2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2016,2017,2018,2019,2020,2021, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
37 * \brief Implements helper routines for PME gather and spline routines.
39 * \author Aleksei Iupinov <a.yupinov@gmail.com>
46 #include "gromacs/gpu_utils/cuda_kernel_utils.cuh"
47 #include "gromacs/gpu_utils/vectype_ops.cuh"
53 * Gets a base of the unique index to an element in a spline parameter buffer (theta/dtheta),
54 * which is laid out for GPU spread/gather kernels. The base only corresponds to the atom index within the execution block.
55 * Feed the result into getSplineParamIndex() to get a full index.
56 * TODO: it's likely that both parameters can be just replaced with a single atom index, as they are derived from it.
57 * Do that, verifying that the generated code is not bloated, and/or revise the spline indexing scheme.
58 * Removing warp dependency would also be nice (and would probably coincide with removing c_pmeSpreadGatherAtomsPerWarp).
60 * \tparam order PME order
61 * \tparam atomsPerWarp Number of atoms processed by a warp
62 * \param[in] warpIndex Warp index wrt the block.
63 * \param[in] atomWarpIndex Atom index wrt the warp (from 0 to atomsPerWarp - 1).
65 * \returns Index into theta or dtheta array using GPU layout.
67 template<int order, int atomsPerWarp>
68 int __device__ __forceinline__ getSplineParamIndexBase(int warpIndex, int atomWarpIndex)
70 assert((atomWarpIndex >= 0) && (atomWarpIndex < atomsPerWarp));
71 const int dimIndex = 0;
72 const int splineIndex = 0;
73 // The zeroes are here to preserve the full index formula for reference
74 return (((splineIndex + order * warpIndex) * DIM + dimIndex) * atomsPerWarp + atomWarpIndex);
78 * Gets a unique index to an element in a spline parameter buffer (theta/dtheta),
79 * which is laid out for GPU spread/gather kernels. The index is wrt to the execution block,
80 * in range(0, atomsPerBlock * order * DIM).
81 * This function consumes result of getSplineParamIndexBase() and adjusts it for \p dimIndex and \p splineIndex.
83 * \tparam order PME order
84 * \tparam atomsPerWarp Number of atoms processed by a warp
85 * \param[in] paramIndexBase Must be result of getSplineParamIndexBase().
86 * \param[in] dimIndex Dimension index (from 0 to 2)
87 * \param[in] splineIndex Spline contribution index (from 0 to \p order - 1)
89 * \returns Index into theta or dtheta array using GPU layout.
91 template<int order, int atomsPerWarp>
92 int __device__ __forceinline__ getSplineParamIndex(int paramIndexBase, int dimIndex, int splineIndex)
94 assert((dimIndex >= XX) && (dimIndex < DIM));
95 assert((splineIndex >= 0) && (splineIndex < order));
96 return (paramIndexBase + (splineIndex * DIM + dimIndex) * atomsPerWarp);
100 * An inline CUDA function for skipping the zero-charge atoms.
102 * \returns Non-0 if atom should be processed, 0 otherwise.
103 * \param[in] coefficient The atom charge.
105 * This is called from the spline_and_spread and gather PME kernels.
107 bool __device__ __forceinline__ pme_gpu_check_atom_charge(const float coefficient)
109 assert(isfinite(coefficient));
110 return c_skipNeutralAtoms ? (coefficient != 0.0F) : true;
113 //! Controls if the atom and charge data is prefeched into shared memory or loaded per thread from global
114 static const bool c_useAtomDataPrefetch = true;
116 /*! \brief Asserts if the argument is finite.
118 * The function works for any data type, that can be casted to float. Note that there is also
119 * a specialized implementation for float3 data type.
121 * \param[in] arg Argument to check.
124 __device__ inline void assertIsFinite(T arg);
127 __device__ inline void assertIsFinite(float3 gmx_unused arg)
129 assert(isfinite(static_cast<float>(arg.x)));
130 assert(isfinite(static_cast<float>(arg.y)));
131 assert(isfinite(static_cast<float>(arg.z)));
135 __device__ inline void assertIsFinite(T gmx_unused arg)
137 assert(isfinite(static_cast<float>(arg)));
141 * General purpose function for loading atom-related data from global to shared memory.
143 * \tparam[in] T Data type (float/int/...)
144 * \tparam[in] atomsPerBlock Number of atoms processed by a block - should be accounted for in
145 * the size of the shared memory array.
146 * \tparam[in] dataCountPerAtom Number of data elements per single atom (e.g. DIM for an rvec
147 * coordinates array).
148 * \param[out] sm_destination Shared memory array for output.
149 * \param[in] gm_source Global memory array for input.
151 template<typename T, int atomsPerBlock, int dataCountPerAtom>
152 __device__ __forceinline__ void pme_gpu_stage_atom_data(T* __restrict__ sm_destination,
153 const T* __restrict__ gm_source)
155 const int blockIndex = blockIdx.y * gridDim.x + blockIdx.x;
156 const int threadLocalIndex = ((threadIdx.z * blockDim.y + threadIdx.y) * blockDim.x) + threadIdx.x;
157 const int localIndex = threadLocalIndex;
158 const int globalIndexBase = blockIndex * atomsPerBlock * dataCountPerAtom;
159 const int globalIndex = globalIndexBase + localIndex;
160 if (localIndex < atomsPerBlock * dataCountPerAtom)
162 assertIsFinite(gm_source[globalIndex]);
163 sm_destination[localIndex] = gm_source[globalIndex];
168 * PME GPU spline parameter and gridline indices calculation.
169 * This corresponds to the CPU functions calc_interpolation_idx() and make_bsplines().
170 * First stage of the whole kernel.
172 * \tparam[in] order PME interpolation order.
173 * \tparam[in] atomsPerBlock Number of atoms processed by a block - should be accounted for
174 * in the sizes of the shared memory arrays.
175 * \tparam[in] atomsPerWarp Number of atoms processed by a warp
176 * \tparam[in] writeSmDtheta Bool controlling if the theta derivative should be written to shared memory. Enables calculation of dtheta if set.
177 * \tparam[in] writeGlobal A boolean which tells if the theta values and gridlines should be written to global memory. Enables calculation of dtheta if set.
178 * \param[in] kernelParams Input PME CUDA data in constant memory.
179 * \param[in] atomIndexOffset Starting atom index for the execution block w.r.t. global memory.
180 * \param[in] atomX Atom coordinate of atom processed by thread.
181 * \param[in] atomCharge Atom charge/coefficient of atom processed by thread.
182 * \param[out] sm_theta Atom spline values in the shared memory.
183 * \param[out] sm_dtheta Derivative of atom spline values in shared memory.
184 * \param[out] sm_gridlineIndices Atom gridline indices in the shared memory.
187 template<int order, int atomsPerBlock, int atomsPerWarp, bool writeSmDtheta, bool writeGlobal>
188 __device__ __forceinline__ void calculate_splines(const PmeGpuCudaKernelParams kernelParams,
189 const int atomIndexOffset,
191 const float atomCharge,
192 float* __restrict__ sm_theta,
193 float* __restrict__ sm_dtheta,
194 int* __restrict__ sm_gridlineIndices)
196 /* Global memory pointers for output */
197 float* __restrict__ gm_theta = kernelParams.atoms.d_theta;
198 float* __restrict__ gm_dtheta = kernelParams.atoms.d_dtheta;
199 int* __restrict__ gm_gridlineIndices = kernelParams.atoms.d_gridlineIndices;
201 /* Fractional coordinates */
202 __shared__ float sm_fractCoords[atomsPerBlock * DIM];
204 /* Thread index w.r.t. block */
205 const int threadLocalId =
206 (threadIdx.z * (blockDim.x * blockDim.y)) + (threadIdx.y * blockDim.x) + threadIdx.x;
207 /* Warp index w.r.t. block - could probably be obtained easier? */
208 const int warpIndex = threadLocalId / warp_size;
209 /* Atom index w.r.t. warp - alternating 0 1 0 1 .. */
210 const int atomWarpIndex = threadIdx.z % atomsPerWarp;
211 /* Atom index w.r.t. block/shared memory */
212 const int atomIndexLocal = warpIndex * atomsPerWarp + atomWarpIndex;
214 /* Spline contribution index in one dimension */
215 const int threadLocalIdXY = (threadIdx.y * blockDim.x) + threadIdx.x;
216 const int orderIndex = threadLocalIdXY / DIM;
217 /* Dimension index */
218 const int dimIndex = threadLocalIdXY % DIM;
220 /* Multi-purpose index of rvec/ivec atom data */
221 const int sharedMemoryIndex = atomIndexLocal * DIM + dimIndex;
223 float splineData[order];
225 const int localCheck = (dimIndex < DIM) && (orderIndex < 1);
227 /* we have 4 threads per atom, but can only use 3 here for the dimensions */
230 /* Indices interpolation */
234 int tableIndex, tInt;
236 assert(atomIndexLocal < DIM * atomsPerBlock);
237 /* Accessing fields in fshOffset/nXYZ/recipbox/... with dimIndex offset
238 * puts them into local memory(!) instead of accessing the constant memory directly.
239 * That's the reason for the switch, to unroll explicitly.
240 * The commented parts correspond to the 0 components of the recipbox.
245 tableIndex = kernelParams.grid.tablesOffsets[XX];
246 n = kernelParams.grid.realGridSizeFP[XX];
247 t = atomX.x * kernelParams.current.recipBox[dimIndex][XX]
248 + atomX.y * kernelParams.current.recipBox[dimIndex][YY]
249 + atomX.z * kernelParams.current.recipBox[dimIndex][ZZ];
253 tableIndex = kernelParams.grid.tablesOffsets[YY];
254 n = kernelParams.grid.realGridSizeFP[YY];
255 t = /*atomX.x * kernelParams.current.recipBox[dimIndex][XX] + */ atomX.y
256 * kernelParams.current.recipBox[dimIndex][YY]
257 + atomX.z * kernelParams.current.recipBox[dimIndex][ZZ];
261 tableIndex = kernelParams.grid.tablesOffsets[ZZ];
262 n = kernelParams.grid.realGridSizeFP[ZZ];
263 t = /*atomX.x * kernelParams.current.recipBox[dimIndex][XX] + atomX.y * kernelParams.current.recipBox[dimIndex][YY] + */ atomX
265 * kernelParams.current.recipBox[dimIndex][ZZ];
268 const float shift = c_pmeMaxUnitcellShift;
269 /* Fractional coordinates along box vectors, adding a positive shift to ensure t is positive for triclinic boxes */
271 tInt = static_cast<int>(t);
272 assert(sharedMemoryIndex < atomsPerBlock * DIM);
273 sm_fractCoords[sharedMemoryIndex] = t - tInt;
276 assert(tInt < c_pmeNeighborUnitcellCount * n);
278 // TODO have shared table for both parameters to share the fetch, as index is always same?
279 // TODO compare texture/LDG performance
280 sm_fractCoords[sharedMemoryIndex] += fetchFromParamLookupTable(
281 kernelParams.grid.d_fractShiftsTable, kernelParams.fractShiftsTableTexture, tableIndex);
282 sm_gridlineIndices[sharedMemoryIndex] =
283 fetchFromParamLookupTable(kernelParams.grid.d_gridlineIndicesTable,
284 kernelParams.gridlineIndicesTableTexture,
288 gm_gridlineIndices[atomIndexOffset * DIM + sharedMemoryIndex] =
289 sm_gridlineIndices[sharedMemoryIndex];
293 /* B-spline calculation */
295 const int chargeCheck = pme_gpu_check_atom_charge(atomCharge);
299 int o = orderIndex; // This is an index that is set once for PME_GPU_PARALLEL_SPLINE == 1
301 const float dr = sm_fractCoords[sharedMemoryIndex];
302 assert(isfinite(dr));
304 /* dr is relative offset from lower cell limit */
305 splineData[order - 1] = 0.0F;
307 splineData[0] = 1.0F - dr;
310 for (int k = 3; k < order; k++)
312 div = 1.0F / (k - 1.0F);
313 splineData[k - 1] = div * dr * splineData[k - 2];
315 for (int l = 1; l < (k - 1); l++)
317 splineData[k - l - 1] =
318 div * ((dr + l) * splineData[k - l - 2] + (k - l - dr) * splineData[k - l - 1]);
320 splineData[0] = div * (1.0F - dr) * splineData[0];
323 const int thetaIndexBase =
324 getSplineParamIndexBase<order, atomsPerWarp>(warpIndex, atomWarpIndex);
325 const int thetaGlobalOffsetBase = atomIndexOffset * DIM * order;
326 /* only calculate dtheta if we are saving it to shared or global memory */
327 if (writeSmDtheta || writeGlobal)
329 /* Differentiation and storing the spline derivatives (dtheta) */
331 for (o = 0; o < order; o++)
333 const int thetaIndex =
334 getSplineParamIndex<order, atomsPerWarp>(thetaIndexBase, dimIndex, o);
336 const float dtheta = ((o > 0) ? splineData[o - 1] : 0.0F) - splineData[o];
337 assert(isfinite(dtheta));
338 assert(thetaIndex < order * DIM * atomsPerBlock);
341 sm_dtheta[thetaIndex] = dtheta;
345 const int thetaGlobalIndex = thetaGlobalOffsetBase + thetaIndex;
346 gm_dtheta[thetaGlobalIndex] = dtheta;
351 div = 1.0F / (order - 1.0F);
352 splineData[order - 1] = div * dr * splineData[order - 2];
354 for (int k = 1; k < (order - 1); k++)
356 splineData[order - k - 1] = div
357 * ((dr + k) * splineData[order - k - 2]
358 + (order - k - dr) * splineData[order - k - 1]);
360 splineData[0] = div * (1.0F - dr) * splineData[0];
362 /* Storing the spline values (theta) */
364 for (o = 0; o < order; o++)
366 const int thetaIndex =
367 getSplineParamIndex<order, atomsPerWarp>(thetaIndexBase, dimIndex, o);
368 assert(thetaIndex < order * DIM * atomsPerBlock);
369 sm_theta[thetaIndex] = splineData[o];
370 assert(isfinite(sm_theta[thetaIndex]));
373 const int thetaGlobalIndex = thetaGlobalOffsetBase + thetaIndex;
374 gm_theta[thetaGlobalIndex] = splineData[o];