Skip to content

GitLab

Commit 4949017b authored by Peter Eastman's avatar Peter Eastman
Browse files

Continuing CUDA implementation of parameter derivatives

parent eae8def5
Expand all Show whitespace changes
Inline Side-by-side
Showing with 218 additions and 49 deletions
platforms/cuda/include/CudaKernels.h
View file @ 4949017b
...@@ -826,13 +826,15 @@ public: ...@@ -826,13 +826,15 @@ public:
void copyParametersToContext(ContextImpl& context, const CustomGBForce& force); void copyParametersToContext(ContextImpl& context, const CustomGBForce& force);
private: private:
double cutoff; double cutoff;
bool hasInitializedKernels, needParameterGradient; bool hasInitializedKernels, needParameterGradient, needEnergyParamDerivs;
int maxTiles, numComputedValues; int maxTiles, numComputedValues;
CudaContext& cu; CudaContext& cu;
CudaParameterSet* params; CudaParameterSet* params;
CudaParameterSet* computedValues; CudaParameterSet* computedValues;
CudaParameterSet* energyDerivs; CudaParameterSet* energyDerivs;
CudaParameterSet* energyDerivChain; CudaParameterSet* energyDerivChain;
std::vector<CudaParameterSet*> dValuedParam;
std::vector<CudaArray*> dValue0dParam;
CudaArray* longEnergyDerivs; CudaArray* longEnergyDerivs;
CudaArray* globals; CudaArray* globals;
CudaArray* valueBuffers; CudaArray* valueBuffers;
......
platforms/cuda/src/CudaKernels.cpp
View file @ 4949017b
This diff is collapsed.
platforms/cuda/src/kernels/customGBEnergyN2.cu
View file @ 4949017b
...@@ -28,6 +28,7 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc ...@@ -28,6 +28,7 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc
const unsigned int tgx = threadIdx.x & (TILE_SIZE-1); const unsigned int tgx = threadIdx.x & (TILE_SIZE-1);
const unsigned int tbx = threadIdx.x - tgx; const unsigned int tbx = threadIdx.x - tgx;
mixed energy = 0; mixed energy = 0;
INIT_PARAM_DERIVS
__shared__ AtomData localData[THREAD_BLOCK_SIZE]; __shared__ AtomData localData[THREAD_BLOCK_SIZE];
// First loop: process tiles that contain exclusions. // First loop: process tiles that contain exclusions.
...@@ -69,6 +70,7 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc ...@@ -69,6 +70,7 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc
atom2 = y*TILE_SIZE+j; atom2 = y*TILE_SIZE+j;
real dEdR = 0; real dEdR = 0;
real tempEnergy = 0; real tempEnergy = 0;
const real interactionScale = 0.5f;
#ifdef USE_EXCLUSIONS #ifdef USE_EXCLUSIONS
bool isExcluded = !(excl & 0x1); bool isExcluded = !(excl & 0x1);
#endif #endif
...@@ -120,6 +122,7 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc ...@@ -120,6 +122,7 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc
atom2 = y*TILE_SIZE+tj; atom2 = y*TILE_SIZE+tj;
real dEdR = 0; real dEdR = 0;
real tempEnergy = 0; real tempEnergy = 0;
const real interactionScale = 1;
#ifdef USE_EXCLUSIONS #ifdef USE_EXCLUSIONS
bool isExcluded = !(excl & 0x1); bool isExcluded = !(excl & 0x1);
#endif #endif
...@@ -266,6 +269,7 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc ...@@ -266,6 +269,7 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc
atom2 = atomIndices[tbx+tj]; atom2 = atomIndices[tbx+tj];
real dEdR = 0; real dEdR = 0;
real tempEnergy = 0; real tempEnergy = 0;
const real interactionScale = 1;
if (atom1 < NUM_ATOMS && atom2 < NUM_ATOMS) { if (atom1 < NUM_ATOMS && atom2 < NUM_ATOMS) {
COMPUTE_INTERACTION COMPUTE_INTERACTION
dEdR /= -r; dEdR /= -r;
...@@ -309,6 +313,7 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc ...@@ -309,6 +313,7 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc
atom2 = atomIndices[tbx+tj]; atom2 = atomIndices[tbx+tj];
real dEdR = 0; real dEdR = 0;
real tempEnergy = 0; real tempEnergy = 0;
const real interactionScale = 1;
if (atom1 < NUM_ATOMS && atom2 < NUM_ATOMS) { if (atom1 < NUM_ATOMS && atom2 < NUM_ATOMS) {
COMPUTE_INTERACTION COMPUTE_INTERACTION
dEdR /= -r; dEdR /= -r;
...@@ -353,4 +358,5 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc ...@@ -353,4 +358,5 @@ extern "C" __global__ void computeN2Energy(unsigned long long* __restrict__ forc
pos++; pos++;
} }
energyBuffer[blockIdx.x*blockDim.x+threadIdx.x] += energy; energyBuffer[blockIdx.x*blockDim.x+threadIdx.x] += energy;
SAVE_PARAM_DERIVS
} }
platforms/cuda/src/kernels/customGBEnergyPerParticle.cu
View file @ 4949017b
...@@ -5,6 +5,7 @@ ...@@ -5,6 +5,7 @@
extern "C" __global__ void computePerParticleEnergy(long long* __restrict__ forceBuffers, mixed* __restrict__ energyBuffer, const real4* __restrict__ posq extern "C" __global__ void computePerParticleEnergy(long long* __restrict__ forceBuffers, mixed* __restrict__ energyBuffer, const real4* __restrict__ posq
PARAMETER_ARGUMENTS) { PARAMETER_ARGUMENTS) {
mixed energy = 0; mixed energy = 0;
INIT_PARAM_DERIVS
for (unsigned int index = blockIdx.x*blockDim.x+threadIdx.x; index < NUM_ATOMS; index += blockDim.x*gridDim.x) { for (unsigned int index = blockIdx.x*blockDim.x+threadIdx.x; index < NUM_ATOMS; index += blockDim.x*gridDim.x) {
// Load the derivatives // Load the derivatives
...@@ -17,4 +18,5 @@ extern "C" __global__ void computePerParticleEnergy(long long* __restrict__ forc ...@@ -17,4 +18,5 @@ extern "C" __global__ void computePerParticleEnergy(long long* __restrict__ forc
COMPUTE_ENERGY COMPUTE_ENERGY
} }
energyBuffer[blockIdx.x*blockDim.x+threadIdx.x] += energy; energyBuffer[blockIdx.x*blockDim.x+threadIdx.x] += energy;
SAVE_PARAM_DERIVS
} }
platforms/cuda/src/kernels/customGBGradientChainRule.cu
View file @ 4949017b
...@@ -4,6 +4,7 @@ ...@@ -4,6 +4,7 @@
extern "C" __global__ void computeGradientChainRuleTerms(long long* __restrict__ forceBuffers, const real4* __restrict__ posq extern "C" __global__ void computeGradientChainRuleTerms(long long* __restrict__ forceBuffers, const real4* __restrict__ posq
PARAMETER_ARGUMENTS) { PARAMETER_ARGUMENTS) {
INIT_PARAM_DERIVS
const real scale = RECIP((real) 0x100000000); const real scale = RECIP((real) 0x100000000);
for (unsigned int index = blockIdx.x*blockDim.x+threadIdx.x; index < NUM_ATOMS; index += blockDim.x*gridDim.x) { for (unsigned int index = blockIdx.x*blockDim.x+threadIdx.x; index < NUM_ATOMS; index += blockDim.x*gridDim.x) {
real4 pos = posq[index]; real4 pos = posq[index];
...@@ -13,4 +14,5 @@ extern "C" __global__ void computeGradientChainRuleTerms(long long* __restrict__ ...@@ -13,4 +14,5 @@ extern "C" __global__ void computeGradientChainRuleTerms(long long* __restrict__
forceBuffers[index+PADDED_NUM_ATOMS] = (long long) (force.y*0x100000000); forceBuffers[index+PADDED_NUM_ATOMS] = (long long) (force.y*0x100000000);
forceBuffers[index+PADDED_NUM_ATOMS*2] = (long long) (force.z*0x100000000); forceBuffers[index+PADDED_NUM_ATOMS*2] = (long long) (force.z*0x100000000);
} }
SAVE_PARAM_DERIVS
} }
platforms/cuda/src/kernels/customGBValueN2.cu
View file @ 4949017b
...@@ -73,6 +73,7 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const ...@@ -73,6 +73,7 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
COMPUTE_VALUE COMPUTE_VALUE
} }
value += tempValue1; value += tempValue1;
ADD_TEMP_DERIVS1
#ifdef USE_CUTOFF #ifdef USE_CUTOFF
} }
#endif #endif
...@@ -121,6 +122,8 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const ...@@ -121,6 +122,8 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
} }
value += tempValue1; value += tempValue1;
localData[tbx+tj].value += tempValue2; localData[tbx+tj].value += tempValue2;
ADD_TEMP_DERIVS1
ADD_TEMP_DERIVS2
#ifdef USE_CUTOFF #ifdef USE_CUTOFF
} }
#endif #endif
...@@ -133,11 +136,13 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const ...@@ -133,11 +136,13 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
// Write results. // Write results.
unsigned int offset = x*TILE_SIZE + tgx; unsigned int offset1 = x*TILE_SIZE + tgx;
atomicAdd(&global_value[offset], static_cast<unsigned long long>((long long) (value*0x100000000))); atomicAdd(&global_value[offset1], static_cast<unsigned long long>((long long) (value*0x100000000)));
STORE_PARAM_DERIVS1
if (x != y) { if (x != y) {
offset = y*TILE_SIZE + tgx; unsigned int offset2 = y*TILE_SIZE + tgx;
atomicAdd(&global_value[offset], static_cast<unsigned long long>((long long) (localData[threadIdx.x].value*0x100000000))); atomicAdd(&global_value[offset2], static_cast<unsigned long long>((long long) (localData[threadIdx.x].value*0x100000000)));
STORE_PARAM_DERIVS2
} }
} }
...@@ -244,6 +249,8 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const ...@@ -244,6 +249,8 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
} }
value += tempValue1; value += tempValue1;
localData[tbx+tj].value += tempValue2; localData[tbx+tj].value += tempValue2;
ADD_TEMP_DERIVS1
ADD_TEMP_DERIVS2
} }
tj = (tj + 1) & (TILE_SIZE - 1); tj = (tj + 1) & (TILE_SIZE - 1);
} }
...@@ -276,6 +283,8 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const ...@@ -276,6 +283,8 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
} }
value += tempValue1; value += tempValue1;
localData[tbx+tj].value += tempValue2; localData[tbx+tj].value += tempValue2;
ADD_TEMP_DERIVS1
ADD_TEMP_DERIVS2
#ifdef USE_CUTOFF #ifdef USE_CUTOFF
} }
#endif #endif
...@@ -285,14 +294,19 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const ...@@ -285,14 +294,19 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
// Write results. // Write results.
atomicAdd(&global_value[atom1], static_cast<unsigned long long>((long long) (value*0x100000000))); unsigned int offset1 = atom1;
atomicAdd(&global_value[offset1], static_cast<unsigned long long>((long long) (value*0x100000000)));
STORE_PARAM_DERIVS1
#ifdef USE_CUTOFF #ifdef USE_CUTOFF
unsigned int atom2 = atomIndices[threadIdx.x]; unsigned int atom2 = atomIndices[threadIdx.x];
#else #else
unsigned int atom2 = y*TILE_SIZE + tgx; unsigned int atom2 = y*TILE_SIZE + tgx;
#endif #endif
if (atom2 < PADDED_NUM_ATOMS) if (atom2 < PADDED_NUM_ATOMS) {
atomicAdd(&global_value[atom2], static_cast<unsigned long long>((long long) (localData[threadIdx.x].value*0x100000000))); unsigned int offset2 = atom2;
atomicAdd(&global_value[offset2], static_cast<unsigned long long>((long long) (localData[threadIdx.x].value*0x100000000)));
STORE_PARAM_DERIVS2
}
} }
pos++; pos++;
} }
......
platforms/cuda/src/kernels/customGBValuePerParticle.cu
View file @ 4949017b
...@@ -8,6 +8,7 @@ extern "C" __global__ void computePerParticleValues(real4* posq, long long* valu ...@@ -8,6 +8,7 @@ extern "C" __global__ void computePerParticleValues(real4* posq, long long* valu
// Load the pairwise value // Load the pairwise value
real sum = valueBuffers[index]/(real) 0x100000000; real sum = valueBuffers[index]/(real) 0x100000000;
REDUCE_PARAM0_DERIV
// Now calculate other values // Now calculate other values
......
platforms/opencl/src/OpenCLKernels.cpp
View file @ 4949017b
...@@ -3301,7 +3301,7 @@ void OpenCLCalcCustomGBForceKernel::initialize(const System& system, const Custo ...@@ -3301,7 +3301,7 @@ void OpenCLCalcCustomGBForceKernel::initialize(const System& system, const Custo
string variableName = "dValuedParam_0_"+cl.intToString(i); string variableName = "dValuedParam_0_"+cl.intToString(i);
if (useLong) { if (useLong) {
extraArgs << ", __global const long* restrict dValue0dParam" << i; extraArgs << ", __global const long* restrict dValue0dParam" << i;
deriv0 << "real " << variableName << " = (1.0f/0x100000000)*dValue0dParam[index];\n"; deriv0 << "real " << variableName << " = (1.0f/0x100000000)*dValue0dParam" << i << "[index];\n";
} }
else { else {
extraArgs << ", __global const real* restrict dValue0dParam" << i; extraArgs << ", __global const real* restrict dValue0dParam" << i;
......
platforms/opencl/src/kernels/customGBValueN2.cl
View file @ 4949017b
...@@ -320,7 +320,7 @@ __kernel void computeN2Value(__global const real4* restrict posq, __local real4* ...@@ -320,7 +320,7 @@ __kernel void computeN2Value(__global const real4* restrict posq, __local real4*
unsigned int atom2 = y*TILE_SIZE + tgx; unsigned int atom2 = y*TILE_SIZE + tgx;
#endif #endif
#ifdef SUPPORTS_64_BIT_ATOMICS #ifdef SUPPORTS_64_BIT_ATOMICS
unsigned in offset1 = atom1; unsigned int offset1 = atom1;
atom_add(&global_value[offset1], (long) (value*0x100000000)); atom_add(&global_value[offset1], (long) (value*0x100000000));
STORE_PARAM_DERIVS1 STORE_PARAM_DERIVS1
if (atom2 < PADDED_NUM_ATOMS) { if (atom2 < PADDED_NUM_ATOMS) {
......
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