GH-49266: [C++][Parquet] Optimize delta bit-packed decoding when bit-width = 0

[pitrou]

Rationale for this change

DELTA_BINARY_PACKED decoding has limited performance due to a back-to-back dependency between the computations of value N and value N+1.

However, we can do better if we know that all deltas are 0 in a miniblock. This happens when a miniblock's delta bit width.

What changes are included in this PR?

Avoid reading and accumulating deltas when we the delta bit width is 0. Instead, use a condensed formula that allows to compute a value without waiting for the previous one.

Benchmark results on constant ranges of integers (on my local machine, AMD Zen 2 CPU):

                                                                 benchmark        baseline        contender  change %                                                                                                                                                                                                               counters
                                 BM_DeltaBitPackingDecode_Int32_Fixed/4096   3.821 GiB/sec   12.164 GiB/sec   218.323                              {'family_index': 11, 'per_family_instance_index': 1, 'run_name': 'BM_DeltaBitPackingDecode_Int32_Fixed/4096', 'repetitions': 1, 'repetition_index': 0, 'threads': 1, 'iterations': 70160}
                                BM_DeltaBitPackingDecode_Int32_Fixed/65536   3.897 GiB/sec   12.378 GiB/sec   217.678                              {'family_index': 11, 'per_family_instance_index': 3, 'run_name': 'BM_DeltaBitPackingDecode_Int32_Fixed/65536', 'repetitions': 1, 'repetition_index': 0, 'threads': 1, 'iterations': 4487}
                                BM_DeltaBitPackingDecode_Int32_Fixed/32768   3.909 GiB/sec   12.325 GiB/sec   215.309                              {'family_index': 11, 'per_family_instance_index': 2, 'run_name': 'BM_DeltaBitPackingDecode_Int32_Fixed/32768', 'repetitions': 1, 'repetition_index': 0, 'threads': 1, 'iterations': 9004}
                                 BM_DeltaBitPackingDecode_Int32_Fixed/1024   3.542 GiB/sec   10.468 GiB/sec   195.538                             {'family_index': 11, 'per_family_instance_index': 0, 'run_name': 'BM_DeltaBitPackingDecode_Int32_Fixed/1024', 'repetitions': 1, 'repetition_index': 0, 'threads': 1, 'iterations': 259535}
                                BM_DeltaBitPackingDecode_Int64_Fixed/32768   9.761 GiB/sec   14.040 GiB/sec    43.847                             {'family_index': 12, 'per_family_instance_index': 2, 'run_name': 'BM_DeltaBitPackingDecode_Int64_Fixed/32768', 'repetitions': 1, 'repetition_index': 0, 'threads': 1, 'iterations': 11192}
                                BM_DeltaBitPackingDecode_Int64_Fixed/65536   9.814 GiB/sec   14.056 GiB/sec    43.222                              {'family_index': 12, 'per_family_instance_index': 3, 'run_name': 'BM_DeltaBitPackingDecode_Int64_Fixed/65536', 'repetitions': 1, 'repetition_index': 0, 'threads': 1, 'iterations': 5617}
                                 BM_DeltaBitPackingDecode_Int64_Fixed/4096   9.672 GiB/sec   13.543 GiB/sec    40.014                              {'family_index': 12, 'per_family_instance_index': 1, 'run_name': 'BM_DeltaBitPackingDecode_Int64_Fixed/4096', 'repetitions': 1, 'repetition_index': 0, 'threads': 1, 'iterations': 88891}
                                 BM_DeltaBitPackingDecode_Int64_Fixed/1024   8.850 GiB/sec   12.207 GiB/sec    37.923                             {'family_index': 12, 'per_family_instance_index': 0, 'run_name': 'BM_DeltaBitPackingDecode_Int64_Fixed/1024', 'repetitions': 1, 'repetition_index': 0, 'threads': 1, 'iterations': 323720}

Are these changes tested?

Yes, by an additional test meant to stress this specific situation.

Are there any user-facing changes?

No.

• GitHub Issue: [C++][Parquet] Optimize delta bit-packed decoding when bit-width = 0 #49266

[pitrou]

FTR @AntoinePrv :)