Rooba
/
boinc
mirror of https://github.com/BOINC/boinc.git
1
0
Fork 0
Code Issues 1 Packages Projects Releases Wiki Activity
boinc/client/gpu_detect.cpp

741 lines
22 KiB
C++
Raw Blame History

// This file is part of BOINC.
// http://boinc.berkeley.edu
// Copyright (C) 2009 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 <http://www.gnu.org/licenses/>.
// client-specific GPU code. Mostly GPU detection
//
// theory of operation:
// there are two ways of detecting GPUs:
// - vendor-specific libraries like CUDA and CAL,
// which detect only that vendor's GPUs
// - OpenCL, which can detect multiple types of GPUs,
// including nvidia/amd/intel as well was new types
// such as ARM integrated GPUs
//
// These libraries sometimes crash,
// and we've been unable to trap these via signal and exception handlers.
// So we do GPU detection in a separate process (boinc --detect_gpus)
// This process writes an XML file "coproc_info.xml" containing
// - lists of GPU detected via CUDA and CAL
// - lists of nvidia/amd/intel GPUs detected via OpenCL
// - a list of other GPUs detected via OpenCL
//
// Also, some dual-GPU laptops (e.g., Macbook Pro) don't power
// down the more powerful GPU until all applications which used them exit.
// Doing GPU detection in a second, short-lived process
// saves battery life on these laptops.
//
// When the process finishes, the client parses the info file.
// Then for each vendor it "correlates" the GPUs, which includes:
// - matching up the OpenCL and vendor-specific descriptions, if both exist
// - finding the most capable GPU, and seeing which other GPUs
// are similar to it in hardware and RAM.
// Other GPUs are not used.
// - copy these to the COPROCS structure
//
// GPUs can also be explicitly described in cc_config.xml
// comment out to do GPU detection in same process (for debugging)
//
#define USE_CHILD_PROCESS_TO_DETECT_GPUS 1
#include "cpp.h"
#ifdef _WIN32
#include "boinc_win.h"
#include "win_util.h"
#else
#include "config.h"
#include <setjmp.h>
#include <signal.h>
#include <sys/types.h>
#include <sys/wait.h>
#endif
#include "coproc.h"
#include "gpu_detect.h"
#include "file_names.h"
#include "util.h"
#include "str_replace.h"
#include "client_msgs.h"
#include "client_state.h"
using std::string;
using std::vector;
#ifndef _WIN32
jmp_buf resume;
void segv_handler(int) {
longjmp(resume, 1);
}
#endif
// the following store GPU instances during initialization
//
vector<COPROC_ATI> ati_gpus;
vector<COPROC_NVIDIA> nvidia_gpus;
vector<COPROC_INTEL> intel_gpus;
vector<OPENCL_DEVICE_PROP> ati_opencls;
vector<OPENCL_DEVICE_PROP> nvidia_opencls;
vector<OPENCL_DEVICE_PROP> intel_gpu_opencls;
vector<OPENCL_DEVICE_PROP> other_opencls;
vector<OPENCL_CPU_PROP> cpu_opencls;
static char* client_path;
// argv[0] from the command used to run client.
// May be absolute or relative.
static char client_dir[MAXPATHLEN];
// current directory at start of client
// find GPU instances, then correlate (merge) them
//
void COPROCS::get(
bool use_all, vector<string>&descs, vector<string>&warnings,
IGNORE_GPU_INSTANCE& ignore_gpu_instance
) {
#if USE_CHILD_PROCESS_TO_DETECT_GPUS
// detect_gpus() can cause crashes even with try/catch,
// so do it in a separate process that writes a file
//
int retval = 0;
char buf[256];
retval = launch_child_process_to_detect_gpus();
if (retval) {
snprintf(buf, sizeof(buf),
"launch_child_process_to_detect_gpus() returned error %d",
retval
);
warnings.push_back(buf);
}
retval = read_coproc_info_file(warnings);
if (retval) {
snprintf(buf, sizeof(buf),
"read_coproc_info_file() returned error %d",
retval
);
warnings.push_back(buf);
}
#else
detect_gpus(warnings);
#endif
correlate_gpus(use_all, descs, ignore_gpu_instance);
}
// populate the global vectors
// ati_gpus, nvidia_gpus, intel_gpus,
// nvidia_opencls, etc.
//
void COPROCS::detect_gpus(vector<string> &warnings) {
#ifdef _WIN32
try {
nvidia.get(warnings);
}
catch (...) {
warnings.push_back("Caught SIGSEGV in NVIDIA GPU detection");
}
try {
ati.get(warnings);
}
catch (...) {
warnings.push_back("Caught SIGSEGV in ATI GPU detection");
}
try {
intel_gpu.get(warnings);
}
catch (...) {
warnings.push_back("Caught SIGSEGV in INTEL GPU detection");
}
try {
// OpenCL detection must come last
get_opencl(warnings);
}
catch (...) {
warnings.push_back("Caught SIGSEGV in OpenCL detection");
}
#else
void (*old_sig)(int) = signal(SIGSEGV, segv_handler);
if (setjmp(resume)) {
warnings.push_back("Caught SIGSEGV in NVIDIA GPU detection");
} else {
nvidia.get(warnings);
}
#ifndef __APPLE__ // ATI does not yet support CAL on Macs
if (setjmp(resume)) {
warnings.push_back("Caught SIGSEGV in ATI GPU detection");
} else {
ati.get(warnings);
}
#endif
if (setjmp(resume)) {
warnings.push_back("Caught SIGSEGV in INTEL GPU detection");
} else {
intel_gpu.get(warnings);
}
if (setjmp(resume)) {
warnings.push_back("Caught SIGSEGV in OpenCL detection");
} else {
// OpenCL detection must come last
get_opencl(warnings);
}
signal(SIGSEGV, old_sig);
#endif
}
// for each GPU type, scan the GPUs we detected
// (e.g. the vector nvidia_gpus).
// Find the most capable one, and the ones equivalent to it.
// Also correlate OpenCL GPUs with CUDA/CAL GPUs.
// Then create a single COPROC (with appropriate count)
//
void COPROCS::correlate_gpus(
bool use_all,
vector<string> &descs,
IGNORE_GPU_INSTANCE &ignore_gpu_instance
) {
unsigned int i;
char buf[256], buf2[1024];
nvidia.correlate(use_all, ignore_gpu_instance[PROC_TYPE_NVIDIA_GPU]);
ati.correlate(use_all, ignore_gpu_instance[PROC_TYPE_AMD_GPU]);
intel_gpu.correlate(use_all, ignore_gpu_instance[PROC_TYPE_INTEL_GPU]);
correlate_opencl(use_all, ignore_gpu_instance);
// NOTE: OpenCL can report a max of only 4GB.
//
for (i=0; i<cpu_opencls.size(); i++) {
gstate.host_info.opencl_cpu_prop[gstate.host_info.num_opencl_cpu_platforms++] = cpu_opencls[i];
}
for (i=0; i<nvidia_gpus.size(); i++) {
// This is really CUDA description
nvidia_gpus[i].description(buf, sizeof(buf));
switch(nvidia_gpus[i].is_used) {
case COPROC_IGNORED:
snprintf(buf2, sizeof(buf2),
"CUDA: NVIDIA GPU %d (ignored by config): %s",
nvidia_gpus[i].device_num, buf
);
break;
case COPROC_USED:
snprintf(buf2, sizeof(buf2),
"CUDA: NVIDIA GPU %d: %s",
nvidia_gpus[i].device_num, buf
);
break;
case COPROC_UNUSED:
default:
snprintf(buf2, sizeof(buf2),
"CUDA: NVIDIA GPU %d (not used): %s",
nvidia_gpus[i].device_num, buf
);
#ifdef __APPLE__
if ((nvidia_gpus[i].cuda_version >= 6050) &&
nvidia_gpus[i].prop.major < 2) {
// This will be called only if CUDA recognized and reported the GPU
msg_printf(NULL, MSG_USER_ALERT, "NVIDIA GPU %d: %s %s",
nvidia_gpus[i].device_num, nvidia_gpus[i].prop.name,
_("cannot be used for CUDA or OpenCL computation with CUDA driver 6.5 or later")
);
}
#endif
break;
}
descs.push_back(string(buf2));
}
for (i=0; i<ati_gpus.size(); i++) {
// This is really CAL description
ati_gpus[i].description(buf, sizeof(buf));
switch(ati_gpus[i].is_used) {
case COPROC_IGNORED:
snprintf(buf2, sizeof(buf2),
"CAL: ATI GPU %d (ignored by config): %s",
ati_gpus[i].device_num, buf
);
break;
case COPROC_USED:
snprintf(buf2, sizeof(buf2),
"CAL: ATI GPU %d: %s",
ati_gpus[i].device_num, buf
);
break;
case COPROC_UNUSED:
default:
snprintf(buf2, sizeof(buf2),
"CAL: ATI GPU %d: (not used) %s",
ati_gpus[i].device_num, buf
);
break;
}
descs.push_back(string(buf2));
}
// Create descriptions for OpenCL NVIDIA GPUs
//
for (i=0; i<nvidia_opencls.size(); i++) {
if (nvidia_opencls[i].warn_bad_cuda) {
// This will be called only if CUDA did _not_ recognize and report the GPU
msg_printf(NULL, MSG_USER_ALERT, "NVIDIA GPU %d: %s %s",
nvidia_opencls[i].device_num, nvidia_opencls[i].name,
_("cannot be used for CUDA or OpenCL computation with CUDA driver 6.5 or later")
);
}
nvidia_opencls[i].description(buf, sizeof(buf), proc_type_name(PROC_TYPE_NVIDIA_GPU));
descs.push_back(string(buf));
}
// Create descriptions for OpenCL ATI GPUs
//
for (i=0; i<ati_opencls.size(); i++) {
ati_opencls[i].description(buf, sizeof(buf), proc_type_name(PROC_TYPE_AMD_GPU));
descs.push_back(string(buf));
}
// Create descriptions for OpenCL Intel GPUs
//
for (i=0; i<intel_gpu_opencls.size(); i++) {
intel_gpu_opencls[i].description(buf, sizeof(buf), proc_type_name(PROC_TYPE_INTEL_GPU));
descs.push_back(string(buf));
}
// Create descriptions for other OpenCL GPUs
//
int max_other_coprocs = MAX_RSC-1; // coprocs[0] is reserved for CPU
if (have_nvidia()) max_other_coprocs--;
if (have_ati()) max_other_coprocs--;
if (have_intel_gpu()) max_other_coprocs--;
// TODO: Should we implement cc_config ignore vectors for other (future) OpenCL coprocessors?
for (i=0; i<other_opencls.size(); i++) {
if ((int)i > max_other_coprocs) {
other_opencls[i].is_used = COPROC_UNUSED;
}
other_opencls[i].description(buf, sizeof(buf), other_opencls[i].name);
descs.push_back(string(buf));
}
// Create descriptions for OpenCL CPUs
//
for (i=0; i<cpu_opencls.size(); i++) {
cpu_opencls[i].description(buf, sizeof(buf));
descs.push_back(string(buf));
}
ati_gpus.clear();
nvidia_gpus.clear();
intel_gpus.clear();
ati_opencls.clear();
nvidia_opencls.clear();
intel_gpu_opencls.clear();
cpu_opencls.clear();
}
// This is called from CLIENT_STATE::init()
// after adding NVIDIA, ATI and Intel GPUs
// If we don't care about the order of GPUs in COPROCS::coprocs[],
// this code could be included at the end of COPROCS::correlate_gpus().
//
int COPROCS::add_other_coproc_types() {
int retval = 0;
for (unsigned int i=0; i<other_opencls.size(); i++) {
if (other_opencls[i].is_used != COPROC_USED) continue;
if (n_rsc >= MAX_RSC) {
retval = ERR_BUFFER_OVERFLOW;
break;
}
COPROC c;
// For device types other than NVIDIA, ATI or Intel GPU.
// we put each instance into a separate other_opencls element,
// so count=1.
//
c.count = 1;
c.opencl_device_count = 1;
c.opencl_prop = other_opencls[i];
c.available_ram = c.opencl_prop.global_mem_size;
c.device_num = c.opencl_prop.device_num;
c.peak_flops = c.opencl_prop.peak_flops;
c.have_opencl = true;
c.opencl_device_indexes[0] = c.opencl_prop.opencl_device_index;
c.opencl_device_ids[0] = c.opencl_prop.device_id;
c.instance_has_opencl[0] = true;
c.clear_usage();
safe_strcpy(c.type, other_opencls[i].name);
// Don't call COPROCS::add() because duplicate type is legal here
coprocs[n_rsc++] = c;
}
other_opencls.clear();
return retval;
}
void COPROCS::set_path_to_client(char *path) {
client_path = path;
// The path may be relative to the current directory
boinc_getcwd(client_dir);
}
int COPROCS::write_coproc_info_file(vector<string> &warnings) {
MIOFILE mf;
unsigned int i, temp;
FILE* f;
f = boinc_fopen(COPROC_INFO_FILENAME, "wb");
if (!f) return ERR_FOPEN;
mf.init_file(f);
mf.printf(" <coprocs>\n");
if (nvidia.have_cuda) {
mf.printf(" <have_cuda>1</have_cuda>\n");
mf.printf(" <cuda_version>%d</cuda_version>\n", nvidia.cuda_version);
}
for (i=0; i<ati_gpus.size(); ++i) {
ati_gpus[i].write_xml(mf, false);
}
for (i=0; i<nvidia_gpus.size(); ++i) {
temp = nvidia_gpus[i].count;
nvidia_gpus[i].count = 1;
nvidia_gpus[i].pci_infos[0] = nvidia_gpus[i].pci_info;
nvidia_gpus[i].write_xml(mf, false);
nvidia_gpus[i].count = temp;
}
for (i=0; i<intel_gpus.size(); ++i) {
intel_gpus[i].write_xml(mf, false);
}
for (i=0; i<ati_opencls.size(); ++i) {
ati_opencls[i].write_xml(mf, "ati_opencl", true);
}
for (i=0; i<nvidia_opencls.size(); ++i) {
nvidia_opencls[i].write_xml(mf, "nvidia_opencl", true);
}
for (i=0; i<intel_gpu_opencls.size(); ++i) {
intel_gpu_opencls[i].write_xml(mf, "intel_gpu_opencl", true);
}
for (i=0; i<other_opencls.size(); i++) {
other_opencls[i].write_xml(mf, "other_opencl", true);
}
for (i=0; i<cpu_opencls.size(); i++) {
cpu_opencls[i].write_xml(mf);
}
for (i=0; i<warnings.size(); ++i) {
mf.printf("<warning>%s</warning>\n", warnings[i].c_str());
}
mf.printf(" </coprocs>\n");
fclose(f);
return 0;
}
// if we're using a separate process to find GPUs,
// read its output file and create data structures
//
int COPROCS::read_coproc_info_file(vector<string> &warnings) {
MIOFILE mf;
int retval;
FILE* f;
string s;
COPROC_ATI my_ati_gpu;
COPROC_NVIDIA my_nvidia_gpu;
COPROC_INTEL my_intel_gpu;
OPENCL_DEVICE_PROP ati_opencl;
OPENCL_DEVICE_PROP nvidia_opencl;
OPENCL_DEVICE_PROP intel_gpu_opencl;
OPENCL_DEVICE_PROP other_opencl;
OPENCL_CPU_PROP cpu_opencl;
ati_gpus.clear();
nvidia_gpus.clear();
intel_gpus.clear();
ati_opencls.clear();
nvidia_opencls.clear();
intel_gpu_opencls.clear();
other_opencls.clear();
cpu_opencls.clear();
f = boinc_fopen(COPROC_INFO_FILENAME, "r");
if (!f) return ERR_FOPEN;
XML_PARSER xp(&mf);
mf.init_file(f);
if (!xp.parse_start("coprocs")) {
fclose(f);
return ERR_XML_PARSE;
}
while (!xp.get_tag()) {
if (xp.match_tag("/coprocs")) {
fclose(f);
return 0;
}
if (xp.parse_bool("have_cuda", nvidia.have_cuda)) continue;
if (xp.parse_int("cuda_version", nvidia.cuda_version)) {
continue;
}
if (xp.match_tag("coproc_ati")) {
retval = my_ati_gpu.parse(xp);
if (retval) {
my_ati_gpu.clear();
} else {
my_ati_gpu.device_num = (int)ati_gpus.size();
ati_gpus.push_back(my_ati_gpu);
}
continue;
}
if (xp.match_tag("coproc_cuda")) {
retval = my_nvidia_gpu.parse(xp);
if (retval) {
my_nvidia_gpu.clear();
} else {
my_nvidia_gpu.device_num = (int)nvidia_gpus.size();
my_nvidia_gpu.pci_info = my_nvidia_gpu.pci_infos[0];
my_nvidia_gpu.pci_infos[0].clear();
nvidia_gpus.push_back(my_nvidia_gpu);
}
continue;
}
if (xp.match_tag("coproc_intel_gpu")) {
retval = my_intel_gpu.parse(xp);
if (retval) {
my_intel_gpu.clear();
} else {
my_intel_gpu.device_num = (int)intel_gpus.size();
intel_gpus.push_back(my_intel_gpu);
}
continue;
}
if (xp.match_tag("ati_opencl")) {
ati_opencl.clear();
retval = ati_opencl.parse(xp, "/ati_opencl");
if (retval) {
ati_opencl.clear();
} else {
ati_opencl.is_used = COPROC_IGNORED;
ati_opencls.push_back(ati_opencl);
}
continue;
}
if (xp.match_tag("nvidia_opencl")) {
nvidia_opencl.clear();
retval = nvidia_opencl.parse(xp, "/nvidia_opencl");
if (retval) {
nvidia_opencl.clear();
} else {
nvidia_opencl.is_used = COPROC_IGNORED;
nvidia_opencls.push_back(nvidia_opencl);
}
continue;
}
if (xp.match_tag("intel_gpu_opencl")) {
intel_gpu_opencl.clear();
retval = intel_gpu_opencl.parse(xp, "/intel_gpu_opencl");
if (retval) {
intel_gpu_opencl.clear();
} else {
intel_gpu_opencl.is_used = COPROC_IGNORED;
intel_gpu_opencls.push_back(intel_gpu_opencl);
}
continue;
}
if (xp.match_tag("other_opencl")) {
other_opencl.clear();
retval = other_opencl.parse(xp, "/other_opencl");
if (retval) {
other_opencl.clear();
} else {
other_opencl.is_used = COPROC_USED;
other_opencls.push_back(other_opencl);
}
continue;
}
if (xp.match_tag("opencl_cpu_prop")) {
cpu_opencl.clear();
retval = cpu_opencl.parse(xp);
if (retval) {
cpu_opencl.clear();
} else {
cpu_opencl.opencl_prop.is_used = COPROC_IGNORED;
cpu_opencls.push_back(cpu_opencl);
}
continue;
}
if (xp.parse_string("warning", s)) {
warnings.push_back(s);
continue;
}
// TODO: parse OpenCL info for CPU when implemented:
// gstate.host_info.have_cpu_opencl
// gstate.host_info.cpu_opencl_prop
}
fclose(f);
return ERR_XML_PARSE;
}
int COPROCS::launch_child_process_to_detect_gpus() {
#ifdef _WIN32
HANDLE prog;
#else
int prog;
#endif
char quoted_data_dir[MAXPATHLEN+2];
char data_dir[MAXPATHLEN];
int retval = 0;
retval = boinc_delete_file(COPROC_INFO_FILENAME);
if (retval) {
msg_printf(0, MSG_INFO,
"Failed to delete old %s. error code %d",
COPROC_INFO_FILENAME, retval
);
} else {
for (;;) {
if (!boinc_file_exists(COPROC_INFO_FILENAME)) break;
boinc_sleep(0.01);
}
}
// use full path to exe if possible, otherwise keep using argv[0]
char execpath[MAXPATHLEN];
if (!get_real_executable_path(execpath, sizeof(execpath))) {
client_path = execpath;
}
boinc_getcwd(data_dir);
#ifdef _WIN32
strlcpy(quoted_data_dir, "\"", sizeof(quoted_data_dir));
strlcat(quoted_data_dir, data_dir, sizeof(quoted_data_dir));
strlcat(quoted_data_dir, "\"", sizeof(quoted_data_dir));
#else
strlcpy(quoted_data_dir, data_dir, sizeof(quoted_data_dir));
#endif
if (log_flags.coproc_debug) {
msg_printf(0, MSG_INFO,
"[coproc] launching child process at %s",
client_path
);
if (!is_path_absolute(client_path)) {
msg_printf(0, MSG_INFO,
"[coproc] relative to directory %s",
client_dir
);
}
msg_printf(0, MSG_INFO,
"[coproc] with data directory %s",
quoted_data_dir
);
}
int argc = 4;
char* const argv[5] = {
const_cast<char *>(client_path),
const_cast<char *>("--detect_gpus"),
const_cast<char *>("--dir"),
const_cast<char *>(quoted_data_dir),
NULL
};
retval = run_program(
client_dir,
client_path,
argc,
argv,
0,
prog
);
if (retval) {
if (log_flags.coproc_debug) {
msg_printf(0, MSG_INFO,
"[coproc] run_program of child process returned error %d",
retval
);
}
return retval;
}
retval = get_exit_status(prog);
if (retval) {
char buf[200];
#ifdef _WIN32
char buf2[200];
windows_format_error_string(retval, buf2, sizeof(buf2));
snprintf(buf, sizeof(buf), "process exited with status 0x%x: %s", retval, buf2);
#else
if (WIFEXITED(retval)) {
int code = WEXITSTATUS(retval);
snprintf(buf, sizeof(buf), "process exited with status %d: %s", code, strerror(code));
} else if (WIFSIGNALED(retval)) {
int sig = WTERMSIG(retval);
snprintf(buf, sizeof(buf), "process was terminated by signal %d", sig);
} else {
snprintf(buf, sizeof(buf), "unknown status %d", retval);
}
#endif
msg_printf(0, MSG_INFO,
"GPU detection failed: %s",
buf
);
}
return 0;
}
// print descriptions of coprocs specified in cc_config.xml,
// and make sure counts are <= 64
//
void COPROCS::bound_counts() {
for (int j=1; j<n_rsc; j++) {
msg_printf(NULL, MSG_INFO, "Coprocessor specified in cc_config.xml. Type %s (%s); count %d",
coprocs[j].type,
coprocs[j].non_gpu?"non-GPU":"GPU",
coprocs[j].count
);
if (coprocs[j].count > MAX_COPROC_INSTANCES) {
msg_printf(NULL, MSG_USER_ALERT,
"%d instances of %s specified in cc_config.xml; max is %d",
coprocs[j].count,
coprocs[j].type,
MAX_COPROC_INSTANCES
);
coprocs[j].count = MAX_COPROC_INSTANCES;
}
}
}
Reference in New Issue View Git Blame Copy Permalink