2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2018,2019,2020, 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 PME OpenCL force gathering kernel.
38 * When including this and other PME OpenCL kernel files, plenty of common
39 * constants/macros are expected to be defined (such as "order" which is PME interpolation order).
40 * For details, please see how pme_program.cl is compiled in pme_gpu_program_impl_ocl.cpp.
42 * This file's kernels specifically expect the following definitions:
44 * - atomsPerBlock which expresses how many atoms are processed by a single work group
45 * - order which is a PME interpolation order
46 * - wrapX and wrapY must evaluate to either true or false to specify whether the grid overlap
47 * in dimension X/Y is to be used
49 * \author Aleksei Iupinov <a.yupinov@gmail.com>
52 #include "pme_gpu_types.h"
53 #include "pme_gpu_calculate_splines.clh"
55 #ifndef COMPILE_GATHER_HELPERS_ONCE
56 # define COMPILE_GATHER_HELPERS_ONCE
59 * Unrolls the dynamic index accesses to the constant grid sizes to avoid local memory operations.
61 inline float read_grid_size(const float* realGridSizeFP, const int dimIndex)
65 case XX: return realGridSizeFP[XX];
66 case YY: return realGridSizeFP[YY];
67 case ZZ: return realGridSizeFP[ZZ];
68 default: assert(false); break;
74 /*! \brief Reduce the partial force contributions.
76 * FIXME: this reduction should be simplified and improved, it does 3x16 force component
77 * reduction per 16 threads so no extra shared mem should be needed for intermediates
78 * or passing results back.
80 * \param[out] sm_forces Local memory array with the output forces (rvec).
81 * \param[in] atomIndexLocal Local atom index
82 * \param[in] splineIndex Spline index
83 * \param[in] lineIndex Line index (same as threadLocalId)
84 * \param[in] realGridSizeFP Local grid size constant
85 * \param[in] fx Input force partial component X
86 * \param[in] fy Input force partial component Y
87 * \param[in] fz Input force partial component Z
88 * \param[in,out] sm_forceReduction Reduction working buffer
89 * \param[in] sm_forceTemp Convenience pointers into \p sm_forceReduction
91 inline void reduce_atom_forces(__local float* __restrict__ sm_forces,
92 const int atomIndexLocal,
93 const int splineIndex,
95 const float* realGridSizeFP,
99 __local float* __restrict__ sm_forceReduction,
100 __local float** __restrict__ sm_forceTemp)
103 // TODO: implement AMD intrinsics reduction, like with shuffles in CUDA version. #2514
105 /* Number of data components and threads for a single atom */
106 # define atomDataSize threadsPerAtom
107 // We use blockSize local memory elements to read fx, or fy, or fz, and then reduce them to fit into smemPerDim elements
108 // All those guys are defines and not consts, because they go into the local memory array size.
109 # define blockSize (atomsPerBlock * atomDataSize)
110 # define smemPerDim warp_size
111 # define smemReserved (DIM * smemPerDim)
113 const int numWarps = blockSize / smemPerDim;
114 const int minStride = max(1, atomDataSize / numWarps);
117 for (int dimIndex = 0; dimIndex < DIM; dimIndex++)
119 int elementIndex = smemReserved + lineIndex;
120 // Store input force contributions
121 sm_forceReduction[elementIndex] = (dimIndex == XX) ? fx : (dimIndex == YY) ? fy : fz;
122 # if (warp_size < 48)
123 // sync here when exec width is smaller than the size of the sm_forceReduction
124 // buffer flushed to local mem above (size 3*16) as different warps will consume
126 barrier(CLK_LOCAL_MEM_FENCE);
129 // Reduce to fit into smemPerDim (warp size)
131 for (int redStride = atomDataSize >> 1; redStride > minStride; redStride >>= 1)
133 if (splineIndex < redStride)
135 sm_forceReduction[elementIndex] += sm_forceReduction[elementIndex + redStride];
138 barrier(CLK_LOCAL_MEM_FENCE);
139 // Last iteration - packing everything to be nearby, storing convenience pointer
140 sm_forceTemp[dimIndex] = sm_forceReduction + dimIndex * smemPerDim;
141 int redStride = minStride;
142 if (splineIndex < redStride)
144 const int packedIndex = atomIndexLocal * redStride + splineIndex;
145 sm_forceTemp[dimIndex][packedIndex] =
146 sm_forceReduction[elementIndex] + sm_forceReduction[elementIndex + redStride];
149 // barrier only needed for the last iteration on hardware with >=64-wide execution (e.g. AMD)
150 # if (warp_size < 64)
151 barrier(CLK_LOCAL_MEM_FENCE);
155 # if (warp_size >= 64)
156 barrier(CLK_LOCAL_MEM_FENCE);
159 assert((blockSize / warp_size) >= DIM);
161 const int warpIndex = lineIndex / warp_size;
162 const int dimIndex = warpIndex;
164 // First 3 warps can now process 1 dimension each
167 const int sourceIndex = lineIndex % warp_size;
169 for (int redStride = minStride >> 1; redStride > 1; redStride >>= 1)
171 if (!(splineIndex & redStride))
173 sm_forceTemp[dimIndex][sourceIndex] += sm_forceTemp[dimIndex][sourceIndex + redStride];
177 const float n = read_grid_size(realGridSizeFP, dimIndex);
178 const int atomIndex = sourceIndex / minStride;
179 if (sourceIndex == minStride * atomIndex)
181 sm_forces[atomIndex * DIM + dimIndex] =
182 (sm_forceTemp[dimIndex][sourceIndex] + sm_forceTemp[dimIndex][sourceIndex + 1]) * n;
187 #endif // COMPILE_GATHER_HELPERS_ONCE
190 * An OpenCL kernel which gathers the atom forces from the grid.
191 * The grid is assumed to be wrapped in dimension Z.
192 * Please see the file description for additional defines which this kernel expects.
194 * \param[in] kernelParams All the PME GPU data.
195 * \param[in] gm_coefficients Atom charges/coefficients.
196 * \param[in] gm_grid Global 3D grid.
197 * \param[in] gm_theta Atom spline parameter values
198 * \param[in] gm_dtheta Atom spline parameter derivatives
199 * \param[in] gm_gridlineIndices Atom gridline indices (ivec)
200 * \param[in,out] gm_forces Atom forces (rvec)
202 __attribute__((reqd_work_group_size(order, order, atomsPerBlock)))
203 __kernel void CUSTOMIZED_KERNEL_NAME(pme_gather_kernel)(const struct PmeOpenCLKernelParams kernelParams,
204 __global const float* __restrict__ gm_coefficients,
205 __global const float* __restrict__ gm_grid,
206 __global const float* __restrict__ gm_theta,
207 __global const float* __restrict__ gm_dtheta,
208 __global const int* __restrict__ gm_gridlineIndices,
209 __global float* __restrict__ gm_forces)
211 /* These are the atom indices - for the shared and global memory */
212 const int atomIndexLocal = get_local_id(ZZ);
213 const int atomIndexOffset = (int)get_group_id(XX) * atomsPerBlock;
214 const int atomIndexGlobal = atomIndexOffset + atomIndexLocal;
216 /* Some sizes which are defines and not consts because they go into the array size */
217 #define blockSize (atomsPerBlock * atomDataSize)
218 assert(blockSize == (get_local_size(0) * get_local_size(1) * get_local_size(2)));
219 #define smemPerDim warp_size
220 #define smemReserved (DIM * smemPerDim)
221 #define totalSharedMemory (smemReserved + blockSize)
222 #define gridlineIndicesSize (atomsPerBlock * DIM)
223 #define splineParamsSize (atomsPerBlock * DIM * order)
225 __local int sm_gridlineIndices[gridlineIndicesSize];
226 __local float2 sm_splineParams[splineParamsSize]; /* Theta/dtheta pairs as .x/.y */
228 /* Spline Y/Z coordinates */
229 const int ithy = get_local_id(YY);
230 const int ithz = get_local_id(XX);
232 assert((get_local_id(2) * get_local_size(1) + get_local_id(1)) * get_local_size(0) + get_local_id(0)
234 const int threadLocalId =
235 (int)((get_local_id(2) * get_local_size(1) + get_local_id(1)) * get_local_size(0)
238 /* These are the spline contribution indices in shared memory */
239 assert((get_local_id(1) * get_local_size(0) + get_local_id(0)) <= MAX_INT);
240 const int splineIndex =
241 (int)(get_local_id(1) * get_local_size(0)
242 + get_local_id(0)); /* Relative to the current particle , 0..15 for order 4 */
243 const int lineIndex = threadLocalId; /* And to all the block's particles */
245 /* Staging the atom gridline indices, DIM * atomsPerBlock threads */
246 const int localGridlineIndicesIndex = threadLocalId;
247 const int globalGridlineIndicesIndex =
248 (int)get_group_id(XX) * gridlineIndicesSize + localGridlineIndicesIndex;
249 if (localGridlineIndicesIndex < gridlineIndicesSize)
251 sm_gridlineIndices[localGridlineIndicesIndex] = gm_gridlineIndices[globalGridlineIndicesIndex];
252 assert(sm_gridlineIndices[localGridlineIndicesIndex] >= 0);
254 /* Staging the spline parameters, DIM * order * atomsPerBlock threads */
255 const int localSplineParamsIndex = threadLocalId;
256 const int globalSplineParamsIndex = (int)get_group_id(XX) * splineParamsSize + localSplineParamsIndex;
257 if (localSplineParamsIndex < splineParamsSize)
259 sm_splineParams[localSplineParamsIndex].x = gm_theta[globalSplineParamsIndex];
260 sm_splineParams[localSplineParamsIndex].y = gm_dtheta[globalSplineParamsIndex];
261 assert(isfinite(sm_splineParams[localSplineParamsIndex].x));
262 assert(isfinite(sm_splineParams[localSplineParamsIndex].y));
264 barrier(CLK_LOCAL_MEM_FENCE);
270 const int chargeCheck = pme_gpu_check_atom_charge(gm_coefficients[atomIndexGlobal]);
274 const int nx = kernelParams.grid.realGridSize[XX];
275 const int ny = kernelParams.grid.realGridSize[YY];
276 const int nz = kernelParams.grid.realGridSize[ZZ];
277 const int pny = kernelParams.grid.realGridSizePadded[YY];
278 const int pnz = kernelParams.grid.realGridSizePadded[ZZ];
280 const int atomWarpIndex = atomIndexLocal % atomsPerWarp;
281 const int warpIndex = atomIndexLocal / atomsPerWarp;
283 const int splineIndexBase = getSplineParamIndexBase(warpIndex, atomWarpIndex);
284 const int splineIndexY = getSplineParamIndex(splineIndexBase, YY, ithy);
285 const float2 tdy = sm_splineParams[splineIndexY];
286 const int splineIndexZ = getSplineParamIndex(splineIndexBase, ZZ, ithz);
287 const float2 tdz = sm_splineParams[splineIndexZ];
289 const int ixBase = sm_gridlineIndices[atomIndexLocal * DIM + XX];
290 int iy = sm_gridlineIndices[atomIndexLocal * DIM + YY] + ithy;
291 if (wrapY & (iy >= ny))
295 int iz = sm_gridlineIndices[atomIndexLocal * DIM + ZZ] + ithz;
300 const int constOffset = iy * pnz + iz;
303 for (int ithx = 0; (ithx < order); ithx++)
305 int ix = ixBase + ithx;
306 if (wrapX & (ix >= nx))
310 const int gridIndexGlobal = ix * pny * pnz + constOffset;
311 assert(gridIndexGlobal >= 0);
312 const float gridValue = gm_grid[gridIndexGlobal];
313 assert(isfinite(gridValue));
314 const int splineIndexX = getSplineParamIndex(splineIndexBase, XX, ithx);
315 const float2 tdx = sm_splineParams[splineIndexX];
316 const float fxy1 = tdz.x * gridValue;
317 const float fz1 = tdz.y * gridValue;
318 fx += tdx.y * tdy.x * fxy1;
319 fy += tdx.x * tdy.y * fxy1;
320 fz += tdx.x * tdy.x * fz1;
324 // Reduction of partial force contributions
325 __local float sm_forces[atomsPerBlock * DIM];
327 __local float sm_forceReduction[totalSharedMemory];
328 __local float* sm_forceTemp[DIM];
330 reduce_atom_forces(sm_forces, atomIndexLocal, splineIndex, lineIndex,
331 kernelParams.grid.realGridSizeFP, fx, fy, fz, sm_forceReduction, sm_forceTemp);
332 barrier(CLK_LOCAL_MEM_FENCE);
334 /* Calculating the final forces with no component branching, atomsPerBlock threads */
335 const int forceIndexLocal = threadLocalId;
336 const int forceIndexGlobal = atomIndexOffset + forceIndexLocal;
337 if (forceIndexLocal < atomsPerBlock)
339 const float3 atomForces = vload3(forceIndexLocal, sm_forces);
340 const float negCoefficient = -gm_coefficients[forceIndexGlobal];
342 result.x = negCoefficient * kernelParams.current.recipBox[XX][XX] * atomForces.x;
343 result.y = negCoefficient
344 * (kernelParams.current.recipBox[XX][YY] * atomForces.x
345 + kernelParams.current.recipBox[YY][YY] * atomForces.y);
346 result.z = negCoefficient
347 * (kernelParams.current.recipBox[XX][ZZ] * atomForces.x
348 + kernelParams.current.recipBox[YY][ZZ] * atomForces.y
349 + kernelParams.current.recipBox[ZZ][ZZ] * atomForces.z);
350 vstore3(result, forceIndexLocal, sm_forces);
353 #if !defined(_AMD_SOURCE_) && !defined(_NVIDIA_SOURCE_)
354 /* This is only here for execution of e.g. 32-sized warps on 16-wide hardware; this was
355 * __syncwarp() in CUDA. #2519
357 barrier(CLK_LOCAL_MEM_FENCE);
360 assert(atomsPerBlock <= warp_size);
362 /* Writing or adding the final forces component-wise, single warp */
363 const int blockForcesSize = atomsPerBlock * DIM;
364 const int numIter = (blockForcesSize + warp_size - 1) / warp_size;
365 const int iterThreads = blockForcesSize / numIter;
366 if (threadLocalId < iterThreads)
369 for (int i = 0; i < numIter; i++)
371 const int outputIndexLocal = i * iterThreads + threadLocalId;
372 const int outputIndexGlobal = (int)get_group_id(XX) * blockForcesSize + outputIndexLocal;
373 const float outputForceComponent = sm_forces[outputIndexLocal];
374 gm_forces[outputIndexGlobal] = outputForceComponent;