Chunk compress() {
// Check for basic mode info: all missing or all strings or mixed stuff
byte mode = type();
if (mode == AppendableVec.NA) // ALL NAs, nothing to do
return new C0DChunk(Double.NaN, _len);
for (int i = 0; i < _len; i++)
if (mode == AppendableVec.ENUM && !isEnum(i) || mode == AppendableVec.NUMBER && isEnum(i))
setNA_impl(i);
_naCnt = -1;
type(); // Re-run rollups after dropping all numbers/enums
// If the data was set8 as doubles, we do a quick check to see if it's
// plain longs. If not, we give up and use doubles.
if (_ds != null) {
int i = 0;
for (; i < _len; i++) // Attempt to inject all doubles into longs
if (!Double.isNaN(_ds[i]) && (double) (long) _ds[i] != _ds[i]) break;
if (i < _len) return chunkD();
_ls = new long[_ds.length]; // Else flip to longs
_xs = new int[_ds.length];
for (i = 0; i < _len; i++) // Inject all doubles into longs
if (Double.isNaN(_ds[i])) _xs[i] = Integer.MIN_VALUE;
else _ls[i] = (long) _ds[i];
_ds = null;
}
// IF (_len2 > _len) THEN Sparse
// Check for compressed *during appends*. Here we know:
// - No specials; _xs[]==0.
// - No floats; _ds==null
// - NZ length in _len, actual length in _len2.
// - Huge ratio between _len2 and _len, and we do NOT want to inflate to
// the larger size; we need to keep it all small all the time.
// - Rows in _xs
// Data in some fixed-point format, not doubles
// See if we can sanely normalize all the data to the same fixed-point.
int xmin = Integer.MAX_VALUE; // min exponent found
long lemin = 0, lemax = lemin; // min/max at xmin fixed-point
boolean overflow = false;
boolean floatOverflow = false;
boolean first = true;
double min = _len2 == _len ? Double.MAX_VALUE : 0;
double max = _len2 == _len ? -Double.MAX_VALUE : 0;
for (int i = 0; i < _len; i++) {
if (isNA(i)) continue;
long l = _ls[i];
int x = _xs[i];
if (x == Integer.MIN_VALUE + 1 || _len2 != _len) x = 0; // Replace enum flag with no scaling
assert l != 0 || x == 0; // Exponent of zero is always zero
// Compute per-chunk min/max
double d = l * DParseTask.pow10(x);
if (d < min) min = d;
if (d > max) max = d;
long t; // Remove extra scaling
while (l != 0 && (t = l / 10) * 10 == l) {
l = t;
x++;
}
floatOverflow = Math.abs(l) > MAX_FLOAT_MANTISSA;
if (first) {
first = false;
xmin = x;
lemin = lemax = l;
continue;
}
// Remove any trailing zeros / powers-of-10
if (overflow || (overflow = (Math.abs(xmin - x)) >= 10)) continue;
// Track largest/smallest values at xmin scale. Note overflow.
if (x < xmin) {
lemin *= DParseTask.pow10i(xmin - x);
lemax *= DParseTask.pow10i(xmin - x);
xmin = x; // Smaller xmin
}
// *this* value, as a long scaled at the smallest scale
long le = l * DParseTask.pow10i(x - xmin);
if (le < lemin) lemin = le;
if (le > lemax) lemax = le;
}
if (_len2 != _len) { // sparse? compare xmin/lemin/lemax with 0
lemin = Math.min(0, lemin);
lemax = Math.max(0, lemax);
}
// Constant column?
if (_naCnt == 0 && min == max) {
return ((long) min == min) ? new C0LChunk((long) min, _len2) : new C0DChunk(min, _len2);
}
// Boolean column?
if (max == 1 && min == 0 && xmin == 0) {
if (_nzCnt * 32 < _len2
&& _naCnt == 0
&& _len2 < 65535
&& xmin == 0) // Very sparse? (and not too big?)
if (_len2 == _len) return new CX0Chunk(_ls, _len2, _nzCnt); // Dense constructor
else return new CX0Chunk(_xs, _len2, _len); // Sparse constructor
int bpv = _strCnt + _naCnt > 0 ? 2 : 1; // Bit-vector
byte[] cbuf = bufB(bpv);
return new CBSChunk(cbuf, cbuf[0], cbuf[1]);
}
final boolean fpoint = xmin < 0 || min < Long.MIN_VALUE || max > Long.MAX_VALUE;
// Result column must hold floats?
// Highly sparse but not a bitvector or constant?
if (!fpoint
&& (_nzCnt + _naCnt) * 8 < _len2
&& _len2 < 65535
&& xmin == 0
&& // (and not too big?)
lemin > Short.MIN_VALUE
&& lemax <= Short.MAX_VALUE) // Only handling unbiased shorts here
if (_len2 == _len) return new CX2Chunk(_ls, _xs, _len2, _nzCnt, _naCnt); // Sparse byte chunk
else return new CX2Chunk(_ls, _xs, _len2, _len);
// Exponent scaling: replacing numbers like 1.3 with 13e-1. '13' fits in a
// byte and we scale the column by 0.1. A set of numbers like
// {1.2,23,0.34} then is normalized to always be represented with 2 digits
// to the right: {1.20,23.00,0.34} and we scale by 100: {120,2300,34}.
// This set fits in a 2-byte short.
// We use exponent-scaling for bytes & shorts only; it's uncommon (and not
// worth it) for larger numbers. We need to get the exponents to be
// uniform, so we scale up the largest lmax by the largest scale we need
// and if that fits in a byte/short - then it's worth compressing. Other
// wise we just flip to a float or double representation.
if (overflow || fpoint && floatOverflow || -35 > xmin || xmin > 35) return chunkD();
if (fpoint) {
if (lemax - lemin < 255) // Fits in scaled biased byte?
return new C1SChunk(bufX(lemin, xmin, C1SChunk.OFF, 0), (int) lemin, DParseTask.pow10(xmin));
if (lemax - lemin < 65535) { // we use signed 2B short, add -32k to the bias!
long bias = 32767 + lemin;
return new C2SChunk(bufX(bias, xmin, C2SChunk.OFF, 1), (int) bias, DParseTask.pow10(xmin));
}
if (lemax - lemin < Integer.MAX_VALUE)
return new C4SChunk(
bufX(lemin, xmin, C4SChunk.OFF, 2), (int) lemin, DParseTask.pow10(xmin));
return chunkD();
} // else an integer column
// Compress column into a byte
if (xmin == 0 && 0 <= lemin && lemax <= 255 && ((_naCnt + _strCnt) == 0))
return new C1NChunk(bufX(0, 0, C1NChunk.OFF, 0));
if (lemax - lemin < 255) { // Span fits in a byte?
if (0 <= min && max < 255) // Span fits in an unbiased byte?
return new C1Chunk(bufX(0, 0, C1Chunk.OFF, 0));
return new C1SChunk(bufX(lemin, xmin, C1SChunk.OFF, 0), (int) lemin, DParseTask.pow10i(xmin));
}
// Compress column into a short
if (lemax - lemin < 65535) { // Span fits in a biased short?
if (xmin == 0
&& Short.MIN_VALUE < lemin
&& lemax <= Short.MAX_VALUE) // Span fits in an unbiased short?
return new C2Chunk(bufX(0, 0, C2Chunk.OFF, 1));
int bias = (int) (lemin - (Short.MIN_VALUE + 1));
return new C2SChunk(bufX(bias, xmin, C2SChunk.OFF, 1), bias, DParseTask.pow10i(xmin));
}
// Compress column into ints
if (Integer.MIN_VALUE < min && max <= Integer.MAX_VALUE) return new C4Chunk(bufX(0, 0, 0, 2));
return new C8Chunk(bufX(0, 0, 0, 3));
}
private Chunk compress2() {
// Check for basic mode info: all missing or all strings or mixed stuff
byte mode = type();
if (mode == Vec.T_BAD) // ALL NAs, nothing to do
return new C0DChunk(Double.NaN, sparseLen());
if (mode == Vec.T_STR) return new CStrChunk(_sslen, _ss, sparseLen(), _len, _is, _isAllASCII);
boolean rerun = false;
if (mode == Vec.T_CAT) {
for (int i = 0; i < sparseLen(); i++)
if (isCategorical2(i)) _xs[i] = 0;
else if (!isNA2(i)) {
setNA_impl2(i);
++_naCnt;
}
// Smack any mismatched string/numbers
} else if (mode == Vec.T_NUM) {
for (int i = 0; i < sparseLen(); i++)
if (isCategorical2(i)) {
setNA_impl2(i);
rerun = true;
}
}
if (rerun) {
_naCnt = -1;
type();
} // Re-run rollups after dropping all numbers/categoricals
boolean sparse = false;
// sparse? treat as sparse iff we have at least MIN_SPARSE_RATIOx more zeros than nonzeros
if (_sparseRatio * (_naCnt + _nzCnt) < _len) {
set_sparse(_naCnt + _nzCnt);
sparse = true;
} else if (sparseLen() != _len) cancel_sparse();
// If the data is UUIDs there's not much compression going on
if (_ds != null && _ls != null) return chunkUUID();
// cut out the easy all NaNs case
if (_naCnt == _len) return new C0DChunk(Double.NaN, _len);
// If the data was set8 as doubles, we do a quick check to see if it's
// plain longs. If not, we give up and use doubles.
if (_ds != null) {
int i; // check if we can flip to ints
for (i = 0; i < sparseLen(); ++i)
if (!Double.isNaN(_ds[i]) && (double) (long) _ds[i] != _ds[i]) break;
boolean isInteger = i == sparseLen();
boolean isConstant = !sparse || sparseLen() == 0;
double constVal = 0;
if (!sparse) { // check the values, sparse with some nonzeros can not be constant - has 0s and
// (at least 1) nonzero
constVal = _ds[0];
for (int j = 1; j < _len; ++j)
if (_ds[j] != constVal) {
isConstant = false;
break;
}
}
if (isConstant)
return isInteger ? new C0LChunk((long) constVal, _len) : new C0DChunk(constVal, _len);
if (!isInteger) return sparse ? new CXDChunk(_len, sparseLen(), 8, bufD(8)) : chunkD();
// Else flip to longs
_ls = new long[_ds.length];
_xs = new int[_ds.length];
double[] ds = _ds;
_ds = null;
final int naCnt = _naCnt;
for (i = 0; i < sparseLen(); i++) // Inject all doubles into longs
if (Double.isNaN(ds[i])) setNA_impl2(i);
else _ls[i] = (long) ds[i];
// setNA_impl2 will set _naCnt to -1!
// we already know what the naCnt is (it did not change!) so set it back to correct value
_naCnt = naCnt;
}
// IF (_len > _sparseLen) THEN Sparse
// Check for compressed *during appends*. Here we know:
// - No specials; _xs[]==0.
// - No floats; _ds==null
// - NZ length in _sparseLen, actual length in _len.
// - Huge ratio between _len and _sparseLen, and we do NOT want to inflate to
// the larger size; we need to keep it all small all the time.
// - Rows in _xs
// Data in some fixed-point format, not doubles
// See if we can sanely normalize all the data to the same fixed-point.
int xmin = Integer.MAX_VALUE; // min exponent found
boolean floatOverflow = false;
double min = Double.POSITIVE_INFINITY;
double max = Double.NEGATIVE_INFINITY;
int p10iLength = PrettyPrint.powers10i.length;
long llo = Long.MAX_VALUE, lhi = Long.MIN_VALUE;
int xlo = Integer.MAX_VALUE, xhi = Integer.MIN_VALUE;
for (int i = 0; i < sparseLen(); i++) {
if (isNA2(i)) continue;
long l = _ls[i];
int x = _xs[i];
assert x != Integer.MIN_VALUE : "l = " + l + ", x = " + x;
if (x == Integer.MIN_VALUE + 1) x = 0; // Replace categorical flag with no scaling
assert l != 0 || x == 0
: "l == 0 while x = "
+ x
+ " ls = "
+ Arrays.toString(_ls); // Exponent of zero is always zero
long t; // Remove extra scaling
while (l != 0 && (t = l / 10) * 10 == l) {
l = t;
x++;
}
// Compute per-chunk min/max
double d = l * PrettyPrint.pow10(x);
if (d < min) {
min = d;
llo = l;
xlo = x;
}
if (d > max) {
max = d;
lhi = l;
xhi = x;
}
floatOverflow = l < Integer.MIN_VALUE + 1 || l > Integer.MAX_VALUE;
xmin = Math.min(xmin, x);
}
if (sparse) { // sparse? then compare vs implied 0s
if (min > 0) {
min = 0;
llo = 0;
xlo = 0;
}
if (max < 0) {
max = 0;
lhi = 0;
xhi = 0;
}
xmin = Math.min(xmin, 0);
}
// Constant column?
if (_naCnt == 0 && (min == max)) {
if (llo == lhi && xlo == 0 && xhi == 0) return new C0LChunk(llo, _len);
else if ((long) min == min) return new C0LChunk((long) min, _len);
else return new C0DChunk(min, _len);
}
// Compute min & max, as scaled integers in the xmin scale.
// Check for overflow along the way
boolean overflow = ((xhi - xmin) >= p10iLength) || ((xlo - xmin) >= p10iLength);
long lemax = 0, lemin = 0;
if (!overflow) { // Can at least get the power-of-10 without overflow
long pow10 = PrettyPrint.pow10i(xhi - xmin);
lemax = lhi * pow10;
// Hacker's Delight, Section 2-13, checking overflow.
// Note that the power-10 is always positive, so the test devolves this:
if ((lemax / pow10) != lhi) overflow = true;
// Note that xlo might be > xmin; e.g. { 101e-49 , 1e-48}.
long pow10lo = PrettyPrint.pow10i(xlo - xmin);
lemin = llo * pow10lo;
if ((lemin / pow10lo) != llo) overflow = true;
}
// Boolean column?
if (max == 1 && min == 0 && xmin == 0 && !overflow) {
if (sparse) { // Very sparse?
return _naCnt == 0
? new CX0Chunk(_len, sparseLen(), bufS(0)) // No NAs, can store as sparse bitvector
: new CXIChunk(_len, sparseLen(), 1, bufS(1)); // have NAs, store as sparse 1byte values
}
int bpv = _catCnt + _naCnt > 0 ? 2 : 1; // Bit-vector
byte[] cbuf = bufB(bpv);
return new CBSChunk(cbuf, cbuf[0], cbuf[1]);
}
final boolean fpoint = xmin < 0 || min < Long.MIN_VALUE || max > Long.MAX_VALUE;
if (sparse) {
if (fpoint) return new CXDChunk(_len, sparseLen(), 8, bufD(8));
int sz = 8;
if (Short.MIN_VALUE <= min && max <= Short.MAX_VALUE) sz = 2;
else if (Integer.MIN_VALUE <= min && max <= Integer.MAX_VALUE) sz = 4;
return new CXIChunk(_len, sparseLen(), sz, bufS(sz));
}
// Exponent scaling: replacing numbers like 1.3 with 13e-1. '13' fits in a
// byte and we scale the column by 0.1. A set of numbers like
// {1.2,23,0.34} then is normalized to always be represented with 2 digits
// to the right: {1.20,23.00,0.34} and we scale by 100: {120,2300,34}.
// This set fits in a 2-byte short.
// We use exponent-scaling for bytes & shorts only; it's uncommon (and not
// worth it) for larger numbers. We need to get the exponents to be
// uniform, so we scale up the largest lmax by the largest scale we need
// and if that fits in a byte/short - then it's worth compressing. Other
// wise we just flip to a float or double representation.
if (overflow || (fpoint && floatOverflow) || -35 > xmin || xmin > 35) return chunkD();
final long leRange = leRange(lemin, lemax);
if (fpoint) {
if ((int) lemin == lemin && (int) lemax == lemax) {
if (leRange < 255) // Fits in scaled biased byte?
return new C1SChunk(bufX(lemin, xmin, C1SChunk._OFF, 0), lemin, PrettyPrint.pow10(xmin));
if (leRange < 65535) { // we use signed 2B short, add -32k to the bias!
long bias = 32767 + lemin;
return new C2SChunk(bufX(bias, xmin, C2SChunk._OFF, 1), bias, PrettyPrint.pow10(xmin));
}
}
if (leRange < 4294967295l) {
long bias = 2147483647l + lemin;
return new C4SChunk(bufX(bias, xmin, C4SChunk._OFF, 2), bias, PrettyPrint.pow10(xmin));
}
return chunkD();
} // else an integer column
// Compress column into a byte
if (xmin == 0 && 0 <= lemin && lemax <= 255 && ((_naCnt + _catCnt) == 0))
return new C1NChunk(bufX(0, 0, C1NChunk._OFF, 0));
if (lemin < Integer.MIN_VALUE) return new C8Chunk(bufX(0, 0, 0, 3));
if (leRange < 255) { // Span fits in a byte?
if (0 <= min && max < 255) // Span fits in an unbiased byte?
return new C1Chunk(bufX(0, 0, C1Chunk._OFF, 0));
return new C1SChunk(bufX(lemin, xmin, C1SChunk._OFF, 0), lemin, PrettyPrint.pow10i(xmin));
}
// Compress column into a short
if (leRange < 65535) { // Span fits in a biased short?
if (xmin == 0
&& Short.MIN_VALUE < lemin
&& lemax <= Short.MAX_VALUE) // Span fits in an unbiased short?
return new C2Chunk(bufX(0, 0, C2Chunk._OFF, 1));
long bias = (lemin - (Short.MIN_VALUE + 1));
return new C2SChunk(bufX(bias, xmin, C2SChunk._OFF, 1), bias, PrettyPrint.pow10i(xmin));
}
// Compress column into ints
if (Integer.MIN_VALUE < min && max <= Integer.MAX_VALUE) return new C4Chunk(bufX(0, 0, 0, 2));
return new C8Chunk(bufX(0, 0, 0, 3));
}
