[Intel MKL] Parallelize UnsortedSegmentSum on CPU deivce

[Zantares]

UnsortedSegmentSum is a single thread op on CPU, here use the Eigen device to parallelize it. It's updated from an old PR #26427 .

I resubmit this PR because the original one got an internal crash and was reverted. Because I can't get the failure case, I just went through the old code and found a potential rounding issue here: https://github.com/Intel-tensorflow/tensorflow/blob/52a6cfddef9b51b608b4a554b77a10e1522d56ec/tensorflow/core/kernels/segment_reduction_ops.cc#L419-L425

This PR uses a simple and reasonable way to balance workload like what TF usually did:

    // Each output row may contain different size of reduction from inputs,
    // balance the workload to a task group. Each task is output row index.
    const int64 kMaxTaskBlock =
        std::min(num_reductions, (int64)cpu_device.numThreads());
    const int64 kAverTaskSize = (N + kMaxTaskBlock - 1) / kMaxTaskBlock

Other parts are the same as the original one.

Performance improvement:

• Origin

BM_UnsortedSegmentSum_4096_128_1_ 300315 2279 6983.2MB/s
BM_UnsortedSegmentSum_4096_128_128_ 306408 2268 6844.3MB/s

• Optimized

BM_UnsortedSegmentSum_4096_128_1_ 302769 2268 6926.6MB/s
BM_UnsortedSegmentSum_4096_128_128_ 116179 6112 18051.1MB/s

This patch tries to parallel UnsortSegmentReduction with num_segments, so the performance has no change if num_segments is 1, otherwise it has ~3X improvement in the benchmark.

[Zantares]

Here's a simple introduce for the optimization:

1. row_counter stores how many rows will be reduce to each output row, num_reductions is the real num_segments which excluded 0 count rows, it also decides the degree of parallelism.
2. Loop row_counter to pick each task to block_range. block_range records task index for shard function.
   

[Zantares]

@ezhulenev can you help to review this again?

[ezhulenev]

It was rolled back because of internal failures, looks like it produces nans. Any ideas? Will try to get better error messages tomorrow.

Error message:

tensorflow.python.framework.errors_impl.InvalidArgumentError: assertion failed: [Condition x >= y did not hold element-wise:] [x (<...internal_node_name...>/Sum_2:0) = ] [-nan -nan -nan...]

[Zantares]

Thank for your reminding, I found it may cause accuracy issue(including NaN) when task_index is 0 but exited from here: https://github.com/tensorflow/tensorflow/pull/49152/files#diff-d29812caca45335066ddacf8fb06afc416f1f557bc446068ebaa163495e0b03bR388-R390. The older PR started from index 1 but the new PR started from index 0 to simplify the code, and seems it caused new issue.
I will make a related case to test locally first, and update it later.