Ctrl+K

Creating and Using Indices

Creating and Using Indices#

Computing the Size of the Index Array#

The expected number of index values can be obtained from:

uint32_t num_index_values(uint32_t buf_size, uint32_t idx_size, uint32_t blk_size);

where the arguments are the size of the uncompressed data, the size of the mini-block, and the size of the block. It assumes all the blocks to have the same (uncompressed) size (as specified) or this value is 0. If the value is 0, all of the input is compressed into a single block (whose size is the same as buf_size).

One index value is stored per mini-block, two values are stored per block, and one extra index value. In the first example (multi-block usage), 3 blocks and 11 mini-blocks present, so the total number of indices is 3*2+11+1 = 18. In the second example (single-block usage), 1 block and 11 mini-blocks present, so the total number of indices is 1*2+11+1 = 14.

Creating Indices#

To generate indices, the application needs to allocate an array large enough to hold all of the index values. The application then defines the mini-block size using the qpl_job.mini_block_size field, points to the index table memory with the qpl_job.idx_array field, and specifies the size of the table with the qpl_job.idx_max_size field.

The index generation is done as part of the verification pass, so the application cannot specify the QPL_FLAG_OMIT_VERIFY flag.

Intel® Query Processing Library (Intel® QPL) implements two slightly different approaches:

  • with multiple blocks of fixed size,

  • with a single block.

To allow for CRC checking, the entire mini-block needs to be decompressed. If the desired data straddles two mini-blocks, then the application separately decompresses each of the mini-blocks, and then extracts the data from both.

To use this feature, specify the ignore_start_bits flag and the ignore_end_bits flag. Both flags should contain values from 0 through 7. They indicate the number of bits at the start and end of the first byte to be skipped/ignored. This allows the application to start and end decompression on a bit rather than a byte boundary.

In general, for each job one would get a START bit-offset and an END bit-offset from the index array. More particularly, the START bit is the first bit to be processed. The END bit is the last bit to be processed plus one. So the total number of bits to be processed is just (END - START).

To convert these to job structure parameters, one needs to calculate:

  • the offsets of the byte containing the START bit,

  • the number of bytes to process (i.e. the number of bytes starting with the START byte to include the byte containing the last bit to be processed (END - 1)),

  • the number of bits to ignore at the start of the first byte,

  • the number of bits to ignore at the end of the last byte.

One way to do this is:

job->next_in_ptr = comp_buffer + START/8;
job->ignore_start_bits = START & 7;
job->available_in = ((END+7)/8) - (START/8);
job->ignore_end_bits = 7 & (0 - END);

Note that ((END+7)/8) - (START/8) is not the same as ((END+7-START)/8).

Multi-Block Usage#

For multi-block case, there is a “block_size”, where all blocks but the last are exactly that size, and the last block is no larger than that size. The block_size is a multiple of the mini-block size.

During the compression, each block of input data is submitted as a single job. Each job is flagged with either QPL_FLAG_DYNAMIC_HUFFMAN or QPL_FLAG_START_NEW_BLOCK.

Using the indices requires up to two jobs. The first parses the block header, and the second decompresses the specified mini-block. To decompress multiple mini-blocks within a given block, the header only needs to be parsed once. For example, the application can maintain a “block cache”. Then to decode a given mini-block (if its header is in the cache), a single job decompresses the mini-block. If its header is not in the cache, then some entry in the cache is replaced with the decompressed data from the new header, and the second job decompresses the mini-block.

The application needs a derived value: mb_per_b, which is the number of mini-blocks per block. This is the ratio of the two sizes.

To access the mini-block containing byte “B”, compute which mini-block needs to be decompressed first. This is “M”:

M = B / mini-block_size

Then compute which block and mini-block within that block contains “B”:

blk_num = M / mb_per_b
mblk_num = M % mb_per_b

Note that if the mb_per_b is a power of two, then this can be done as a SHIFT and an AND rather than using actual divides.

Decompressing the Block Header#

The index of the start of the block header is (blk_num \* (mb_per_b + 2)). The index of the end is one greater than this.

The job to parse the header is a decompress job with QPL_FLAG_FIRST and QPL_FLAG_RND_ACCESS. The START bit offset and the END bit offset are the low 32-bits of the two entries identified earlier. Those offsets are converted to job parameters as previously described.

Decompressing the Mini-block#

The index of the start of the block header is (1 + blk_num \* (mb_per_b + 2) + mb_num), and the index of the end is one greater. That is, the start of the mini-block is found by adding the mini-block number within the block to the end of the block header.

For example, assuming the user wants to decompress mini-block 6. In this case, there are 4 mb_per_b. So the blk_num is 1, and the mb_num is 2. The header can be found between indices 6 and 7, and the mini-block between indices 9 and 10.

The job uses the flag QPL_FLAG_RND_ACCESS.

The initial CRC (i.e. the value written to job crc before the job is executed) is the CRC value (the high 32-bits of the index array) from the start of the mini-block. In the previous example, that is from index 1. After the job completes, the returned CRC value (in job crc) should match the CRC value from the end of the mini-block, in this example 10. If it does not match, then the mini-block is not decompressed properly.

Note

The state of decompression with mini-blocks is not saved from job to job. Namely the qpl_job.total_out will not store the cumulative number of bytes written, but only that job’s particular mini-block’s number of bytes written.

Single Block Usage#

This scenario is similar to the Multi-block case, except that there is only one block, and the mb_per_b is not used.

In this case, the header is always between indices 0 and 1, and the start of the mini-block is index M+1. So using the example earlier, if the user wants to decompress mini-block 6, the start of that mini-block would be index 7.

Contents