;
atomicAdd(&forceBuffers[offset], static_cast<unsigned long long>((long long) (force.x*0x100000000)));
atomicAdd(&forceBuffers[offset+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (force.y*0x100000000)));
atomicAdd(&forceBuffers[offset+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (force.z*0x100000000)));
//if (x != y) {
// offset = y*TILE_SIZE + tgx;
// atomicAdd(&forceBuffers[offset], static_cast<unsigned long long>((long long) (tempForces.x*0x100000000)));
// atomicAdd(&forceBuffers[offset+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (tempForces.y*0x100000000)));
// atomicAdd(&forceBuffers[offset+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (tempForces.z*0x100000000)));
//}
}
// Second loop: tiles without exclusions, either from the neighbor list (with cutoff) or just enumerating all
// of them (no cutoff).
#ifdef USE_CUTOFF
const unsigned int numTiles = interactionCount[0];
int pos = (numTiles > maxTiles ? startTileIndex+warp*numTileIndices/totalWarps : warp*numTiles/totalWarps);
......@@ -267,6 +310,8 @@ extern "C" __global__ void computeNonbonded(
#endif
int skipBase = 0;
int currentSkipIndex = tbx;
// atomIndices can probably be shuffled as well
// but it probably wouldn't make things any faster
__shared__ int atomIndices[THREAD_BLOCK_SIZE];
__shared__ volatile int skipTiles[THREAD_BLOCK_SIZE];
skipTiles[threadIdx.x] = -1;
......@@ -277,7 +322,6 @@ extern "C" __global__ void computeNonbonded(
bool includeTile = true;
// Extract the coordinates of this tile.
unsigned int x, y;
bool singlePeriodicCopy = false;
#ifdef USE_CUTOFF
......@@ -317,99 +361,118 @@ extern "C" __global__ void computeNonbonded(
}
if (includeTile) {
unsigned int atom1 = x*TILE_SIZE + tgx;
// Load atom data for this tile.
real4 posq1 = posq[atom1];
LOAD_ATOM1_PARAMETERS
const unsigned int localAtomIndex = threadIdx.x;
//const unsigned int localAtomIndex = threadIdx.x;
#ifdef USE_CUTOFF
unsigned int j = (numTiles <= maxTiles ? interactingAtoms[pos*TILE_SIZE+tgx] : y*TILE_SIZE + tgx);
#else
unsigned int j = y*TILE_SIZE + tgx;
#endif
atomIndices[threadIdx.x] = j;
real4 tempPosq;
real3 tempForces;
tempForces.x = 0.0f;
tempForces.y = 0.0f;
tempForces.z = 0.0f;
#ifdef ENABLE_SHUFFLE
DECLARE_LOCAL_PARAMETERS
real4 shflPosq;
real3 shflForce;
shflForce.x = 0.0f;
shflForce.y = 0.0f;
shflForce.z = 0.0f;
#endif
if (j < PADDED_NUM_ATOMS) {
// Load position of atom j from from global memory
tempPosq = posq[j];
//localData[localAtomIndex].x = tempPosq.x;
//localData[localAtomIndex].y = tempPosq.y;
//localData[localAtomIndex].z = tempPosq.z;
//localData[localAtomIndex].q = tempPosq.w;
#ifdef ENABLE_SHUFFLE
shflPosq = posq[j];
#elif
localData[threadIdx.x].x = posq[j].x;
localData[threadIdx.x].y = posq[j].y;
localData[threadIdx.x].z = posq[j].z;
localData[threadIdx.x].q = posq[j].w;
localData[threadIdx.x].fx = 0.0f;
localData[threadIdx.x].fy = 0.0f;
localData[threadIdx.x].fz = 0.0f;
#endif
LOAD_LOCAL_PARAMETERS_FROM_GLOBAL
//localData[localAtomIndex].fx = 0.0f;
//localData[localAtomIndex].fy = 0.0f;
//localData[localAtomIndex].fz = 0.0f;
}
#ifdef USE_PERIODIC
if (singlePeriodicCopy) {
// The box is small enough that we can just translate all the atoms into a single periodic
// box, then skip having to apply periodic boundary conditions later.
real4 blockCenterX = blockCenter[x];
posq1.x -= floor((posq1.x-blockCenterX.x)*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
posq1.y -= floor((posq1.y-blockCenterX.y)*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
posq1.z -= floor((posq1.z-blockCenterX.z)*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
//localData[localAtomIndex].x -= floor((localData[localAtomIndex].x-blockCenterX.x)*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
//localData[localAtomIndex].y -= floor((localData[localAtomIndex].y-blockCenterX.y)*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
//localData[localAtomIndex].z -= floor((localData[localAtomIndex].z-blockCenterX.z)*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
tempPosq.x -= floor((tempPosq.x-blockCenterX.x)*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
tempPosq.y -= floor((tempPosq.y-blockCenterX.y)*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
tempPosq.z -= floor((tempPosq.z-blockCenterX.z)*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
#ifdef ENABLE_SHUFFLE
shflPosq.x -= floor((shflPosq.x-blockCenterX.x)*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
shflPosq.y -= floor((shflPosq.y-blockCenterX.y)*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
shflPosq.z -= floor((shflPosq.z-blockCenterX.z)*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
#elif
localData[threadIdx.x].x -= floor((localData[threadIdx.x].x-blockCenterX.x)*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
localData[threadIdx.x].y -= floor((localData[threadIdx.x].y-blockCenterX.y)*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
localData[threadIdx.x].z -= floor((localData[threadIdx.x].z-blockCenterX.z)*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
#endif
unsigned int tj = tgx;
for (j = 0; j < TILE_SIZE; j++) {
int atom2 = tbx+tj;
real4 posq2 = tempPosq;
#ifdef ENABLE_SHUFFLE
real4 posq2 = shflPosq;
#elif
real4 posq2 = make_real4(localData[atom2].x, localData[atom2].y, localData[atom2].z, localData[atom2].q);
#endif
real3 delta = make_real3(posq2.x-posq1.x, posq2.y-posq1.y, posq2.z-posq1.z);
real r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
if (r2 < CUTOFF_SQUARED) {
real invR = RSQRT(r2);
real r = RECIP(invR);
LOAD_ATOM2_PARAMETERS
atom2 = atomIndices[tbx+tj];
#ifdef USE_SYMMETRIC
real dEdR = 0.0f;
#else
real3 dEdR1 = make_real3(0);
real3 dEdR2 = make_real3(0);
#endif
#ifdef USE_EXCLUSIONS
bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS);
#endif
real tempEnergy = 0.0f;
COMPUTE_INTERACTION
energy += tempEnergy;
#ifdef USE_SYMMETRIC
delta *= dEdR;
force.x -= delta.x;
force.y -= delta.y;
force.z -= delta.z;
tempForces.x += delta.x;
tempForces.y += delta.y;
tempForces.z += delta.z;
#else
force.x -= dEdR1.x;
force.y -= dEdR1.y;
force.z -= dEdR1.z;
tempForces.x += dEdR2.x;
tempForces.y += dEdR2.y;
tempForces.z += dEdR2.z;
#endif
if (r2 < CUTOFF_SQUARED) {
real invR = RSQRT(r2);
real r = RECIP(invR);
LOAD_ATOM2_PARAMETERS
atom2 = atomIndices[tbx+tj];
#ifdef USE_SYMMETRIC
real dEdR = 0.0f;
#else
real3 dEdR1 = make_real3(0);
real3 dEdR2 = make_real3(0);
#endif
#ifdef USE_EXCLUSIONS
bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS);
#endif
real tempEnergy = 0.0f;
COMPUTE_INTERACTION
energy += tempEnergy;
#ifdef USE_SYMMETRIC
delta *= dEdR;
force.x -= delta.x;
force.y -= delta.y;
force.z -= delta.z;
#ifdef ENABLE_SHUFFLE
shflForce.x += delta.x;
shflForce.y += delta.y;
shflForce.z += delta.z;
#elif
localData[tbx+tj].fx += delta.x;
localData[tbx+tj].fy += delta.y;
localData[tbx+tj].fz += delta.z;
#endif
#else // !USE_SYMMETRIC
force.x -= dEdR1.x;
force.y -= dEdR1.y;
force.z -= dEdR1.z;
#ifdef ENABLE_SHUFFLE
shflForce.x += dEdR2.x;
shflForce.y += dEdR2.y;
shflForce.z += dEdR2.z;
#elif
localData[tbx+tj].fx += dEdR2.x;
localData[tbx+tj].fy += dEdR2.y;
localData[tbx+tj].fz += dEdR2.z;
#endif
#endif // end USE_SYMMETRIC
}
#ifdef ENABLE_SHUFFLE
SHUFFLE_WARP_DATA
#endif
tj = (tj + 1) & (TILE_SIZE - 1);
}
}
else
#endif
......@@ -418,7 +481,11 @@ extern "C" __global__ void computeNonbonded(
unsigned int tj = tgx;
for (j = 0; j < TILE_SIZE; j++) {
int atom2 = tbx+tj;
real4 posq2 = tempPosq;
#ifdef ENABLE_SHUFFLE
real4 posq2 = shflPosq;
#elif
real4 posq2 = make_real4(localData[atom2].x, localData[atom2].y, localData[atom2].z, localData[atom2].q);
#endif
real3 delta = make_real3(posq2.x-posq1.x, posq2.y-posq1.y, posq2.z-posq1.z);
#ifdef USE_PERIODIC
delta.x -= floor(delta.x*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
......@@ -426,52 +493,65 @@ extern "C" __global__ void computeNonbonded(
delta.z -= floor(delta.z*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
#endif
real r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
#ifdef USE_CUTOFF
if (r2 < CUTOFF_SQUARED) {
#endif
real invR = RSQRT(r2);
real r = RECIP(invR);
LOAD_ATOM2_PARAMETERS
atom2 = atomIndices[tbx+tj];
#ifdef USE_SYMMETRIC
real dEdR = 0.0f;
#else
real3 dEdR1 = make_real3(0);
real3 dEdR2 = make_real3(0);
#endif
#ifdef USE_EXCLUSIONS
bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS);
#endif
real tempEnergy = 0.0f;
COMPUTE_INTERACTION
energy += tempEnergy;
#ifdef USE_SYMMETRIC
delta *= dEdR;
force.x -= delta.x;
force.y -= delta.y;
force.z -= delta.z;
#ifdef ENABLE_SHUFFLE
shflForce.x += delta.x;
shflForce.y += delta.y;
shflForce.z += delta.z;
#ifdef USE_CUTOFF
if (r2 < CUTOFF_SQUARED) {
#endif
real invR = RSQRT(r2);
real r = RECIP(invR);
LOAD_ATOM2_PARAMETERS
atom2 = atomIndices[tbx+tj];
#ifdef USE_SYMMETRIC
real dEdR = 0.0f;
#else
real3 dEdR1 = make_real3(0);
real3 dEdR2 = make_real3(0);
#endif
#ifdef USE_EXCLUSIONS
bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS);
#endif
real tempEnergy = 0.0f;
COMPUTE_INTERACTION
energy += tempEnergy;
#ifdef USE_SYMMETRIC
delta *= dEdR;
force.x -= delta.x;
force.y -= delta.y;
force.z -= delta.z;
tempForces.x += delta.x;
tempForces.y += delta.y;
tempForces.z += delta.z;
#else
force.x -= dEdR1.x;
force.y -= dEdR1.y;
force.z -= dEdR1.z;
tempForces.x += dEdR2.x;
tempForces.y += dEdR2.y;
tempForces.z += dEdR2.z;
#endif
#ifdef USE_CUTOFF
}
#elif
localData[tbx+tj].fx += delta.x;
localData[tbx+tj].fy += delta.y;
localData[tbx+tj].fz += delta.z;
#endif
#else // !USE_SYMMETRIC
force.x -= dEdR1.x;
force.y -= dEdR1.y;
force.z -= dEdR1.z;
#ifdef ENABLE_SHUFFLE
shflForce.x += dEdR2.x;
shflForce.y += dEdR2.y;
shflForce.z += dEdR2.z;
#elif
localData[tbx+tj].fx += dEdR2.x;
localData[tbx+tj].fy += dEdR2.y;
localData[tbx+tj].fz += dEdR2.z;
#endif
#endif // end USE_SYMMETRIC
#ifdef USE_CUTOFF
}
#endif
#ifdef ENABLE_SHUFFLE
SHUFFLE_WARP_DATA
#endif
tj = (tj + 1) & (TILE_SIZE - 1);
}
}
// Write results.
atomicAdd(&forceBuffers[atom1], static_cast<unsigned long long>((long long) (force.x*0x100000000)));
atomicAdd(&forceBuffers[atom1+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (force.y*0x100000000)));
atomicAdd(&forceBuffers[atom1+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (force.z*0x100000000)));
......@@ -481,13 +561,18 @@ extern "C" __global__ void computeNonbonded(
unsigned int atom2 = y*TILE_SIZE + tgx;
#endif
if (atom2 < PADDED_NUM_ATOMS) {
atomicAdd(&forceBuffers[atom2], static_cast<unsigned long long>((long long) (tempForces.x*0x100000000)));
atomicAdd(&forceBuffers[atom2+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (tempForces.y*0x100000000)));
atomicAdd(&forceBuffers[atom2+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (tempForces.z*0x100000000)));
#ifdef ENABLE_SHUFFLE
atomicAdd(&forceBuffers[atom2], static_cast<unsigned long long>((long long) (shflForce.x*0x100000000)));
atomicAdd(&forceBuffers[atom2+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (shflForce.y*0x100000000)));
atomicAdd(&forceBuffers[atom2+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (shflForce.z*0x100000000)));
#elif
atomicAdd(&forceBuffers[atom2], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fx*0x100000000)));
atomicAdd(&forceBuffers[atom2+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fy*0x100000000)));
atomicAdd(&forceBuffers[atom2+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fz*0x100000000)));
#endif
}
}
pos++;
}
energyBuffer[blockIdx.x*blockDim.x+threadIdx.x] += energy;
}
}
\ No newline at end of file
platforms/cuda/tests/TestCudaNonbondedForce.cpp
View file @ ec39f6ff
......@@ -872,21 +872,21 @@ int main(int argc, char* argv[]) {
try {
if (argc > 1)
platform.setPropertyDefaultValue("CudaPrecision", string(argv[1]));
//testCoulomb();
//testLJ();
//testExclusionsAnd14();
//testCutoff();
//testCutoff14();
//testPeriodic();
testCoulomb();
testLJ();
testExclusionsAnd14();
testCutoff();
testCutoff14();
testPeriodic();
testLargeSystem();
//testBlockInteractions(false);
//testBlockInteractions(true);
//testDispersionCorrection();
//testChangingParameters();
//testParallelComputation(false);
//testParallelComputation(true);
//testSwitchingFunction(NonbondedForce::CutoffNonPeriodic);
//testSwitchingFunction(NonbondedForce::PME);
testDispersionCorrection();
testChangingParameters();
testParallelComputation(false);
testParallelComputation(true);
testSwitchingFunction(NonbondedForce::CutoffNonPeriodic);
testSwitchingFunction(NonbondedForce::PME);
}
catch(const exception& e) {
cout << "exception: " << e.what() << endl;
......
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