update specification and comments

This commit is contained in:
Yann Collet 2016-07-01 20:55:28 +02:00
parent 406dcca98a
commit 2fa9904844
4 changed files with 200 additions and 94 deletions

View File

@ -51,7 +51,7 @@
/*-*************************************
* Common constants
***************************************/
#define ZSTD_OPT_DEBUG 0 // 3 = compression stats; 5 = check encoded sequences; 9 = full logs
#define ZSTD_OPT_DEBUG 0 /* 3 = compression stats; 5 = check encoded sequences; 9 = full logs */
#include <stdio.h>
#if defined(ZSTD_OPT_DEBUG) && ZSTD_OPT_DEBUG>=9
#define ZSTD_LOG_PARSER(...) printf(__VA_ARGS__)

View File

@ -427,21 +427,8 @@ static void ZSTD_reduceIndex (ZSTD_CCtx* zc, const U32 reducerValue)
*/
/* Frame descriptor
/* Frame header :
// old
1 byte - Alloc :
bit 0-3 : windowLog - ZSTD_WINDOWLOG_ABSOLUTEMIN (see zstd_internal.h)
bit 4 : reserved for windowLog (must be zero)
bit 5 : reserved (must be zero)
bit 6-7 : Frame content size : unknown, 1 byte, 2 bytes, 8 bytes
1 byte - checker :
bit 0-1 : dictID (0, 1, 2 or 4 bytes)
bit 2-7 : reserved (must be zero)
// new
1 byte - FrameHeaderDescription :
bit 0-1 : dictID (0, 1, 2 or 4 bytes)
bit 2-4 : reserved (must be zero)
@ -453,24 +440,24 @@ static void ZSTD_reduceIndex (ZSTD_CCtx* zc, const U32 reducerValue)
bit 0-2 : octal Fractional (1/8th)
bit 3-7 : Power of 2, with 0 = 1 KB (up to 2 TB)
Optional : content size (0, 1, 2, 4 or 8 bytes)
0 : unknown
1 : 0-255 bytes
2 : 256 - 65535+256
8 : up to 16 exa
Optional : dictID (0, 1, 2 or 4 bytes)
Automatic adaptation
0 : no dictID
1 : 1 - 255
2 : 256 - 65535
4 : all other values
Optional : content size (0, 1, 2, 4 or 8 bytes)
0 : unknown
1 : 0-255 bytes
2 : 256 - 65535+256
8 : up to 16 exa
*/
/* Block format description
Block = Literal Section - Sequences Section
Block = Literals Section - Sequences Section
Prerequisite : size of (compressed) block, maximum size of regenerated data
1) Literal Section
@ -478,7 +465,7 @@ static void ZSTD_reduceIndex (ZSTD_CCtx* zc, const U32 reducerValue)
1.1) Header : 1-5 bytes
flags: 2 bits
00 compressed by Huff0
01 unused
01 repeat
10 is Raw (uncompressed)
11 is Rle
Note : using 01 => Huff0 with precomputed table ?
@ -514,7 +501,7 @@ static void ZSTD_reduceIndex (ZSTD_CCtx* zc, const U32 reducerValue)
else => 5 bytes (2-2-18-18)
big endian convention
1- CTable available (stored into workspace ?)
1- CTable available (stored into workspace)
2- Small input (fast heuristic ? Full comparison ? depend on clevel ?)

View File

@ -207,20 +207,8 @@ void ZSTD_copyDCtx(ZSTD_DCtx* dstDCtx, const ZSTD_DCtx* srcDCtx)
*/
/* Frame descriptor
/* Frame Header :
// old
1 byte - Alloc :
bit 0-3 : windowLog - ZSTD_WINDOWLOG_ABSOLUTEMIN (see zstd_internal.h)
bit 4 : reserved for windowLog (must be zero)
bit 5 : reserved (must be zero)
bit 6-7 : Frame content size : unknown, 1 byte, 2 bytes, 8 bytes
1 byte - checker :
bit 0-1 : dictID (0, 1, 2 or 4 bytes)
bit 2-7 : reserved (must be zero)
// new
1 byte - FrameHeaderDescription :
bit 0-1 : dictID (0, 1, 2 or 4 bytes)
bit 2 : checksumFlag
@ -454,16 +442,14 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
const void* src, size_t srcSize) /* note : srcSize < BLOCKSIZE */
{
const BYTE* const istart = (const BYTE*) src;
litBlockType_t lbt;
if (srcSize < MIN_CBLOCK_SIZE) return ERROR(corruption_detected);
lbt = (litBlockType_t)(istart[0]>> 6);
switch(lbt)
switch((litBlockType_t)(istart[0]>> 6))
{
case lbt_huffman:
{ size_t litSize, litCSize, singleStream=0;
U32 lhSize = ((istart[0]) >> 4) & 3;
U32 lhSize = (istart[0] >> 4) & 3;
if (srcSize < 5) return ERROR(corruption_detected); /* srcSize >= MIN_CBLOCK_SIZE == 3; here we need up to 5 for lhSize, + cSize (+nbSeq) */
switch(lhSize)
{

View File

@ -1,5 +1,5 @@
Zstandard Frame Format Description
==================================
Zstandard Compression Format Description
========================================
### Notices
@ -16,7 +16,7 @@ Distribution of this document is unlimited.
### Version
0.1.0 (30/06/2016)
0.1.0 (30/06/2016 - unfinished)
Introduction
@ -53,17 +53,32 @@ A compliant decompressor must be able to decompress
at least one working set of parameters
that conforms to the specifications presented here.
It may also ignore informative fields, such as checksum.
Whenever it does not support a specific parameter within the compressed stream,
it must produce a non-ambiguous error code
and associated error message explaining which parameter is unsupported.
Whenever it does not support a parameter defined in the compressed stream,
it must produce a non-ambiguous error code and associated error message
explaining which parameter is unsupported.
Definitions
-----------
A content compressed by Zstandard is transformed into a Zstandard __frame__.
Multiple frames can be appended into a single file or stream.
A frame is totally independent, has a defined beginning and end,
and a set of parameters which tells the decoder how to decompress it.
A frame encapsulates one or multiple __blocks__.
Each block can be compressed or not,
and has a guaranteed maximum content size, which depends on frame parameters.
Unlike frames, each block depends on previous blocks for proper decoding.
However, each block can be decompressed without waiting for its successor,
allowing streaming operations.
General Structure of Zstandard Frame format
-------------------------------------------
| MagicNb | F. Header | Block | (...) | EndMark |
|:-------:|:----------:| ----- | ----- | ------- |
| 4 bytes | 2-14 bytes | | | 3 bytes |
| MagicNb | F. Header | Block | (More blocks) | EndMark |
|:-------:|:----------:| ----- | ------------- | ------- |
| 4 bytes | 2-14 bytes | | | 3 bytes |
__Magic Number__
@ -73,7 +88,6 @@ Value : 0xFD2FB527
__Frame Header__
2 to 14 Bytes, to be detailed in the next part.
Most important part of the spec.
__Data Blocks__
@ -87,11 +101,11 @@ This last block header may optionally host a __Content Checksum__ .
__Content Checksum__
Content Checksum verify that the full content has been decoded correctly.
Content Checksum verify that frame content has been regenrated correctly.
The content checksum is the result
of [xxh64() hash function](https://www.xxHash.com)
digesting the original (decoded) data as input, and a seed of zero.
Bits from 11 to 32 (included) are extracted to form the 22 bits checksum
Bits from 11 to 32 (included) are extracted to form a 22 bits checksum
stored into the last block header.
```
contentChecksum = (XXH64(content, size, 0) >> 11) & (1<<22)-1);
@ -114,52 +128,64 @@ The ability to decode multiple concatenated frames
within a single stream or file is left outside of this specification.
As an example, the reference `zstd` command line utility is able
to decode all concatenated frames in their sequential order,
presenting the final decompressed result as if it was a single frame.
delivering the final decompressed result as if it was a single content.
Frame Header
----------------
-------------
| FHD | (WD) | (Content Size) | (dictID) |
| ------- | --------- |:--------------:| --------- |
| 1 byte | 0-1 byte | 0 - 8 bytes | 0-4 bytes |
Frame header uses a minimum of 2 bytes,
Frame header has a variable size, which uses a minimum of 2 bytes,
and up to 14 bytes depending on optional parameters.
__FHD byte__ (Frame Header Descriptor)
The first Header's byte is called the Frame Header Descriptor.
It tells which other fields are present.
Decoding this byte is enough to get the full size of the Frame Header.
| BitNb | 7-6 | 5 | 4 | 3 | 2 | 1-0 |
| ------- | ------ | ------- | ------ | -------- | -------- | -------- |
|FieldName| FCSize | Segment | Unused | Reserved | Checksum | dictID |
In the table, bit 7 is highest bit, while bit 0 is lowest.
In this table, bit 7 is highest bit, while bit 0 is lowest.
__Frame Content Size flag__
This is a 2-bits flag (`= FHD >> 6`),
telling if original data size is provided within the header
specifying if decompressed data size is provided within the header.
| Value | 0 | 1 | 2 | 3 |
| ------- | --- | --- | --- | --- |
|FieldSize| 0-1 | 2 | 4 | 8 |
Value 0 is special : it means `0` (data size not provided)
_if_ the `WD` byte is present.
Otherwise, it means `1` byte (data size <= 255 bytes).
Value 0 has a double meaning :
it either means `0` (size not provided) _if_ the `WD` byte is present,
or it means `1` byte (size <= 255 bytes).
__Single Segment__
If this flag is set,
data shall be regenerated within a single continuous memory segment.
In which case, `WD` byte is not present,
In which case, `WD` byte __is not present__,
but `Frame Content Size` field necessarily is.
The size of the memory segment must be at least `>= Frame Content Size`.
In order to preserve decoder from unreasonable memory requirement,
As a consequence, the decoder must allocate a memory segment
of size `>= Frame Content Size`.
In order to preserve the decoder from unreasonable memory requirement,
a decoder can refuse a compressed frame
which requests a memory size beyond decoder's authorized range.
For broader compatibility, decoders are recommended to support
memory sizes of 8 MB at least.
However, this is merely a recommendation,
and each decoder is free to support higher or lower limits,
depending on local limitations.
__Unused bit__
The value of this bit is unimportant
@ -170,7 +196,7 @@ to signal a property which is not required to properly decode the frame.
__Reserved bit__
This bit is reserved for some future feature.
Its value must be zero.
Its value _must be zero_.
A decoder compliant with this specification version must ensure it is not set.
This bit may be used in a future revision,
to signal a feature that must be interpreted in order to decode the frame.
@ -193,7 +219,7 @@ telling if a dictionary ID is provided within the header
__WD byte__ (Window Descriptor)
Provides guarantees on maximum back-reference distance
that will be used within compressed data.
that will be present within compressed data.
This information is useful for decoders to allocate enough memory.
| BitNb | 7-3 | 0-2 |
@ -208,16 +234,25 @@ windowAdd = (windowBase / 8) * Mantissa;
windowSize = windowBase + windowAdd;
```
The minimum window size is 1 KB.
The maximum value is (15*(2^38))-1 bytes, which is almost 1.875 TB.
The maximum size is (15*(2^38))-1 bytes, which is almost 1.875 TB.
`WD` byte is optional. It's not present in `single segment` mode.
In which case, the maximum back-reference distance is the content size itself, which can be any value from 1 to 2^64-1 bytes (16 EB).
To properly decode compressed data,
a decoder will need to allocate a buffer of at least `windowSize` bytes.
Note that `WD` byte is optional. It's not present in `single segment` mode.
In which case, the maximum back-reference distance is the content size itself,
which can be any value from 1 to 2^64-1 bytes (16 EB).
In order to preserve decoder from unreasonable memory requirements,
a decoder can refuse a compressed frame
which requests a memory size beyond decoder's authorized range.
For better interoperability, decoders are recommended to be compatible with window sizes up to 8 MB. Encoders are recommended to not request more than 8 MB. It's just a recommendation, decoders are free to accept or refuse larger or lower values.
For better interoperability,
decoders are recommended to be compatible with window sizes of 8 MB.
Encoders are recommended to not request more than 8 MB.
It's merely a recommendation though,
decoders are free to support larger or lower limits,
depending on local limitations.
__Frame Content Size__
@ -235,11 +270,11 @@ Format is Little endian.
| 8 | 0 - 2^64-1 |
When field size is 1, 4 or 8 bytes, the value is read directly.
When field size is 2, an offset of 256 is added.
It's possible to represent a small size of `18` using the 8-bytes variant.
A size of `0` means `data size is unknown`.
In which case, the `WD` byte will be the only hint
to determine memory allocation.
When field size is 2, _an offset of 256 is added_.
It's allowed to represent a small size (ex: `18`) using the 8-bytes variant.
A size of `0` means `content size is unknown`.
In which case, the `WD` byte will necessarily be present,
and becomes the only hint to determine memory allocation.
In order to preserve decoder from unreasonable memory requirement,
a decoder can refuse a compressed frame
@ -257,7 +292,8 @@ Field size depends on __Dictionary ID flag__.
2 bytes can represent an ID 0-65535.
4 bytes can represent an ID 0-(2^32-1).
It's possible to represent a small ID (for example `13`) with a large 4-bytes dictionary ID, losing some efficiency in the process.
It's allowed to represent a small ID (for example `13`)
with a large 4-bytes dictionary ID, losing some efficiency in the process.
Data Blocks
@ -270,10 +306,10 @@ Data Blocks
__Block Header__
This field uses 3-bytes, format is big-endian.
This field uses 3-bytes, format is __big-endian__.
The 2 highest bits represent the `block type`,
while the remaining 22 bits represent the block size.
while the remaining 22 bits represent the (compressed) block size.
There are 4 block types :
@ -281,24 +317,24 @@ There are 4 block types :
| ---------- | ---------- | --- | --- | ------- |
| Block Type | Compressed | Raw | RLE | EndMark |
- Compressed : this is a compressed block,
following Zstandard's block format specification.
The "block size" is the compressed size.
- Compressed : this is a Zstandard compressed block,
detailed in a later part of this specification.
"block size" is the compressed size.
Decompressed size is unknown,
but its maximum possible value is guaranteed (see later)
but its maximum possible value is guaranteed (see below)
- Raw : this is an uncompressed block.
"block size" is the number of bytes to read and copy.
- RLE : this is a single byte, repeated N times.
In which case, the size of the "compressed" block is always 1,
and the "block size" is the size to regenerate.
In which case, "block size" is the size to regenerate,
while the "compressed" block is just 1 byte (the byte to repeat).
- EndMark : this is not a block. Signal the end of the frame.
The rest of the field may be optionally filled by a checksum
(see frame checksum).
Block Size shall never be larger than Block Maximum Size.
Block Maximum Size is the smallest of :
- Max back-reference distance
- 128 KB
Block sizes must respect a few rules :
- In compressed mode, compressed size if always strictly `< contentSize`.
- Block decompressed size is necessarily <= maximum back-reference distance .
- Block decompressed size is necessarily <= 128 KB
__Data__
@ -306,8 +342,8 @@ __Data__
Where the actual data to decode stands.
It might be compressed or not, depending on previous field indications.
A data block is not necessarily "full" :
an arbitrary “flush” may happen anytime. Any block can be “partially filled”.
Therefore, data can have any size, up to Block Maximum Size.
since an arbitrary “flush” may happen anytime,
block content can be any size, up to Block Maximum Size.
Block Maximum Size is the smallest of :
- Max back-reference distance
- 128 KB
@ -329,25 +365,122 @@ over user-defined data and continue decoding.
Skippable frames defined in this specification are compatible with LZ4 ones.
__Magic Number__
__Magic Number__ :
4 Bytes, Little endian format.
Value : 0x184D2A5X, which means any value from 0x184D2A50 to 0x184D2A5F.
All 16 values are valid to identify a skippable frame.
__Frame Size__
__Frame Size__ :
This is the size, in bytes, of the following User Data
(without including the magic number nor the size field itself).
4 Bytes, Little endian format, unsigned 32-bits.
This means User Data cant be bigger than (2^32-1) Bytes.
__User Data__
__User Data__ :
User Data can be anything. Data will just be skipped by the decoder.
Compressed block format
-----------------------
This specification details the content of a _compressed block_.
A compressed block has a size, which must be known in order to decode it.
It also has a guaranteed maximum regenerated size,
in order to properly allocate destination buffer.
See "Frame format" for more details.
A compressed block consists of 2 sections :
- Literals section
- Sequences section
### Compressed Literals
Literals are compressed using order-0 huffman compression.
During sequence phase, literals will be entangled with match copy operations.
All literals are regrouped in the first part of the block.
They can be decoded first, and then copied during sequence operations,
or they can be decoded on the flow, as needed by sequences.
| Header | (Tree Description) | Stream1 | (Stream2) | (Stream3) | (Stream4) |
| ------ | ------------------ | ------- | --------- | --------- | --------- |
Literals can be compressed, or uncompressed.
When compressed, an optional tree description can be present,
followed by 1 or 4 streams.
#### Block Literal Header
Header is in charge of describing precisely how literals are packed.
It's a byte-aligned variable-size bitfield, ranging from 1 to 5 bytes,
using big-endian convention.
| BlockType | sizes format | (compressed size) | regenerated size |
| --------- | ------------ | ----------------- | ---------------- |
| 2 bits | 1 - 2 bits | 0 - 18 bits | 5 - 20 bits |
__Block Type__ :
This is a 2-bits field, describing 4 different block types :
| Value | 0 | 1 | 2 | 3 |
| ---------- | ---------- | ------ | --- | ------- |
| Block Type | Compressed | Repeat | Raw | RLE |
- Compressed : This is a standard huffman-compressed block,
starting with a huffman tree description.
See details below.
- Repeat Stats : This is a huffman-compressed block,
using huffman tree from previous huffman-compressed block.
Huffman tree description will be skipped.
Compressed stream is equivalent to "compressed" block type.
- Raw : Literals are stored uncompressed.
- RLE : Literals consist of a single byte value repeated N times.
__Sizes format__ :
Sizes format are divided into 2 families :
- For compressed block, it requires to decode both the compressed size
and the decompressed size. It will also decode the number of streams.
- For Raw or RLE blocks, it's enough to decode the size to regenerate.
For values spanning several bytes, convention is Big-endian.
__Sizes format for Raw or RLE block__ :
- Value : 0x : Regenerated size uses 5 bits (0-31).
Total literal header size is 1 byte.
`size = h[0] & 31;`
- Value : 10 : Regenerated size uses 12 bits (0-4095).
Total literal header size is 2 bytes.
`size = ((h[0] & 15) << 8) + h[1];`
- Value : 11 : Regenerated size uses 20 bits (0-1048575).
Total literal header size is 2 bytes.
`size = ((h[0] & 15) << 16) + (h[1]<<8) + h[2];`
Note : it's allowed to represent a short value (ex : `13`)
using a long format, accepting the reduced compacity.
__Sizes format for Compressed Block__ :
Note : also applicable to "repeat-stats" blocks.
- Value : 00 : 4 streams
Compressed and regenerated sizes use 10 bits (0-1023)
Total literal header size is 3 bytes
- Value : 01 : _Single stream_
Compressed and regenerated sizes use 10 bits (0-1023)
Total literal header size is 3 bytes
- Value : 10 : 4 streams
Compressed and regenerated sizes use 14 bits (0-16383)
Total literal header size is 4 bytes
- Value : 10 : 4 streams
Compressed and regenerated sizes use 18 bits (0-262143)
Total literal header size is 5 bytes
Version changes
---------------
0.1 : initial release