void
airMopUnMem(airArray *arr, void *_ptrP) {
void **ptrP;
if (!(arr && _ptrP)) {
return;
}
ptrP = (void **)_ptrP;
airMopSub(arr, ptrP, (airMopper)airSetNull);
airMopSub(arr, *ptrP, airFree);
return;
}
/*
******** nrrdSpatialResample()
**
** general-purpose array-resampler: resamples a nrrd of any type
** (except block) and any dimension along any or all of its axes, with
** any combination of up- or down-sampling along the axes, with any
** kernel (specified by callback), with potentially a different kernel
** for each axis. Whether or not to resample along axis d is
** controlled by the non-NULL-ity of info->kernel[ai]. Where to sample
** on the axis is controlled by info->min[ai] and info->max[ai]; these
** specify a range of "positions" aka "world space" positions, as
** determined by the per-axis min and max of the input nrrd, which must
** be set for every resampled axis.
**
** we cyclically permute those axes being resampled, and never touch
** the position (in axis ordering) of axes along which we are not
** resampling. This strategy is certainly not the most intelligent
** one possible, but it does mean that the axis along which we're
** currently resampling-- the one along which we'll have to look at
** multiple adjecent samples-- is that resampling axis which is
** currently most contiguous in memory. It may make sense to precede
** the resampling with an axis permutation which bubbles all the
** resampled axes to the front (most contiguous) end of the axis list,
** and then puts them back in place afterwards, depending on the cost
** of such axis permutation overhead.
*/
int
nrrdSpatialResample(Nrrd *nout, const Nrrd *nin,
const NrrdResampleInfo *info) {
char me[]="nrrdSpatialResample", func[]="resample", err[BIFF_STRLEN];
nrrdResample_t
*array[NRRD_DIM_MAX], /* intermediate copies of the input data
undergoing resampling; we don't need a full-
fledged nrrd for these. Only about two of
these arrays will be allocated at a time;
intermediate results will be free()d when not
needed */
*_inVec, /* current input vector being resampled;
not necessarily contiguous in memory
(if strideIn != 1) */
*inVec, /* buffer for input vector; contiguous */
*_outVec; /* output vector in context of volume;
never contiguous */
double tmpF;
double ratio, /* factor by which or up or downsampled */
ratios[NRRD_DIM_MAX]; /* record of "ratio" for all resampled axes,
used to compute new spacing in output */
Nrrd *floatNin; /* if the input nrrd type is not nrrdResample_t,
then we convert it and keep it here */
unsigned int ai,
pi, /* current pass */
topLax,
permute[NRRD_DIM_MAX], /* how to permute axes of last pass to get
axes for current pass */
ax[NRRD_DIM_MAX+1][NRRD_DIM_MAX], /* axis ordering on each pass */
passes; /* # of passes needed to resample all axes */
int i, s, e,
topRax, /* the lowest index of an axis which is
resampled. If all axes are being resampled,
then this is 0. If for some reason the
"x" axis (fastest stride) is not being
resampled, but "y" is, then topRax is 1 */
botRax, /* index of highest axis being resampled */
typeIn, typeOut; /* types of input and output of resampling */
size_t sz[NRRD_DIM_MAX+1][NRRD_DIM_MAX];
/* how many samples along each
axis, changing on each pass */
/* all these variables have to do with the spacing of elements in
memory for the current pass of resampling, and they (except
strideIn) are re-set at the beginning of each pass */
nrrdResample_t
*weight; /* sample weights */
unsigned int ci[NRRD_DIM_MAX+1],
co[NRRD_DIM_MAX+1];
int
sizeIn, sizeOut, /* lengths of input and output vectors */
dotLen, /* # input samples to dot with weights to get
one output sample */
doRound, /* actually do rounding on output: we DO NOT
round when info->round but the output
type is not integral */
*index; /* dotLen*sizeOut 2D array of input indices */
size_t
I, /* swiss-army int */
strideIn, /* the stride between samples in the input
"scanline" being resampled */
strideOut, /* stride between samples in output
"scanline" from resampling */
L, LI, LO, numLines, /* top secret */
numOut; /* # of _samples_, total, in output volume;
this is for allocating the output */
airArray *mop; /* for cleaning up */
if (!(nout && nin && info)) {
sprintf(err, "%s: got NULL pointer", me);
biffAdd(NRRD, err); return 1;
}
if (nrrdBoundaryUnknown == info->boundary) {
sprintf(err, "%s: need to specify a boundary behavior", me);
biffAdd(NRRD, err); return 1;
}
typeIn = nin->type;
typeOut = nrrdTypeDefault == info->type ? typeIn : info->type;
if (_nrrdResampleCheckInfo(nin, info)) {
sprintf(err, "%s: problem with arguments", me);
biffAdd(NRRD, err); return 1;
}
_nrrdResampleComputePermute(permute, ax, sz,
&topRax, &botRax, &passes,
nin, info);
topLax = topRax ? 0 : 1;
/* not sure where else to put this:
(want to put it before 0 == passes branch)
We have to assume some centering when doing resampling, and it would
be stupid to not record it in the outgoing nrrd, since the value of
nrrdDefaultCenter could always change. */
for (ai=0; ai<nin->dim; ai++) {
if (info->kernel[ai]) {
nout->axis[ai].center = _nrrdCenter(nin->axis[ai].center);
}
}
if (0 == passes) {
/* actually, no resampling was desired. Copy input to output,
but with the clamping that we normally do at the end of resampling */
nrrdAxisInfoGet_nva(nin, nrrdAxisInfoSize, sz[0]);
if (nrrdMaybeAlloc_nva(nout, typeOut, nin->dim, sz[0])) {
sprintf(err, "%s: couldn't allocate output", me);
biffAdd(NRRD, err); return 1;
}
numOut = nrrdElementNumber(nout);
for (I=0; I<numOut; I++) {
tmpF = nrrdDLookup[nin->type](nin->data, I);
tmpF = nrrdDClamp[typeOut](tmpF);
nrrdDInsert[typeOut](nout->data, I, tmpF);
}
nrrdAxisInfoCopy(nout, nin, NULL, NRRD_AXIS_INFO_NONE);
/* HEY: need to create textual representation of resampling parameters */
if (nrrdContentSet_va(nout, func, nin, "")) {
sprintf(err, "%s:", me);
biffAdd(NRRD, err); return 1;
}
if (nrrdBasicInfoCopy(nout, nin,
NRRD_BASIC_INFO_DATA_BIT
| NRRD_BASIC_INFO_TYPE_BIT
| NRRD_BASIC_INFO_BLOCKSIZE_BIT
| NRRD_BASIC_INFO_DIMENSION_BIT
| NRRD_BASIC_INFO_CONTENT_BIT
| NRRD_BASIC_INFO_COMMENTS_BIT
| (nrrdStateKeyValuePairsPropagate
? 0
: NRRD_BASIC_INFO_KEYVALUEPAIRS_BIT))) {
sprintf(err, "%s:", me);
biffAdd(NRRD, err); return 1;
}
return 0;
}
mop = airMopNew();
/* convert input nrrd to nrrdResample_t if necessary */
if (nrrdResample_nrrdType != typeIn) {
if (nrrdConvert(floatNin = nrrdNew(), nin, nrrdResample_nrrdType)) {
sprintf(err, "%s: couldn't create float copy of input", me);
biffAdd(NRRD, err); airMopError(mop); return 1;
}
array[0] = (nrrdResample_t*)floatNin->data;
airMopAdd(mop, floatNin, (airMopper)nrrdNuke, airMopAlways);
} else {
floatNin = NULL;
array[0] = (nrrdResample_t*)nin->data;
}
/* compute strideIn; this is actually the same for every pass
because (strictly speaking) in every pass we are resampling
the same axis, and axes with lower indices are constant length */
strideIn = 1;
for (ai=0; ai<(unsigned int)topRax; ai++) { /* HEY scrutinize casts */
strideIn *= nin->axis[ai].size;
}
/*
printf("%s: strideIn = " _AIR_SIZE_T_CNV "\n", me, strideIn);
*/
/* go! */
for (pi=0; pi<passes; pi++) {
/*
printf("%s: --- pass %d --- \n", me, pi);
*/
numLines = strideOut = 1;
for (ai=0; ai<nin->dim; ai++) {
if (ai < (unsigned int)botRax) { /* HEY scrutinize cast */
strideOut *= sz[pi+1][ai];
}
if (ai != (unsigned int)topRax) { /* HEY scrutinize cast */
numLines *= sz[pi][ai];
}
}
sizeIn = sz[pi][topRax];
sizeOut = sz[pi+1][botRax];
numOut = numLines*sizeOut;
/* for the rest of the loop body, d is the original "dimension"
for the axis being resampled */
ai = ax[pi][topRax];
/*
printf("%s(%d): numOut = " _AIR_SIZE_T_CNV "\n", me, pi, numOut);
printf("%s(%d): numLines = " _AIR_SIZE_T_CNV "\n", me, pi, numLines);
printf("%s(%d): stride: In=%d, Out=%d\n", me, pi,
(int)strideIn, (int)strideOut);
printf("%s(%d): sizeIn = %d\n", me, pi, sizeIn);
printf("%s(%d): sizeOut = %d\n", me, pi, sizeOut);
*/
/* we can free the input to the previous pass
(if its not the given data) */
if (pi > 0) {
if (pi == 1) {
if (array[0] != nin->data) {
airMopSub(mop, floatNin, (airMopper)nrrdNuke);
floatNin = nrrdNuke(floatNin);
array[0] = NULL;
/*
printf("%s: pi %d: freeing array[0]\n", me, pi);
*/
}
} else {
airMopSub(mop, array[pi-1], airFree);
array[pi-1] = (nrrdResample_t*)airFree(array[pi-1]);
/*
printf("%s: pi %d: freeing array[%d]\n", me, pi, pi-1);
*/
}
}
/* allocate output volume */
array[pi+1] = (nrrdResample_t*)calloc(numOut, sizeof(nrrdResample_t));
if (!array[pi+1]) {
sprintf(err, "%s: couldn't create array of " _AIR_SIZE_T_CNV
" nrrdResample_t's for output of pass %d",
me, numOut, pi);
biffAdd(NRRD, err); airMopError(mop); return 1;
}
airMopAdd(mop, array[pi+1], airFree, airMopAlways);
/*
printf("%s: allocated array[%d]\n", me, pi+1);
*/
/* allocate contiguous input scanline buffer, we alloc one more
than needed to provide a place for the pad value. That is, in
fact, the over-riding reason to copy a scanline to a local
array: so that there is a simple consistent (non-branchy) way
to incorporate the pad values */
inVec = (nrrdResample_t *)calloc(sizeIn+1, sizeof(nrrdResample_t));
airMopAdd(mop, inVec, airFree, airMopAlways);
inVec[sizeIn] = AIR_CAST(nrrdResample_t, info->padValue);
dotLen = _nrrdResampleMakeWeightIndex(&weight, &index, &ratio,
nin, info, ai);
if (!dotLen) {
sprintf(err, "%s: trouble creating weight and index vector arrays", me);
biffAdd(NRRD, err); airMopError(mop); return 1;
}
ratios[ai] = ratio;
airMopAdd(mop, weight, airFree, airMopAlways);
airMopAdd(mop, index, airFree, airMopAlways);
/* the skinny: resample all the scanlines */
_inVec = array[pi];
_outVec = array[pi+1];
memset(ci, 0, (NRRD_DIM_MAX+1)*sizeof(int));
memset(co, 0, (NRRD_DIM_MAX+1)*sizeof(int));
for (L=0; L<numLines; L++) {
/* calculate the index to get to input and output scanlines,
according the coordinates of the start of the scanline */
NRRD_INDEX_GEN(LI, ci, sz[pi], nin->dim);
NRRD_INDEX_GEN(LO, co, sz[pi+1], nin->dim);
_inVec = array[pi] + LI;
_outVec = array[pi+1] + LO;
/* read input scanline into contiguous array */
for (i=0; i<sizeIn; i++) {
inVec[i] = _inVec[i*strideIn];
}
/* do the weighting */
for (i=0; i<sizeOut; i++) {
tmpF = 0.0;
/*
fprintf(stderr, "%s: i = %d (tmpF=0)\n", me, (int)i);
*/
for (s=0; s<dotLen; s++) {
tmpF += inVec[index[s + dotLen*i]]*weight[s + dotLen*i];
/*
fprintf(stderr, " tmpF += %g*%g == %g\n",
inVec[index[s + dotLen*i]], weight[s + dotLen*i], tmpF);
*/
}
_outVec[i*strideOut] = tmpF;
/*
fprintf(stderr, "--> out[%d] = %g\n",
i*strideOut, _outVec[i*strideOut]);
*/
}
/* update the coordinates for the scanline starts. We don't
use the usual NRRD_COORD macros because we're subject to
the unusual constraint that ci[topRax] and co[permute[topRax]]
must stay exactly zero */
e = topLax;
ci[e]++;
co[permute[e]]++;
while (L < numLines-1 && ci[e] == sz[pi][e]) {
ci[e] = co[permute[e]] = 0;
e++;
e += e == topRax;
ci[e]++;
co[permute[e]]++;
}
}
/* pass-specific clean up */
airMopSub(mop, weight, airFree);
airMopSub(mop, index, airFree);
airMopSub(mop, inVec, airFree);
weight = (nrrdResample_t*)airFree(weight);
index = (int*)airFree(index);
inVec = (nrrdResample_t*)airFree(inVec);
}
/* clean up second-to-last array and scanline buffers */
if (passes > 1) {
airMopSub(mop, array[passes-1], airFree);
array[passes-1] = (nrrdResample_t*)airFree(array[passes-1]);
/*
printf("%s: now freeing array[%d]\n", me, passes-1);
*/
} else if (array[passes-1] != nin->data) {
airMopSub(mop, floatNin, (airMopper)nrrdNuke);
floatNin = nrrdNuke(floatNin);
}
array[passes-1] = NULL;
/* create output nrrd and set axis info */
if (nrrdMaybeAlloc_nva(nout, typeOut, nin->dim, sz[passes])) {
sprintf(err, "%s: couldn't allocate final output nrrd", me);
biffAdd(NRRD, err); airMopError(mop); return 1;
}
airMopAdd(mop, nout, (airMopper)nrrdNuke, airMopOnError);
nrrdAxisInfoCopy(nout, nin, NULL,
(NRRD_AXIS_INFO_SIZE_BIT
| NRRD_AXIS_INFO_MIN_BIT
| NRRD_AXIS_INFO_MAX_BIT
| NRRD_AXIS_INFO_SPACING_BIT
| NRRD_AXIS_INFO_SPACEDIRECTION_BIT /* see below */
| NRRD_AXIS_INFO_THICKNESS_BIT
| NRRD_AXIS_INFO_KIND_BIT));
for (ai=0; ai<nin->dim; ai++) {
if (info->kernel[ai]) {
/* we do resample this axis */
nout->axis[ai].spacing = nin->axis[ai].spacing/ratios[ai];
/* no way to usefully update thickness: we could be doing blurring
but maintaining the number of samples: thickness increases, or
we could be downsampling, in which the relationship between the
sampled and the skipped regions of space becomes complicated:
no single scalar can represent it, or we could be upsampling,
in which the notion of "skip" could be rendered meaningless */
nout->axis[ai].thickness = AIR_NAN;
nout->axis[ai].min = info->min[ai];
nout->axis[ai].max = info->max[ai];
/*
HEY: this is currently a bug: all this code was written long
before there were space directions, so min/max are always
set, regardless of whethere there are incoming space directions
which then disallows output space directions on the same axes
_nrrdSpaceVecScale(nout->axis[ai].spaceDirection,
1.0/ratios[ai], nin->axis[ai].spaceDirection);
*/
nout->axis[ai].kind = _nrrdKindAltered(nin->axis[ai].kind, AIR_TRUE);
} else {
/* this axis remains untouched */
nout->axis[ai].min = nin->axis[ai].min;
nout->axis[ai].max = nin->axis[ai].max;
nout->axis[ai].spacing = nin->axis[ai].spacing;
nout->axis[ai].thickness = nin->axis[ai].thickness;
nout->axis[ai].kind = nin->axis[ai].kind;
}
}
/* HEY: need to create textual representation of resampling parameters */
if (nrrdContentSet_va(nout, func, nin, "")) {
sprintf(err, "%s:", me);
biffAdd(NRRD, err); return 1;
}
/* copy the resampling final result into the output nrrd, maybe
rounding as we go to make sure that 254.9999 is saved as 255
in uchar output, and maybe clamping as we go to insure that
integral results don't have unexpected wrap-around. */
if (info->round) {
if (nrrdTypeInt == typeOut ||
nrrdTypeUInt == typeOut ||
nrrdTypeLLong == typeOut ||
nrrdTypeULLong == typeOut) {
fprintf(stderr, "%s: WARNING: possible erroneous output with "
"rounding of %s output type due to int-based implementation "
"of rounding\n", me, airEnumStr(nrrdType, typeOut));
}
doRound = nrrdTypeIsIntegral[typeOut];
} else {
doRound = AIR_FALSE;
}
numOut = nrrdElementNumber(nout);
for (I=0; I<numOut; I++) {
tmpF = array[passes][I];
if (doRound) {
tmpF = AIR_CAST(nrrdResample_t, AIR_ROUNDUP(tmpF));
}
if (info->clamp) {
tmpF = nrrdDClamp[typeOut](tmpF);
}
nrrdDInsert[typeOut](nout->data, I, tmpF);
}
if (nrrdBasicInfoCopy(nout, nin,
NRRD_BASIC_INFO_DATA_BIT
| NRRD_BASIC_INFO_TYPE_BIT
| NRRD_BASIC_INFO_BLOCKSIZE_BIT
| NRRD_BASIC_INFO_DIMENSION_BIT
| NRRD_BASIC_INFO_CONTENT_BIT
| NRRD_BASIC_INFO_COMMENTS_BIT
| (nrrdStateKeyValuePairsPropagate
? 0
: NRRD_BASIC_INFO_KEYVALUEPAIRS_BIT))) {
sprintf(err, "%s:", me);
biffAdd(NRRD, err); return 1;
}
/* enough already */
airMopOkay(mop);
return 0;
}
int
unrrdu_fftMain(int argc, const char **argv, const char *me,
hestParm *hparm) {
hestOpt *opt = NULL;
char *out, *err;
Nrrd *nin, *_nin, *nout;
int pret;
airArray *mop;
int sign, rigor, rescale, realInput;
char *wispath;
FILE *fwise;
unsigned int *axes, axesLen;
hestOptAdd(&opt, NULL, "dir", airTypeEnum, 1, 1, &sign, NULL,
"forward (\"forw\", \"f\") or backward/inverse "
"(\"back\", \"b\") transform ", NULL, direction_enm);
hestOptAdd(&opt, "a,axes", "ax0", airTypeUInt, 1, -1, &axes, NULL,
"the one or more axes that should be transformed", &axesLen);
hestOptAdd(&opt, "pr,planrigor", "pr", airTypeEnum, 1, 1, &rigor, "est",
"rigor with which fftw plan is constructed. Options include:\n "
"\b\bo \"e\", \"est\", \"estimate\": only an estimate\n "
"\b\bo \"m\", \"meas\", \"measure\": standard amount of "
"measurements of system properties\n "
"\b\bo \"p\", \"pat\", \"patient\": slower, more measurements\n "
"\b\bo \"x\", \"ex\", \"exhaustive\": slowest, most measurements",
NULL, nrrdFFTWPlanRigor);
hestOptAdd(&opt, "r,rescale", "bool", airTypeBool, 1, 1, &rescale, "true",
"scale fftw output (by sqrt(1/N)) so that forward and backward "
"transforms will get back to original values");
hestOptAdd(&opt, "w,wisdom", "filename", airTypeString, 1, 1, &wispath, "",
"A filename here is used to read in fftw wisdom (if the file "
"exists already), and is used to save out updated wisdom "
"after the transform. By default (not using this option), "
"no wisdom is read or saved. Note: no wisdom is gained "
"(that is, learned by FFTW) with planning rigor \"estimate\".");
OPT_ADD_NIN(_nin, "input nrrd");
hestOptAdd(&opt, "ri,realinput", NULL, airTypeInt, 0, 0, &realInput, NULL,
"input is real-valued, so insert new length-2 axis 0 "
"and set complex component to 0.0. Axes to transform "
"(indicated by \"-a\") will be incremented accordingly.");
OPT_ADD_NOUT(out, "output nrrd");
mop = airMopNew();
airMopAdd(mop, opt, (airMopper)hestOptFree, airMopAlways);
if (nrrdFFTWEnabled) {
USAGE(_unrrdu_fftInfoL_yes);
} else {
USAGE(_unrrdu_fftInfoL_no);
}
PARSE();
airMopAdd(mop, opt, (airMopper)hestParseFree, airMopAlways);
nout = nrrdNew();
airMopAdd(mop, nout, (airMopper)nrrdNuke, airMopAlways);
if (realInput) {
ptrdiff_t minPad[NRRD_DIM_MAX], maxPad[NRRD_DIM_MAX];
unsigned int axi;
Nrrd *ntmp;
ntmp = nrrdNew();
airMopAdd(mop, ntmp, (airMopper)nrrdNuke, airMopAlways);
if (nrrdAxesInsert(ntmp, _nin, 0)) {
airMopAdd(mop, err = biffGetDone(NRRD), airFree, airMopAlways);
fprintf(stderr, "%s: error creating complex axis:\n%s", me, err);
airMopError(mop);
return 1;
}
nin = nrrdNew();
airMopAdd(mop, nin, (airMopper)nrrdNuke, airMopAlways);
minPad[0] = 0;
maxPad[0] = 1;
for (axi=1; axi<ntmp->dim; axi++) {
minPad[axi] = 0;
maxPad[axi] = AIR_CAST(ptrdiff_t, ntmp->axis[axi].size-1);
}
if (nrrdPad_nva(nin, ntmp, minPad, maxPad, nrrdBoundaryPad, 0.0)) {
airMopAdd(mop, err = biffGetDone(NRRD), airFree, airMopAlways);
fprintf(stderr, "%s: error padding out complex axis:\n%s", me, err);
airMopError(mop);
return 1;
}
/* increment specified axes to transform */
for (axi=0; axi<axesLen; axi++) {
axes[axi]++;
}
/* ntmp is really done with, we can free up the space now; this
is one of the rare times we want airMopSub */
airMopSub(mop, ntmp, (airMopper)nrrdNuke);
nrrdNuke(ntmp);
} else {
/* input is apparently already complex */
nin = _nin;
}
if (airStrlen(wispath) && nrrdFFTWEnabled) {
fwise = fopen(wispath, "r");
if (fwise) {
if (nrrdFFTWWisdomRead(fwise)) {
airMopAdd(mop, err = biffGetDone(NRRD), airFree, airMopAlways);
fprintf(stderr, "%s: error with fft wisdom:\n%s", me, err);
airMopError(mop);
return 1;
}
fclose(fwise);
} else {
fprintf(stderr, "%s: (\"%s\" couldn't be opened, will try to save "
"wisdom afterwards)", me, wispath);
}
}
if (nrrdFFT(nout, nin, axes, axesLen, sign, rescale, rigor)) {
airMopAdd(mop, err = biffGetDone(NRRD), airFree, airMopAlways);
fprintf(stderr, "%s: error with fft:\n%s", me, err);
airMopError(mop);
return 1;
}
if (airStrlen(wispath) && nrrdFFTWEnabled) {
if (!(fwise = fopen(wispath, "w"))) {
fprintf(stderr, "%s: couldn't open %s for writing: %s\n",
me, wispath, strerror(errno));
airMopError(mop);
return 1;
}
if (nrrdFFTWWisdomWrite(fwise)) {
airMopAdd(mop, err = biffGetDone(NRRD), airFree, airMopAlways);
fprintf(stderr, "%s: error with fft wisdom:\n%s", me, err);
airMopError(mop);
return 1;
}
fclose(fwise);
}
SAVE(out, nout, NULL);
airMopOkay(mop);
return 0;
}
int
main(int argc, char *argv[]) {
#if TEEM_DIO == 0
AIR_UNUSED(argc);
fprintf(stderr, "%s: no direct-io testing for you\n", argv[0]);
return 1;
#else
char *me, *fname, *multS, *data;
FILE *file;
double time0, time1, time2;
int fd, align, mult, min, max, ret;
size_t size;
airArray *mop;
me = argv[0];
if (3 != argc) {
/* 0 1 2 (3) */
fprintf(stderr, "usage: %s <filename> <mult>\n", me);
return 1;
}
fname = argv[1];
multS = argv[2];
if (1 != sscanf(multS, "%d", &mult)) {
fprintf(stderr, "%s: couln't parse mult %s as int\n", me, multS);
return 1;
}
mop = airMopNew();
if (!(file = fopen(fname, "w"))) {
fprintf(stderr, "%s: couldn't open %s for writing\n", me, fname);
airMopError(mop); return 1;
}
airMopAdd(mop, file, (airMopper)airFclose, airMopAlways);
fd = fileno(file);
if (-1 == fd) {
fprintf(stderr, "%s: couldn't get underlying descriptor\n", me);
airMopError(mop); return 1;
}
fprintf(stderr, "%s: fd(%s) = %d\n", me, fname, fd);
ret = airDioTest(fd, NULL, 0);
if (airNoDio_okay != ret) {
fprintf(stderr, "%s: no good: \"%s\"\n", me, airNoDioErr(ret));
airMopError(mop); return 1;
}
airDioInfo(&align, &min, &max, fd);
fprintf(stderr, "%s: --> align=%d, min=%d, max=%d\n", me, align, min, max);
size = (size_t)max*mult;
data = airDioMalloc(size, fd);
if (!data) {
fprintf(stderr, "%s: airDioMalloc(" _AIR_SIZE_T_CNV ") failed\n", me,
size);
airMopError(mop); return 1;
}
airMopAdd(mop, data, airFree, airMopAlways);
fprintf(stderr, "\ndata size = %g MB\n", (double)size/(1024*1024));
/* -------------------------------------------------------------- */
fprintf(stderr, "(1) non-aligned memory, regular write:\n");
time0 = airTime();
if (size-1 != write(fd, data+1, size-1)) {
fprintf(stderr, "%s: write failed\n", me);
airMopError(mop); return 1;
}
time1 = airTime();
fsync(fd);
time2 = airTime();
fprintf(stderr, " time = %g + %g = %g (%g MB/sec)\n",
time1 - time0, time2 - time1, time2 - time0,
(size/(1024*1024)) / (time2 - time0));
airMopSub(mop, file, (airMopper)airFclose);
fclose(file);
/* -------------------------------------------------------------- */
/* -------------------------------------------------------------- */
fprintf(stderr, "(2) aligned memory, regular write:\n");
file = fopen(fname, "w");
airMopAdd(mop, file, (airMopper)airFclose, airMopAlways);
fd = fileno(file);
time0 = airTime();
if (size != write(fd, data, size)) {
fprintf(stderr, "%s: write failed\n", me);
airMopError(mop); return 1;
}
time1 = airTime();
fsync(fd);
time2 = airTime();
fprintf(stderr, " time = %g + %g = %g (%g MB/sec)\n",
time1 - time0, time2 - time1, time2 - time0,
(size/(1024*1024)) / (time2 - time0));
airMopSub(mop, file, (airMopper)airFclose);
fclose(file);
/* -------------------------------------------------------------- */
/* -------------------------------------------------------------- */
fprintf(stderr, "(3) aligned memory, air's direct IO:\n");
file = fopen(fname, "w");
airMopAdd(mop, file, (airMopper)airFclose, airMopAlways);
fd = fileno(file);
time0 = airTime();
if (size != airDioWrite(fd, data, size)) {
fprintf(stderr, "%s: write failed\n", me);
airMopError(mop); return 1;
}
time1 = airTime();
fsync(fd);
time2 = airTime();
fprintf(stderr, " time = %g + %g = %g (%g MB/sec)\n",
time1 - time0, time2 - time1, time2 - time0,
(size/(1024*1024)) / (time2 - time0));
airMopSub(mop, file, (airMopper)airFclose);
fclose(file);
/* -------------------------------------------------------------- */
/* -------------------------------------------------------------- */
fprintf(stderr, "(4) aligned memory, direct IO by hand:\n");
{
/* "input": fname, size, data */
int flags;
struct dioattr dio;
char *ptr;
size_t remain, totalrit, rit, part;
file = fopen(fname, "w");
if (-1 == (fd = fileno(file))) {
fprintf(stderr, "%s: couldn't get underlying descriptor\n", me);
airMopError(mop); return 1;
}
airMopAdd(mop, file, (airMopper)airFclose, airMopAlways);
flags = fcntl(fd, F_GETFL);
if (-1 == fcntl(fd, F_SETFL, flags | FDIRECT)) {
fprintf(stderr, "%s: couldn't turn on direct IO\n", me);
airMopError(mop); return 1;
}
if (0 != fcntl(fd, F_DIOINFO, &dio)) {
fprintf(stderr, "%s: couldn't learn direct IO specifics", me);
airMopError(mop); return 1;
}
remain = size;
totalrit = 0;
ptr = data;
time0 = airTime();
do {
part = remain > dio.d_maxiosz ? dio.d_maxiosz : remain;
rit = write(fd, ptr, part);
if (rit != part) {
fprintf(stderr, "%s: write failed\n", me);
airMopError(mop); return 1;
}
totalrit += rit;
ptr += rit;
remain -= rit;
} while (remain);
time1 = airTime();
fsync(fd);
time2 = airTime();
fprintf(stderr, " time = %g + %g = %g (%g MB/sec)\n",
time1 - time0, time2 - time1, time2 - time0,
(size/(1024*1024)) / (time2 - time0));
airMopSub(mop, file, (airMopper)airFclose);
fclose(file);
}
/* -------------------------------------------------------------- */
airMopError(mop);
exit(0);
#endif
}
