Skip to content

GitLab

Commit 93c467b2 authored by Peter Eastman's avatar Peter Eastman
Browse files

Merged 5.1Optimizations branch back to trunk

parent f6d4557d
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1083 additions and 892 deletions
platforms/cuda/src/kernels/langevin.cu
View file @ 93c467b2
......@@ -95,8 +95,8 @@ extern "C" __global__ void selectLangevinStepSize(mixed maxStepSize, mixed error
if (blockIdx.x*blockDim.x+threadIdx.x == 0) {
// Select the new step size.
mixed totalError = sqrt(error[0]/(NUM_ATOMS*3));
mixed newStepSize = sqrt(errorTol/totalError);
mixed totalError = SQRT(error[0]/(NUM_ATOMS*3));
mixed newStepSize = SQRT(errorTol/totalError);
mixed oldStepSize = dt[0].y;
if (oldStepSize > 0.0f)
newStepSize = min(newStepSize, oldStepSize*2.0f); // For safety, limit how quickly dt can increase.
......@@ -108,9 +108,9 @@ extern "C" __global__ void selectLangevinStepSize(mixed maxStepSize, mixed error
// Recalculate the integration parameters.
mixed vscale = exp(-newStepSize/tau);
mixed vscale = EXP(-newStepSize/tau);
mixed fscale = (1-vscale)*tau;
mixed noisescale = sqrt(2*kT/tau)*sqrt(0.5f*(1-vscale*vscale)*tau);
mixed noisescale = SQRT(2*kT/tau)*SQRT(0.5f*(1-vscale*vscale)*tau);
params[VelScale] = vscale;
params[ForceScale] = fscale;
params[NoiseScale] = noisescale;
......
platforms/cuda/src/kernels/nonbonded.cu
View file @ 93c467b2
This diff is collapsed.
platforms/cuda/src/kernels/pme.cu
View file @ 93c467b2
extern "C" __global__ void updateBsplines(const real4* __restrict__ posq, real4* __restrict__ pmeBsplineTheta, int2* __restrict__ pmeAtomGridIndex,
extern "C" __global__ void findAtomGridIndex(const real4* __restrict__ posq, int2* __restrict__ pmeAtomGridIndex,
real4 periodicBoxSize, real4 invPeriodicBoxSize) {
extern __shared__ real3 bsplinesCache[];
real3* data = &bsplinesCache[threadIdx.x*PME_ORDER];
const real3 scale = make_real3(RECIP(PME_ORDER-1));
// Compute the index of the grid point each atom is associated with.
for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < NUM_ATOMS; i += blockDim.x*gridDim.x) {
real4 pos = posq[i];
pos.x -= floor(pos.x*invPeriodicBoxSize.x)*periodicBoxSize.x;
......@@ -11,11 +10,40 @@ extern "C" __global__ void updateBsplines(const real4* __restrict__ posq, real4*
real3 t = make_real3((pos.x*invPeriodicBoxSize.x)*GRID_SIZE_X,
(pos.y*invPeriodicBoxSize.y)*GRID_SIZE_Y,
(pos.z*invPeriodicBoxSize.z)*GRID_SIZE_Z);
real3 dr = make_real3(t.x-(int) t.x, t.y-(int) t.y, t.z-(int) t.z);
int3 gridIndex = make_int3(((int) t.x) % GRID_SIZE_X,
((int) t.y) % GRID_SIZE_Y,
((int) t.z) % GRID_SIZE_Z);
pmeAtomGridIndex[i] = make_int2(i, gridIndex.x*GRID_SIZE_Y*GRID_SIZE_Z+gridIndex.y*GRID_SIZE_Z+gridIndex.z);
}
}
extern "C" __global__ void gridSpreadCharge(const real4* __restrict__ posq, real* __restrict__ originalPmeGrid,
real4 periodicBoxSize, real4 invPeriodicBoxSize, const int2* __restrict__ pmeAtomGridIndex) {
real3 data[PME_ORDER];
const real scale = RECIP(PME_ORDER-1);
// Process the atoms in spatially sorted order. This improves efficiency when writing
// the grid values.
for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < NUM_ATOMS; i += blockDim.x*gridDim.x) {
int atom = pmeAtomGridIndex[i].x;
real charge = posq[atom].w;
real3 force = make_real3(0);
real4 pos = posq[atom];
pos.x -= floor(pos.x*invPeriodicBoxSize.x)*periodicBoxSize.x;
pos.y -= floor(pos.y*invPeriodicBoxSize.y)*periodicBoxSize.y;
pos.z -= floor(pos.z*invPeriodicBoxSize.z)*periodicBoxSize.z;
real3 t = make_real3((pos.x*invPeriodicBoxSize.x)*GRID_SIZE_X,
(pos.y*invPeriodicBoxSize.y)*GRID_SIZE_Y,
(pos.z*invPeriodicBoxSize.z)*GRID_SIZE_Z);
int3 gridIndex = make_int3(((int) t.x) % GRID_SIZE_X,
((int) t.y) % GRID_SIZE_Y,
((int) t.z) % GRID_SIZE_Z);
// Since we need the full set of thetas, it's faster to compute them here than load them
// from global memory.
real3 dr = make_real3(t.x-(int) t.x, t.y-(int) t.y, t.z-(int) t.z);
data[PME_ORDER-1] = make_real3(0);
data[1] = dr;
data[0] = make_real3(1)-dr;
......@@ -23,98 +51,46 @@ extern "C" __global__ void updateBsplines(const real4* __restrict__ posq, real4*
real div = RECIP(j-1);
data[j-1] = div*dr*data[j-2];
for (int k = 1; k < (j-1); k++)
data[j-k-1] = div*((dr+make_real3(k)) *data[j-k-2] + (make_real3(j-k)-dr)*data[j-k-1]);
data[j-k-1] = div*((dr+make_real3(k))*data[j-k-2] + (make_real3(j-k)-dr)*data[j-k-1]);
data[0] = div*(make_real3(1)-dr)*data[0];
}
data[PME_ORDER-1] = scale*dr*data[PME_ORDER-2];
for (int j = 1; j < (PME_ORDER-1); j++)
data[PME_ORDER-j-1] = scale*((dr+make_real3(j))*data[PME_ORDER-j-2] + (make_real3(PME_ORDER-j)-dr)*data[PME_ORDER-j-1]);
data[0] = scale*(make_real3(1)-dr)*data[0];
for (int j = 0; j < PME_ORDER; j++) {
real3 d = data[j]; // Copy it as a workaround for a bug in CUDA 5.0
pmeBsplineTheta[i+j*NUM_ATOMS] = make_real4(d.x, d.y, d.z, pos.w); // Storing the charge here improves cache coherency in the charge spreading kernel
}
}
}
/**
* For each grid point, find the range of sorted atoms associated with that point.
*/
extern "C" __global__ void findAtomRangeForGrid(int2* __restrict__ pmeAtomGridIndex, int* __restrict__ pmeAtomRange, const real4* __restrict__ posq, real4 periodicBoxSize, real4 invPeriodicBoxSize) {
int start = (NUM_ATOMS*(blockIdx.x*blockDim.x+threadIdx.x))/(blockDim.x*gridDim.x);
int end = (NUM_ATOMS*(blockIdx.x*blockDim.x+threadIdx.x+1))/(blockDim.x*gridDim.x);
int last = (start == 0 ? -1 : pmeAtomGridIndex[start-1].y);
for (int i = start; i < end; ++i) {
int2 atomData = pmeAtomGridIndex[i];
int gridIndex = atomData.y;
if (gridIndex != last) {
for (int j = last+1; j <= gridIndex; ++j)
pmeAtomRange[j] = i;
last = gridIndex;
}
}
// Fill in values beyond the last atom.
if (blockIdx.x == gridDim.x-1 && threadIdx.x == blockDim.x-1) {
int gridSize = GRID_SIZE_X*GRID_SIZE_Y*GRID_SIZE_Z;
for (int j = last+1; j <= gridSize; ++j)
pmeAtomRange[j] = NUM_ATOMS;
}
}
#define BUFFER_SIZE (PME_ORDER*PME_ORDER*PME_ORDER)
extern "C" __global__ void gridSpreadCharge(const real4* __restrict__ posq, real* __restrict__ originalPmeGrid,
const real4* __restrict__ pmeBsplineTheta, real4 periodicBoxSize, real4 invPeriodicBoxSize) {
int ix = threadIdx.x/(PME_ORDER*PME_ORDER);
int remainder = threadIdx.x-ix*PME_ORDER*PME_ORDER;
int iy = remainder/PME_ORDER;
int iz = remainder-iy*PME_ORDER;
__shared__ real4 theta[PME_ORDER];
__shared__ real charge[BUFFER_SIZE];
__shared__ int basex[BUFFER_SIZE];
__shared__ int basey[BUFFER_SIZE];
__shared__ int basez[BUFFER_SIZE];
if (ix < PME_ORDER) {
for (int baseIndex = blockIdx.x*BUFFER_SIZE; baseIndex < NUM_ATOMS; baseIndex += gridDim.x*BUFFER_SIZE) {
// Load the next block of atoms into the buffers.
// Spread the charge from this atom onto each grid point.
for (int ix = 0; ix < PME_ORDER; ix++) {
int xbase = gridIndex.x+ix;
xbase -= (xbase >= GRID_SIZE_X ? GRID_SIZE_X : 0);
xbase = xbase*GRID_SIZE_Y*GRID_SIZE_Z;
real dx = data[ix].x;
for (int iy = 0; iy < PME_ORDER; iy++) {
int ybase = gridIndex.y+iy;
ybase -= (ybase >= GRID_SIZE_Y ? GRID_SIZE_Y : 0);
ybase = xbase + ybase*GRID_SIZE_Z;
real dy = data[iy].y;
for (int iz = 0; iz < PME_ORDER; iz++) {
int zindex = gridIndex.z+iz;
zindex -= (zindex >= GRID_SIZE_Z ? GRID_SIZE_Z : 0);
int index = ybase + zindex;
int atomIndex = baseIndex+threadIdx.x;
if (atomIndex < NUM_ATOMS) {
real4 pos = posq[atomIndex];
charge[threadIdx.x] = pos.w;
pos.x -= floor(pos.x*invPeriodicBoxSize.x)*periodicBoxSize.x;
pos.y -= floor(pos.y*invPeriodicBoxSize.y)*periodicBoxSize.y;
pos.z -= floor(pos.z*invPeriodicBoxSize.z)*periodicBoxSize.z;
basex[threadIdx.x] = (int) ((pos.x*invPeriodicBoxSize.x)*GRID_SIZE_X);
basey[threadIdx.x] = (int) ((pos.y*invPeriodicBoxSize.y)*GRID_SIZE_Y);
basez[threadIdx.x] = (int) ((pos.z*invPeriodicBoxSize.z)*GRID_SIZE_Z);
}
__syncthreads();
int lastIndex = min(BUFFER_SIZE, NUM_ATOMS-baseIndex);
for (int index = 0; index < lastIndex; index++) {
int atomIndex = index+baseIndex;
if (threadIdx.x < PME_ORDER)
theta[threadIdx.x] = pmeBsplineTheta[atomIndex+threadIdx.x*NUM_ATOMS];
__syncthreads();
real add = charge[index]*theta[ix].x*theta[iy].y*theta[iz].z;
int x = basex[index]+ix;
int y = basey[index]+iy;
int z = basez[index]+iz;
x -= (x >= GRID_SIZE_X ? GRID_SIZE_X : 0);
y -= (y >= GRID_SIZE_Y ? GRID_SIZE_Y : 0);
z -= (z >= GRID_SIZE_Z ? GRID_SIZE_Z : 0);
real add = charge*dx*dy*data[iz].z;
#ifdef USE_DOUBLE_PRECISION
unsigned long long * ulonglong_p = (unsigned long long *) originalPmeGrid;
atomicAdd(&ulonglong_p[x*GRID_SIZE_Y*GRID_SIZE_Z+y*GRID_SIZE_Z+z], static_cast<unsigned long long>((long long) (add*0x100000000)));
unsigned long long * ulonglong_p = (unsigned long long *) originalPmeGrid;
atomicAdd(&ulonglong_p[index], static_cast<unsigned long long>((long long) (add*0x100000000)));
#elif __CUDA_ARCH__ < 200
unsigned long long * ulonglong_p = (unsigned long long *) originalPmeGrid;
int gridIndex = x*GRID_SIZE_Y*GRID_SIZE_Z+y*GRID_SIZE_Z+z;
gridIndex = (gridIndex%2 == 0 ? gridIndex/2 : (gridIndex+GRID_SIZE_X*GRID_SIZE_Y*GRID_SIZE_Z)/2);
atomicAdd(&ulonglong_p[gridIndex], static_cast<unsigned long long>((long long) (add*0x100000000)));
unsigned long long * ulonglong_p = (unsigned long long *) originalPmeGrid;
int gridIndex = index;
gridIndex = (gridIndex%2 == 0 ? gridIndex/2 : (gridIndex+GRID_SIZE_X*GRID_SIZE_Y*GRID_SIZE_Z)/2);
atomicAdd(&ulonglong_p[gridIndex], static_cast<unsigned long long>((long long) (add*0x100000000)));
#else
atomicAdd(&originalPmeGrid[x*GRID_SIZE_Y*GRID_SIZE_Z+y*GRID_SIZE_Z+z], add*EPSILON_FACTOR);
atomicAdd(&originalPmeGrid[index], add*EPSILON_FACTOR);
#endif
}
}
}
}
......@@ -182,48 +158,52 @@ gridEvaluateEnergy(real2* __restrict__ halfcomplex_pmeGrid, real* __restrict__ e
const unsigned int gridSize = GRID_SIZE_X*GRID_SIZE_Y*GRID_SIZE_Z;
const real recipScaleFactor = RECIP(M_PI*periodicBoxSize.x*periodicBoxSize.y*periodicBoxSize.z);
real energy = 0;
real energy = 0;
for (int index = blockIdx.x*blockDim.x+threadIdx.x; index < gridSize; index += blockDim.x*gridDim.x) {
// real indices
int kx = index/(GRID_SIZE_Y*(GRID_SIZE_Z));
int remainder = index-kx*GRID_SIZE_Y*(GRID_SIZE_Z);
int ky = remainder/(GRID_SIZE_Z);
int kz = remainder-ky*(GRID_SIZE_Z);
int mx = (kx < (GRID_SIZE_X+1)/2) ? kx : (kx-GRID_SIZE_X);
int my = (ky < (GRID_SIZE_Y+1)/2) ? ky : (ky-GRID_SIZE_Y);
int mz = (kz < (GRID_SIZE_Z+1)/2) ? kz : (kz-GRID_SIZE_Z);
real mhx = mx*invPeriodicBoxSize.x;
real mhy = my*invPeriodicBoxSize.y;
real mhz = mz*invPeriodicBoxSize.z;
real m2 = mhx*mhx+mhy*mhy+mhz*mhz;
real bx = pmeBsplineModuliX[kx];
real by = pmeBsplineModuliY[ky];
real bz = pmeBsplineModuliZ[kz];
real denom = m2*bx*by*bz;
real eterm = recipScaleFactor*EXP(-RECIP_EXP_FACTOR*m2)/denom;
int mx = (kx < (GRID_SIZE_X+1)/2) ? kx : (kx-GRID_SIZE_X);
int my = (ky < (GRID_SIZE_Y+1)/2) ? ky : (ky-GRID_SIZE_Y);
int mz = (kz < (GRID_SIZE_Z+1)/2) ? kz : (kz-GRID_SIZE_Z);
real mhx = mx*invPeriodicBoxSize.x;
real mhy = my*invPeriodicBoxSize.y;
real mhz = mz*invPeriodicBoxSize.z;
real m2 = mhx*mhx+mhy*mhy+mhz*mhz;
real bx = pmeBsplineModuliX[kx];
real by = pmeBsplineModuliY[ky];
real bz = pmeBsplineModuliZ[kz];
real denom = m2*bx*by*bz;
real eterm = recipScaleFactor*EXP(-RECIP_EXP_FACTOR*m2)/denom;
if(kz >= (GRID_SIZE_Z/2+1)) {
kx = ((kx == 0) ? kx : GRID_SIZE_X-kx);
ky = ((ky == 0) ? ky : GRID_SIZE_Y-ky);
kz = GRID_SIZE_Z-kz;
}
int indexInHalfComplexGrid = kz + ky*(GRID_SIZE_Z/2+1)+kx*(GRID_SIZE_Y*(GRID_SIZE_Z/2+1));
real2 grid = halfcomplex_pmeGrid[indexInHalfComplexGrid];
if (kx != 0 || ky != 0 || kz != 0) {
energy += eterm*(grid.x*grid.x + grid.y*grid.y);
}
if(kz >= (GRID_SIZE_Z/2+1)) {
kx = ((kx == 0) ? kx : GRID_SIZE_X-kx);
ky = ((ky == 0) ? ky : GRID_SIZE_Y-ky);
kz = GRID_SIZE_Z-kz;
}
int indexInHalfComplexGrid = kz + ky*(GRID_SIZE_Z/2+1)+kx*(GRID_SIZE_Y*(GRID_SIZE_Z/2+1));
real2 grid = halfcomplex_pmeGrid[indexInHalfComplexGrid];
if (kx != 0 || ky != 0 || kz != 0) {
energy += eterm*(grid.x*grid.x + grid.y*grid.y);
}
}
energyBuffer[blockIdx.x*blockDim.x+threadIdx.x] += 0.5f*energy;
}
extern "C" __global__
void gridInterpolateForce(const real4* __restrict__ posq, unsigned long long* __restrict__ forceBuffers, const real* __restrict__ originalPmeGrid,
real4 periodicBoxSize, real4 invPeriodicBoxSize) {
real4 periodicBoxSize, real4 invPeriodicBoxSize, const int2* __restrict__ pmeAtomGridIndex) {
real3 data[PME_ORDER];
real3 ddata[PME_ORDER];
const real scale = RECIP(PME_ORDER-1);
for (int atom = blockIdx.x*blockDim.x+threadIdx.x; atom < NUM_ATOMS; atom += blockDim.x*gridDim.x) {
// Process the atoms in spatially sorted order. This improves cache performance when loading
// the grid values.
for (int i = blockIdx.x*blockDim.x+threadIdx.x; i < NUM_ATOMS; i += blockDim.x*gridDim.x) {
int atom = pmeAtomGridIndex[i].x;
real3 force = make_real3(0);
real4 pos = posq[atom];
pos.x -= floor(pos.x*invPeriodicBoxSize.x)*periodicBoxSize.x;
......@@ -233,8 +213,8 @@ void gridInterpolateForce(const real4* __restrict__ posq, unsigned long long* __
(pos.y*invPeriodicBoxSize.y)*GRID_SIZE_Y,
(pos.z*invPeriodicBoxSize.z)*GRID_SIZE_Z);
int3 gridIndex = make_int3(((int) t.x) % GRID_SIZE_X,
((int) t.y) % GRID_SIZE_Y,
((int) t.z) % GRID_SIZE_Z);
((int) t.y) % GRID_SIZE_Y,
((int) t.z) % GRID_SIZE_Z);
// Since we need the full set of thetas, it's faster to compute them here than load them
// from global memory.
......@@ -243,7 +223,6 @@ void gridInterpolateForce(const real4* __restrict__ posq, unsigned long long* __
data[PME_ORDER-1] = make_real3(0);
data[1] = dr;
data[0] = make_real3(1)-dr;
for (int j = 3; j < PME_ORDER; j++) {
real div = RECIP(j-1);
data[j-1] = div*dr*data[j-2];
......@@ -252,15 +231,13 @@ void gridInterpolateForce(const real4* __restrict__ posq, unsigned long long* __
data[0] = div*(make_real3(1)-dr)*data[0];
}
ddata[0] = -data[0];
for (int j = 1; j < PME_ORDER; j++)
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] = scale*((dr+make_real3(j))*data[PME_ORDER-j-2] + (make_real3(PME_ORDER-j)-dr)*data[PME_ORDER-j-1]);
data[0] = scale*(make_real3(1)-dr)*data[0];
// Compute the force on this atom.
for (int ix = 0; ix < PME_ORDER; ix++) {
......
platforms/cuda/src/kernels/sort.cu
View file @ 93c467b2
......@@ -3,7 +3,49 @@ __device__ KEY_TYPE getValue(DATA_TYPE value) {
}
extern "C" {
/**
* Sort a list that is short enough to entirely fit in local memory. This is executed as
* a single thread block.
*/
__global__ void sortShortList(DATA_TYPE* __restrict__ data, unsigned int length) {
// Load the data into local memory.
extern __shared__ DATA_TYPE dataBuffer[];
for (int index = threadIdx.x; index < length; index += blockDim.x)
dataBuffer[index] = data[index];
__syncthreads();
// Perform a bitonic sort in local memory.
for (unsigned int k = 2; k < 2*length; k *= 2) {
for (unsigned int j = k/2; j > 0; j /= 2) {
for (unsigned int i = threadIdx.x; i < length; i += blockDim.x) {
int ixj = i^j;
if (ixj > i && ixj < length) {
DATA_TYPE value1 = dataBuffer[i];
DATA_TYPE value2 = dataBuffer[ixj];
bool ascending = ((i&k) == 0);
for (unsigned int mask = k*2; mask < 2*length; mask *= 2)
ascending = ((i&mask) == 0 ? !ascending : ascending);
KEY_TYPE lowKey = (ascending ? getValue(value1) : getValue(value2));
KEY_TYPE highKey = (ascending ? getValue(value2) : getValue(value1));
if (lowKey > highKey) {
dataBuffer[i] = value2;
dataBuffer[ixj] = value1;
}
}
}
__syncthreads();
}
}
// Write the data back to global memory.
for (int index = threadIdx.x; index < length; index += blockDim.x)
data[index] = dataBuffer[index];
}
/**
* Calculate the minimum and maximum value in the array to be sorted. This kernel
* is executed as a single work group.
......
platforms/cuda/src/kernels/torsionForce.cu
View file @ 93c467b2
......@@ -16,12 +16,12 @@ if (cosangle > 0.99f || cosangle < -0.99f) {
theta = PI-theta;
}
else
theta = acos(cosangle);
theta = ACOS(cosangle);
theta = (dot(v0, cp1) >= 0 ? theta : -theta);
COMPUTE_FORCE
real normCross1 = dot(cp0, cp0);
real normSqrBC = dot(v1, v1);
real normBC = sqrt(normSqrBC);
real normBC = SQRT(normSqrBC);
real normCross2 = dot(cp1, cp1);
real dp = RECIP(normSqrBC);
real4 ff = make_real4((-dEdAngle*normBC)/normCross1, dot(v0, v1)*dp, dot(v2, v1)*dp, (dEdAngle*normBC)/normCross2);
......
platforms/cuda/src/kernels/verlet.cu
View file @ 93c467b2
......@@ -93,8 +93,8 @@ extern "C" __global__ void selectVerletStepSize(mixed maxStepSize, mixed errorTo
__syncthreads();
}
if (threadIdx.x == 0) {
mixed totalError = sqrt(error[0]/(NUM_ATOMS*3));
mixed newStepSize = sqrt(errorTol/totalError);
mixed totalError = SQRT(error[0]/(NUM_ATOMS*3));
mixed newStepSize = SQRT(errorTol/totalError);
mixed oldStepSize = dt[0].y;
if (oldStepSize > 0.0f)
newStepSize = min(newStepSize, oldStepSize*2.0f); // For safety, limit how quickly dt can increase.
......
platforms/cuda/tests/TestCudaSort.cpp
View file @ 93c467b2
This diff is collapsed.
platforms/opencl/src/OpenCLContext.cpp
View file @ 93c467b2
This diff is collapsed.
platforms/opencl/src/OpenCLFFT3D.cpp
View file @ 93c467b2
This diff is collapsed.
platforms/opencl/src/OpenCLFFT3D.h
View file @ 93c467b2
......@@ -81,8 +81,9 @@ public:
*/
static int findLegalDimension(int minimum);
private:
cl::Kernel createKernel(int xsize, int ysize, int zsize);
cl::Kernel createKernel(int xsize, int ysize, int zsize, int& threads);
int xsize, ysize, zsize;
int xthreads, ythreads, zthreads;
OpenCLContext& context;
cl::Kernel xkernel, ykernel, zkernel;
};
......
platforms/opencl/src/OpenCLIntegrationUtilities.h
View file @ 93c467b2
......@@ -141,8 +141,6 @@ private:
OpenCLArray* ccmaDelta1;
OpenCLArray* ccmaDelta2;
OpenCLArray* ccmaConverged;
cl::Buffer* ccmaConvergedBuffer;
cl_int* ccmaConvergedMemory;
OpenCLArray* vsite2AvgAtoms;
OpenCLArray* vsite2AvgWeights;
OpenCLArray* vsite3AvgAtoms;
......
platforms/opencl/src/OpenCLKernels.cpp
View file @ 93c467b2
This diff is collapsed.
platforms/opencl/src/OpenCLKernels.h
View file @ 93c467b2
This diff is collapsed.
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