Related
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.)
I have a classic problem about the output of sobel filter using CUDA.
this is a main class (main.cpp)
/*main class */
int main(int argc, char** argv)
{
IplImage* image_source = cvLoadImage("test.jpg",
CV_LOAD_IMAGE_GRAYSCALE);
IplImage* image_input = cvCreateImage(cvGetSize(image_source),
IPL_DEPTH_8U,image_source->nChannels);
IplImage* image_output = cvCreateImage(cvGetSize(image_source),
IPL_DEPTH_8U,image_source->nChannels);
/* Convert from IplImage tofloat */
cvConvert(image_source,image_input);
unsigned char *h_out = (unsigned char*)image_output->imageData;
unsigned char *h_in = (unsigned char*)image_input->imageData;
width = image_input->width;
height = image_input->height;
widthStep = image_input->widthStep;
sobel_parallel(h_in, h_out, width, height, widthStep);
cvShowImage( "CPU", image_output );
cvReleaseImage( &image_output );
waitKey(0);
}
And this is the CUDA file (kernel_gpu.cu)
__global__ void kernel ( unsigned char *d_in , unsigned char *d_out , int width ,
int height, int widthStep ) {
int col = blockIdx . x * blockDim . x + threadIdx . x ;
int row = blockIdx . y * blockDim . y + threadIdx . y ;
int dx [3][3] = { -1 , 0 , 1 ,
-2 , 0 , 2 ,
-1 , 0 , 1};
int dy [3][3] = {1 ,2 ,1 ,
0 ,0 ,0 ,
-1 , -2 , -1};
int s;
if( col < width && row < height)
{
int i = row;
int j = col;
// apply kernel in X direction
int sum_x=0;
for(int m=-1; m<=1; m++)
for(int n=-1; n<=1; n++)
{
s=d_in[(i+m)*widthStep+j+n]; // get the (i,j) pixel value
sum_x+=s*dx[m+1][n+1];
}
// apply kernel in Y direction
int sum_y=0;
for(int m=-1; m<=1; m++)
for(int n=-1; n<=1; n++)
{
s=d_in[(i+m)*widthStep+j+n]; // get the (i,j) pixel value
sum_y+=s*dy[m+1][n+1];
}
int sum=abs(sum_x)+abs(sum_y);
if (sum>255)
sum=255;
d_out[i*widthStep+j]=sum; // set the (i,j) pixel value
}
}
// Kernel Calling Function
extern "C" void sobel_parallel( unsigned char* h_in, unsigned char* h_out,
int rows, int cols, int widthStep){
unsigned char* d_in;
unsigned char* d_out;
cudaMalloc((void**) &d_in, rows*cols);
cudaMalloc((void**) &d_out, rows*cols);
cudaMemcpy(d_in, h_in, rows*cols*sizeof( unsigned char), cudaMemcpyHostToDevice);
dim3 block (16,16);
dim3 grid ((rows * cols) / 256.0);
kernel<<<grid,block>>>(d_in, d_out, rows, cols, widthStep);
cudaMemcpy(h_out, d_out, rows*cols*sizeof( unsigned char), cudaMemcpyDeviceToHost);
cudaFree(d_in);
cudaFree(d_out);
}
Error :
the result image does not appear in their entirety, only part of the image.
Why is the result(GPU) like this?? (I tried to make CPU computation using the same function and no problem).
You are creating 1 Dimensional grid, while using 2D indexing inside the kernel which will cover only the x direction and only the top 16 rows of the image will be filtered (because the height of the block is 16).
dim3 grid ((rows * cols) / 256.0); //This is incorrect in current case
Consider creating 2 dimensional grid, so that it spans all the rows of the image.
dim3 grid ((cols + 15)/16, (rows + 15)/16);
Check the width and widthStep variables to see if they are actually equal or not because in your sobel_parallel function you are implicitly assuming this (which might not be true since your data is aligned). If this is not true the code
cudaMalloc((void**) &d_in, rows*cols);
will actually allocate less memory than necessary and hence you will only process part of your image. It would be better to use
cudaMalloc((void**) &d_in, rows*widthStep);
And of course adjust the rest of your code as necessary.
You are also calling
void sobel_parallel( unsigned char* h_in, unsigned char* h_out,
int rows, int cols, int widthStep)
with
sobel_parallel(h_in, h_out, width, height, widthStep);
which exchanges rows with cols and this is again exchanged when you are calling your kernel. This will cause a problem when you use the above suggestion.
I am currently writing a code, that calculates a integral Histogram on the GPU using the Nvidia thrust library.
Therefore I allocate a continuous Block of device memory which I update with a custom functor all the time.
The problem is, that the write to the device memory is veeery slow, but the reads are actually ok.
The basic setup is the following:
struct HistogramCreation
{
HistogramCreation(
...
// pointer to memory
...
){}
/// The actual summation operator
__device__ void operator()(int index){
.. do the calculations ..
for(int j=0;j<30;j++){
(1) *_memoryPointer = values (also using reads to such locations) ;
}
}
}
void foo(){
cudaMalloc(_pointer,size);
HistogramCreation initialCreation( ... _pointer ...);
thrust::for_each(
thrust::make_counting_iterator(0),
thrust::make_counting_iterator(_imageSize),
initialCreation);
}
if I change the writing in (1) to the following>
unsigned int val = values;
The performance is much better. THis is the only global memory write I have.
Using the memory write I get about 2s for HD Footage.
using the local variable it takes about 50 ms so about a factor of 40 less.
Why is this so slow? how could I improve it?
Just as #OlegTitov said, frequent load/store with global
memory should be avoided as much as possible. When there's a
situation where it's inevitable, then coalesced memory
access can help the execution process not to get too slow;
however in most cases, histogram calculation is pretty tough
to realize the coalesced access.
While most of the above is basically just restating
#OlegTitov's answer, i'd just like to share about an
investigation i did about finding summation with NVIDIA
CUDA. Actually the result is pretty interesting and i hope
it'll be a helpful information for other xcuda developers.
The experiment was basically to run a speed test of finding
summation with various memory access patterns: using global
memory (1 thread), L2 cache (atomic ops - 128 threads), and
L1 cache (shared mem - 128 threads)
This experiment used:
Kepler GTX 680,
1546 cores # 1.06GHz
GDDR5 256-bit # 3GHz
Here are the kernels:
__global__
void glob(float *h) {
float* hist = h;
uint sd = SEEDRND;
uint random;
for (int i = 0; i < NUMLOOP; i++) {
if (i%NTHREADS==0) random = rnd(sd);
int rind = random % NBIN;
float randval = (float)(random % 10)*1.0f ;
hist[rind] += randval;
}
}
__global__
void atom(float *h) {
float* hist = h;
uint sd = SEEDRND;
for (int i = threadIdx.x; i < NUMLOOP; i+=NTHREADS) {
uint random = rnd(sd);
int rind = random % NBIN;
float randval = (float)(random % 10)*1.0f ;
atomicAdd(&hist[rind], randval);
}
}
__global__
void shm(float *h) {
int lid = threadIdx.x;
uint sd = SEEDRND;
__shared__ float shm[NTHREADS][NBIN];
for (int i = 0; i < NBIN; i++) shm[lid][i] = h[i];
for (int i = lid; i < NUMLOOP; i+=NTHREADS) {
uint random = rnd(sd);
int rind = random % NBIN;
float randval = (float)(random % 10)*1.0f ;
shm[lid][rind] += randval;
}
/* reduction here */
for (int i = 0; i < NBIN; i++) {
__syncthreads();
if (threadIdx.x < 64) {
shm[threadIdx.x][i] += shm[threadIdx.x+64][i];
}
__syncthreads();
if (threadIdx.x < 32) {
shm[threadIdx.x][i] += shm[threadIdx.x+32][i];
}
__syncthreads();
if (threadIdx.x < 16) {
shm[threadIdx.x][i] += shm[threadIdx.x+16][i];
}
__syncthreads();
if (threadIdx.x < 8) {
shm[threadIdx.x][i] += shm[threadIdx.x+8][i];
}
__syncthreads();
if (threadIdx.x < 4) {
shm[threadIdx.x][i] += shm[threadIdx.x+4][i];
}
__syncthreads();
if (threadIdx.x < 2) {
shm[threadIdx.x][i] += shm[threadIdx.x+2][i];
}
__syncthreads();
if (threadIdx.x == 0) {
shm[0][i] += shm[1][i];
}
}
for (int i = 0; i < NBIN; i++) h[i] = shm[0][i];
}
OUTPUT
atom: 102656.00 shm: 102656.00 glob: 102656.00
atom: 122240.00 shm: 122240.00 glob: 122240.00
... blah blah blah ...
One Thread: 126.3919 msec
Atomic: 7.5459 msec
Sh_mem: 2.2207 msec
The ratio between these kernels is 57:17:1. Many things can
be analyzed here, and it truly does not mean that using
L1 or L2 memory spaces will always give you more than 10
times speedup of the whole program.
And here's the main and other funcs:
#include <iostream>
#include <cstdlib>
#include <cstdio>
using namespace std;
#define NUMLOOP 1000000
#define NBIN 36
#define SEEDRND 1
#define NTHREADS 128
#define NBLOCKS 1
__device__ uint rnd(uint & seed) {
#if LONG_MAX > (16807*2147483647)
int const a = 16807;
int const m = 2147483647;
seed = (long(seed * a))%m;
return seed;
#else
double const a = 16807;
double const m = 2147483647;
double temp = seed * a;
seed = (int) (temp - m * floor(temp/m));
return seed;
#endif
}
... the above kernels ...
int main()
{
float *h_hist, *h_hist2, *h_hist3, *d_hist, *d_hist2,
*d_hist3;
h_hist = (float*)malloc(NBIN * sizeof(float));
h_hist2 = (float*)malloc(NBIN * sizeof(float));
h_hist3 = (float*)malloc(NBIN * sizeof(float));
cudaMalloc((void**)&d_hist, NBIN * sizeof(float));
cudaMalloc((void**)&d_hist2, NBIN * sizeof(float));
cudaMalloc((void**)&d_hist3, NBIN * sizeof(float));
for (int i = 0; i < NBIN; i++) h_hist[i] = 0.0f;
cudaMemcpy(d_hist, h_hist, NBIN * sizeof(float),
cudaMemcpyHostToDevice);
cudaMemcpy(d_hist2, h_hist, NBIN * sizeof(float),
cudaMemcpyHostToDevice);
cudaMemcpy(d_hist3, h_hist, NBIN * sizeof(float),
cudaMemcpyHostToDevice);
cudaEvent_t start, end;
float elapsed = 0, elapsed2 = 0, elapsed3;
cudaEventCreate(&start);
cudaEventCreate(&end);
cudaEventRecord(start, 0);
atom<<<NBLOCKS, NTHREADS>>>(d_hist);
cudaThreadSynchronize();
cudaEventRecord(end, 0);
cudaEventSynchronize(start);
cudaEventSynchronize(end);
cudaEventElapsedTime(&elapsed, start, end);
cudaEventRecord(start, 0);
shm<<<NBLOCKS, NTHREADS>>>(d_hist2);
cudaThreadSynchronize();
cudaEventRecord(end, 0);
cudaEventSynchronize(start);
cudaEventSynchronize(end);
cudaEventElapsedTime(&elapsed2, start, end);
cudaEventRecord(start, 0);
glob<<<1, 1>>>(d_hist3);
cudaThreadSynchronize();
cudaEventRecord(end, 0);
cudaEventSynchronize(start);
cudaEventSynchronize(end);
cudaEventElapsedTime(&elapsed3, start, end);
cudaMemcpy(h_hist, d_hist, NBIN * sizeof(float),
cudaMemcpyDeviceToHost);
cudaMemcpy(h_hist2, d_hist2, NBIN * sizeof(float),
cudaMemcpyDeviceToHost);
cudaMemcpy(h_hist3, d_hist3, NBIN * sizeof(float),
cudaMemcpyDeviceToHost);
/* print output */
for (int i = 0; i < NBIN; i++) {
printf("atom: %10.2f shm: %10.2f glob:
%10.2f¥n",h_hist[i],h_hist2[i],h_hist3[i]);
}
printf("%12s: %8.4f msec¥n", "One Thread", elapsed3);
printf("%12s: %8.4f msec¥n", "Atomic", elapsed);
printf("%12s: %8.4f msec¥n", "Sh_mem", elapsed2);
return 0;
}
When writing GPU code you should avoid reading and writing to/from global memory. Global memory is very slow on GPU. That's the hardware feature. The only thing you can do is to make neighboring treads read/write in neighboring adresses in global memory. This will cause coalescing and speed up the process. But in general read your data once, process it and write it out once.
Note that NVCC might optimize out a lot of your code after you make the modification - it detects that no write to global memory is made and just removes the "unneeded" code. So this speedup may not be coming out of the global writer per ce.
I would recommend using profiler on your actual code (the one with global write) to see if there's anything like unaligned access or other perf problem.
OK, so lets say I have an ( N x N ) matrix that I would like to process. This matrix is quite large for my computer, and if I try to send it to the device all at once I get a 'out of memory error.'
So is there a way to send sections of the matrix to the device? One way I can see to do it is copy portions of the matrix on the host, and then send these manageable copied portions from the host to the device, and then put them back together at the end.
Here is something I have tried, but the cudaMemcpy in the for loop returns error code 11, 'invalid argument.'
int h_N = 10000;
size_t h_size_m = h_N*sizeof(float);
h_A = (float*)malloc(h_size_m*h_size_m);
int d_N = 2500;
size_t d_size_m = d_N*sizeof(float);
InitializeMatrices(h_N);
int i;
int iterations = (h_N*h_N)/(d_N*d_N);
for( i = 0; i < iterations; i++ )
{
float* h_array_ref = h_A+(i*d_N*d_N);
cudasafe( cudaMemcpy(d_A, h_array_ref, d_size_m*d_size_m, cudaMemcpyHostToDevice), "cudaMemcpy");
cudasafe( cudaFree(d_A), "cudaFree(d_A)" );
}
What I'm trying to accomplish with the above code is this: instead of send the entire matrix to the device, I simply send a pointer to a place within that matrix and reserve enough space on the device to do the work, and then with the next iteration of the loop move the pointer forward within the matrix, etc. etc.
Not only can you do this (assuming your problem is easily decomposed this way into sub-arrays), it can be a very useful thing to do for performance; once you get the basic approach you've described working, you can start using asynchronous memory copies and double-buffering to overlap some of the memory transfer time with the time spent computing what is already on-card.
But first one gets the simple thing working. Below is a 1d example (multiplying a vector by a scalar and adding another scalar) but using a linearized 2d array would be the same; the key part is
CHK_CUDA( cudaMalloc(&xd, batchsize*sizeof(float)) );
CHK_CUDA( cudaMalloc(&yd, batchsize*sizeof(float)) );
tick(&gputimer);
int nbatches = 0;
for (int nstart=0; nstart < n; nstart+=batchsize) {
int size=batchsize;
if ((nstart + batchsize) > n) size = n - nstart;
CHK_CUDA( cudaMemcpy(xd, &(x[nstart]), size*sizeof(float), cudaMemcpyHostToDevice) );
blocksize = (size+nblocks-1)/nblocks;
cuda_saxpb<<<nblocks, blocksize>>>(xd, a, b, yd, size);
CHK_CUDA( cudaMemcpy(&(ycuda[nstart]), yd, size*sizeof(float), cudaMemcpyDeviceToHost) );
nbatches++;
}
gputime = tock(&gputimer);
CHK_CUDA( cudaFree(xd) );
CHK_CUDA( cudaFree(yd) );
You allocate the buffers at the start, and then loop through until you're done, each time doing the copy, starting the kernel, and then copying back. You free at the end.
The full code is
#include <stdio.h>
#include <stdlib.h>
#include <getopt.h>
#include <cuda.h>
#include <sys/time.h>
#include <math.h>
#define CHK_CUDA(e) {if (e != cudaSuccess) {fprintf(stderr,"Error: %s\n", cudaGetErrorString(e)); exit(-1);}}
__global__ void cuda_saxpb(const float *xd, const float a, const float b,
float *yd, const int n) {
int i = threadIdx.x + blockIdx.x*blockDim.x;
if (i<n) {
yd[i] = a*xd[i]+b;
}
return;
}
void cpu_saxpb(const float *x, float a, float b, float *y, int n) {
int i;
for (i=0;i<n;i++) {
y[i] = a*x[i]+b;
}
return;
}
int get_options(int argc, char **argv, int *n, int *s, int *nb, float *a, float *b);
void tick(struct timeval *timer);
double tock(struct timeval *timer);
int main(int argc, char **argv) {
int n=1000;
int nblocks=10;
int batchsize=100;
float a = 5.;
float b = -1.;
int err;
float *x, *y, *ycuda;
float *xd, *yd;
double abserr;
int blocksize;
int i;
struct timeval cputimer;
struct timeval gputimer;
double cputime, gputime;
err = get_options(argc, argv, &n, &batchsize, &nblocks, &a, &b);
if (batchsize > n) {
fprintf(stderr, "Resetting batchsize to size of vector, %d\n", n);
batchsize = n;
}
if (err) return 0;
x = (float *)malloc(n*sizeof(float));
if (!x) return 1;
y = (float *)malloc(n*sizeof(float));
if (!y) {free(x); return 1;}
ycuda = (float *)malloc(n*sizeof(float));
if (!ycuda) {free(y); free(x); return 1;}
/* run CPU code */
tick(&cputimer);
cpu_saxpb(x, a, b, y, n);
cputime = tock(&cputimer);
/* run GPU code */
/* only have to allocate once */
CHK_CUDA( cudaMalloc(&xd, batchsize*sizeof(float)) );
CHK_CUDA( cudaMalloc(&yd, batchsize*sizeof(float)) );
tick(&gputimer);
int nbatches = 0;
for (int nstart=0; nstart < n; nstart+=batchsize) {
int size=batchsize;
if ((nstart + batchsize) > n) size = n - nstart;
CHK_CUDA( cudaMemcpy(xd, &(x[nstart]), size*sizeof(float), cudaMemcpyHostToDevice) );
blocksize = (size+nblocks-1)/nblocks;
cuda_saxpb<<<nblocks, blocksize>>>(xd, a, b, yd, size);
CHK_CUDA( cudaMemcpy(&(ycuda[nstart]), yd, size*sizeof(float), cudaMemcpyDeviceToHost) );
nbatches++;
}
gputime = tock(&gputimer);
CHK_CUDA( cudaFree(xd) );
CHK_CUDA( cudaFree(yd) );
abserr = 0.;
for (i=0;i<n;i++) {
abserr += fabs(ycuda[i] - y[i]);
}
printf("Y = a*X + b, problemsize = %d\n", n);
printf("CPU time = %lg millisec.\n", cputime*1000.);
printf("GPU time = %lg millisec (done with %d batches of %d).\n",
gputime*1000., nbatches, batchsize);
printf("CUDA and CPU results differ by %lf\n", abserr);
free(x);
free(y);
free(ycuda);
return 0;
}
int get_options(int argc, char **argv, int *n, int *s, int *nb, float *a, float *b) {
const struct option long_options[] = {
{"nvals" , required_argument, 0, 'n'},
{"nblocks" , required_argument, 0, 'B'},
{"batchsize" , required_argument, 0, 's'},
{"a", required_argument, 0, 'a'},
{"b", required_argument, 0, 'b'},
{"help", no_argument, 0, 'h'},
{0, 0, 0, 0}};
char c;
int option_index;
int tempint;
while (1) {
c = getopt_long(argc, argv, "n:B:a:b:s:h", long_options, &option_index);
if (c == -1) break;
switch(c) {
case 'n': tempint = atoi(optarg);
if (tempint < 1 || tempint > 500000) {
fprintf(stderr,"%s: Cannot use number of points %s;\n Using %d\n", argv[0], optarg, *n);
} else {
*n = tempint;
}
break;
case 's': tempint = atoi(optarg);
if (tempint < 1 || tempint > 50000) {
fprintf(stderr,"%s: Cannot use number of points %s;\n Using %d\n", argv[0], optarg, *s);
} else {
*s = tempint;
}
break;
case 'B': tempint = atoi(optarg);
if (tempint < 1 || tempint > 1000 || tempint > *n) {
fprintf(stderr,"%s: Cannot use number of blocks %s;\n Using %d\n", argv[0], optarg, *nb);
} else {
*nb = tempint;
}
break;
case 'a': *a = atof(optarg);
break;
case 'b': *b = atof(optarg);
break;
case 'h':
puts("Calculates y[i] = a*x[i] + b on the GPU.");
puts("Options: ");
puts(" --nvals=N (-n N): Set the number of values in y,x.");
puts(" --batchsize=N (-s N): Set the number of values to transfer at a time.");
puts(" --nblocks=N (-B N): Set the number of blocks used.");
puts(" --a=X (-a X): Set the parameter a.");
puts(" --b=X (-b X): Set the parameter b.");
puts(" --niters=N (-I X): Set number of iterations to calculate.");
puts("");
return +1;
}
}
return 0;
}
void tick(struct timeval *timer) {
gettimeofday(timer, NULL);
}
double tock(struct timeval *timer) {
struct timeval now;
gettimeofday(&now, NULL);
return (now.tv_usec-timer->tv_usec)/1.0e6 + (now.tv_sec - timer->tv_sec);
}
Running this one gets:
$ ./batched-saxpb --nvals=10240 --batchsize=10240 --nblocks=20
Y = a*X + b, problemsize = 10240
CPU time = 0.072 millisec.
GPU time = 0.117 millisec (done with 1 batches of 10240).
CUDA and CPU results differ by 0.000000
$ ./batched-saxpb --nvals=10240 --batchsize=5120 --nblocks=20
Y = a*X + b, problemsize = 10240
CPU time = 0.066 millisec.
GPU time = 0.133 millisec (done with 2 batches of 5120).
CUDA and CPU results differ by 0.000000
$ ./batched-saxpb --nvals=10240 --batchsize=2560 --nblocks=20
Y = a*X + b, problemsize = 10240
CPU time = 0.067 millisec.
GPU time = 0.167 millisec (done with 4 batches of 2560).
CUDA and CPU results differ by 0.000000
The GPU time goes up in this case (we're doing more memory copies) but the answers stay the same.
Edited: The original version of this code had an option for running multiple iterations of the kernel for timing purposes, but that's unnecessarily confusing in this context so it's removed.
The following is the code I wrote for adding 2 2d arrays. Its getting compiled but when I try to run it its showing : error: failed to build program. runtime0.0000
why is it that the prpgram isnt built?
And also why is it that the buildlog that I have queried isnt getting displayed?
Actually since I am just initialising the arrays, I have directly stored to 1d array, not shown the conversion from 2d to 1d.
code:
# include <stdio.h>
#include <stdlib.h>
#ifdef APPLE
#include<OpenCL/opencl.h>
#else
#include <CL/cl.h>
#endif
#define order 1000
#define MAX_SOURCE_SIZE (0x100000)
int main(int argc, char **argv)
{
float *A;
float *B;
float *C;
int n,m,p;
int err;
int szA, szB,szC;
cl_device_id device_id;
cl_context context;
cl_command_queue commands;
cl_program program;
cl_kernel kernel;
cl_uint nd;
cl_mem a_in;
cl_mem b_in;
cl_mem c_out;
int i,j;
n=order;
m=order;
p=order;
size_t global[2];
nd=1;
cl_uint numPlatforms;
cl_platform_id firstPlatformId;
szA=n*p;
szB=p*m;
szC=n*m;
A=(float *)malloc(sizeof(float)*szA);
B=(float *)malloc(sizeof(float)*szB);
C=(float *)malloc(sizeof(float)*szC);
for(i=0; i<order; i++)
for(j=0; j<order; j++)
A[i*m+j]=i;
B[i*m+j]=i;
FILE *fp;
char fileName[] = "./array_add_kernel.cl";
char *source_str;
size_t source_size;
fp = fopen(fileName, "r");
if (!fp) {
fprintf(stderr, "Failed to load kernel.\n");
exit(1);
}
source_str = (char*)malloc(MAX_SOURCE_SIZE);
source_size = fread( source_str, 1, MAX_SOURCE_SIZE, fp);
fclose( fp );
err=clGetPlatformIDs(1, &firstPlatformId, &numPlatforms);
err=clGetDeviceIDs(firstPlatformId, CL_DEVICE_TYPE_GPU, 1, &device_id, NULL);
cl_context_properties conpro[]={ CL_CONTEXT_PLATFORM,(cl_context_properties) firstPlatformId, 0};
context=clCreateContext(conpro, 1, &device_id, NULL, NULL, &err);
commands=clCreateCommandQueue(context, device_id,CL_QUEUE_PROFILING_ENABLE, &err);
a_in= clCreateBuffer(context, CL_MEM_READ_ONLY, sizeof(float)*szA, NULL, NULL);
b_in= clCreateBuffer(context, CL_MEM_READ_ONLY, sizeof(float)*szB, NULL, NULL);
c_out= clCreateBuffer(context, CL_MEM_WRITE_ONLY, sizeof(float)*szC, NULL, NULL);
program= clCreateProgramWithSource(context, 1, (const char**)&source_str,(const size_t *)&source_size, &err);
err= clBuildProgram(program,0, NULL, NULL, NULL, NULL );
if(err!= CL_SUCCESS)
{
size_t len;
char buffer[2048];
printf("Error:Failed to build program executable!");
clGetProgramBuildInfo(program,device_id,CL_PROGRAM_BUILD_LOG,sizeof(buffer),buffer,&len);
printf("%s \n",buffer);
}
kernel= clCreateKernel(program, "array_add_kernel", &err);
err= 0;
err= clSetKernelArg(kernel, 0, sizeof(int), &n);
err|= clSetKernelArg(kernel, 1, sizeof(int), &p);
err|= clSetKernelArg(kernel, 2, sizeof(int), &m);
err|= clSetKernelArg(kernel, 3, sizeof(cl_mem), &a_in);
err|= clSetKernelArg(kernel, 4, sizeof(cl_mem), &b_in);
err|= clSetKernelArg(kernel, 5, sizeof(cl_mem), &c_out);
err=clEnqueueWriteBuffer(commands, a_in, CL_TRUE, 0, sizeof(float)*szA, A, 0, NULL, NULL);
err= clEnqueueWriteBuffer(commands, a_in, CL_TRUE, 0, sizeof(float)*szB, B, 0, NULL, NULL);
cl_event prof_event;
global[0]= (size_t)n;
global[1]=(size_t)m;
err=clEnqueueNDRangeKernel(commands, kernel, nd, NULL, global, NULL, 0, NULL, &prof_event);
clFinish(commands);
cl_ulong ev_start_time=(cl_ulong)0;
cl_ulong ev_end_time=(cl_ulong)0;
size_t ret_size;
err= clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &ev_start_time, NULL);
err= clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &ev_end_time, NULL);
err=clEnqueueReadBuffer(commands,c_out,CL_TRUE,0,sizeof(float)*szC,C,0,NULL,NULL);
cl_float runtime=(ev_end_time-ev_start_time)*1.0e-9;
printf("Runtime:%f ",runtime);
clReleaseProgram(program);
clReleaseKernel(kernel);
clReleaseMemObject(a_in);
clReleaseMemObject(b_in);
clReleaseMemObject(c_out);
clReleaseCommandQueue(commands);
clReleaseContext(context);
}
kernel:
kernel void array_add_kernel(
const int n, const int m, const p, _global const float * A, _global const float * B, , _global float * C )
{
int i= get_global_id(0);
int j= get_global_id(1);
C[i*m + j] = A[i*m + j] + B[i*m + j];
}
Fix your kernel. It's filled with errors.
kernel void array_add_kernel(
const int n,
const int m,
const p, // No type specifier
_global const float * A, // Should be global, not _global
_global const float * B, , // Double comma
_global float * C )
{
int i= get_global_id(0);
int j= get_global_id(1);
C[i*m + j] = A[i*m + j] + B[i*m + j];
}
This is the working kernel.
kernel void array_add_kernel(const int n, const int m, global const float * A, global const float * B, global float * C )
{
int i= get_global_id(0);
int j= get_global_id(1);
C[i*m + j] = A[i*m + j] + B[i*m + j];
}