Optimizations to PME in OpenCL

platforms/opencl/include/OpenCLKernels.h

@@ -663,7 +663,7 @@ private:
     OpenCLArray exceptionOffsetIndices;
     OpenCLArray globalParams;
     OpenCLArray cosSinSums;
-    OpenCLArray pmeGrid;
+    OpenCLArray pmeGrid1;
     OpenCLArray pmeGrid2;
     OpenCLArray pmeBsplineModuliX;
     OpenCLArray pmeBsplineModuliY;

platforms/opencl/src/OpenCLKernels.cpp

@@ -1726,15 +1726,18 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
                     defines["MULTSHIFT6"] = cl.doubleToString(multShift6);
                 }
                 int elementSize = (cl.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
-                int gridElements = gridSizeX*gridSizeY*gridSizeZ;
-                if (doLJPME)
-                    gridElements = max(gridElements, dispersionGridSizeX*dispersionGridSizeY*dispersionGridSizeZ);
-                pmeGrid.initialize(cl, gridElements, 2*elementSize, "pmeGrid");
+                int roundedZSize = PmeOrder*(int) ceil(gridSizeZ/(double) PmeOrder);
+                int gridElements = gridSizeX*gridSizeY*roundedZSize;
+                if (doLJPME) {
+                    roundedZSize = PmeOrder*(int) ceil(dispersionGridSizeZ/(double) PmeOrder);
+                    gridElements = max(gridElements, dispersionGridSizeX*dispersionGridSizeY*roundedZSize);
+                }
+                pmeGrid1.initialize(cl, gridElements, 2*elementSize, "pmeGrid1");
                 pmeGrid2.initialize(cl, gridElements, 2*elementSize, "pmeGrid2");
                 if (cl.getSupports64BitGlobalAtomics())
                     cl.addAutoclearBuffer(pmeGrid2);
                 else
-                    cl.addAutoclearBuffer(pmeGrid);
+                    cl.addAutoclearBuffer(pmeGrid1);
                 pmeBsplineModuliX.initialize(cl, gridSizeX, elementSize, "pmeBsplineModuliX");
                 pmeBsplineModuliY.initialize(cl, gridSizeY, elementSize, "pmeBsplineModuliY");
                 pmeBsplineModuliZ.initialize(cl, gridSizeZ, elementSize, "pmeBsplineModuliZ");
@@ -1755,8 +1758,6 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
                     dispersionFft = new OpenCLFFT3D(cl, dispersionGridSizeX, dispersionGridSizeY, dispersionGridSizeZ, true);
                 string vendor = cl.getDevice().getInfo<CL_DEVICE_VENDOR>();
                 bool isNvidia = (vendor.size() >= 6 && vendor.substr(0, 6) == "NVIDIA");
-                if (isNvidia)
-                    pmeDefines["USE_ALTERNATE_MEMORY_ACCESS_PATTERN"] = "1";
                 usePmeQueue = (!cl.getPlatformData().disablePmeStream && isNvidia);
                 if (usePmeQueue) {
                     pmeDefines["USE_PME_STREAM"] = "1";
@@ -2006,7 +2007,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
             ewaldForcesKernel.setArg<cl::Buffer>(1, cl.getPosq().getDeviceBuffer());
             ewaldForcesKernel.setArg<cl::Buffer>(2, cosSinSums.getDeviceBuffer());
         }
-        if (pmeGrid.isInitialized()) {
+        if (pmeGrid1.isInitialized()) {
             // Create kernels for Coulomb PME.
             
             map<string, string> replacements;
@@ -2036,7 +2037,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
             if (cl.getSupports64BitGlobalAtomics())
                 pmeSpreadChargeKernel.setArg<cl::Buffer>(3, pmeGrid2.getDeviceBuffer());
             else
-                pmeSpreadChargeKernel.setArg<cl::Buffer>(3, pmeGrid.getDeviceBuffer());
+                pmeSpreadChargeKernel.setArg<cl::Buffer>(3, pmeGrid1.getDeviceBuffer());
             pmeSpreadChargeKernel.setArg<cl::Buffer>(4, pmeBsplineTheta.getDeviceBuffer());
             if (deviceIsCpu || cl.getSupports64BitGlobalAtomics())
                 pmeSpreadChargeKernel.setArg<cl::Buffer>(13, charges.getDeviceBuffer());
@@ -2053,13 +2054,13 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
             pmeEvalEnergyKernel.setArg<cl::Buffer>(4, pmeBsplineModuliZ.getDeviceBuffer());
             pmeInterpolateForceKernel.setArg<cl::Buffer>(0, cl.getPosq().getDeviceBuffer());
             pmeInterpolateForceKernel.setArg<cl::Buffer>(1, cl.getForceBuffers().getDeviceBuffer());
-            pmeInterpolateForceKernel.setArg<cl::Buffer>(2, pmeGrid.getDeviceBuffer());
+            pmeInterpolateForceKernel.setArg<cl::Buffer>(2, pmeGrid1.getDeviceBuffer());
             pmeInterpolateForceKernel.setArg<cl::Buffer>(11, pmeAtomGridIndex.getDeviceBuffer());
             pmeInterpolateForceKernel.setArg<cl::Buffer>(12, charges.getDeviceBuffer());
             if (cl.getSupports64BitGlobalAtomics()) {
                 pmeFinishSpreadChargeKernel = cl::Kernel(program, "finishSpreadCharge");
                 pmeFinishSpreadChargeKernel.setArg<cl::Buffer>(0, pmeGrid2.getDeviceBuffer());
-                pmeFinishSpreadChargeKernel.setArg<cl::Buffer>(1, pmeGrid.getDeviceBuffer());
+                pmeFinishSpreadChargeKernel.setArg<cl::Buffer>(1, pmeGrid1.getDeviceBuffer());
             }
             if (usePmeQueue)
                 syncQueue->setKernel(cl::Kernel(program, "addEnergy"));
@@ -2099,7 +2100,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
                 if (cl.getSupports64BitGlobalAtomics())
                     pmeDispersionSpreadChargeKernel.setArg<cl::Buffer>(3, pmeGrid2.getDeviceBuffer());
                 else
-                    pmeDispersionSpreadChargeKernel.setArg<cl::Buffer>(3, pmeGrid.getDeviceBuffer());
+                    pmeDispersionSpreadChargeKernel.setArg<cl::Buffer>(3, pmeGrid1.getDeviceBuffer());
                 pmeDispersionSpreadChargeKernel.setArg<cl::Buffer>(4, pmeBsplineTheta.getDeviceBuffer());
                 if (deviceIsCpu || cl.getSupports64BitGlobalAtomics())
                     pmeDispersionSpreadChargeKernel.setArg<cl::Buffer>(13, sigmaEpsilon.getDeviceBuffer());
@@ -2116,13 +2117,13 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
                 pmeDispersionEvalEnergyKernel.setArg<cl::Buffer>(4, pmeDispersionBsplineModuliZ.getDeviceBuffer());
                 pmeDispersionInterpolateForceKernel.setArg<cl::Buffer>(0, cl.getPosq().getDeviceBuffer());
                 pmeDispersionInterpolateForceKernel.setArg<cl::Buffer>(1, cl.getForceBuffers().getDeviceBuffer());
-                pmeDispersionInterpolateForceKernel.setArg<cl::Buffer>(2, pmeGrid.getDeviceBuffer());
+                pmeDispersionInterpolateForceKernel.setArg<cl::Buffer>(2, pmeGrid1.getDeviceBuffer());
                 pmeDispersionInterpolateForceKernel.setArg<cl::Buffer>(11, pmeAtomGridIndex.getDeviceBuffer());
                 pmeDispersionInterpolateForceKernel.setArg<cl::Buffer>(12, sigmaEpsilon.getDeviceBuffer());
                 if (cl.getSupports64BitGlobalAtomics()) {
                     pmeDispersionFinishSpreadChargeKernel = cl::Kernel(program, "finishSpreadCharge");
                     pmeDispersionFinishSpreadChargeKernel.setArg<cl::Buffer>(0, pmeGrid2.getDeviceBuffer());
-                    pmeDispersionFinishSpreadChargeKernel.setArg<cl::Buffer>(1, pmeGrid.getDeviceBuffer());
+                    pmeDispersionFinishSpreadChargeKernel.setArg<cl::Buffer>(1, pmeGrid1.getDeviceBuffer());
                 }
             }
        }
@@ -2176,7 +2177,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
         cl.executeKernel(ewaldSumsKernel, cosSinSums.getSize());
         cl.executeKernel(ewaldForcesKernel, cl.getNumAtoms());
     }
-    if (pmeGrid.isInitialized() && includeReciprocal) {
+    if (pmeGrid1.isInitialized() && includeReciprocal) {
         if (usePmeQueue && !includeEnergy)
             cl.setQueue(pmeQueue);
         
@@ -2251,7 +2252,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
                     cl.executeKernel(pmeSpreadChargeKernel, cl.getNumAtoms());
                 }
             }
-            fft->execFFT(pmeGrid, pmeGrid2, true);
+            fft->execFFT(pmeGrid1, pmeGrid2, true);
             mm_double4 boxSize = cl.getPeriodicBoxSizeDouble();
             if (cl.getUseDoublePrecision()) {
                 pmeConvolutionKernel.setArg<mm_double4>(4, recipBoxVectors[0]);
@@ -2272,7 +2273,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
             if (includeEnergy)
                 cl.executeKernel(pmeEvalEnergyKernel, gridSizeX*gridSizeY*gridSizeZ);
             cl.executeKernel(pmeConvolutionKernel, gridSizeX*gridSizeY*gridSizeZ);
-            fft->execFFT(pmeGrid2, pmeGrid, false);
+            fft->execFFT(pmeGrid2, pmeGrid1, false);
             setPeriodicBoxArgs(cl, pmeInterpolateForceKernel, 3);
             if (cl.getUseDoublePrecision()) {
                 pmeInterpolateForceKernel.setArg<mm_double4>(8, recipBoxVectors[0]);
@@ -2304,7 +2305,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
             }
             cl.executeKernel(pmeDispersionUpdateBsplinesKernel, cl.getNumAtoms());
             if (deviceIsCpu && !cl.getSupports64BitGlobalAtomics()) {
-                cl.clearBuffer(pmeGrid);
+                cl.clearBuffer(pmeGrid1);
                 setPeriodicBoxArgs(cl, pmeDispersionSpreadChargeKernel, 5);
                 if (cl.getUseDoublePrecision()) {
                     pmeDispersionSpreadChargeKernel.setArg<mm_double4>(10, recipBoxVectors[0]);
@@ -2319,7 +2320,8 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
                 cl.executeKernel(pmeDispersionSpreadChargeKernel, 2*cl.getDevice().getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>(), 1);
             }
             else {
-                sort->sort(pmeAtomGridIndex);
+                if (!hasCoulomb)
+                    sort->sort(pmeAtomGridIndex);
                 if (cl.getSupports64BitGlobalAtomics()) {
                     cl.clearBuffer(pmeGrid2);
                     setPeriodicBoxArgs(cl, pmeDispersionSpreadChargeKernel, 5);
@@ -2337,7 +2339,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
                     cl.executeKernel(pmeDispersionFinishSpreadChargeKernel, gridSizeX*gridSizeY*gridSizeZ);
                 }
                 else {
-                    cl.clearBuffer(pmeGrid);
+                    cl.clearBuffer(pmeGrid1);
                     cl.executeKernel(pmeDispersionAtomRangeKernel, cl.getNumAtoms());
                     setPeriodicBoxSizeArg(cl, pmeDispersionZIndexKernel, 2);
                     if (cl.getUseDoublePrecision())
@@ -2348,7 +2350,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
                     cl.executeKernel(pmeDispersionSpreadChargeKernel, cl.getNumAtoms());
                 }
             }
-            dispersionFft->execFFT(pmeGrid, pmeGrid2, true);
+            dispersionFft->execFFT(pmeGrid1, pmeGrid2, true);
             mm_double4 boxSize = cl.getPeriodicBoxSizeDouble();
             if (cl.getUseDoublePrecision()) {
                 pmeDispersionConvolutionKernel.setArg<mm_double4>(4, recipBoxVectors[0]);
@@ -2369,7 +2371,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
             if (includeEnergy)
                 cl.executeKernel(pmeDispersionEvalEnergyKernel, gridSizeX*gridSizeY*gridSizeZ);
             cl.executeKernel(pmeDispersionConvolutionKernel, gridSizeX*gridSizeY*gridSizeZ);
-            dispersionFft->execFFT(pmeGrid2, pmeGrid, false);
+            dispersionFft->execFFT(pmeGrid2, pmeGrid1, false);
             setPeriodicBoxArgs(cl, pmeDispersionInterpolateForceKernel, 3);
             if (cl.getUseDoublePrecision()) {
                 pmeDispersionInterpolateForceKernel.setArg<mm_double4>(8, recipBoxVectors[0]);

platforms/opencl/src/kernels/pme.cl

@@ -105,12 +105,25 @@ __kernel void gridSpreadCharge(__global const real4* restrict posq, __global con
         , __global const real* restrict charges
 #endif
     ) {
-    const real scale = 1/(real) (PME_ORDER-1);
-    real4 data[PME_ORDER];
+    // To improve memory efficiency, we divide indices along the z axis into
+    // PME_ORDER blocks, where the data for each block is stored together.  We
+    // can ensure that all threads write to the same block at the same time,
+    // which leads to better coalescing of writes.
     
+    __local int zindexTable[GRID_SIZE_Z+PME_ORDER];
+    int blockSize = (int) ceil(GRID_SIZE_Z/(real) PME_ORDER);
+    for (int i = get_local_id(0); i < GRID_SIZE_Z+PME_ORDER; i += get_local_size(0)) {
+        int zindex = i % GRID_SIZE_Z;
+	int block = zindex % PME_ORDER;
+        zindexTable[i] = zindex/PME_ORDER + block*GRID_SIZE_X*GRID_SIZE_Y*blockSize;
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+        
     // Process the atoms in spatially sorted order.  This improves efficiency when writing
     // the grid values.
     
+    const real scale = 1/(real) (PME_ORDER-1);
+    real4 data[PME_ORDER];
     for (int i = get_global_id(0); i < NUM_ATOMS; i += get_global_size(0)) {
         int atom = pmeAtomGridIndex[i].x;
         real4 pos = posq[atom];
@@ -118,7 +131,7 @@ __kernel void gridSpreadCharge(__global const real4* restrict posq, __global con
         const float2 sigEps = sigmaEpsilon[atom];
         const real charge = 8*sigEps.x*sigEps.x*sigEps.x*sigEps.y;
 #else
-        const real charge = CHARGE;
+        const real charge = (CHARGE)*EPSILON_FACTOR;
 #endif
         if (charge == 0)
             continue;
@@ -154,40 +167,47 @@ __kernel void gridSpreadCharge(__global const real4* restrict posq, __global con
 
         // Spread the charge from this atom onto each grid point.
 
+	int izoffset = (PME_ORDER-(gridIndex.z%PME_ORDER)) % PME_ORDER;
         for (int ix = 0; ix < PME_ORDER; ix++) {
-            int xindex = gridIndex.x+ix;
-            xindex -= (xindex >= GRID_SIZE_X ? GRID_SIZE_X : 0);
+            int xbase = gridIndex.x+ix;
+            xbase -= (xbase >= GRID_SIZE_X ? GRID_SIZE_X : 0);
+            xbase = xbase*GRID_SIZE_Y;
+            real dx = charge*data[ix].x;
             for (int iy = 0; iy < PME_ORDER; iy++) {
-                int yindex = gridIndex.y+iy;
-                yindex -= (yindex >= GRID_SIZE_Y ? GRID_SIZE_Y : 0);
-                for (int iz = 0; iz < PME_ORDER; iz++) {
+                int ybase = gridIndex.y+iy;
+                ybase -= (ybase >= GRID_SIZE_Y ? GRID_SIZE_Y : 0);
+                ybase = (xbase+ybase)*blockSize;
+                real dxdy = dx*data[iy].y;
+                for (int i = 0; i < PME_ORDER; i++) {
+		    int iz = (i+izoffset) % PME_ORDER;
                     int zindex = gridIndex.z+iz;
-                    zindex -= (zindex >= GRID_SIZE_Z ? GRID_SIZE_Z : 0);
-                    int index = xindex*GRID_SIZE_Y*GRID_SIZE_Z + yindex*GRID_SIZE_Z + zindex;
-                    real add = charge*data[ix].x*data[iy].y*data[iz].z;
-#ifdef USE_ALTERNATE_MEMORY_ACCESS_PATTERN
-                    // On Nvidia devices (at least Maxwell anyway), this split ordering produces much higher performance.  Why?
-                    // I have no idea!  And of course on AMD it produces slower performance.  GPUs are not meant to be understood.
-                    atom_add(&pmeGrid[index%2 == 0 ? index/2 : (index+GRID_SIZE_X*GRID_SIZE_Y*GRID_SIZE_Z)/2], (long) (add*0x100000000));
-#else
+                    int index = ybase + zindexTable[zindex];
+                    real add = dxdy*data[iz].z;
                     atom_add(&pmeGrid[index], (long) (add*0x100000000));
-#endif
                 }
             }
         }
     }
 }
 
-__kernel void finishSpreadCharge(__global long* restrict fixedGrid, __global real* restrict realGrid) {
+__kernel void finishSpreadCharge(__global long* restrict grid1, __global real* restrict grid2) {
+    // During charge spreading, we shuffled the order of indices along the z
+    // axis to make memory access more efficient.  We now need to unshuffle
+    // them and convert fixed point values to floating point.
+
+    __local int zindexTable[GRID_SIZE_Z];
+    int blockSize = (int) ceil(GRID_SIZE_Z/(real) PME_ORDER);
+    for (int i = get_local_id(0); i < GRID_SIZE_Z; i += get_local_size(0)) {
+	int block = i % PME_ORDER;
+        zindexTable[i] = i/PME_ORDER + block*GRID_SIZE_X*GRID_SIZE_Y*blockSize;
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
     const unsigned int gridSize = GRID_SIZE_X*GRID_SIZE_Y*GRID_SIZE_Z;
-    real scale = EPSILON_FACTOR/(real) 0x100000000;
+    real scale = 1/(real) 0x100000000;
     for (int index = get_global_id(0); index < gridSize; index += get_global_size(0)) {
-#ifdef USE_ALTERNATE_MEMORY_ACCESS_PATTERN
-        long value = fixedGrid[index%2 == 0 ? index/2 : (index+gridSize)/2];
-#else
-        long value = fixedGrid[index];
-#endif
-        realGrid[index] = (real) (value*scale);
+        int zindex = index%GRID_SIZE_Z;
+        int loadIndex = zindexTable[zindex] + blockSize*(int) (index/GRID_SIZE_Z);
+        grid2[index] = scale*grid1[loadIndex];
     }
 }
 #elif defined(DEVICE_IS_CPU)
@@ -230,7 +250,7 @@ __kernel void gridSpreadCharge(__global const real4* restrict posq, __global con
         const float2 sigEps = sigmaEpsilon[atom];
         const real charge = 8*sigEps.x*sigEps.x*sigEps.x*sigEps.y;
 #else
-        const real charge = CHARGE;
+        const real charge = (CHARGE)*EPSILON_FACTOR;
 #endif
         if (charge == 0)
             continue;
@@ -269,7 +289,7 @@ __kernel void gridSpreadCharge(__global const real4* restrict posq, __global con
                     int zindex = gridIndex.z+iz;
                     zindex -= (zindex >= GRID_SIZE_Z ? GRID_SIZE_Z : 0);
                     int index = xindex*GRID_SIZE_Y*GRID_SIZE_Z + yindex*GRID_SIZE_Z + zindex;
-                    pmeGrid[index] += EPSILON_FACTOR*charge*data[ix].x*data[iy].y*data[iz].z;
+                    pmeGrid[index] += charge*data[ix].x*data[iy].y*data[iz].z;
                 }
             }
         }