// This file is part of BOINC.
// http://boinc.berkeley.edu
// Copyright (C) 2008 University of California
//
// BOINC is free software; you can redistribute it and/or modify it
// under the terms of the GNU Lesser General Public License
// as published by the Free Software Foundation,
// either version 3 of the License, or (at your option) any later version.
//
// BOINC is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
// See the GNU Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public License
// along with BOINC. If not, see .
#ifndef _USING_FCGI_
#include "boinc_fcgi.h"
#else
#include
#endif
#include
#include
#ifdef _WIN32
#ifndef SIM
#include
#endif
#else
#ifdef __APPLE__
// Suppress obsolete warning when building for OS 10.3.9
#define DLOPEN_NO_WARN
#endif
#include
#endif
#include "error_numbers.h"
#include "filesys.h"
#include "parse.h"
#include "str_util.h"
#include "coproc.h"
using std::string;
using std::vector;
#ifndef _USING_FCGI_
using std::perror;
#endif
#ifndef _USING_FCGI_
void COPROC::write_xml(MIOFILE& f) {
f.printf(
"\n"
" %s\n"
" %d\n"
"\n",
type, count
);
}
#endif
int COPROC::parse(MIOFILE& fin) {
char buf[1024];
strcpy(type, "");
count = 0;
used = 0;
req_secs = 0;
estimated_delay = 0;
req_instances = 0;
while (fin.fgets(buf, sizeof(buf))) {
if (match_tag(buf, "")) {
if (!strlen(type)) return ERR_XML_PARSE;
return 0;
}
if (parse_str(buf, "", type, sizeof(type))) continue;
if (parse_int(buf, "", count)) continue;
if (parse_double(buf, "", req_secs)) continue;
if (parse_int(buf, "", req_instances)) continue;
if (parse_double(buf, "", estimated_delay)) continue;
}
return ERR_XML_PARSE;
}
void COPROCS::summary_string(char* buf, int len) {
char bigbuf[8192], buf2[1024];
strcpy(bigbuf, "");
for (unsigned int i=0; itype, "CUDA")) {
COPROC_CUDA* cp2 = (COPROC_CUDA*) cp;
int mem = (int)(cp2->prop.dtotalGlobalMem/MEGA);
sprintf(buf2, "[CUDA|%s|%d|%dMB|%d]",
cp2->prop.name, cp2->count, mem, cp2->drvVersion
);
strcat(bigbuf, buf2);
}
}
bigbuf[len-1] = 0;
strcpy(buf, bigbuf);
}
vector COPROCS::get() {
vector strings;
COPROC_CUDA::get(*this, strings);
COPROC_CELL_SPE::get(*this, strings);
return strings;
}
// used only to parse scheduler request messages
//
int COPROCS::parse(FILE* fin) {
char buf[1024];
while (fgets(buf, sizeof(buf), fin)) {
if (match_tag(buf, "")) {
return 0;
}
if (strstr(buf, "")) {
COPROC_CUDA* cc = new COPROC_CUDA;
int retval = cc->parse(fin);
if (!retval) {
coprocs.push_back(cc);
}
}
}
return ERR_XML_PARSE;
}
COPROC* COPROCS::lookup(const char* type) {
for (unsigned int i=0; itype)) return cp;
}
return NULL;
}
#ifdef _WIN32
#endif
// return 1/-1/0 if device 1 is more/less/same capable than device 2
//
int cuda_compare(COPROC_CUDA& c1, COPROC_CUDA& c2) {
if (c1.prop.major > c2.prop.major) return 1;
if (c1.prop.major < c2.prop.major) return -1;
if (c1.prop.minor > c2.prop.minor) return 1;
if (c1.prop.minor < c2.prop.minor) return -1;
if (c1.drvVersion > c2.drvVersion) return 1;
if (c1.drvVersion < c2.drvVersion) return -1;
if (c1.prop.totalGlobalMem > c2.prop.totalGlobalMem) return 1;
if (c1.prop.totalGlobalMem < c2.prop.totalGlobalMem) return -1;
double s1 = c1.flops_estimate();
double s2 = c1.flops_estimate();
if (s1 > s2) return 1;
if (s1 < s2) return -1;
return 0;
}
void COPROC_CUDA::get(COPROCS& coprocs, vector& strings) {
int count;
#ifdef _WIN32
typedef int (__stdcall *PCGDC)(int *count);
typedef int (__stdcall *PCGDP)(struct cudaDeviceProp *prop, int device);
PCGDC __cudaGetDeviceCount = NULL;
PCGDP __cudaGetDeviceProperties = NULL;
HMODULE cudalib = LoadLibrary("cudart.dll");
if (!cudalib) {
strings.push_back("Can't load library cudart.dll");
return;
}
__cudaGetDeviceCount = (PCGDC)GetProcAddress( cudalib, "cudaGetDeviceCount" );
if (!__cudaGetDeviceCount) {
strings.push_back("Library doesn't have cudaGetDeviceCount()");
return;
}
__cudaGetDeviceProperties = (PCGDP)GetProcAddress( cudalib, "cudaGetDeviceProperties" );
if (!__cudaGetDeviceProperties) {
strings.push_back("Library doesn't have cudaGetDeviceProperties()");
return;
}
#ifndef SIM
NvAPI_Status nvapiStatus;
NvDisplayHandle hDisplay;
NV_DISPLAY_DRIVER_VERSION Version;
memset(&Version, 0, sizeof(Version));
Version.version = NV_DISPLAY_DRIVER_VERSION_VER;
NvAPI_Initialize();
for (int i=0; ; i++) {
nvapiStatus = NvAPI_EnumNvidiaDisplayHandle(i, &hDisplay);
if (nvapiStatus != NVAPI_OK) break;
nvapiStatus = NvAPI_GetDisplayDriverVersion(hDisplay, &Version);
if (nvapiStatus == NVAPI_OK) break;
}
#endif
#else
void* cudalib;
void (*__cudaGetDeviceCount)(int*);
void (*__cudaGetDeviceProperties)(cudaDeviceProp*, int);
#ifdef __APPLE__
cudalib = dlopen("/usr/local/cuda/lib/libcudart.dylib", RTLD_NOW);
#else
// libcudart.so is included with the BOINC install for linux,
// so look for it in the current dir.
//
cudalib = dlopen("./libcudart.so", RTLD_NOW);
if (!cudalib) {
// If that fails, look for it in the library search path
//
cudalib = dlopen("libcudart.so", RTLD_NOW);
}
#endif
if (!cudalib) {
strings.push_back("Can't load library libcudart");
return;
}
__cudaGetDeviceCount = (void(*)(int*)) dlsym(cudalib, "cudaGetDeviceCount");
if(!__cudaGetDeviceCount) {
strings.push_back("Library doesn't have cudaGetDeviceCount()");
return;
}
__cudaGetDeviceProperties = (void(*)(cudaDeviceProp*, int)) dlsym( cudalib, "cudaGetDeviceProperties" );
if (!__cudaGetDeviceProperties) {
strings.push_back("Library doesn't have cudaGetDeviceProperties()");
return;
}
#endif
vector gpus;
(*__cudaGetDeviceCount)(&count);
int j;
unsigned int i;
COPROC_CUDA cc;
string s;
for (j=0; j 100) continue; // e.g. 9999 is an error
#if defined(_WIN32) && !defined(SIM)
cc.drvVersion = Version.drvVersion;
#else
cc.drvVersion = 0;
#endif
cc.device_num = j;
gpus.push_back(cc);
}
if (!gpus.size()) {
strings.push_back("No CUDA devices found");
return;
}
// identify the most capable instance
//
COPROC_CUDA best;
for (i=0; i 0) {
best = gpus[i];
}
}
// see which other instances are equivalent,
// and set the "count" and "device_nums" fields
//
best.count = 0;
for (i=0; itype, "CUDA");
cc->count = count;
strcpy(cc->prop.name, "CUDA NVIDIA chip");
cc->prop.totalGlobalMem = 256*1024*1024;
cc->prop.sharedMemPerBlock = 100;
cc->prop.regsPerBlock = 8;
cc->prop.warpSize = 10;
cc->prop.memPitch = 10;
cc->prop.maxThreadsPerBlock = 20;
cc->prop.maxThreadsDim[0] = 2;
cc->prop.maxThreadsDim[1] = 2;
cc->prop.maxThreadsDim[2] = 2;
cc->prop.maxGridSize[0] = 10;
cc->prop.maxGridSize[1] = 10;
cc->prop.maxGridSize[2] = 10;
cc->prop.totalConstMem = 10;
cc->prop.major = 1;
cc->prop.minor = 2;
cc->prop.clockRate = 1250000;
cc->prop.textureAlignment = 1000;
cc->prop.multiProcessorCount = 14;
coprocs.coprocs.push_back(cc);
}
#ifndef _USING_FCGI_
void COPROC_CUDA::write_xml(MIOFILE& f) {
f.printf(
"\n"
" %d\n"
" %s\n"
" %f\n"
" %d\n"
" %f\n"
" %d\n"
" %u\n"
" %u\n"
" %d\n"
" %d\n"
" %u\n"
" %d\n"
" %d %d %d\n"
" %d %d %d\n"
" %u\n"
" %d\n"
" %d\n"
" %d\n"
" %u\n"
" %d\n"
" %d\n"
"\n",
count,
prop.name,
req_secs,
req_instances,
estimated_delay,
drvVersion,
(unsigned int)prop.totalGlobalMem,
(unsigned int)prop.sharedMemPerBlock,
prop.regsPerBlock,
prop.warpSize,
(unsigned int)prop.memPitch,
prop.maxThreadsPerBlock,
prop.maxThreadsDim[0], prop.maxThreadsDim[1], prop.maxThreadsDim[2],
prop.maxGridSize[0], prop.maxGridSize[1], prop.maxGridSize[2],
(unsigned int)prop.totalConstMem,
prop.major,
prop.minor,
prop.clockRate,
(unsigned int)prop.textureAlignment,
prop.deviceOverlap,
prop.multiProcessorCount
);
}
#endif
void COPROC_CUDA::clear() {
count = 0;
used = 0;
req_secs = 0;
req_instances = 0;
estimated_delay = -1; // mark as absent
strcpy(prop.name, "");
prop.totalGlobalMem = 0;
prop.sharedMemPerBlock = 0;
prop.regsPerBlock = 0;
prop.warpSize = 0;
prop.memPitch = 0;
prop.maxThreadsPerBlock = 0;
prop.maxThreadsDim[0] = 0;
prop.maxThreadsDim[1] = 0;
prop.maxThreadsDim[2] = 0;
prop.maxGridSize[0] = 0;
prop.maxGridSize[1] = 0;
prop.maxGridSize[2] = 0;
prop.clockRate = 0;
prop.totalConstMem = 0;
prop.major = 0;
prop.minor = 0;
prop.textureAlignment = 0;
prop.deviceOverlap = 0;
prop.multiProcessorCount = 0;
}
int COPROC_CUDA::parse(FILE* fin) {
char buf[1024], buf2[256];
clear();
while (fgets(buf, sizeof(buf), fin)) {
if (strstr(buf, "")) {
return 0;
}
if (parse_int(buf, "", count)) continue;
if (parse_double(buf, "", req_secs)) continue;
if (parse_int(buf, "", req_instances)) continue;
if (parse_double(buf, "", estimated_delay)) continue;
if (parse_str(buf, "", prop.name, sizeof(prop.name))) continue;
if (parse_int(buf, "", drvVersion)) continue;
if (parse_double(buf, "", prop.dtotalGlobalMem)) continue;
if (parse_int(buf, "", (int&)prop.sharedMemPerBlock)) continue;
if (parse_int(buf, "", prop.regsPerBlock)) continue;
if (parse_int(buf, "", prop.warpSize)) continue;
if (parse_int(buf, "", (int&)prop.memPitch)) continue;
if (parse_int(buf, "", prop.maxThreadsPerBlock)) continue;
if (parse_str(buf, "", buf2, sizeof(buf2))) {
// can't use sscanf here (FCGI)
//
prop.maxThreadsDim[0] = atoi(buf2);
char* p = strchr(buf2, ' ');
if (p) {
p++;
prop.maxThreadsDim[1] = atoi(p);
p = strchr(p, ' ');
if (p) {
p++;
prop.maxThreadsDim[2] = atoi(p);
}
}
continue;
}
if (parse_str(buf, "", buf2, sizeof(buf2))) {
prop.maxGridSize[0] = atoi(buf2);
char* p = strchr(buf2, ' ');
if (p) {
p++;
prop.maxGridSize[1] = atoi(p);
p = strchr(p, ' ');
if (p) {
p++;
prop.maxGridSize[2] = atoi(p);
}
}
continue;
}
if (parse_int(buf, "", prop.clockRate)) continue;
if (parse_int(buf, "", (int&)prop.totalConstMem)) continue;
if (parse_int(buf, "", prop.major)) continue;
if (parse_int(buf, "", prop.minor)) continue;
if (parse_int(buf, "", (int&)prop.textureAlignment)) continue;
if (parse_int(buf, "", prop.deviceOverlap)) continue;
if (parse_int(buf, "", prop.multiProcessorCount)) continue;
}
return ERR_XML_PARSE;
}
void COPROC_CELL_SPE::get(COPROCS&, vector&) {
return;
}