I have a simple multi-threaded program where the thread performs random reads on a given file (in memory) divided evenly amongst the threads. The thread reads from the file to buffer and sets a value. This is really a program designed to test memory bandwidth. This is the following program,
#include <sys/mman.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <getopt.h>
#include <errno.h>
#include <stdbool.h>
#include <ctype.h>
#include <inttypes.h>
#include <pthread.h>
#include <assert.h>
#include <time.h>
#define NS_IN_SECOND 1000000000
uint64_t nano_time(void) {
struct timespec ts;
if( clock_gettime(CLOCK_REALTIME, &ts) == 0)
return ts.tv_sec * NS_IN_SECOND + ts.tv_nsec;
}
// avx512 test
#include <stdint.h>
void *__memmove_chk_avx512_no_vzeroupper(void *dest, void *src, size_t s);
/**
* To create 4 GB file: This will allocate space on disk
* $ dd < /dev/zero bs=1048576 count=4096 > testfile
*
* 100 GiB
* dd if=/dev/zero of=bigmmaptest bs=1M count=102400
* To clear cache:
* $ sync; echo 1 > /proc/sys/vm/drop_caches
*/
//#define SAMPLE_LATENCY 1
#define BYTES_IN_GB (1024*1024*1024)
// Block sized will be used for read and the same will be used for striding
// when iterating over a file in mmap.
#define DEFAULT_BLOCK_SIZE 4096 //8192
#define NANOSECONDS_IN_SECOND 1000000000
const char DEFAULT_NAME[] = "/mnt/tmp/mmaptest";
#define EXIT_MSG(...) \
do { \
printf(__VA_ARGS__); \
_exit(-1); \
} while (0)
uint64_t read_mmap_test(int fd, int tid, size_t block_size, size_t filesize, char* buf,
off_t *offsets, uint64_t *begin, uint64_t *end);
uint64_t write_mmap_test(int fd, int tid, size_t block_size, size_t filesize, char* buf,
off_t *offsets, uint64_t *begin, uint64_t *end);
uint64_t mmap_test(int fd, int tid, size_t block_size, size_t filesize, char *buf,
char optype, off_t *offsets, uint64_t *begin, uint64_t *end);
uint64_t read_syscall_test(int fd, int tid, size_t block_size, size_t filesize,
off_t *offsets, uint64_t *begin, uint64_t *end);
uint64_t write_syscall_test(int fd, int tid, size_t block_size, size_t filesize,
off_t *offsets, uint64_t *begin, uint64_t *end);
uint64_t syscall_test(int fd, int tid, size_t block_size, size_t filesize,
char optype, off_t *offsets, uint64_t *begin, uint64_t *end);
size_t get_filesize(const char* filename);
void print_help_message(const char *progname);
char* map_buffer(int fd, size_t size);
void *run_tests(void *);
static int silent = 0;
typedef struct {
int tid;
int fd;
char *mapped_buffer;
int read_mmap;
int read_syscall;
int write_mmap;
int write_syscall;
off_t *offsets;
size_t block_size;
size_t chunk_size;
int retval;
uint64_t start_time;
uint64_t end_time;
} threadargs_t;
size_t filesize;
int main(int argc, char **argv) {
char *fname = (char*) DEFAULT_NAME;
char *mapped_buffer = NULL;
int c, fd, i, flags = O_RDWR, numthreads = 1, ret, option_index;
static int randomaccess = 0,
read_mmap = 0, read_syscall = 0,
write_mmap = 0, write_syscall = 0,
mixed_mmap = 0, write_tr = 0;
off_t *offsets = 0;
size_t block_size = DEFAULT_BLOCK_SIZE, numblocks,
new_file_size = 0;
uint64_t min_start_time, max_end_time = 0, retval;
// permissions
uint64_t mode = S_IRWXU | S_IRWXG;
pthread_t *threads;
threadargs_t *threadargs;
static struct option long_options[] =
{
// Options set a flag
{"randomaccess", no_argument, &randomaccess, 1},
{"readmmap", no_argument, &read_mmap, 1},
{"readsyscall", no_argument, &read_syscall, 1},
{"silent", no_argument, &silent, 1},
{"writemmap", no_argument, &write_mmap, 1},
{"writesyscall", no_argument, &write_syscall, 1},
{"mixedmmap", no_argument, &mixed_mmap, 1},
// Options take an argument
{"block", required_argument, 0, 'b'},
{"file", required_argument, 0, 'f'},
{"help", no_argument, 0, 'h'},
{"size", no_argument, 0, 's'},
{"threads", required_argument, 0, 't'},
{"writethreads", no_argument, 0, 'w'},
{0, 0, 0, 0}
};
//read operations
while(1) {
c = getopt_long(argc, argv, "b:f:h:s:t:w:",
long_options, &option_index);
// is end of the option
if (c == -1)
break;
switch(c)
{
case 0:
break;
case 'b':
block_size = atoi(optarg);
break;
case 'f':
fname = optarg;
break;
case 'h':
print_help_message(argv[0]);
_exit(0);
case 's':
new_file_size = (size_t)(atoi(optarg)) * BYTES_IN_GB;
break;
case 't':
numthreads = (int) (atoi(optarg));
break;
case 'w':
write_tr = atoi(optarg);
break;
default:
break;
}
}
if(!silent){
printf("PID: %d\n", getpid());
printf("Using file %s \n", fname);
}
if ((filesize = get_filesize(fname)) == -1) {
if (read_mmap || read_syscall) {
printf("Cannot obtain file size for %s: %s"
"File must exist prior to running read tests.\n",
fname, strerror(errno));
_exit(-1);
}
else
filesize = new_file_size;
}
fd = open((const char*)fname, flags, mode);
if(fd <0) {
printf("Clould not open/create file %s: %s\n",
fname, strerror(errno));
_exit(-1);
}
if(block_size < 0 || block_size > filesize){
printf("Invalid block size: %zu for file of size "
"%zu. Block size must be greater than 0 and no"
"greater than the file size.\n",
block_size, filesize);
_exit(-1);
}
/*
* Generate random block number for random file access.
* Sequential for sequential access
*/
numblocks = filesize/block_size;
if(filesize % block_size > 0)
numblocks++;
offsets = (off_t *) malloc(numblocks * sizeof(off_t));
if(offsets == 0){
printf("Failed to allocate memory: %s\n", strerror(errno));
_exit(-1);
}
for (uint64_t i = 0; i < numblocks; i++)
if(randomaccess)
offsets[i] = ((int)random() % numblocks) * block_size;
else
offsets[i] = i*block_size;
if (numblocks % numthreads != 0)
EXIT_MSG("We have %" PRIu64 " blocks and %d threads. "
"Threads must evenly divide blocks. "
"Please fix the args.\n",
(uint_least64_t)numblocks, numthreads);
if( read_mmap || write_mmap || mixed_mmap)
assert((mapped_buffer = map_buffer(fd, filesize)) != NULL);
threads = (pthread_t*)malloc(numthreads * sizeof(pthread_t));
threadargs =
(threadargs_t*)malloc(numthreads * sizeof(threadargs_t));
if (threads == NULL || threadargs == NULL)
EXIT_MSG("Could not allocate thread array for %d threads.\n", numthreads);
for (i = 0; i < numthreads; i++) {
if(mixed_mmap){
if (i < write_tr) {
write_mmap = 1;
} else {
read_mmap = 1;
}
}
threadargs[i].fd = fd;
threadargs[i].tid = i;
threadargs[i].block_size = block_size;
threadargs[i].chunk_size = filesize/numthreads;
threadargs[i].mapped_buffer = mapped_buffer;
threadargs[i].offsets = &offsets[numblocks/numthreads * i];
threadargs[i].read_mmap = read_mmap;
threadargs[i].read_syscall = read_syscall;
threadargs[i].write_mmap = write_mmap;
threadargs[i].write_syscall = write_syscall;
int ret = pthread_create(&threads[i], NULL, run_tests, &threadargs[i]);
if (ret!=0)
EXIT_MSG("pthread_create for %dth thread failed: %s\n",
i, strerror(errno));
}
for (i = 0; i< numthreads; i++){
ret = pthread_join(threads[i], NULL);
if (ret !=0)
EXIT_MSG("Thread %d failed in join: %s\n",
i, strerror(errno));
}
// for mixed mode determine read and write aggregate b/w.
if(mixed_mmap) {
// Write b/w
min_start_time = threadargs[0].start_time;
max_end_time = 0;
// Since tid 0 to write_tr-1 did writes, find it's min and max.
for(i=0; i < write_tr; i++){
min_start_time = (threadargs[i].start_time < min_start_time)?
threadargs[i].start_time:min_start_time;
max_end_time = (threadargs[i].end_time > max_end_time)?
threadargs[i].end_time:max_end_time;
}
printf("Write: %.2f\n",
(double)write_tr*(filesize/numthreads)/(double)(max_end_time-min_start_time)
* NANOSECONDS_IN_SECOND / BYTES_IN_GB);
// Read b/w
min_start_time = threadargs[write_tr].start_time;
max_end_time = 0;
for(i=write_tr; i < numthreads; i++){
min_start_time = (threadargs[i].start_time < min_start_time)?
threadargs[i].start_time:min_start_time;
max_end_time = (threadargs[i].end_time > max_end_time)?
threadargs[i].end_time:max_end_time;
}
printf("Read: %.2f\n",
(double)(numthreads-write_tr)*(filesize/numthreads)/(double)(max_end_time-min_start_time)
* NANOSECONDS_IN_SECOND / BYTES_IN_GB);
}
/**
* For total run time. Find the smallest start time
* and largest end time across all threads.
*/
min_start_time = threadargs[0].start_time;
max_end_time = 0;
for (i=0; i< numthreads; i++){
min_start_time = (threadargs[i].start_time < min_start_time)?
threadargs[i].start_time:min_start_time;
max_end_time = (threadargs[i].end_time > max_end_time)?
threadargs[i].end_time:max_end_time;
}
printf("%.2f\n",
(double)filesize/(double)(max_end_time-min_start_time)
* NANOSECONDS_IN_SECOND / BYTES_IN_GB);
munmap(mapped_buffer, filesize);
close(fd);
}
void * run_tests(void *args) {
uint64_t retval;
threadargs_t t = *(threadargs_t*)args;
if(t.read_mmap) {
if(!silent)
printf("Running read mmap test:\n");
retval = read_mmap_test(t.fd, t.tid, t.block_size, t.chunk_size,
t.mapped_buffer, t.offsets,
&((threadargs_t*)args)->start_time,
&((threadargs_t*)args)->end_time);
}
else if(t.read_syscall) {
if(!silent)
printf("Running read syscall test:\n");
retval = read_syscall_test(t.fd, t.tid, t.block_size, t.chunk_size,
t.offsets,
&((threadargs_t*)args)->start_time,
&((threadargs_t*)args)->end_time);
}
else if(t.write_mmap) {
if(!silent)
printf("Running write mmap test:\n");
retval = write_mmap_test(t.fd, t.tid, t.block_size, t.chunk_size,
t.mapped_buffer, t.offsets,
&((threadargs_t*)args)->start_time,
&((threadargs_t*)args)->end_time);
}
else if(t.write_syscall) {
if(!silent)
printf("Running write syscall test:\n");
retval = write_syscall_test(t.fd, t.tid, t.block_size, t.chunk_size,
t.offsets,
&((threadargs_t*)args)->start_time,
&((threadargs_t*)args)->end_time);
}
return (void*) 0;
}
#define READ 1
#define WRITE 2
/**
********* SYSCALL section
*/
uint64_t read_syscall_test(int fd, int tid, size_t block_size, size_t filesize,
off_t *offsets, uint64_t *begin, uint64_t *end) {
return syscall_test(fd, tid, block_size, filesize, READ, offsets,
begin, end);
}
uint64_t write_syscall_test(int fd, int tid, size_t block_size, size_t filesize,
off_t *offsets, uint64_t *begin, uint64_t *end) {
return syscall_test(fd, tid, block_size, filesize, WRITE, offsets,
begin, end);
}
uint64_t syscall_test(int fd, int tid, size_t block_size, size_t filesize,
char optype, off_t *offsets, uint64_t *begin, uint64_t *end) {
bool done = false;
char * buffer = NULL;
int i = 0;
size_t total_bytes_transferred = 0;
uint64_t begin_time, end_time, ret_token = 0;
buffer = (char*)malloc(block_size);
if(buffer == NULL) {
printf("Failed to allocate memory: %s\n", strerror(errno));
return -1;
}
memset((void*)buffer, 0, block_size);
begin_time= nano_time();
while(!done) {
size_t bytes_transferred = 0;
if(optype == READ)
bytes_transferred = pread(fd, buffer, block_size, offsets[i++]);
else if (optype == WRITE)
bytes_transferred = pwrite(fd, buffer, block_size, offsets[i++]);
if (bytes_transferred == 0)
done = true;
else if(bytes_transferred == -1){
printf("Failed to IO: %s\n", strerror(errno));
return -1;
}
else {
total_bytes_transferred += bytes_transferred;
if (optype == WRITE && total_bytes_transferred == filesize)
done = true;
// Do random operation
ret_token += buffer[0];
}
if (i*block_size >= filesize)
done = true;
}
end_time = nano_time();
if(!silent){
printf("%s: %" PRIu64 " bytes transferred in %" PRIu64 ""
" ns.\n", (optype == READ)?"read-syscall":"write-syscall",
(uint_least64_t)total_bytes_transferred, (end_time-begin_time));
// Throughput in GB/s
printf("(tid %d) %.2f\n", tid,
(double)filesize/(double)(end_time-begin_time)
* NANOSECONDS_IN_SECOND / BYTES_IN_GB);
}
*begin = begin_time;
*end = end_time;
return ret_token;
}
/**
* MMAP tests
*/
uint64_t read_mmap_test(int fd, int tid, size_t block_size, size_t filesize,
char *buf, off_t *offsets, uint64_t *begin, uint64_t *end) {
return mmap_test(fd, tid, block_size, filesize, buf, READ, offsets, begin, end);
}
uint64_t write_mmap_test(int fd, int tid, size_t block_size, size_t filesize,
char *buf, off_t *offsets, uint64_t *begin, uint64_t *end){
return mmap_test(fd, tid, block_size, filesize, buf, WRITE, offsets, begin, end);
}
// Add memory addr
#if SAMPLE_LATENCY
#define BEGIN_LAT_SAMPLE \
if (num_samples < MAX_LAT_SAMPLES && i%LAT_SAMPL_INTERVAL == 0) \
lat_begin_time = nano_time();
#define END_LAT_SAMPLE \
if (num_samples < MAX_LAT_SAMPLES && i%LAT_SAMPL_INTERVAL == 0) { \
lat_end_time = nano_time(); \
latency_samples[i/LAT_SAMPL_INTERVAL % MAX_LAT_SAMPLES] = \
lat_end_time - lat_begin_time; \
num_samples++; \
}
#define MAX_LAT_SAMPLES 50
//#define LAT_SAMPL_INTERVAL (1000*1048576)
#define LAT_SAMPL_INTERVAL block_size
#else
#define BEGIN_LAT_SAMPLE ;
#define END_LAT_SAMPLE
#endif
uint64_t mmap_test(int fd, int tid, size_t block_size, size_t filesize, char *mapped_buffer,
char optype, off_t *offsets, uint64_t *begin, uint64_t *end) {
bool done = false;
char *buffer = NULL;
uint64_t i, j, numblocks, ret;
uint64_t begin_time, end_time, ret_token = 0;
#if SAMPLE_LATENCY
uint64_t lat_begin_time, lat_end_time;
size_t latency_samples[MAX_LAT_SAMPLES];
int num_samples = 0;
memset((void*)latency_samples, 0, sizeof(latency_samples));
#endif
buffer = (char*)malloc(block_size);
if(buffer == NULL) {
printf("Failed to allocate memory: %s\n", strerror(errno));
return -1;
}
memset((void*)buffer, 1, block_size);
begin_time = nano_time();
for(i=0; i<filesize; i+=block_size){
off_t offset = offsets[i/block_size];
BEGIN_LAT_SAMPLE;
if(optype == READ) {
//__memmove_chk_avx512_no_vzeroupper(buffer, &mapped_buffer[offset], block_size);
memcpy(buffer, &mapped_buffer[offset], block_size);
ret_token += buffer[0];
}
else if (optype == WRITE) {
//__memmove_chk_avx512_no_vzeroupper(&mapped_buffer[offset], buffer, block_size);
memcpy(&mapped_buffer[offset], buffer, block_size);
ret_token += mapped_buffer[i];
}
END_LAT_SAMPLE;
}
end_time = nano_time();
if(!silent) {
printf("%s: %" PRIu64 " bytes read in %" PRIu64 " ns.\n",
(optype==READ)?"readmap":"writemap",
(uint_least64_t)filesize, (end_time-begin_time));
// print GB/s
printf("(tid %d) %.2f\n", tid,
(double)filesize/(double)(end_time-begin_time)
* NANOSECONDS_IN_SECOND / BYTES_IN_GB);
}
*begin = begin_time;
*end = end_time;
#if SAMPLE_LATENCY
printf("\nSample latency for %ld byte block:\n", block_size);
for (i = 0; i < MAX_LAT_SAMPLES; i++)
printf("\t%ld: %ld\n", i, latency_samples[i]);
#endif
return ret_token;
}
char* map_buffer(int fd, size_t size) {
char *mapped_buffer = NULL;
// Populate
mapped_buffer = (char*)mmap(NULL, size, PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_POPULATE, fd, 0);
// Shared
// mapped_buffer = (char*)mmap(NULL, size, PROT_READ | PROT_WRITE,
// MAP_SHARED, fd, 0);
// Anon test
// mapped_buffer = (char*)mmap(NULL, size, PROT_READ | PROT_WRITE,
// MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
if(mapped_buffer == MAP_FAILED)
EXIT_MSG("Failed to mmap file of size %zu: %s\n",
size, strerror(errno));
// Might also need to gurantee page aligned - posix_memalign()
// int mret = madvise(mapped_buffer, filesize, MADV_HUGEPAGE);
// if(mret!=0) {
// fprintf(stderr, "failed madvise: %s\n", strerror(errno));
// }
return mapped_buffer;
}
size_t get_filesize(const char* filename){
int retval;
struct stat st;
retval = stat(filename, &st);
if(retval)
return -1;
else
return st.st_size;
}
void print_help_message(const char *progname) {
/* take only the last portion of the path */
const char *basename = strrchr(progname, '/');
basename = basename ? basename + 1 : progname;
printf("usage: %s [OPTION]\n", basename);
printf(" -h, --help\n"
" Print this help and exit.\n");
printf(" -b, --block[=BLOCKSIZE]\n"
" Block size used for read system calls.\n"
" For mmap tests, the size of the stride when iterating\n"
" over the file.\n"
" Defaults to %d.\n", DEFAULT_BLOCK_SIZE);
printf(" -f, --file[=FILENAME]\n"
" Perform all tests on this file (defaults to %s).\n",
DEFAULT_NAME);
printf(" --readsyscall\n"
" Perform a read test using system calls.\n");
printf(" --readmmap\n"
" Perform a read test using mmap.\n");
printf(" --writesyscall\n"
" Perform a write test using system calls.\n");
printf(" --writemmap\n"
" Perform a write test using mmap.\n");
printf(" --randomaccess\n"
" Perform random access.\n");
printf(" --threads\n"
" Number of threads to use. Defaults to one.\n");
printf(" --mixedmmap\n"
" Perfom read and write concurrently at different offsets\n");
printf(" -w, -writethreads[=0]\n"
" Number of threads that should perform write\n");
}
To compile:
$ gcc testm.c -o testm -lpthread -static -O2 -fno-builtin-memcpy
Commands to run the program:
$ dd if=/dev/zero of=bigmmaptest bs=1M count=25600 # 25 GiB file
$ ./testm -b 1024 -f bigmmaptest --threads 16 --randomaccess --readmmap
I am on a 32 core Xeon 5218 2nd Gen. L1d KiB /L2 MiB /L3 MiB -- 512 / 16 / 22
When the memcpy size is 1 KiB I get 21.7 GB/s but when the size is 256B I get 26.68 GB/s and 34.8 GB/s when the size is 4 KiB. Why is there a drop in the middle?
I observe that 2 KiB also performs poorly when compared to 256B and 4 KiB.
What's more interesting is, when I disable the L2 hardware prefetcher and without any other changes my bandwidth automatically increases for 1 KiB and 2 KiB. Without prefetch 2 KiB memcpy gives 34.8 GB/s. All of these are aggregate bandwidth.
With perf, I did measure L2 load-store misses but they turned out to not change drastically. This effect is also not seen for 8 threads and below.
I am on linux 5.0.4. I am using the glibC memcpy (gcc 7.5.0) and even with -O2 I observe the above quirk. Where 1 KiB access size gives 18.76 GiB/s with L2 prefetch and without I get 30.32 GiB/s. For comparison, 256 B access size provides 24.7 GiB/s with prefetch and 24.8 GiB/s without. Clearly, the drop in performance is because of the L2 cache pollution caused by the prefetcher, as this is not observed with smaller thread counts. I was considering if SMT could be the reason for increased pollution but I observe the effect distinctly at 16 threads on 16 physical cores.
Skimming through glibc memcpy code, I can see that any access below the size of 4 KiB uses AVX 256 instructions, so there is nothing changing there.
The smaller 256B size not seeing a drop from the L2 streamer might be due to the sequence of cache misses being too short to activate the streamer and waste bandwidth (and slots in the LFBs and L2 <-> L3 superqueue) on requests that won't be useful.
For aligned 4k, there are no bytes within the same page that you're not fetching, so the L2 prefetcher is positively useful, or at least not harmful. (Demand loads come in pretty quickly for later lines when running memcpy so I'm guessing speeds were about the same with/without HW prefetch enabled, unless HW prefetch helps getting started on a new 4k chunk while still waiting for the end of the previous.)
The L2 only sees physical addresses, and AFAIK it doesn't try to prefetch across a 4k boundary. (Even if its within the same 2M hugepage, because it doesn't know that either.) The "next-page prefetcher" Intel mentions being new in Ivy Bridge is AFAIK just a TLB prefetch, not data.
So with aligned 4k memcpy, HW prefetch stops automatically at the end of the data you're actually going to read, not wasting any bandwidth. Since mmap gives you page-aligned memory, these 4k memcopies are from a single source page.
(The destination is irrelevant as it probably stays hot in L1d cache, with maybe an occasional eviction to L2, and the reload from it after memcpy can come from store-forwarding, not even having to wait for memcpy's store to commit to L1d.)
Prediction: If your smaller memcpy source starts part way into a 4k page, but still end at the end of a 4k page, you'd probably see similar behaviour to prefetch disabled. e.g. generate a random page number, and start at 3072 bytes into it, doing a 1 KiB copy. So all your 1 KiB copies come from the ends of pages, never middles.
(You'd still have more dTLB misses per byte memcpyed, because each TLB entry is only covering 1 K of the data you ever actually read. You did you use MAP_POPULATE so you shouldn't be seeing page faults in the timed region, assuming you have enough RAM.)
L1d KiB /L2 MiB /L3 MiB -- 512 / 16 / 22
Those are aggregate totals, but L1d and L2 are private per-core! You have 32kiB L1d and 1MiB L2 per core, because this is Cascade Lake, same layout as Skylake-X.
And BTW, I'd consider using a fast PRNG like xorshift+ or xorshift* inside the timing loop; that's easily random enough to defeat prefetching; even a simple LFSR or even LCG with a power-of-2 modulo would do that (and be very cheap, just an imul and add). It avoids having to read offsets from another array, if you really want to isolate just the memcpy memory accesses. Probably doesn't make a difference though. One advantage of a very simple PRNG with a period equal to the space you're trying to cover (like an LCG) is that you won't generate the same address twice, giving you a random permutation of the blocks. But with a big enough block of memory, random cache hits even from L3 are unlikely even without that hard-to-achieve property.
Your current array of offsets is fine. (I didn't look at the code super closely, so I'm just assuming there aren't bugs.)
How can an array of structs that has been dynamically allocated on the host be used by a kernel, without passing the array of structs as a kernel argument? This seems like a common procedure with a good amount of documentation online, yet it doesn't work on the following program.
Note: Please note that the following questions have been studied before posting this question:
1) copying host memory to cuda __device__ variable 2) Global variable in CUDA 3) Is there any way to dynamically allocate constant memory? CUDA
So far, unsuccessful attempts have been made to:
Dynamically allocate array of structs with cudaMalloc(), then
Use cudaMemcpyToSymbol() with the pointer returned from cudaMalloc() to copy to a __device__ variable which can be used by the kernel.
Code attempt:
NBody.cu (error checking using cudaStatus has mostly been omitted for better readability, and function to read data from file into dynamic array removed):
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
#include <stdlib.h>
#define BLOCK 256
struct nbody {
float x, y, vx, vy, m;
};
typedef struct nbody nbody;
// Global declarations
nbody* particle;
// Device variables
__device__ unsigned int d_N; // Kernel can successfully access this
__device__ nbody d_particle; // Update: part of problem was here with (*)
// Aim of kernel: to print contents of array of structs without using kernel argument
__global__ void step_cuda_v1() {
int i = threadIdx.x + blockDim.x * blockIdx.x;
if (i < d_N) {
printf("%.f\n", d_particle.x);
}
}
int main() {
unsigned int N = 10;
unsigned int I = 1;
cudaMallocHost((void**)&particle, N * sizeof(nbody)); // Host allocation
cudaError_t cudaStatus;
for (int i = 0; i < N; i++) particle[i].x = i;
nbody* particle_buf; // device buffer
cudaSetDevice(0);
cudaMalloc((void**)&particle_buf, N * sizeof(nbody)); // Allocate device mem
cudaMemcpy(particle_buf, particle, N * sizeof(nbody), cudaMemcpyHostToDevice); // Copy data into device mem
cudaMemcpyToSymbol(d_particle, &particle_buf, sizeof(nbody*)); // Copy pointer to data into __device__ var
cudaMemcpyToSymbol(d_N, &N, sizeof(unsigned int)); // This works fine
int NThreadBlock = (N + BLOCK - 1) / BLOCK;
for (int iteration = 0; iteration <= I; iteration++) {
step_cuda_v1 << <NThreadBlock, BLOCK >> > ();
//step_cuda_v1 << <1, 5 >> > (particle_buf);
cudaDeviceSynchronize();
cudaStatus = cudaGetLastError();
if (cudaStatus != cudaSuccess)
{
fprintf(stderr, "ERROR: %s\n", cudaGetErrorString(cudaStatus));
exit(-1);
}
}
return 0;
}
OUTPUT:
"ERROR: kernel launch failed."
Summary:
How can I print the contents of the array of structs from the kernel, without passing it as a kernel argument?
Coding in C using VS2019 with CUDA 10.2
With the help of #Robert Crovella and #talonmies, here is the solution that outputs a sequence that cycles from 0 to 9 repeatedly.
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
#include <stdlib.h>
#define BLOCK 256
//#include "Nbody.h"
struct nbody {
float x, y, vx, vy, m;
};
typedef struct nbody nbody;
// Global declarations
nbody* particle;
// Device variables
__device__ unsigned int d_N; // Kernel can successfully access this
__device__ nbody* d_particle;
//__device__ nbody d_particle; // Update: part of problem was here with (*)
// Aim of kernel: to print contents of array of structs without using kernel argument
__global__ void step_cuda_v1() {
int i = threadIdx.x + blockDim.x * blockIdx.x;
if (i < d_N) {
printf("%.f\n", d_particle[i].x);
}
}
int main() {
unsigned int N = 10;
unsigned int I = 1;
cudaMallocHost((void**)&particle, N * sizeof(nbody)); // Host allocation
cudaError_t cudaStatus;
for (int i = 0; i < N; i++) particle[i].x = i;
nbody* particle_buf; // device buffer
cudaSetDevice(0);
cudaMalloc((void**)&particle_buf, N * sizeof(nbody)); // Allocate device mem
cudaMemcpy(particle_buf, particle, N * sizeof(nbody), cudaMemcpyHostToDevice); // Copy data into device mem
cudaMemcpyToSymbol(d_particle, &particle_buf, sizeof(nbody*)); // Copy pointer to data into __device__ var
cudaMemcpyToSymbol(d_N, &N, sizeof(unsigned int)); // This works fine
int NThreadBlock = (N + BLOCK - 1) / BLOCK;
for (int iteration = 0; iteration <= I; iteration++) {
step_cuda_v1 << <NThreadBlock, BLOCK >> > ();
//step_cuda_v1 << <1, 5 >> > (particle_buf);
cudaDeviceSynchronize();
cudaStatus = cudaGetLastError();
if (cudaStatus != cudaSuccess)
{
fprintf(stderr, "ERROR: %s\n", cudaGetErrorString(cudaStatus));
exit(-1);
}
}
return 0;
}
I have one structure xyz as given below struct xyz { char a; int32_t b; char c[50]; uint32_t d; uchar e[10];}
I need to broadcast it so I used MPI_Bcast() where i required MPI Datatype corresponding to struct xyz for that I used MPI_Type_creat_struct() function to create a new MPI datatype as MPI_Datatype MPI_my_new_datatype, oldtypes[4]; where I used MPI datatypes corresponding to above structure members datatype as followings
oldtypes[4] = {MPI_CHAR, MPI_INT, MPI_UNSIGNED, MPI_UNSIGNED_CHAR}; and to craete new datatype i used following arguments in the function..
MPI_Type_create_struct(4,blockcounts, offsets, oldtypes, &MPI_my_new_datatype); MPI_Type_commit(&MPI_my_new_datatype);
Now it is compiling but giving run time error as below::
* An error occurred in MPI_Type_create_structon communicator MPI_COMM_WORLD MPI_ERR_ARG: invalid argument of some other kind MPI_ERRORS_ARE_FATAL (goodbye).
Can any one find out where is the problem?
You can't "bundle up" the similar types like that. Each field needs to be addressed seperately, and there are 5 of them, not 4.
Also note that, in general, it's a good idea to actually "measure" the offsets rather than infer them.
The following works:
#include <stdio.h>
#include <mpi.h>
#include <stdint.h>
struct xyz_t {
char a; int32_t b; char c[50]; uint32_t d; unsigned char e[10];
};
int main(int argc, char **argv) {
int rank, size, ierr;
MPI_Datatype oldtypes[5] = {MPI_CHAR, MPI_INT, MPI_CHAR, MPI_UNSIGNED, MPI_UNSIGNED_CHAR};
int blockcounts[5] = {1, 1, 50, 1, 10};
MPI_Datatype my_mpi_struct;
MPI_Aint offsets[5];
struct xyz_t old, new;
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
/* find offsets */
offsets[0] = (char*)&(old.a) - (char*)&old;
offsets[1] = (char*)&(old.b) - (char*)&old;
offsets[2] = (char*)&(old.c) - (char*)&old;
offsets[3] = (char*)&(old.d) - (char*)&old;
offsets[4] = (char*)&(old.e) - (char*)&old;
MPI_Type_create_struct(5, blockcounts, offsets, oldtypes, &my_mpi_struct);
MPI_Type_commit(&my_mpi_struct);
if (rank == 0) {
old.a = 'a';
old.b = (int)'b';
strcpy(old.c,"This is field c");
old.d = (unsigned int)'d';
strcpy(old.e,"Field e");
MPI_Send(&old, 1, my_mpi_struct, 1, 1, MPI_COMM_WORLD);
} else if (rank == 1) {
MPI_Status status;
MPI_Recv(&new, 1, my_mpi_struct, 0, 1, MPI_COMM_WORLD, &status);
printf("new.a = %c\n", new.a);
printf("new.b = %d\n", new.b);
printf("new.e = %s\n", new.e);
}
MPI_Type_free(&my_mpi_struct);
MPI_Finalize();
return 0;
}
Running:
$ mpirun -np 2 ./struct
new.a = a
new.b = 98
new.e = Field e
Updated: As Dave Goodell below points out, the offset calculations would be better done as
#include <stddef.h>
/* ... */
offsets[0] = offsetof(struct xyz_t,a);
offsets[1] = offsetof(struct xyz_t,b);
offsets[2] = offsetof(struct xyz_t,c);
offsets[3] = offsetof(struct xyz_t,d);
offsets[4] = offsetof(struct xyz_t,e);
and if your MPI supports it (most should, though OpenMPI was slow with some of the MPI2.2 types) the MPI_UNSIGNED should be replaced with an MPI_UINT32