torch.sparse.sum() (#12430)

torch.sparse.sum() (#12430)

Summary:
- to fix #12241
- add `_sparse_sum()` to ATen, and expose as `torch.sparse.sum()`, not support `SparseTensor.sum()` currently
- this PR depends on #11253, and will need to be updated upon it lands
- [x] implement forward
- [x] implement backward
- performance [benchmark script](https://gist.github.com/weiyangfb/f4c55c88b6092ef8f7e348f6b9ad8946#file-sparse_sum_benchmark-py):
  - sum all dims is fastest for sparse tensor
  - when input is sparse enough nnz = 0.1%, sum of sparse tensor is faster than dense in CPU, but not necessary in CUDA
  - CUDA backward is comparable (<2x) between `sum several dims` vs `sum all dims` in sparse
  - CPU backward uses binary search is still slow in sparse, takes `5x` time in `sum [0, 2, 3] dims` vs `sum all dims`
    - optimize CUDA backward for now
      - using thrust for sort and binary search, but runtime not improved
  - both of CPU and CUDA forward are slow in sparse (`sum several dims` vs `sum all dims`), at most `20x` slower in CPU, and `10x` in CUDA
    - improve CPU and CUDA forward kernels

(nnz, sizes, sum_dims, keepdim, sum all or dims, bk=backward) | CPU (sparse vs dense) | CUDA(sparse vs dense)
-- | -- | --
(1000,   [1000, 1000, 2, 2], [0, 1], False, sumAll) | 8.77 µs vs 72.9 µs | 42.5 µs vs 108 µs
(1000,   [1000, 1000, 2, 2], [0, 1], False, sumD) | 112 µs vs 4.47 ms | 484 µs vs 407 µs
(1000,   [1000, 1000, 2, 2], [0, 1], False, sumAll, bk) | 141 µs vs 148 µs | 647 µs vs 231 µs
(1000,   [1000, 1000, 2, 2], [0, 1], False, sumD, bk) | 235 µs vs 1.23 ms | 781 µs vs 213 µs
(1000,   [1000, 1000, 2, 2], [2, 3], False, sumD) | 48.5 µs vs 360 µs | 160 µs vs 2.03 ms
(1000,   [1000, 1000, 2, 2], [2, 3], False, sumD, bk) | 258 µs vs 1.22 ms | 798 µs vs 224 µs
(1000,   [1000, 1000, 2, 2], [0, 2, 3], False, sumD) | 204 µs vs 882 µs | 443 µs vs 133 µs
(1000,   [1000, 1000, 2, 2], [0, 2, 3], False, sumD, bk) | 709 µs vs 1.15 ms | 893 µs vs 202 µs
(10000,   [1000, 1000, 2, 2], [0, 1], False, sumAll) | 39.8 µs vs 81 µs | 42.4 µs vs 113 µs
(10000,   [1000, 1000, 2, 2], [0, 1], False, sumD) | 747 µs vs 4.7 ms | 2.4 ms vs 414 µs
(10000,   [1000, 1000, 2, 2], [0, 1], False, sumAll, bk) | 1.04 ms vs 126 µs | 5.03 ms vs 231 µs
(10000,   [1000, 1000, 2, 2], [0, 1], False, sumD, bk) | 1.12 ms vs 1.24 ms | 5.99 ms vs 213 µs
(10000,   [1000, 1000, 2, 2], [2, 3], False, sumD) | 133 µs vs 366 µs | 463 µs vs 2.03 ms
(10000,   [1000, 1000, 2, 2], [2, 3], False, sumD, bk) | 1.56 ms vs 1.22 ms | 6.11 ms vs 229 µs
(10000,   [1000, 1000, 2, 2], [0, 2, 3], False, sumD) | 1.53 ms vs 799 µs | 824 µs vs 134 µs
(10000,   [1000, 1000, 2, 2], [0, 2, 3], False, sumD, bk) | 5.15 ms vs 1.09 ms | 7.02 ms vs 205 µs

- after improving CPU and CUDA forward kernels
  - in `(1000,   [1000, 1000, 2, 2], [0, 2, 3], False, sumD)` forward, CPU takes ~~`171 µs`~~, in which `130 µs` is spent on `coalesce()`, for CUDA, total time is ~~`331 µs`~~, in which `141 µs` is spent on `coalesce()`, we need to reduce time at other places outside `coalesce()`.
  - after a few simple tweaks, now in the forward, it is at most `10x` slower in CPU, and `7x` in CUDA. And time takes in `sum dense dims only [2, 3]` is `~2x` of `sum all dims`. Speed of `sum all sparse dims [0, 1]` is on bar with `sum all dims`

(nnz,   sizes, sum_dims, keepdim, sum all or dims, bk=backward) | CPU (sparse vs dense) | CUDA(sparse vs dense)
-- | -- | --
(1000,   [1000, 1000, 2, 2], [0, 1], False, sumAll) | 7 µs vs 69.5 µs | 31.5 µs vs 61.6 µs
(1000,   [1000, 1000, 2, 2], [0, 1], False, sumD) | 11.3 µs vs 4.72 ms | 35.2 µs vs 285 µs
(1000,   [1000, 1000, 2, 2], [0, 1], False, sumAll, bk) | 197 µs vs 124 µs | 857 µs vs 134 µs
(1000,   [1000, 1000, 2, 2], [0, 1], False, sumD, bk) | 124 µs vs 833 µs | 796 µs vs 106 µs
(1000,   [1000, 1000, 2, 2], [2, 3], False, sumD) | 20.5 µs vs 213 µs | 39.4 µs vs 1.24 ms
(1000,   [1000, 1000, 2, 2], [2, 3], False, sumD, bk) | 131 µs vs 830 µs | 881 µs vs 132 µs
(1000,   [1000, 1000, 2, 2], [0, 2, 3], False, sumD) | 95.8 µs vs 409 µs | 246 µs vs 87.2 µs
(1000,   [1000, 1000, 2, 2], [0, 2, 3], False, sumD, bk) | 624 µs vs 820 µs | 953 µs vs 124 µs
(10000,   [1000, 1000, 2, 2], [0, 1], False, sumAll) | 45.3 µs vs 72.9 µs | 33.9 µs vs 57.2 µs
(10000,   [1000, 1000, 2, 2], [0, 1], False, sumD) | 81.4 µs vs 4.49 ms | 39.7 µs vs 280 µs
(10000,   [1000, 1000, 2, 2], [0, 1], False, sumAll, bk) | 984 µs vs 111 µs | 6.41 ms vs 121 µs
(10000,   [1000, 1000, 2, 2], [0, 1], False, sumD, bk) | 1.45 ms vs 828 µs | 6.77 ms vs 113 µs
(10000,   [1000, 1000, 2, 2], [2, 3], False, sumD) | 74.9 µs vs 209 µs | 37.7 µs vs 1.23 ms
(10000,   [1000, 1000, 2, 2], [2, 3], False, sumD, bk) | 1.48 ms vs 845 µs | 6.96 ms vs 132 µs
(10000,   [1000, 1000, 2, 2], [0, 2, 3], False, sumD) | 1.14 ms vs 411 µs | 252 µs vs 87.8 µs
(10000,   [1000, 1000, 2, 2], [0, 2, 3], False, sumD, bk) | 4.53 ms vs 851 µs | 7.12 ms vs 128 µs

- time takes in CUDA backward of sparse is super long with large variance (in case of nnz=10000, it normally takes 6-7ms). To improve backward of sparse ops, we will need to debug at places other than CUDA kernels. here is a benchmark of `torch.copy_()`:
```
>>> d = [1000, 1000, 2, 2]
>>> nnz = 10000
>>> I = torch.cat([torch.randint(0, d[0], size=(nnz,)),
               torch.randint(0, d[1], size=(nnz,))], 0).reshape(2, nnz)
>>> V = torch.randn(nnz, d[2], d[3])
>>> size = torch.Size(d)
>>> S = torch.sparse_coo_tensor(I, V, size).coalesce().cuda()
>>> S2 = torch.sparse_coo_tensor(I, V, size).coalesce().cuda().requires_grad_()
>>> data = S2.clone()
>>> S.copy_(S2)
>>> y = S * 2
>>> torch.cuda.synchronize()
>>> %timeit y.backward(data, retain_graph=True); torch.cuda.synchronize()
7.07 ms ± 3.06 ms per loop (mean ± std. dev. of 7 runs, 1000 loops each)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12430

Differential Revision: D12878313

Pulled By: weiyangfb

fbshipit-source-id: e16dc7681ba41fdabf4838cf05e491ca9108c6fe