1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
|
#include "scatter_cuda.h"
#include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/detail/IndexUtils.cuh>
#include <ATen/cuda/detail/TensorInfo.cuh>
#include "reducer.cuh"
#include "utils.cuh"
#define THREADS 256
#define BLOCKS(N) (N + THREADS - 1) / THREADS
template <typename scalar_t, ReductionType REDUCE>
__global__ void
scatter_kernel(const scalar_t *src_data,
const at::cuda::detail::TensorInfo<int64_t, int> index_info,
scalar_t *out_data, int E, int K, int N, int numel) {
int thread_idx = blockIdx.x * blockDim.x + threadIdx.x;
int b = thread_idx / (E * K);
int k = thread_idx % K;
if (thread_idx < numel) {
int offset = at::cuda::detail::IndexToOffset<int64_t, int, -1>::get(
thread_idx, index_info);
int64_t idx = index_info.data[offset];
Reducer<scalar_t, REDUCE>::atomic_write(out_data + b * N * K + idx * K + k,
src_data[thread_idx]);
}
}
template <typename scalar_t>
__global__ void
scatter_arg_kernel(const scalar_t *src_data,
const at::cuda::detail::TensorInfo<int64_t, int> index_info,
const scalar_t *out_data, int64_t *arg_out_data, int E,
int K, int N, int numel) {
int thread_idx = blockIdx.x * blockDim.x + threadIdx.x;
int b = thread_idx / (E * K);
int e = (thread_idx / K) % E;
int k = thread_idx % K;
if (thread_idx < numel) {
int offset = at::cuda::detail::IndexToOffset<int64_t, int, -1>::get(
thread_idx, index_info);
int64_t idx = index_info.data[offset];
if (src_data[thread_idx] == out_data[b * N * K + idx * K + k]) {
arg_out_data[b * N * K + idx * K + k] = e;
}
}
}
std::tuple<torch::Tensor, torch::optional<torch::Tensor>>
scatter_cuda(torch::Tensor src, torch::Tensor index, int64_t dim,
torch::optional<torch::Tensor> optional_out,
torch::optional<int64_t> dim_size, std::string reduce) {
CHECK_CUDA(src);
CHECK_CUDA(index);
if (optional_out.has_value())
CHECK_CUDA(optional_out.value());
cudaSetDevice(src.get_device());
CHECK_INPUT(src.dim() == index.dim());
for (auto i = 0; i < index.dim() - 1; i++)
CHECK_INPUT(src.size(i) >= index.size(i));
src = src.contiguous();
torch::Tensor out;
if (optional_out.has_value()) {
out = optional_out.value().contiguous();
for (auto i = 0; i < out.dim(); i++)
if (i != dim)
CHECK_INPUT(src.size(i) == out.size(i));
} else {
auto sizes = src.sizes().vec();
if (dim_size.has_value())
sizes[dim] = dim_size.value();
else if (index.numel() == 0)
sizes[dim] = 0;
else {
sizes[dim] = 1 + index.max().cpu().data_ptr<int64_t>()[0];
}
out = torch::empty(sizes, src.options());
}
torch::optional<torch::Tensor> arg_out = torch::nullopt;
int64_t *arg_out_data = nullptr;
if (reduce2REDUCE.at(reduce) == MIN || reduce2REDUCE.at(reduce) == MAX) {
arg_out = torch::full_like(out, src.size(dim), index.options());
arg_out_data = arg_out.value().data_ptr<int64_t>();
}
if (src.numel() == 0) {
if (!optional_out.has_value())
out.fill_(0);
return std::make_tuple(out, arg_out);
}
auto B = 1;
for (auto i = 0; i < dim; i++)
B *= src.size(i);
auto E = src.size(dim);
auto K = src.numel() / (B * E);
auto N = out.size(dim);
auto index_info = at::cuda::detail::getTensorInfo<int64_t, int>(index);
auto stream = at::cuda::getCurrentCUDAStream();
AT_DISPATCH_ALL_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16, src.scalar_type(), "_", [&] {
auto src_data = src.data_ptr<scalar_t>();
auto out_data = out.data_ptr<scalar_t>();
AT_DISPATCH_REDUCTION_TYPES(reduce, [&] {
if (!optional_out.has_value())
out.fill_(Reducer<scalar_t, REDUCE>::init());
scatter_kernel<scalar_t, REDUCE>
<<<BLOCKS(src.numel()), THREADS, 0, stream>>>(
src_data, index_info, out_data, E, K, N, src.numel());
if (!optional_out.has_value() && (REDUCE == MIN || REDUCE == MAX))
out.masked_fill_(out == Reducer<scalar_t, REDUCE>::init(), (scalar_t)0);
if (REDUCE == MIN || REDUCE == MAX)
scatter_arg_kernel<scalar_t>
<<<BLOCKS(src.numel()), THREADS, 0, stream>>>(
src_data, index_info, out_data, arg_out_data, E, K, N,
src.numel());
});
});
return std::make_tuple(out, arg_out);
}
|
