RFC1: serialized format (storage) for RASTER type

Author: Sandro Santilli <strk@…>
Date: 2009-01-28
Status: Adopted

Revisions

2011-01-24 by Jorge Arévalo

• Adds isNodataValue bit to band flags

Goals

The goals of the serialized version for RASTER type are:

• Small memory footprint on deserialization This means that the amount of allocated memory required for deserialization is minimal

• Fast access Access to band data must be aligned, saving from memory copies on full scan.

• Ease of format switch On-disk format must be allowed to change w/out need for dump-reload of the whole database.

The first two goals boil down to forcing alignment of band data in the serialized format itself, which in turn will require variable padding based on pixeltype of each band.

For simplicity we will ensure that each band of the raster starts at the 8-byte boundary and thus pad previous structures in the stream accordingly.

The structure will then look like this:

[HEADER]  [BAND0]    [BAND1]    [BAND2]
          ^aligned   ^aligned   ^aligned

The third goal can be accomplished by adding a version number to the serialized format so that in case of changes the deserializer can pick the correct parsing procedure based on that.

The HEADER

PostgreSQL forces a 4-byte size field a the start of the detoasted datum, and ensure this start of structure is aligned to 8-byte. We'll add version number right after it, and then make sure the total size is a multiple of 8 bytes.

The following structure is composed by 8 slots of 8-bytes, totaling 64 bytes:

struct rt_raster_serialized_t {

    /*---[ 8 byte boundary ]---{ */
    uint32_t size;    /* required by postgresql: 4 bytes */
    uint16_t version; /* format version (this is version 0): 2 bytes */
    uint16_t numBands; /* Number of bands: 2 bytes */

    /* }---[ 8 byte boundary ]---{ */
    double scaleX; /* pixel width: 8 bytes */

    /* }---[ 8 byte boundary ]---{ */
    double scaleY; /* pixel height: 8 bytes */

    /* }---[ 8 byte boundary ]---{ */
    double ipX; /* insertion point X: 8 bytes */

    /* }---[ 8 byte boundary ]---{ */
    double ipY; /* insertion point Y: 8 bytes */

    /* }---[ 8 byte boundary ]---{ */
    double skewX;  /* rotation about the X axis: 8 bytes */

    /* }---[ 8 byte boundary ]---{ */
    double skewY;  /* rotation about the Y axis: 8 bytes */

    /* }---[ 8 byte boundary ]--- */
    int32_t srid; /* Spatial reference id: 4 bytes */
    uint16_t width;  /* pixel columns: 2 bytes */
    uint16_t height; /* pixel rows: 2 bytes */
};

The BANDS

Given the serialized raster header structure above, it is guaranteed that a serialized band always start at 8-bytes boundary, so it's simpler to compute padding required at the end of each band to ensure next band will be guaranteed the same assumption.

We'll need to take 2 padding spots into account: the first is to ensure actual band data is aligned accordingly to the pixel type (and storage flag) needs, the second is to ensure next band (if any) will also be aligned to 8-bytes:

[PIXELTYPE+STORAGE_FLAG] [DATA_PADDING] [DATA] [TRAILING_PADDING]

The total size of a band's serialized form in bytes must be a multiple of 8.

The maximum required data padding size will be of 7 bytes (64bit pixel type). The maximum required trailing padding size will be of 7 bytes.

Pixel type and storage flag

Pixel type specifies type of pixel values in a band. Storage flag specifies whether the band data is stored as part of the datum or is to be found on the server's filesytem.

There are currently 11 supported pixel value types, so 4 bits are enough to account for all. We'll reserve the upper 4 bits for generic flags and define upmost as storage flag:

#define BANDTYPE_FLAGS_MASK 0xF0
#define BANDTYPE_PIXTYPE_MASK 0x0F

#define BANDTYPE_FLAG_OFFDB     (1<<7)
#define BANDTYPE_FLAG_HASNODATA (1<<6)
#define BANDTYPE_FLAG_ISNODATA  (1<<5)
#define BANDTYPE_FLAG_RESERVED3 (1<<4)

Data padding

Band alignment depends on pixeltypes, as follows:

• PT_1BB, PT_2BUI, PT_4BUI, PT_8BSI, PT_8BUI: No alignment required, each value is 1 byte.

• PT_16BSI, PT_16BUI: Data must be aligned to 2-bytes boundary.

• PT_32BSI, PT_32BUI, PT_32BF: Data must be aligned to 4-bytes boundary.

• PT_64BF: Data must be aligned to 8-bytes boundary.

Accordingly we can then define the following structures:

struct rt_band8_serialized_t {
    uint8_t pixeltype;
    uint8_t data[1]; /* no data padding */
}

struct rt_band16_serialized_t {
    uint8_t pixeltype;
    uint8_t padding; /* 1-byte padding */
    uint8_t data[1];
}

struct rt_band32_serialized_t {
    uint8_t pixeltype;
    uint8_t padding[3]; /* 3-bytes padding */
    uint8_t data[1];
}

struct rt_band64_serialized_t {
    uint8_t pixeltype;
    uint8_t padding[7]; /* 7-bytes padding */
    uint8_t data[1];
}

And an abstract base class:

struct rt_band_serialized_t {
    uint8_t pixeltype
}

Data

The band data - guaranteed to be always aligned as required by pixeltype - will start with the nodata value.

After that we may have pixel values or off-db raster reference depending on OFFDB flag in the pixeltype field:

• For in-db bands the nodata value is followed by a value for each column in first row, then in second row and so on. For example, a 2x2 raster band data will have this form:

  [nodata] [x:0,y:0] [x:1,y:0] [x:0,y:1] [x:1,y:1]
  

  Where the size of the […] blocks is 1,2,4 or 8 bytes depending on pixeltype. Endiannes of multi-bytes value is the host endiannes.

• For off-db bands the nodata value is followed by a band number followed by a null-terminated string expressing the path to the raster file:

  [nodata] [bandno] [path]
  

  Where the size of the [nodata] block is 1,2,4 or 8 bytes depending on pixeltype (endiannes of multi-bytes value is the host endiannes), size of [bandno] is 1 byte, and [path] is null-terminated.

Trailing padding

The trailing band padding is used to ensure next band (if any) will start on the 8-bytes boundary. It is both dependent on raster dimensions (number of values) and band data pixel type (size of each value).

In order to obtain the required padding size for a band we'll need to compute the minimum size required to hold the band data, add the data padding and pixeltype sizes, and then grow the resulting size to reach a multiple of 8 bytes:

size_t 
rt_band_serialized_size(rt_context ctx, rt_band band)
{
    rt_pixtype pixtype = rt_band_get_pixtype(ctx, band);
    size_t sz;

    /* pixeltype + data padding */
    sz = rt_pixtype_alignment(ctx, pixtype);

    /* add data size */
    sz += rt_band_get_data_size(ctx, band);

    /* grow size to reach a multiple of 8 bytes */
    sz = TYPEALIGN(sz, 8);

    assert( !(sz%8) );

    return sz;
}

Example sizes

255x255 single band PT_16BUI:
    header size:                        64 +
    pixeltype+data_padding:              2 +
    data size:     (255*255+1)*2 == 130052 =
                                    130118 +
    trailing padding:                    2 =
    total size:                     130120 (~127k)

255x255 single band PT_8BUI:
    header size:                        64 +
    pixeltype+data_padding:              1 +
    data size:        (255*255+1) == 65026 =
                                     65091 + 
    trailing padding:                    5 =
    total size:                      65096 (~63k)

64x64 single band PT_16BSI:
    header size:                        64 +
    pixeltype+data_padding:              2 +
    data size:         (64*64+1)*2 == 8194 =
                                      8260 + 
    trailing padding:                    4 =
    total size:                       8264 (~8k -- >page size)

64x64 single band PT_8BUI:
    header size:                        64 +
    pixeltype+data_padding:              1 +
    data size:           (64*64+1) == 4097 =
                                      4162 + 
    trailing padding:                    6 =
    total size:                       4168 (~4k)