Files
vortex/benchmarks/old_opencl/cutcp/cutoff.c
2020-04-14 06:35:20 -04:00

500 lines
16 KiB
C

/***************************************************************************
*cr
*cr (C) Copyright 2008-2010 The Board of Trustees of the
*cr University of Illinois
*cr All Rights Reserved
*cr
***************************************************************************/
#include <CL/cl.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <math.h>
#include <parboil.h>
#include "atom.h"
#include "cutoff.h"
#include "macros.h"
#include "ocl.h"
// OpenCL 1.1 support for int3 is not uniform on all implementations, so
// we use int4 instead. Only the 'x', 'y', and 'z' fields of xyz are used.
typedef cl_int4 xyz;
//extern "C" int gpu_compute_cutoff_potential_lattice(
int gpu_compute_cutoff_potential_lattice(
struct pb_TimerSet *timers,
Lattice *lattice, /* the lattice */
float cutoff, /* cutoff distance */
Atoms *atoms, /* array of atoms */
int verbose, /* print info/debug messages */
struct pb_Parameters *parameters
)
{
int nx = lattice->dim.nx;
int ny = lattice->dim.ny;
int nz = lattice->dim.nz;
float xlo = lattice->dim.lo.x;
float ylo = lattice->dim.lo.y;
float zlo = lattice->dim.lo.z;
float h = lattice->dim.h;
int natoms = atoms->size;
Atom *atom = atoms->atoms;
xyz nbrlist[NBRLIST_MAXLEN];
int nbrlistlen = 0;
int binHistoFull[BIN_DEPTH+1] = { 0 }; /* clear every array element */
int binHistoCover[BIN_DEPTH+1] = { 0 }; /* clear every array element */
int num_excluded = 0;
int xRegionDim, yRegionDim, zRegionDim;
int xRegionIndex, yRegionIndex, zRegionIndex;
int xOffset, yOffset, zOffset;
int lnx, lny, lnz, lnall;
float *regionZeroAddr, *thisRegion;
cl_mem regionZeroCl;
int index, indexRegion;
int c;
xyz binDim;
int nbins;
cl_float4 *binBaseAddr, *binZeroAddr;
cl_mem binBaseCl, binZeroCl;
int *bincntBaseAddr, *bincntZeroAddr;
Atoms *extra = NULL;
cl_mem NbrListLen;
cl_mem NbrList;
int i, j, k, n;
int sum, total;
float avgFillFull, avgFillCover;
const float cutoff2 = cutoff * cutoff;
const float inv_cutoff2 = 1.f / cutoff2;
size_t gridDim[3], blockDim[3];
// The "compute" timer should be active upon entry to this function
/* pad lattice to be factor of 8 in each dimension */
xRegionDim = (int) ceilf(nx/8.f);
yRegionDim = (int) ceilf(ny/8.f);
zRegionDim = (int) ceilf(nz/8.f);
lnx = 8 * xRegionDim;
lny = 8 * yRegionDim;
lnz = 8 * zRegionDim;
lnall = lnx * lny * lnz;
/* will receive energies from OpenCL */
regionZeroAddr = (float *) malloc(lnall * sizeof(float));
/* create bins */
c = (int) ceil(cutoff * BIN_INVLEN); /* count extra bins around lattice */
binDim.x = (int) ceil(lnx * h * BIN_INVLEN) + 2*c;
binDim.y = (int) ceil(lny * h * BIN_INVLEN) + 2*c;
binDim.z = (int) ceil(lnz * h * BIN_INVLEN) + 2*c;
nbins = binDim.x * binDim.y * binDim.z;
binBaseAddr = (cl_float4 *) calloc(nbins * BIN_DEPTH, sizeof(cl_float4));
binZeroAddr = binBaseAddr + ((c * binDim.y + c) * binDim.x + c) * BIN_DEPTH;
bincntBaseAddr = (int *) calloc(nbins, sizeof(int));
bincntZeroAddr = bincntBaseAddr + (c * binDim.y + c) * binDim.x + c;
/* create neighbor list */
if (ceilf(BIN_LENGTH / (8*h)) == floorf(BIN_LENGTH / (8*h))) {
float s = sqrtf(3);
float r2 = (cutoff + s*BIN_LENGTH) * (cutoff + s*BIN_LENGTH);
int cnt = 0;
/* develop neighbor list around 1 cell */
if (2*c + 1 > NBRLIST_DIM) {
fprintf(stderr, "must have cutoff <= %f\n",
(NBRLIST_DIM-1)/2 * BIN_LENGTH);
return -1;
}
for (k = -c; k <= c; k++) {
for (j = -c; j <= c; j++) {
for (i = -c; i <= c; i++) {
if ((i*i + j*j + k*k)*BIN_LENGTH*BIN_LENGTH >= r2) continue;
nbrlist[cnt].x = i;
nbrlist[cnt].y = j;
nbrlist[cnt].z = k;
cnt++;
}
}
}
nbrlistlen = cnt;
}
else if (8*h <= 2*BIN_LENGTH) {
float s = 2.f*sqrtf(3);
float r2 = (cutoff + s*BIN_LENGTH) * (cutoff + s*BIN_LENGTH);
int cnt = 0;
/* develop neighbor list around 3-cube of cells */
if (2*c + 3 > NBRLIST_DIM) {
fprintf(stderr, "must have cutoff <= %f\n",
(NBRLIST_DIM-3)/2 * BIN_LENGTH);
return -1;
}
for (k = -c; k <= c; k++) {
for (j = -c; j <= c; j++) {
for (i = -c; i <= c; i++) {
if ((i*i + j*j + k*k)*BIN_LENGTH*BIN_LENGTH >= r2) continue;
nbrlist[cnt].x = i;
nbrlist[cnt].y = j;
nbrlist[cnt].z = k;
cnt++;
}
}
}
nbrlistlen = cnt;
}
else {
fprintf(stderr, "must have h <= %f\n", 0.25 * BIN_LENGTH);
return -1;
}
/* perform geometric hashing of atoms into bins */
{
/* array of extra atoms, permit average of one extra per bin */
Atom *extra_atoms = (Atom *) calloc(nbins, sizeof(Atom));
int extra_len = 0;
for (n = 0; n < natoms; n++) {
cl_float4 p;
p.x = atom[n].x - xlo;
p.y = atom[n].y - ylo;
p.z = atom[n].z - zlo;
p.w = atom[n].q;
i = (int) floorf(p.x * BIN_INVLEN);
j = (int) floorf(p.y * BIN_INVLEN);
k = (int) floorf(p.z * BIN_INVLEN);
if (i >= -c && i < binDim.x - c &&
j >= -c && j < binDim.y - c &&
k >= -c && k < binDim.z - c &&
atom[n].q != 0) {
int index = (k * binDim.y + j) * binDim.x + i;
cl_float4 *bin = binZeroAddr + index * BIN_DEPTH;
int bindex = bincntZeroAddr[index];
if (bindex < BIN_DEPTH) {
/* copy atom into bin and increase counter for this bin */
bin[bindex] = p;
bincntZeroAddr[index]++;
}
else {
/* add index to array of extra atoms to be computed with CPU */
if (extra_len >= nbins) {
fprintf(stderr, "exceeded space for storing extra atoms\n");
return -1;
}
extra_atoms[extra_len] = atom[n];
extra_len++;
}
}
else {
/* excluded atoms are either outside bins or neutrally charged */
num_excluded++;
}
}
/* Save result */
extra = (Atoms *)malloc(sizeof(Atoms));
extra->atoms = extra_atoms;
extra->size = extra_len;
}
/* bin stats */
sum = total = 0;
for (n = 0; n < nbins; n++) {
binHistoFull[ bincntBaseAddr[n] ]++;
sum += bincntBaseAddr[n];
total += BIN_DEPTH;
}
avgFillFull = sum / (float) total;
sum = total = 0;
for (k = 0; k < binDim.z - 2*c; k++) {
for (j = 0; j < binDim.y - 2*c; j++) {
for (i = 0; i < binDim.x - 2*c; i++) {
int index = (k * binDim.y + j) * binDim.x + i;
binHistoCover[ bincntZeroAddr[index] ]++;
sum += bincntZeroAddr[index];
total += BIN_DEPTH;
}
}
}
avgFillCover = sum / (float) total;
if (verbose) {
/* report */
printf("number of atoms = %d\n", natoms);
printf("lattice spacing = %g\n", h);
printf("cutoff distance = %g\n", cutoff);
printf("\n");
printf("requested lattice dimensions = %d %d %d\n", nx, ny, nz);
printf("requested space dimensions = %g %g %g\n", nx*h, ny*h, nz*h);
printf("expanded lattice dimensions = %d %d %d\n", lnx, lny, lnz);
printf("expanded space dimensions = %g %g %g\n", lnx*h, lny*h, lnz*h);
printf("number of bytes for lattice data = %u\n", (unsigned int) (lnall*sizeof(float)));
printf("\n");
printf("bin padding thickness = %d\n", c);
printf("bin cover dimensions = %d %d %d\n",
binDim.x - 2*c, binDim.y - 2*c, binDim.z - 2*c);
printf("bin full dimensions = %d %d %d\n", binDim.x, binDim.y, binDim.z);
printf("number of bins = %d\n", nbins);
printf("total number of atom slots = %d\n", nbins * BIN_DEPTH);
printf("%% overhead space = %g\n",
(natoms / (double) (nbins * BIN_DEPTH)) * 100);
printf("number of bytes for bin data = %u\n",
(unsigned int)(nbins * BIN_DEPTH * sizeof(cl_float4)));
printf("\n");
printf("bin histogram with padding:\n");
sum = 0;
for (n = 0; n <= BIN_DEPTH; n++) {
printf(" number of bins with %d atoms: %d\n", n, binHistoFull[n]);
sum += binHistoFull[n];
}
printf(" total number of bins: %d\n", sum);
printf(" %% average fill: %g\n", avgFillFull * 100);
printf("\n");
printf("bin histogram excluding padding:\n");
sum = 0;
for (n = 0; n <= BIN_DEPTH; n++) {
printf(" number of bins with %d atoms: %d\n", n, binHistoCover[n]);
sum += binHistoCover[n];
}
printf(" total number of bins: %d\n", sum);
printf(" %% average fill: %g\n", avgFillCover * 100);
printf("\n");
printf("number of extra atoms = %d\n", extra->size);
printf("%% atoms that are extra = %g\n", (extra->size / (double) natoms) * 100);
printf("\n");
/* sanity check on bins */
sum = 0;
for (n = 0; n <= BIN_DEPTH; n++) {
sum += n * binHistoFull[n];
}
sum += extra->size + num_excluded;
printf("sanity check on bin histogram with edges: "
"sum + others = %d\n", sum);
sum = 0;
for (n = 0; n <= BIN_DEPTH; n++) {
sum += n * binHistoCover[n];
}
sum += extra->size + num_excluded;
printf("sanity check on bin histogram excluding edges: "
"sum + others = %d\n", sum);
printf("\n");
/* neighbor list */
printf("neighbor list length = %d\n", nbrlistlen);
printf("\n");
}
printf("Ok!\n");
pb_Context* pb_context;
pb_context = pb_InitOpenCLContext(parameters);
if (pb_context == NULL) {
fprintf (stderr, "Error: No OpenCL platform/device can be found.");
return -1;
}
printf("Ok!\n");
cl_int clStatus;
cl_device_id clDevice = (cl_device_id) pb_context->clDeviceId;
cl_platform_id clPlatform = (cl_platform_id) pb_context->clPlatformId;
cl_context clContext = (cl_context) pb_context->clContext;
cl_command_queue clCommandQueue = clCreateCommandQueue(clContext,clDevice,CL_QUEUE_PROFILING_ENABLE,&clStatus);
CHECK_ERROR("clCreateCommandQueue")
pb_SetOpenCL(&clContext, &clCommandQueue);
//const char* clSource[] = {readFile("src/opencl_base/kernel.cl")};
//cl_program clProgram = clCreateProgramWithSource(clContext,1,clSource,NULL,&clStatus);
cl_program clProgram = clCreateProgramWithBuiltInKernels(
clContext, 1, &clDevice, "opencl_cutoff_potential_lattice", &clStatus);
CHECK_ERROR("clCreateProgramWithSource")
char clOptions[50];
sprintf(clOptions,"-I src/opencl_base"); //-cl-nv-verbose
clStatus = clBuildProgram(clProgram,1,&clDevice,clOptions,NULL,NULL);
if (clStatus != CL_SUCCESS) {
size_t string_size = 0;
clGetProgramBuildInfo(clProgram, clDevice, CL_PROGRAM_BUILD_LOG,
0, NULL, &string_size);
char* string = (char*)malloc(string_size*sizeof(char));
clGetProgramBuildInfo(clProgram, clDevice, CL_PROGRAM_BUILD_LOG,
string_size, string, NULL);
puts(string);
}
CHECK_ERROR("clBuildProgram")
cl_kernel clKernel = clCreateKernel(clProgram,"opencl_cutoff_potential_lattice",&clStatus);
CHECK_ERROR("clCreateKernel")
/* setup OpenCL kernel parameters */
blockDim[0] = 8;
blockDim[1] = 8;
blockDim[2] = 2;
gridDim[0] = 4 * xRegionDim * blockDim[0];
gridDim[1] = yRegionDim * blockDim[1];
gridDim[2] = 1 * blockDim[2];
/* allocate and initialize memory on OpenCL device */
pb_SwitchToTimer(timers, pb_TimerID_COPY);
if (verbose) {
printf("Allocating %.2fMB on OpenCL device for potentials\n",
lnall * sizeof(float) / (double) (1024*1024));
}
regionZeroCl = clCreateBuffer(clContext,CL_MEM_WRITE_ONLY,lnall*sizeof(float),NULL,&clStatus);
CHECK_ERROR("clCreateBuffer")
// clMemSet(clCommandQueue,regionZeroCl,0,lnall*sizeof(float));
if (verbose) {
printf("Allocating %.2fMB on OpenCL device for atom bins\n",
nbins * BIN_DEPTH * sizeof(cl_float4) / (double) (1024*1024));
}
binBaseCl = clCreateBuffer(clContext,CL_MEM_READ_ONLY,nbins*BIN_DEPTH*sizeof(cl_float4),NULL,&clStatus);
CHECK_ERROR("clCreateBuffer")
clStatus = clEnqueueWriteBuffer(clCommandQueue,binBaseCl,CL_TRUE,0,nbins*BIN_DEPTH*sizeof(cl_float4),binBaseAddr,0,NULL,NULL);
CHECK_ERROR("clEnqueueWriteBuffer")
//Sub buffers are not supported in OpenCL v1.0
int offset = ((c * binDim.y + c) * binDim.x + c) * BIN_DEPTH;
NbrListLen = clCreateBuffer(clContext,CL_MEM_READ_ONLY,sizeof(int),NULL,&clStatus);
CHECK_ERROR("clCreateBuffer")
clStatus = clEnqueueWriteBuffer(clCommandQueue,NbrListLen,CL_TRUE,0,sizeof(int),&nbrlistlen,0,NULL,NULL);
CHECK_ERROR("clEnqueueWriteBuffer")
NbrList = clCreateBuffer(clContext,CL_MEM_READ_ONLY,NBRLIST_MAXLEN*sizeof(xyz),NULL,&clStatus);
CHECK_ERROR("clCreateBuffer")
clStatus = clEnqueueWriteBuffer(clCommandQueue,NbrList,CL_TRUE,0,nbrlistlen*sizeof(xyz),nbrlist,0,NULL,NULL);
CHECK_ERROR("clEnqueueWriteBuffer")
if (verbose)
printf("\n");
clStatus = clSetKernelArg(clKernel,0,sizeof(int),&(binDim.x));
clStatus = clSetKernelArg(clKernel,1,sizeof(int),&(binDim.y));
clStatus = clSetKernelArg(clKernel,2,sizeof(cl_mem),&binBaseCl);
clStatus = clSetKernelArg(clKernel,3,sizeof(int),&offset);
clStatus = clSetKernelArg(clKernel,4,sizeof(float),&h);
clStatus = clSetKernelArg(clKernel,5,sizeof(float),&cutoff2);
clStatus = clSetKernelArg(clKernel,6,sizeof(float),&inv_cutoff2);
clStatus = clSetKernelArg(clKernel,7,sizeof(cl_mem),&regionZeroCl);
clStatus = clSetKernelArg(clKernel,9,sizeof(cl_mem),&NbrListLen);
clStatus = clSetKernelArg(clKernel,10,sizeof(cl_mem),&NbrList);
CHECK_ERROR("clSetKernelArg")
printf("Ok!!\n");
/* loop over z-dimension, invoke OpenCL kernel for each x-y plane */
pb_SwitchToTimer(timers, pb_TimerID_KERNEL);
printf("Invoking OpenCL kernel on %d region planes...\n", zRegionDim);
for (zRegionIndex = 0; zRegionIndex < zRegionDim; zRegionIndex++) {
printf(" computing plane %d\r", zRegionIndex);
fflush(stdout);
clStatus = clSetKernelArg(clKernel,8,sizeof(int),&zRegionIndex);
CHECK_ERROR("clSetKernelArg")
printf("Ok**!2\n");
clStatus = clEnqueueNDRangeKernel(clCommandQueue,clKernel,3,NULL,gridDim,blockDim,0,NULL,NULL);
printf("Ok**!2\n");
CHECK_ERROR("clEnqueueNDRangeKernel")
printf("Ok**!2\n");
clStatus = clFinish(clCommandQueue);
printf("Ok**!2\n");
CHECK_ERROR("clFinish")
}
printf("Ok++!\n");
printf("Finished OpenCL kernel calls \n");
/* copy result regions from OpenCL device */
pb_SwitchToTimer(timers, pb_TimerID_COPY);
clStatus = clEnqueueReadBuffer(clCommandQueue,regionZeroCl,CL_TRUE,0,lnall*sizeof(float),regionZeroAddr,0,NULL,NULL);
CHECK_ERROR("clEnqueueReadBuffer")
/* free OpenCL memory allocations */
clStatus = clReleaseMemObject(regionZeroCl);
clStatus = clReleaseMemObject(binBaseCl);
clStatus = clReleaseMemObject(NbrListLen);
clStatus = clReleaseMemObject(NbrList);
CHECK_ERROR("clReleaseMemObject")
clStatus = clReleaseKernel(clKernel);
clStatus = clReleaseProgram(clProgram);
clStatus = clReleaseCommandQueue(clCommandQueue);
clStatus = clReleaseContext(clContext);
//free((void*)clSource[0]);
/* transpose regions back into lattice */
pb_SwitchToTimer(timers, pb_TimerID_COMPUTE);
for (k = 0; k < nz; k++) {
zRegionIndex = (k >> 3);
zOffset = (k & 7);
for (j = 0; j < ny; j++) {
yRegionIndex = (j >> 3);
yOffset = (j & 7);
for (i = 0; i < nx; i++) {
xRegionIndex = (i >> 3);
xOffset = (i & 7);
thisRegion = regionZeroAddr
+ ((zRegionIndex * yRegionDim + yRegionIndex) * xRegionDim
+ xRegionIndex) * REGION_SIZE;
indexRegion = (zOffset * 8 + yOffset) * 8 + xOffset;
index = (k * ny + j) * nx + i;
lattice->lattice[index] = thisRegion[indexRegion];
}
}
}
/* handle extra atoms */
if (extra->size > 0) {
printf("computing extra atoms on CPU\n");
if (cpu_compute_cutoff_potential_lattice(lattice, cutoff, extra)) {
fprintf(stderr, "cpu_compute_cutoff_potential_lattice() failed "
"for extra atoms\n");
return -1;
}
printf("\n");
}
/* cleanup memory allocations */
free(regionZeroAddr);
free(binBaseAddr);
free(bincntBaseAddr);
free_atom(extra);
return 0;
}