func (this *FastPFOR) Compress(in []int32, inpos *cursor.Cursor, inlength int, out []int32, outpos *cursor.Cursor) error {
	inlength = encoding.FloorBy(inlength, DefaultBlockSize)

	if inlength == 0 {
		return errors.New("fastpfor/Compress: inlength = 0. No work done.")
	}

	out[outpos.Get()] = int32(inlength)
	outpos.Increment()

	initoffset := cursor.New()

	copy(this.dataPointers, zeroDataPointers)
	copy(this.freqs, zeroFreqs)

	finalInpos := inpos.Get() + inlength

	for inpos.Get() != finalInpos {
		thissize := int(math.Min(float64(this.pageSize), float64(finalInpos-inpos.Get())))

		if err := this.encodePage(in, inpos, thissize, out, outpos, initoffset); err != nil {
			return errors.New("fastpfor/Compress: " + err.Error())
		}
	}

	return nil
}

func (this *Composition) Uncompress(in []int32, inpos *cursor.Cursor, inlength int, out []int32, outpos *cursor.Cursor) error {
	if inlength == 0 {
		return errors.New("composition/Uncompress: inlength = 0. No work done.")
	}

	init := inpos.Get()
	this.f1.Uncompress(in, inpos, inlength, out, outpos)
	//log.Printf("composition/Uncompress: f1 inpos = %d, outpos = %d, inlength = %d\n", inpos.Get(), outpos.Get(), inlength)
	inlength -= inpos.Get() - init
	this.f2.Uncompress(in, inpos, inlength, out, outpos)
	//log.Printf("composition/Uncompress: f2 inpos = %d, outpos = %d, inlength = %d\n", inpos.Get(), outpos.Get(), inlength)

	return nil
}

func (this *FastPFOR) Uncompress(in []int32, inpos *cursor.Cursor, inlength int, out []int32, outpos *cursor.Cursor) error {
	if inlength == 0 {
		return errors.New("fastpfor/Uncompress: inlength = 0. No work done.")
	}

	mynvalue := in[inpos.Get()]
	inpos.Increment()

	copy(this.dataPointers, zeroDataPointers)

	finalout := outpos.Get() + int(mynvalue)
	for outpos.Get() != finalout {
		thissize := int(math.Min(float64(this.pageSize), float64(finalout-outpos.Get())))
		if err := this.decodePage(in, inpos, out, outpos, thissize); err != nil {
			return errors.New("fastpfor/Uncompress: " + err.Error())
		}
	}
	return nil
}

func (this *VariableByte) Uncompress(in []int32, inpos *cursor.Cursor, inlength int, out []int32, outpos *cursor.Cursor) error {
	if inlength == 0 {
		return errors.New("variablebyte/Uncompress: inlength = 0. No work done.")
	}

	//fmt.Printf("variablebyte/Uncompress: after inlength = %d\n", inlength)

	s := uint(0)
	p := inpos.Get()
	finalp := inpos.Get() + inlength
	tmpoutpos := outpos.Get()
	initoffset := int32(0)
	v := int32(0)
	shift := uint(0)

	for p < finalp {
		c := in[p] >> (24 - s)
		s += 8

		if s == 32 {
			s = 0
			p += 1
		}

		v += ((c & 127) << shift)
		if c&128 == 0 {
			out[tmpoutpos] = v + initoffset
			initoffset = out[tmpoutpos]
			tmpoutpos += 1
			v = 0
			shift = 0
		} else {
			shift += 7
		}

		outpos.Set(tmpoutpos)
		inpos.Add(inlength)
	}

	return nil
}

func (this *VariableByte) Compress(in []int32, inpos *cursor.Cursor, inlength int, out []int32, outpos *cursor.Cursor) error {
	if inlength == 0 {
		return errors.New("variablebyte/Compress: inlength = 0. No work done.")
	}

	//fmt.Printf("variablebyte/Compress: after inlength = %d\n", inlength)

	buf := bytebuffer.NewByteBuffer(inlength * 8)
	initoffset := int32(0)

	tmpinpos := inpos.Get()
	for _, v := range in[tmpinpos : tmpinpos+inlength] {
		val := uint32(v - initoffset)
		initoffset = v

		for val >= 0x80 {
			buf.Put(byte(val) | 0x80)
			val >>= 7
		}
		buf.Put(byte(val))
	}

	for buf.Position()%4 != 0 {
		//fmt.Printf("variablebyte/Compress: putting 128\n")
		buf.Put(128)
	}

	length := buf.Position()
	buf.Flip()
	ibuf := buf.AsInt32Buffer()
	//fmt.Printf("variablebyte/Compress: l = %d, outpos = %d, ibuf = %v, buf = %v\n", length/4, outpos.Get(), ibuf, buf)
	err := ibuf.GetInt32s(out, outpos.Get(), length/4)
	if err != nil {
		//fmt.Printf("variablebyte/Compress: error with GetUint32s - %v\n", err)
		return err
	}
	outpos.Add(length / 4)
	inpos.Add(inlength)
	//fmt.Printf("variablebyte/Compress: out = %v\n", out)

	return nil
}

func (this *FastPFOR) decodePage(in []int32, inpos *cursor.Cursor, out []int32, outpos *cursor.Cursor, thissize int, initoffset *cursor.Cursor) error {
	initpos := int32(inpos.Get())
	wheremeta := in[initpos]
	inpos.Increment()

	inexcept := initpos + wheremeta
	bytesize := in[inexcept]
	inexcept += 1

	this.byteContainer.Clear()
	if err := this.byteContainer.AsInt32Buffer().PutInt32s(in, int(inexcept), int(bytesize/4)); err != nil {
		return err
	}

	inexcept += bytesize / 4
	bitmap := in[inexcept]
	inexcept += 1

	for k := int32(1); k < 33; k++ {
		if bitmap&(1<<uint32(k-1)) != 0 {
			size := in[inexcept]
			inexcept += 1

			if int32(len(this.dataToBePacked[k])) < size {
				this.dataToBePacked[k] = make([]int32, encoding.CeilBy(int(size), 32))
			}

			for j := int32(0); j < size; j += 32 {
				bitpacking.FastUnpack(in, int(inexcept), this.dataToBePacked[k], int(j), int(k))
				inexcept += k
			}
		}
	}

	copy(this.dataPointers, zeroDataPointers)
	tmpoutpos := int32(outpos.Get())
	tmpinpos := int32(inpos.Get())

	delta := make([]int32, DefaultBlockSize)

	run := 0
	run_end := thissize / DefaultBlockSize
	for run < run_end {
		var err error
		var bestb int32
		if bestb, err = this.byteContainer.GetAsInt32(); err != nil {
			return err
		}

		var cexcept int32
		if cexcept, err = this.byteContainer.GetAsInt32(); err != nil {
			return err
		}

		for k := int32(0); k < 128; k += 32 {
			//bitpacking.FastUnpack(in, int(tmpinpos), out, int(tmpoutpos+k), int(bestb))
			bitpacking.FastUnpack(in, int(tmpinpos), delta, int(k), int(bestb))
			tmpinpos += bestb
		}

		if cexcept > 0 {
			var maxbits int32
			if maxbits, err = this.byteContainer.GetAsInt32(); err != nil {
				return err
			}

			index := maxbits - bestb

			for k := int32(0); k < cexcept; k++ {
				var pos int32
				if pos, err = this.byteContainer.GetAsInt32(); err != nil {
					return err
				}

				exceptvalue := this.dataToBePacked[index][this.dataPointers[index]]
				this.dataPointers[index] += 1
				//out[pos + tmpoutpos] |= exceptvalue << uint(bestb)
				delta[pos] |= exceptvalue << uint(bestb)
			}
		}

		// Calculate the original from the deltas, inlining to gain a bit of performance
		offset := int32(initoffset.Get())
		for i, v := range delta {
			n := int32(uint32(v)>>1) ^ ((v << 31) >> 31)
			out[int(tmpoutpos)+i] = n + offset
			offset += n
		}
		initoffset.Set(int(out[tmpoutpos+DefaultBlockSize-1]))

		run += 1
		tmpoutpos += DefaultBlockSize
	}

	outpos.Set(int(tmpoutpos))
	inpos.Set(int(inexcept))

	return nil
}

func (this *FastPFOR) encodePage(in []int32, inpos *cursor.Cursor, thissize int, out []int32, outpos *cursor.Cursor, initoffset *cursor.Cursor) error {
	headerpos := int32(outpos.Get())
	outpos.Increment()
	tmpoutpos := int32(outpos.Get())

	// Clear working area
	copy(this.dataPointers, zeroDataPointers)
	this.byteContainer.Clear()

	tmpinpos := int32(inpos.Get())
	var delta [DefaultBlockSize]int32

	for finalInpos := tmpinpos + int32(thissize) - DefaultBlockSize; tmpinpos <= finalInpos; tmpinpos += DefaultBlockSize {
		// Calculate the deltas, inlining to gain a bit of performance
		offset := int32(initoffset.Get())
		for i, v := range in[tmpinpos : tmpinpos+DefaultBlockSize] {
			n := v - offset
			delta[i] = (n << 1) ^ (n >> 31)
			offset = v
		}

		initoffset.Set(int(in[tmpinpos+DefaultBlockSize-1]))

		//bestb, bestc, maxb := this.getBestBFromData(in[tmpinpos:tmpinpos+DefaultBlockSize])
		bestb, bestc, maxb := this.getBestBFromData(delta[:])
		tmpbestb := bestb
		this.byteContainer.Put(byte(bestb))
		this.byteContainer.Put(byte(bestc))

		if bestc > 0 {
			this.byteContainer.Put(byte(maxb))
			index := maxb - bestb

			if int(this.dataPointers[index]+bestc) >= len(this.dataToBePacked[index]) {
				newSize := int(2 * (this.dataPointers[index] + bestc))

				// make sure it is a multiple of 32.
				// there might be a better way to do this
				newSize = encoding.CeilBy(newSize, 32)
				newSlice := make([]int32, newSize)
				copy(newSlice, this.dataToBePacked[index])
				this.dataToBePacked[index] = newSlice
			}

			for k := int32(0); k < DefaultBlockSize; k++ {
				if uint32(delta[k])>>uint(bestb) != 0 {
					// we have an exception
					this.byteContainer.Put(byte(k))
					this.dataToBePacked[index][this.dataPointers[index]] = int32(uint32(delta[k]) >> uint(tmpbestb))
					this.dataPointers[index] += 1
				}
			}
		}

		for k := int32(0); k < 128; k += 32 {
			bitpacking.FastPack(delta[:], int(k), out, int(tmpoutpos), int(tmpbestb))
			tmpoutpos += tmpbestb
		}
	}

	inpos.Set(int(tmpinpos))
	out[headerpos] = tmpoutpos - headerpos

	for this.byteContainer.Position()&3 != 0 {
		this.byteContainer.Put(0)
	}

	bytesize := int32(this.byteContainer.Position())
	out[tmpoutpos] = bytesize
	tmpoutpos += 1
	howmanyints := bytesize / 4

	this.byteContainer.Flip()
	this.byteContainer.AsInt32Buffer().GetInt32s(out, int(tmpoutpos), int(howmanyints))
	tmpoutpos += howmanyints

	bitmap := int32(0)
	for k := 1; k <= 32; k++ {
		v := this.dataPointers[k]
		if v != 0 {
			bitmap |= (1 << uint(k-1))
		}
	}

	out[tmpoutpos] = bitmap
	tmpoutpos += 1

	for k := 1; k < 33; k++ {
		v := this.dataPointers[k]
		if v != 0 {
			out[tmpoutpos] = v // size
			tmpoutpos += 1
			for j := 0; j < int(v); j += 32 {
				bitpacking.FastPack(this.dataToBePacked[k], j, out, int(tmpoutpos), k)
				tmpoutpos += int32(k)
			}
		}
	}

	outpos.Set(int(tmpoutpos))
//.........这里部分代码省略.........

func (this *BP32) Uncompress(in []int32, inpos *cursor.Cursor, inlength int, out []int32, outpos *cursor.Cursor) error {
	if inlength == 0 {
		return errors.New("BP32/Uncompress: Length is 0. No work done.")
	}

	outlength := in[inpos.Get()]
	inpos.Increment()

	tmpinpos := inpos.Get()
	initoffset := int32(0)

	//log.Printf("bp32/Uncompress: outlength = %d, inpos = %d, outpos = %d\n", outlength, inpos.Get(), outpos.Get())
	for s := outpos.Get(); s < outpos.Get()+int(outlength); s += 32 * 4 {
		tmp := in[tmpinpos]
		mbits1 := tmp >> 24
		mbits2 := (tmp >> 16) & 0xFF
		mbits3 := (tmp >> 8) & 0xFF
		mbits4 := (tmp) & 0xFF

		//log.Printf("bp32/Uncopmress: mbits1 = %d, mbits2 = %d, mbits3 = %d, mbits4 = %d, s = %d\n", mbits1, mbits2, mbits3, mbits4, s)
		tmpinpos += 1

		bitpacking.DeltaUnpack(initoffset, in, tmpinpos, out, s, int(mbits1))
		tmpinpos += int(mbits1)
		initoffset = out[s+31]
		//log.Printf("bp32/Uncompress: out = %v\n", out)

		bitpacking.DeltaUnpack(initoffset, in, tmpinpos, out, s+32, int(mbits2))
		tmpinpos += int(mbits2)
		initoffset = out[s+32+31]
		//log.Printf("bp32/Uncompress: out = %v\n", out)

		bitpacking.DeltaUnpack(initoffset, in, tmpinpos, out, s+2*32, int(mbits3))
		tmpinpos += int(mbits3)
		initoffset = out[s+2*32+31]
		//log.Printf("bp32/Uncompress: out = %v\n", out)

		bitpacking.DeltaUnpack(initoffset, in, tmpinpos, out, s+3*32, int(mbits4))
		tmpinpos += int(mbits4)
		initoffset = out[s+3*32+31]
		//log.Printf("bp32/Uncompress: out = %v\n", out)
	}

	outpos.Add(int(outlength))
	inpos.Set(tmpinpos)

	return nil
}

func (this *BP32) Compress(in []int32, inpos *cursor.Cursor, inlength int, out []int32, outpos *cursor.Cursor) error {
	//log.Printf("bp32/Compress: before inlength = %d\n", inlength)

	inlength = encoding.FloorBy(inlength, DefaultBlockSize)

	if inlength == 0 {
		return errors.New("BP32/Compress: block size less than 128. No work done.")
	}

	//log.Printf("bp32/Compress: after inlength = %d, len(in) = %d\n", inlength, len(in))

	out[outpos.Get()] = int32(inlength)
	outpos.Increment()

	tmpoutpos := outpos.Get()
	initoffset := int32(0)
	s := inpos.Get()
	finalinpos := s + inlength

	for ; s < finalinpos; s += DefaultBlockSize {
		mbits1 := encoding.DeltaMaxBits(initoffset, in[s:s+32])
		initoffset2 := in[s+31]
		mbits2 := encoding.DeltaMaxBits(initoffset2, in[s+32:s+2*32])
		initoffset3 := in[s+32+31]
		mbits3 := encoding.DeltaMaxBits(initoffset3, in[s+2*32:s+3*32])
		initoffset4 := in[s+2*32+31]
		mbits4 := encoding.DeltaMaxBits(initoffset4, in[s+3*32:s+4*32])

		//log.Printf("bp32/Compress: tmpoutpos = %d, s = %d\n", tmpoutpos, s)

		out[tmpoutpos] = (mbits1 << 24) | (mbits2 << 16) | (mbits3 << 8) | mbits4
		tmpoutpos += 1

		//log.Printf("bp32/Compress: mbits1 = %d, mbits2 = %d, mbits3 = %d, mbits4 = %d, s = %d\n", mbits1, mbits2, mbits3, mbits4, out[tmpoutpos-1])

		bitpacking.DeltaPack(initoffset, in, s, out, tmpoutpos, int(mbits1))
		//encoding.PrintUint32sInBits(in, s, 32)
		//encoding.PrintUint32sInBits(out, tmpoutpos, int(mbits1))
		tmpoutpos += int(mbits1)

		bitpacking.DeltaPack(initoffset2, in, s+32, out, tmpoutpos, int(mbits2))
		//encoding.PrintUint32sInBits(in, s+32, 32)
		//encoding.PrintUint32sInBits(out, tmpoutpos, int(mbits2))
		tmpoutpos += int(mbits2)

		bitpacking.DeltaPack(initoffset3, in, s+2*32, out, tmpoutpos, int(mbits3))
		//encoding.PrintUint32sInBits(in, s+2*32, 32)
		//encoding.PrintUint32sInBits(out, tmpoutpos, int(mbits3))
		tmpoutpos += int(mbits3)

		bitpacking.DeltaPack(initoffset4, in, s+3*32, out, tmpoutpos, int(mbits4))
		//encoding.PrintUint32sInBits(in, s+3*32, 32)
		//encoding.PrintUint32sInBits(out, tmpoutpos, int(mbits4))
		tmpoutpos += int(mbits4)

		initoffset = in[s+3*32+31]
	}

	inpos.Add(inlength)
	outpos.Set(tmpoutpos)

	return nil
}

func (this *BP32) Uncompress(in []int32, inpos *cursor.Cursor, inlength int, out []int32, outpos *cursor.Cursor) error {
	if inlength == 0 {
		return errors.New("BP32/Uncompress: Length is 0. No work done.")
	}

	outlength := int(in[inpos.Get()])
	inpos.Increment()

	tmpinpos := inpos.Get()

	for s := outpos.Get(); s < outpos.Get()+outlength; s += 32 * 4 {
		tmp := in[tmpinpos]
		mbits1 := tmp >> 24
		mbits2 := (tmp >> 16) & 0xFF
		mbits3 := (tmp >> 8) & 0xFF
		mbits4 := (tmp) & 0xFF

		tmpinpos += 1

		bitpacking.FastUnpack(in, tmpinpos, out, s, int(mbits1))
		tmpinpos += int(mbits1)

		bitpacking.FastUnpack(in, tmpinpos, out, s+32, int(mbits2))
		tmpinpos += int(mbits2)

		bitpacking.FastUnpack(in, tmpinpos, out, s+2*32, int(mbits3))
		tmpinpos += int(mbits3)

		bitpacking.FastUnpack(in, tmpinpos, out, s+3*32, int(mbits4))
		tmpinpos += int(mbits4)
	}

	outpos.Add(outlength)
	inpos.Set(tmpinpos)

	return nil
}

func (this *BP32) Compress(in []int32, inpos *cursor.Cursor, inlength int, out []int32, outpos *cursor.Cursor) error {

	inlength = encoding.FloorBy(inlength, DefaultBlockSize)

	if inlength == 0 {
		return errors.New("BP32/Compress: block size less than 128. No work done.")
	}

	out[outpos.Get()] = int32(inlength)
	outpos.Increment()

	tmpoutpos := outpos.Get()
	s := inpos.Get()
	finalinpos := s + inlength

	for ; s < finalinpos; s += DefaultBlockSize {
		mbits1 := encoding.MaxBits(in[s : s+32])
		mbits2 := encoding.MaxBits(in[s+32 : s+2*32])
		mbits3 := encoding.MaxBits(in[s+2*32 : s+3*32])
		mbits4 := encoding.MaxBits(in[s+3*32 : s+4*32])

		out[tmpoutpos] = (mbits1 << 24) | (mbits2 << 16) | (mbits3 << 8) | mbits4
		tmpoutpos += 1
		bitpacking.FastPackWithoutMask(in, s, out, tmpoutpos, int(mbits1))
		tmpoutpos += int(mbits1)
		bitpacking.FastPackWithoutMask(in, s+32, out, tmpoutpos, int(mbits2))
		tmpoutpos += int(mbits2)
		bitpacking.FastPackWithoutMask(in, s+2*32, out, tmpoutpos, int(mbits3))
		tmpoutpos += int(mbits3)
		bitpacking.FastPackWithoutMask(in, s+3*32, out, tmpoutpos, int(mbits4))
		tmpoutpos += int(mbits4)
	}

	inpos.Add(inlength)
	outpos.Set(tmpoutpos)

	return nil
}

func (this *FastPFOR) decodePage(in []int32, inpos *cursor.Cursor, out []int32, outpos *cursor.Cursor, thissize int) error {
	initpos := int32(inpos.Get())
	wheremeta := in[initpos]
	inpos.Increment()

	inexcept := initpos + wheremeta
	bytesize := in[inexcept]
	inexcept += 1
	mybytearray := in[inexcept:]
	mybp := uint(0)

	inexcept += (bytesize + 3) / 4
	bitmap := in[inexcept]
	inexcept += 1

	for k := int32(1); k < 33; k++ {
		if bitmap&(1<<uint32(k-1)) != 0 {
			size := in[inexcept]
			inexcept += 1

			if int32(len(this.dataToBePacked[k])) < size {
				this.dataToBePacked[k] = make([]int32, encoding.CeilBy(int(size), 32))
			}
			for j := int32(0); j < size; j += 32 {
				bitpacking.FastUnpack(in, int(inexcept), this.dataToBePacked[k], int(j), int(k))
				inexcept += k
			}
		}
	}

	copy(this.dataPointers, zeroDataPointers)
	tmpoutpos := uint32(outpos.Get())
	tmpinpos := uint32(inpos.Get())

	run := 0
	run_end := thissize / DefaultBlockSize
	for run < run_end {
		bestb := uint32(grapByte(mybytearray, mybp))
		mybp++
		cexcept := int32(grapByte(mybytearray, mybp))
		mybp++
		for k := uint32(0); k < 128; k += 32 {
			bitpacking.FastUnpack(in, int(tmpinpos), out, int(tmpoutpos+k), int(bestb))
			tmpinpos += bestb
		}

		if cexcept > 0 {
			maxbits := uint32(grapByte(mybytearray, mybp))
			mybp++
			index := maxbits - bestb
			// assuming that the Go compiler is bad, we move everything that is indexed outside the upcoming loop
			packedexceptions := this.dataToBePacked[index]
			myindex := this.dataPointers[index]

			for k := int32(0); k < cexcept; k++ {
				pos := uint32(grapByte(mybytearray, mybp))
				mybp++
				exceptvalue := packedexceptions[myindex]
				myindex++
				out[pos+tmpoutpos] |= exceptvalue << bestb
			}
			this.dataPointers[index] = myindex
		}

		run += 1
		tmpoutpos += DefaultBlockSize
	}

	outpos.Set(int(tmpoutpos))
	inpos.Set(int(inexcept))

	return nil
}

func (this *BP32) Uncompress(in []int32, inpos *cursor.Cursor, inlength int, out []int32, outpos *cursor.Cursor) error {
	if inlength == 0 {
		return errors.New("zigzag_bp32/Uncompress: Length is 0. No work done.")
	}

	outlength := int(in[inpos.Get()])
	inpos.Increment()

	tmpinpos := inpos.Get()
	s := outpos.Get()
	finalinpos := s + outlength
	delta := make([]int32, DefaultBlockSize)

	//log.Printf("zigzag_bp32/Uncompress: outlength = %d, inpos = %d, outpos = %d\n", outlength, inpos.Get(), outpos.Get())
	for ; s < finalinpos; s += DefaultBlockSize {
		tmp := in[tmpinpos]
		mbits1 := tmp >> 24
		mbits2 := (tmp >> 16) & 0xFF
		mbits3 := (tmp >> 8) & 0xFF
		mbits4 := (tmp) & 0xFF

		//log.Printf("zigzag_bp32/Uncopmress: mbits1 = %d, mbits2 = %d, mbits3 = %d, mbits4 = %d, s = %d\n", mbits1, mbits2, mbits3, mbits4, s)
		tmpinpos += 1

		bitpacking.FastUnpack(in, tmpinpos, delta, 0, int(mbits1))
		tmpinpos += int(mbits1)
		//log.Printf("zigzag_bp32/Uncompress: delta = %v\n", out)

		bitpacking.FastUnpack(in, tmpinpos, delta, 32, int(mbits2))
		tmpinpos += int(mbits2)
		//log.Printf("zigzag_bp32/Uncompress: delta = %v\n", out)

		bitpacking.FastUnpack(in, tmpinpos, delta, 64, int(mbits3))
		tmpinpos += int(mbits3)
		//log.Printf("zigzag_bp32/Uncompress: delta = %v\n", out)

		bitpacking.FastUnpack(in, tmpinpos, delta, 96, int(mbits4))
		tmpinpos += int(mbits4)

		encoding.InverseZigZagDelta(delta, out[s:s+DefaultBlockSize])

		//log.Printf("zigzag_bp32/Uncompress: delta = %v\n", delta)
		//log.Printf("zigzag_bp32/Uncompress: out = %v\n", out[s:s+DefaultBlockSize])

	}

	outpos.Add(outlength)
	inpos.Set(tmpinpos)

	return nil
}

func (this *BP32) Compress(in []int32, inpos *cursor.Cursor, inlength int, out []int32, outpos *cursor.Cursor) error {
	//log.Printf("zigzag_bp32/Compress: before inlength = %d\n", inlength)

	inlength = encoding.FloorBy(inlength, DefaultBlockSize)

	if inlength == 0 {
		return errors.New("zigzag_bp32/Compress: block size less than 128. No work done.")
	}

	//log.Printf("zigzag_bp32/Compress: after inlength = %d, len(in) = %d\n", inlength, len(in))

	out[outpos.Get()] = int32(inlength)
	outpos.Increment()

	tmpoutpos := outpos.Get()
	s := inpos.Get()
	finalinpos := s + inlength
	delta := make([]int32, DefaultBlockSize)

	for ; s < finalinpos; s += DefaultBlockSize {
		encoding.ZigZagDelta(in[s:s+DefaultBlockSize], delta)
		//log.Printf("zigzag_bp32/Compress: in = %v\n", in[s:s+DefaultBlockSize])
		//log.Printf("zigzag_bp32/Compress: delta = %v\n", delta)

		mbits1 := encoding.MaxBits(delta[0:32])
		mbits2 := encoding.MaxBits(delta[32:64])
		mbits3 := encoding.MaxBits(delta[64:96])
		mbits4 := encoding.MaxBits(delta[96:128])

		//log.Printf("zigzag_bp32/Compress: tmpoutpos = %d, s = %d\n", tmpoutpos, s)

		out[tmpoutpos] = (mbits1 << 24) | (mbits2 << 16) | (mbits3 << 8) | mbits4
		tmpoutpos += 1

		//log.Printf("zigzag_bp32/Compress: mbits1 = %d, mbits2 = %d, mbits3 = %d, mbits4 = %d, s = %d\n", mbits1, mbits2, mbits3, mbits4, out[tmpoutpos-1])

		bitpacking.FastPackWithoutMask(delta, 0, out, tmpoutpos, int(mbits1))
		//encoding.PrintUint32sInBits(in[s:s+32])
		//encoding.PrintUint32sInBits(out[tmpoutpos:tmpoutpos+int(mbits1]))
		tmpoutpos += int(mbits1)

		bitpacking.FastPackWithoutMask(delta, 32, out, tmpoutpos, int(mbits2))
		//encoding.PrintUint32sInBits(in, s+32, 32)
		//encoding.PrintUint32sInBits(out, tmpoutpos, int(mbits2))
		tmpoutpos += int(mbits2)

		bitpacking.FastPackWithoutMask(delta, 64, out, tmpoutpos, int(mbits3))
		//encoding.PrintUint32sInBits(in, s+2*32, 32)
		//encoding.PrintUint32sInBits(out, tmpoutpos, int(mbits3))
		tmpoutpos += int(mbits3)

		bitpacking.FastPackWithoutMask(delta, 96, out, tmpoutpos, int(mbits4))
		//encoding.PrintUint32sInBits(in, s+3*32, 32)
		//encoding.PrintUint32sInBits(out, tmpoutpos, int(mbits4))
		tmpoutpos += int(mbits4)

		//log.Printf("zigzag_bp32/Compress: out = %v\n", out[s:s+DefaultBlockSize])
	}

	inpos.Add(inlength)
	outpos.Set(tmpoutpos)

	return nil
}