unsigned int readmailbox(unsigned int channel) {
unsigned int count = 0;
unsigned int data;
/* Loop until something is received from channel
* If nothing recieved, it eventually give up and returns 0xffffffff
*/
while(1) {
while (*MAILBOX0STATUS & MAILBOX_EMPTY) {
/* Need to check if this is the right thing to do */
flushcache();
/* This is an arbritarily large number */
if(count++ >(1<<25)) {
return 0xffffffff;
}
}
/* Read the data
* Data memory barriers as we've switched peripheral
*/
dmb();
data = *MAILBOX0READ;
dmb();
if ((data & 15) == channel)
return data;
}
return 0;
}
static int
aoerio(SDreq *r)
{
int i, count, rw;
uvlong lba;
Ctlr *c;
SDunit *u;
u = r->unit;
c = u->dev->ctlr;
// if(c->feat & Datapi)
// return aoeriopkt(r, d);
if(r->cmd[0] == 0x35 || r->cmd[0] == 0x91) {
qlock(c);
i = flushcache(c);
qunlock(c);
if(i == 0)
return sdsetsense(r, SDok, 0, 0, 0);
return sdsetsense(r, SDcheck, 3, 0xc, 2);
}
if((i = sdfakescsi(r)) != SDnostatus) {
r->status = i;
return i;
}
if((i = sdfakescsirw(r, &lba, &count, &rw)) != SDnostatus)
return i;
r->rlen = aoebio(u, r->lun, rw == SDwrite, r->data, count, lba);
return r->status = SDok;
}
int main(){
unsigned long max_offset;
unsigned long buffer_size=1;
int fd = open("10GigFile",O_RDONLY|O_LARGEFILE);
if(fd == -1)
{
perror("file could not be opened for reading\n");
exit(1);
}
char* buffer;
flushcache();
max_offset = (10240*1024/buffer_size)-1;
buffer = (char*) malloc(buffer_size*1024);
run_experiment(fd,buffer,buffer_size*1024,max_offset);
free(buffer);
flushcache();
srand(time(NULL)%INT_MAX);
}
void writemailbox(unsigned int channel, unsigned int data) {
/* Wait for mailbox to be not full */
while (*MAILBOX0STATUS & MAILBOX_FULL) {
/* Need to check if this is the right thing to do */
flushcache();
}
dmb();
*MAILBOX0WRITE = (data | channel);
}
static void acpi_test(int ac, char **av)
{
unsigned int i;
unsigned int k;
unsigned int sleep_gpio_save[13];
ls1a_save_gpio_val(sleep_gpio_save);
suspend_save();
i = strtoul(av[1], 0, 0);
// printf ("you set %d for test !\n", i);
if (i & 1)
*(volatile unsigned int *)0xbfe7c004 = 1 << 8; //power button
k = *(volatile unsigned int *)0xbfe7c024;
k &= ~((1 << 8) | (1 << 9));
*(volatile unsigned int *)0xbfe7c024 = (i & 0x3) << 8; //[8:9]RI_EN、PME_EN
// *(volatile unsigned int *)0xbfe7c008 = (1 << 13) | (5 << 10); //sleep to ram
flushcache();
__asm__ volatile (
"la $2, 2f\n\t"
"li $3, 0xa01ffc00\n\t"
"sw $2, 0x0($3)\n\t" //save return address
"li $2, 0xaffffe34\n\t"
"lw $3, 0x0($2)\n\t"
"or $3, 0x1\n\t"
"sw $3, 0x0($2)\n\t" //enable ddr autorefresh
"li $2, 0xbfe7c008\n\t"
"li $3, (1<<13) | (5<<10)\n\t"
"sw $3, 0x0($2)\n\t" //go to sleep
"1:\n\t"
"bal 1b\n\t"
"nop\n\t"
"2:\n\t"
::
: "$2","$3","memory"
// "move %0,$2\n\t"
// : "=r" (p)
// : "0" (p), "r" (len), "r" (1)
// : "$2","$3","$4","$5"
);
ls1a_restore_gpio_val(sleep_gpio_save);
dc_init();
outstring("acpi resume back here !\n");
// printf ("acpi resume back here !\n");
}
/*
* Pseudo (chained) interrupt from the esp driver to kick the
* current running DMA transfer. I am replying on espintr() to
* pickup and clean errors for now
*
* return 1 if it was a DMA continue.
*/
int
espdmaintr(struct esp_softc *sc)
{
struct ncr53c9x_softc *nsc = (struct ncr53c9x_softc *)sc;
int trans, resid;
u_long csr = sc->sc_dma_direction;
#if 0
if (csr & D_ERR_PEND) {
DMACSR(sc) &= ~D_EN_DMA; /* Stop DMA */
DMACSR(sc) |= D_INVALIDATE;
printf("%s: error: csr=%s\n", device_xname(nsc->sc_dev),
bitmask_snprintf(csr, DMACSRBITS, bits, sizeof(bits)));
return -1;
}
#endif
/* This is an "assertion" :) */
if (sc->sc_dmaactive == 0)
panic("%s: DMA wasn't active", __func__);
/* dbdma_flush(sc->sc_dmareg); */
/* DMA has stopped */
dbdma_stop(sc->sc_dmareg);
sc->sc_dmaactive = 0;
if (sc->sc_dmasize == 0) {
/* A "Transfer Pad" operation completed */
NCR_DMA(("dmaintr: discarded %d bytes (tcl=%d, tcm=%d)\n",
NCR_READ_REG(nsc, NCR_TCL) |
(NCR_READ_REG(nsc, NCR_TCM) << 8),
NCR_READ_REG(nsc, NCR_TCL),
NCR_READ_REG(nsc, NCR_TCM)));
return 0;
}
resid = 0;
/*
* If a transfer onto the SCSI bus gets interrupted by the device
* (e.g. for a SAVEPOINTER message), the data in the FIFO counts
* as residual since the ESP counter registers get decremented as
* bytes are clocked into the FIFO.
*/
if (!(csr & D_WRITE) &&
(resid = (NCR_READ_REG(nsc, NCR_FFLAG) & NCRFIFO_FF)) != 0) {
NCR_DMA(("dmaintr: empty esp FIFO of %d ", resid));
}
if ((nsc->sc_espstat & NCRSTAT_TC) == 0) {
/*
* `Terminal count' is off, so read the residue
* out of the ESP counter registers.
*/
resid += (NCR_READ_REG(nsc, NCR_TCL) |
(NCR_READ_REG(nsc, NCR_TCM) << 8) |
((nsc->sc_cfg2 & NCRCFG2_FE)
? (NCR_READ_REG(nsc, NCR_TCH) << 16)
: 0));
if (resid == 0 && sc->sc_dmasize == 65536 &&
(nsc->sc_cfg2 & NCRCFG2_FE) == 0)
/* A transfer of 64K is encoded as `TCL=TCM=0' */
resid = 65536;
}
trans = sc->sc_dmasize - resid;
if (trans < 0) { /* transferred < 0 ? */
#if 0
/*
* This situation can happen in perfectly normal operation
* if the ESP is reselected while using DMA to select
* another target. As such, don't print the warning.
*/
printf("%s: xfer (%d) > req (%d)\n",
device_xname(nsc->sc_dev), trans, sc->sc_dmasize);
#endif
trans = sc->sc_dmasize;
}
NCR_DMA(("dmaintr: tcl=%d, tcm=%d, tch=%d; trans=%d, resid=%d\n",
NCR_READ_REG(nsc, NCR_TCL),
NCR_READ_REG(nsc, NCR_TCM),
(nsc->sc_cfg2 & NCRCFG2_FE)
? NCR_READ_REG(nsc, NCR_TCH) : 0,
trans, resid));
#if 0
if (csr & D_WRITE)
flushcache(*sc->sc_dmaaddr, trans);
#endif
*sc->sc_dmalen -= trans;
*sc->sc_dmaaddr += trans;
#if 0 /* this is not normal operation just yet */
if (*sc->sc_dmalen == 0 ||
nsc->sc_phase != nsc->sc_prevphase)
return 0;
/* and again */
dma_start(sc, sc->sc_dmaaddr, sc->sc_dmalen, DMACSR(sc) & D_WRITE);
return 1;
#endif
return 0;
}
int main(int argc, char **argv)
{
FILE *out; /* Output data file */
char s[255],s2[255],delim[255],*pstr; /* Generic strings */
int *memcache; /* used to flush cache */
int len_buf_align, /* meaningful when args.cache is 0. buflen */
/* rounded up to be divisible by 8 */
num_buf_align; /* meaningful when args.cache is 0. number */
/* of aligned buffers in memtmp */
int c, /* option index */
i, j, n, nq, /* Loop indices */
asyncReceive=0, /* Pre-post a receive buffer? */
bufalign=16*1024,/* Boundary to align buffer to */
errFlag, /* Error occurred in inner testing loop */
nrepeat, /* Number of time to do the transmission */
nrepeat_const=0,/* Set if we are using a constant nrepeat */
len, /* Number of bytes to be transmitted */
inc=0, /* Increment value */
perturbation=DEFPERT, /* Perturbation value */
pert,
start= 1, /* Starting value for signature curve */
end=MAXINT, /* Ending value for signature curve */
streamopt=0, /* Streaming mode flag */
reset_connection;/* Reset the connection between trials */
ArgStruct args; /* Arguments for all the calls */
double t, t0, t1, t2, /* Time variables */
tlast, /* Time for the last transmission */
latency; /* Network message latency */
Data bwdata[NSAMP]; /* Bandwidth curve data */
int integCheck=0; /* Integrity check */
/* Initialize vars that may change from default due to arguments */
strcpy(s, "np.out"); /* Default output file */
/* Let modules initialize related vars, and possibly call a library init
function that requires argc and argv */
Init(&args, &argc, &argv); /* This will set args.tr and args.rcv */
args.preburst = 0; /* Default to not bursting preposted receives */
args.bidir = 0; /* Turn bi-directional mode off initially */
args.cache = 1; /* Default to use cache */
args.upper = end;
args.host = NULL;
args.soffset=0; /* default to no offsets */
args.roffset=0;
args.syncflag=0; /* use normal mpi_send */
args.port = DEFPORT; /* just in case the user doesn't set this. */
/* TCGMSG launches NPtcgmsg with a -master master_hostname
* argument, so ignore all arguments and set them manually
* in netpipe.c instead.
*/
#if ! defined(TCGMSG)
/* Parse the arguments. See Usage for description */
while ((c = getopt(argc, argv, "SO:rIiszgfaB2h:p:o:l:u:b:m:n:t:c:d:D:P:")) != -1)
{
switch(c)
{
case 'O':
strcpy(s2,optarg);
strcpy(delim,",");
if((pstr=strtok(s2,delim))!=NULL) {
args.soffset=atoi(pstr);
if((pstr=strtok((char *)NULL,delim))!=NULL)
args.roffset=atoi(pstr);
else /* only got one token */
args.roffset=args.soffset;
} else {
args.soffset=0; args.roffset=0;
}
printf("Transmit buffer offset: %d\nReceive buffer offset: %d\n",args.soffset,args.roffset);
break;
case 'p': perturbation = atoi(optarg);
if( perturbation > 0 ) {
printf("Using a perturbation value of %d\n\n", perturbation);
} else {
perturbation = 0;
printf("Using no perturbations\n\n");
}
break;
case 'B': if(integCheck == 1) {
fprintf(stderr, "Integrity check not supported with prepost burst\n");
exit(-1);
}
args.preburst = 1;
asyncReceive = 1;
printf("Preposting all receives before a timed run.\n");
printf("Some would consider this cheating,\n");
printf("but it is needed to match some vendor tests.\n"); fflush(stdout);
break;
case 'I': args.cache = 0;
printf("Performance measured without cache effects\n\n"); fflush(stdout);
break;
case 'o': strcpy(s,optarg);
printf("Sending output to %s\n", s); fflush(stdout);
break;
case 's': streamopt = 1;
printf("Streaming in one direction only.\n\n");
#if defined(TCP) && ! defined(INFINIBAND)
printf("Sockets are reset between trials to avoid\n");
printf("degradation from a collapsing window size.\n\n");
#endif
args.reset_conn = 1;
printf("Streaming does not provide an accurate\n");
printf("measurement of the latency since small\n");
printf("messages may get bundled together.\n\n");
if( args.bidir == 1 ) {
printf("You can't use -s and -2 together\n");
exit(0);
}
fflush(stdout);
break;
case 'l': start = atoi(optarg);
if (start < 1)
{
fprintf(stderr,"Need a starting value >= 1\n");
exit(0);
}
break;
case 'u': end = atoi(optarg);
break;
#if defined(TCP) && ! defined(INFINIBAND)
case 'b': /* -b # resets the buffer size, -b 0 keeps system defs */
args.prot.sndbufsz = args.prot.rcvbufsz = atoi(optarg);
break;
#endif
case '2': args.bidir = 1; /* Both procs are transmitters */
/* end will be maxed at sndbufsz+rcvbufsz */
printf("Passing data in both directions simultaneously.\n");
printf("Output is for the combined bandwidth.\n");
#if defined(TCP) && ! defined(INFINIBAND)
printf("The socket buffer size limits the maximum test size.\n\n");
#endif
if( streamopt ) {
printf("You can't use -s and -2 together\n");
exit(0);
}
break;
case 'h': args.tr = 1; /* -h implies transmit node */
args.rcv = 0;
args.host = (char *)malloc(strlen(optarg)+1);
strcpy(args.host, optarg);
break;
#ifdef DISK
case 'd': args.tr = 1; /* -d to specify input/output file */
args.rcv = 0;
args.prot.read = 0;
args.prot.read_type = 'c';
args.prot.dfile_name = (char *)malloc(strlen(optarg)+1);
strcpy(args.prot.dfile_name, optarg);
break;
case 'D': if( optarg[0] == 'r' )
args.prot.read = 1;
else
args.prot.read = 0;
args.prot.read_type = optarg[1];
break;
#endif
case 'i': if(args.preburst == 1) {
fprintf(stderr, "Integrity check not supported with prepost burst\n");
exit(-1);
}
integCheck = 1;
perturbation = 0;
start = sizeof(int)+1; /* Start with integer size */
printf("Doing an integrity check instead of measuring performance\n"); fflush(stdout);
break;
#if defined(MPI)
case 'z': args.source_node = -1;
printf("Receive using the ANY_SOURCE flag\n"); fflush(stdout);
break;
case 'a': asyncReceive = 1;
printf("Preposting asynchronous receives\n"); fflush(stdout);
break;
case 'S': args.syncflag=1;
fprintf(stderr,"Using synchronous sends\n");
break;
#endif
#if defined(MPI2)
case 'g': if(args.prot.no_fence == 1) {
fprintf(stderr, "-f cannot be used with -g\n");
exit(-1);
}
args.prot.use_get = 1;
printf("Using MPI-2 Get instead of Put\n");
break;
case 'f': if(args.prot.use_get == 1) {
fprintf(stderr, "-f cannot be used with -g\n");
exit(-1);
}
args.prot.no_fence = 1;
bufalign = 0;
printf("Buffer alignment off (Required for no fence)\n");
break;
#endif /* MPI2 */
#if defined(INFINIBAND)
case 'm': switch(atoi(optarg)) {
case 256: args.prot.ib_mtu = MTU256;
break;
case 512: args.prot.ib_mtu = MTU512;
break;
case 1024: args.prot.ib_mtu = MTU1024;
break;
case 2048: args.prot.ib_mtu = MTU2048;
break;
case 4096: args.prot.ib_mtu = MTU4096;
break;
default:
fprintf(stderr, "Invalid MTU size, must be one of "
"256, 512, 1024, 2048, 4096\n");
exit(-1);
}
break;
case 't': if( !strcmp(optarg, "send_recv") ) {
printf("Using Send/Receive communications\n");
args.prot.commtype = NP_COMM_SENDRECV;
} else if( !strcmp(optarg, "send_recv_with_imm") ) {
printf("Using Send/Receive communications with immediate data\n");
args.prot.commtype = NP_COMM_SENDRECV_WITH_IMM;
} else if( !strcmp(optarg, "rdma_write") ) {
printf("Using RDMA Write communications\n");
args.prot.commtype = NP_COMM_RDMAWRITE;
} else if( !strcmp(optarg, "rdma_write_with_imm") ) {
printf("Using RDMA Write communications with immediate data\n");
args.prot.commtype = NP_COMM_RDMAWRITE_WITH_IMM;
} else {
fprintf(stderr, "Invalid transfer type "
"specified, please choose one of:\n\n"
"\tsend_recv\t\tUse Send/Receive communications\t(default)\n"
"\tsend_recv_with_imm\tSame as above with immediate data\n"
"\trdma_write\t\tUse RDMA Write communications\n"
"\trdma_write_with_imm\tSame as above with immediate data\n\n");
exit(-1);
}
break;
case 'c': if( !strcmp(optarg, "local_poll") ) {
printf("Using local polling completion\n");
args.prot.comptype = NP_COMP_LOCALPOLL;
} else if( !strcmp(optarg, "vapi_poll") ) {
printf("Using VAPI polling completion\n");
args.prot.comptype = NP_COMP_VAPIPOLL;
} else if( !strcmp(optarg, "event") ) {
printf("Using VAPI event completion\n");
args.prot.comptype = NP_COMP_EVENT;
} else {
fprintf(stderr, "Invalid completion type specified, "
"please choose one of:\n\n"
"\tlocal_poll\tWait for last byte of data\t(default)\n"
"\tvapi_poll\tUse VAPI polling function\n"
"\tevent\t\tUse VAPI event handling function\n\n");
exit(-1);
}
break;
#endif
case 'n': nrepeat_const = atoi(optarg);
break;
#if defined(TCP) && ! defined(INFINIBAND)
case 'r': args.reset_conn = 1;
printf("Resetting connection after every trial\n");
break;
#endif
case 'P':
args.port = atoi(optarg);
break;
default:
PrintUsage();
exit(-12);
}
}
#endif /* ! defined TCGMSG */
#if defined(INFINIBAND)
asyncReceive = 1;
fprintf(stderr, "Preposting asynchronous receives (required for Infiniband)\n");
if(args.bidir && (
(args.cache && args.prot.commtype == NP_COMM_RDMAWRITE) || /* rdma_write only works with no-cache mode */
(!args.preburst && args.prot.commtype != NP_COMM_RDMAWRITE) || /* anything besides rdma_write requires prepost burst */
(args.preburst && args.prot.comptype == NP_COMP_LOCALPOLL && args.cache) || /* preburst with local polling in cache mode doesn't work */
0)) {
fprintf(stderr,
"\n"
"Bi-directional mode currently only works with a subset of the\n"
"Infiniband options. Restrictions are:\n"
"\n"
" RDMA write (-t rdma_write) requires no-cache mode (-I).\n"
"\n"
" Local polling (-c local_poll, default if no -c given) requires\n"
" no-cache mode (-I), and if not using RDMA write communication,\n"
" burst mode (-B).\n"
"\n"
" Any other communication type and any other completion type\n"
" require burst mode (-B). No-cache mode (-I) may be used\n"
" optionally.\n"
"\n"
" All other option combinations will fail.\n"
"\n");
exit(-1);
}
#endif
if (start > end)
{
fprintf(stderr, "Start MUST be LESS than end\n");
exit(420132);
}
args.nbuff = TRIALS;
Setup(&args);
if( args.bidir && end > args.upper ) {
end = args.upper;
if( args.tr ) {
printf("The upper limit is being set to %d Bytes\n", end);
#if defined(TCP) && ! defined(INFINIBAND)
printf("due to socket buffer size limitations\n\n");
#endif
} }
#if defined(GM)
if(streamopt && (!nrepeat_const || nrepeat_const > args.prot.num_stokens)) {
printf("\nGM is currently limited by the driver software to %d\n",
args.prot.num_stokens);
printf("outstanding sends. The number of repeats will be set\n");
printf("to this limit for every trial in streaming mode. You\n");
printf("may use the -n switch to set a smaller number of repeats\n\n");
nrepeat_const = args.prot.num_stokens;
}
#endif
if( args.tr ) /* Primary transmitter */
{
if ((out = fopen(s, "w")) == NULL)
{
fprintf(stderr,"Can't open %s for output\n", s);
exit(1);
}
}
else out = stdout;
/* Set a starting value for the message size increment. */
inc = (start > 1) ? start / 2 : 1;
nq = (start > 1) ? 1 : 0;
/* Test the timing to set tlast for the first test */
args.bufflen = start;
MyMalloc(&args, args.bufflen, 0, 0);
InitBufferData(&args, args.bufflen, 0, 0);
if(args.cache) args.s_buff = args.r_buff;
args.r_ptr = args.r_buff_orig = args.r_buff;
args.s_ptr = args.s_buff_orig = args.s_buff;
AfterAlignmentInit(&args); /* MPI-2 needs this to create a window */
/* Infiniband requires use of asynchronous communications, so we need
* the PrepareToReceive calls below
*/
if( asyncReceive )
PrepareToReceive(&args);
Sync(&args); /* Sync to prevent race condition in armci module */
/* For simplicity's sake, even if the real test below will be done in
* bi-directional mode, we still do the ping-pong one-way-at-a-time test
* here to estimate the one-way latency. Unless it takes significantly
* longer to send data in both directions at once than it does to send data
* one way at a time, this shouldn't be too far off anyway.
*/
t0 = When();
for( n=0; n<100; n++) {
if( args.tr) {
SendData(&args);
RecvData(&args);
if( asyncReceive && n<99 )
PrepareToReceive(&args);
} else if( args.rcv) {
RecvData(&args);
if( asyncReceive && n<99 )
PrepareToReceive(&args);
SendData(&args);
}
}
tlast = (When() - t0)/200;
/* Sync up and Reset before freeing the buffers */
Sync(&args);
Reset(&args);
/* Free the buffers and any other module-specific resources. */
if(args.cache)
FreeBuff(args.r_buff_orig, NULL);
else
FreeBuff(args.r_buff_orig, args.s_buff_orig);
/* Do setup for no-cache mode, using two distinct buffers. */
if (!args.cache)
{
/* Allocate dummy pool of memory to flush cache with */
if ( (memcache = (int *)malloc(MEMSIZE)) == NULL)
{
perror("malloc");
exit(1);
}
mymemset(memcache, 0, MEMSIZE/sizeof(int));
/* Allocate large memory pools */
MyMalloc(&args, MEMSIZE+bufalign, args.soffset, args.roffset);
/* Save buffer addresses */
args.s_buff_orig = args.s_buff;
args.r_buff_orig = args.r_buff;
/* Align buffers */
args.s_buff = AlignBuffer(args.s_buff, bufalign);
args.r_buff = AlignBuffer(args.r_buff, bufalign);
/* Post alignment initialization */
AfterAlignmentInit(&args);
/* Initialize send buffer pointer */
/* both soffset and roffset should be zero if we don't have any offset stuff, so this should be fine */
args.s_ptr = args.s_buff+args.soffset;
args.r_ptr = args.r_buff+args.roffset;
}
/**************************
* Main loop of benchmark *
**************************/
if( args.tr ) fprintf(stderr,"Now starting the main loop\n");
for ( n = 0, len = start, errFlag = 0;
n < NSAMP - 3 && tlast < STOPTM && len <= end && !errFlag;
len = len + inc, nq++ )
{
/* Exponentially increase the block size. */
if (nq > 2) inc = ((nq % 2))? inc + inc: inc;
/* This is a perturbation loop to test nearby values */
for (pert = ((perturbation > 0) && (inc > perturbation+1)) ? -perturbation : 0;
pert <= perturbation;
n++, pert += ((perturbation > 0) && (inc > perturbation+1)) ? perturbation : perturbation+1)
{
Sync(&args); /* Sync to prevent race condition in armci module */
/* Calculate how many times to repeat the experiment. */
if( args.tr )
{
if (nrepeat_const) {
nrepeat = nrepeat_const;
/* } else if (len == start) {*/
/* nrepeat = MAX( RUNTM/( 0.000020 + start/(8*1000) ), TRIALS);*/
} else {
nrepeat = MAX((RUNTM / ((double)args.bufflen /
(args.bufflen - inc + 1.0) * tlast)),TRIALS);
}
SendRepeat(&args, nrepeat);
}
else if( args.rcv )
{
RecvRepeat(&args, &nrepeat);
}
args.bufflen = len + pert;
if( args.tr )
fprintf(stderr,"%3d: %7d bytes %6d times --> ",
n,args.bufflen,nrepeat);
if (args.cache) /* Allow cache effects. We use only one buffer */
{
/* Allocate the buffer with room for alignment*/
MyMalloc(&args, args.bufflen+bufalign, args.soffset, args.roffset);
/* Save buffer address */
args.r_buff_orig = args.r_buff;
args.s_buff_orig = args.r_buff;
/* Align buffer */
args.r_buff = AlignBuffer(args.r_buff, bufalign);
args.s_buff = args.r_buff;
/* Initialize buffer with data
*
* NOTE: The buffers should be initialized with some sort of
* valid data, whether it is actually used for anything else,
* to get accurate results. Performance increases noticeably
* if the buffers are left uninitialized, but this does not
* give very useful results as realworld apps tend to actually
* have data stored in memory. We are not sure what causes
* the difference in performance at this time.
*/
InitBufferData(&args, args.bufflen, args.soffset, args.roffset);
/* Post-alignment initialization */
AfterAlignmentInit(&args);
/* Initialize buffer pointers (We use r_ptr and s_ptr for
* compatibility with no-cache mode, as this makes the code
* simpler)
*/
/* offsets are zero by default so this saves an #ifdef */
args.r_ptr = args.r_buff+args.roffset;
args.s_ptr = args.r_buff+args.soffset;
}
else /* Eliminate cache effects. We use two distinct buffers */
{
/* this isn't truly set up for offsets yet */
/* Size of an aligned memory block including trailing padding */
len_buf_align = args.bufflen;
if(bufalign != 0)
len_buf_align += bufalign - args.bufflen % bufalign;
/* Initialize the buffers with data
*
* See NOTE above.
*/
InitBufferData(&args, MEMSIZE, args.soffset, args.roffset);
/* Reset buffer pointers to beginning of pools */
args.r_ptr = args.r_buff+args.roffset;
args.s_ptr = args.s_buff+args.soffset;
}
bwdata[n].t = LONGTIME;
/* t2 = t1 = 0;*/
/* Finally, we get to transmit or receive and time */
/* NOTE: If a module is running that uses only one process (e.g.
* memcpy), we assume that it will always have the args.tr flag
* set. Thus we make some special allowances in the transmit
* section that are not in the receive section.
*/
if( args.tr || args.bidir )
{
/*
This is the transmitter: send the block TRIALS times, and
if we are not streaming, expect the receiver to return each
block.
*/
for (i = 0; i < (integCheck ? 1 : TRIALS); i++)
{
if(args.preburst && asyncReceive && !streamopt)
{
/* We need to save the value of the recv ptr so
* we can reset it after we do the preposts, in case
* the module needs to use the same ptr values again
* so it can wait on the last byte to change to indicate
* the recv is finished.
*/
SaveRecvPtr(&args);
for(j=0; j<nrepeat; j++)
{
PrepareToReceive(&args);
if(!args.cache)
AdvanceRecvPtr(&args, len_buf_align);
}
ResetRecvPtr(&args);
}
/* Flush the cache using the dummy buffer */
if (!args.cache)
flushcache(memcache, MEMSIZE/sizeof(int));
Sync(&args);
t0 = When();
for (j = 0; j < nrepeat; j++)
{
if (!args.preburst && asyncReceive && !streamopt)
{
PrepareToReceive(&args);
}
if (integCheck) SetIntegrityData(&args);
SendData(&args);
if (!streamopt)
{
RecvData(&args);
if (integCheck) VerifyIntegrity(&args);
if(!args.cache)
AdvanceRecvPtr(&args, len_buf_align);
}
/* Wait to advance send pointer in case RecvData uses
* it (e.g. memcpy module).
*/
if (!args.cache)
AdvanceSendPtr(&args, len_buf_align);
}
/* t is the 1-directional trasmission time */
t = (When() - t0)/ nrepeat;
if( !streamopt && !args.bidir) t /= 2; /* Normal ping-pong */
Reset(&args);
/* NOTE: NetPIPE does each data point TRIALS times, bouncing the message
* nrepeats times for each trial, then reports the lowest of the TRIALS
* times. -Dave Turner
*/
bwdata[n].t = MIN(bwdata[n].t, t);
/* t1 += t;*/
/* t2 += t*t;*/
}
if (streamopt){ /* Get time info from Recv node */
RecvTime(&args, &bwdata[n].t);
/* RecvTime(&args, &t1);*/
/* RecvTime(&args, &t2);*/
}
/* Calculate variance after completing this set of trials */
/* bwdata[n].variance = t2/TRIALS - t1/TRIALS * t1/TRIALS;*/
}
else if( args.rcv )
{
/*
This is the receiver: receive the block TRIALS times, and
if we are not streaming, send the block back to the
sender.
*/
for (i = 0; i < (integCheck ? 1 : TRIALS); i++)
{
if (asyncReceive)
{
if (args.preburst)
{
/* We need to save the value of the recv ptr so
* we can reset it after we do the preposts, in case
* the module needs to use the same ptr values again
* so it can wait on the last byte to change to
* indicate the recv is finished.
*/
SaveRecvPtr(&args);
for (j=0; j < nrepeat; j++)
{
PrepareToReceive(&args);
if (!args.cache)
AdvanceRecvPtr(&args, len_buf_align);
}
ResetRecvPtr(&args);
}
else
{
PrepareToReceive(&args);
}
}
/* Flush the cache using the dummy buffer */
if (!args.cache)
flushcache(memcache, MEMSIZE/sizeof(int));
Sync(&args);
t0 = When();
for (j = 0; j < nrepeat; j++)
{
RecvData(&args);
if (integCheck) VerifyIntegrity(&args);
if (!args.cache)
{
AdvanceRecvPtr(&args, len_buf_align);
}
if (!args.preburst && asyncReceive && (j < nrepeat-1))
{
PrepareToReceive(&args);
}
if (!streamopt)
{
if (integCheck) SetIntegrityData(&args);
SendData(&args);
if(!args.cache)
AdvanceSendPtr(&args, len_buf_align);
}
}
t = (When() - t0)/ nrepeat;
if( !streamopt && !args.bidir) t /= 2; /* Normal ping-pong */
Reset(&args);
bwdata[n].t = MIN(bwdata[n].t, t);
/* t1 += t;*/
/* t2 += t*t;*/
}
if (streamopt){ /* Recv proc calcs time and sends to Trans */
SendTime(&args, &bwdata[n].t);
/* SendTime(&args, &t1);*/
/* SendTime(&args, &t2);*/
}
}
else /* Just going along for the ride */
{
for (i = 0; i < (integCheck ? 1 : TRIALS); i++)
{
Sync(&args);
}
}
/* Streaming mode doesn't really calculate correct latencies
* for small message sizes, and on some nics we can get
* zero second latency after doing the math. Protect against
* this.
*/
if(bwdata[n].t == 0.0) {
bwdata[n].t = 0.000001;
}
tlast = bwdata[n].t;
bwdata[n].bits = args.bufflen * CHARSIZE * (1+args.bidir);
bwdata[n].bps = bwdata[n].bits / (bwdata[n].t * 1024 * 1024);
bwdata[n].repeat = nrepeat;
if (args.tr)
{
if(integCheck) {
fprintf(out,"%8d %d", bwdata[n].bits / 8, nrepeat);
} else {
fprintf(out,"%8d %lf %12.8lf",
bwdata[n].bits / 8, bwdata[n].bps, bwdata[n].t);
}
fprintf(out, "\n");
fflush(out);
}
/* Free using original buffer addresses since we may have aligned
r_buff and s_buff */
if (args.cache)
FreeBuff(args.r_buff_orig, NULL);
if ( args.tr ) {
if(integCheck) {
fprintf(stderr, " Integrity check passed\n");
} else {
fprintf(stderr," %8.2lf Mbps in %10.2lf usec\n",
bwdata[n].bps, tlast*1.0e6);
}
}
} /* End of perturbation loop */
} /* End of main loop */
/* Free using original buffer addresses since we may have aligned
r_buff and s_buff */
if (!args.cache) {
FreeBuff(args.s_buff_orig, args.r_buff_orig);
}
if (args.tr) fclose(out);
CleanUp(&args);
return 0;
}
