1 #define TORCH_ASSERT_NO_OPERATORS
2 #include <ATen/native/sparse/SparseFactories.h>
3
4 #include <ATen/Dispatch.h>
5 #include <ATen/TensorIterator.h>
6 #include <ATen/core/TensorBase.h>
7 #include <ATen/native/cpu/Loops.h>
8 #include <c10/core/ScalarType.h>
9 #include <c10/util/Exception.h>
10
11 namespace at::native {
12
13 namespace {
_spdiags_kernel_cpu(TensorIterator & iter,const TensorBase & diagonals,TensorBase & values,TensorBase & indices)14 void _spdiags_kernel_cpu(
15 TensorIterator& iter,
16 const TensorBase& diagonals,
17 TensorBase& values,
18 TensorBase& indices) {
19 auto* row_index_write_ptr = indices.data_ptr<int64_t>();
20 auto* col_index_write_ptr = row_index_write_ptr ? row_index_write_ptr + indices.stride(0) : nullptr;
21 const int64_t diagonals_index_stride = diagonals.stride(0);
22 const int64_t diagonals_read_stride = diagonals.stride(1);
23 AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND4(
24 at::ScalarType::BFloat16,
25 at::ScalarType::Half,
26 at::ScalarType::Bool,
27 at::ScalarType::ComplexHalf,
28 diagonals.scalar_type(),
29 "spdiags_cpu",
30 [&] {
31 auto* const values_write_ptr = values.data_ptr<scalar_t>();
32 const auto* const diagonals_ptr = diagonals.const_data_ptr<scalar_t>();
33
34 cpu_kernel(
35 iter,
36 [&](int64_t diag_index,
37 int64_t diag_offset,
38 int64_t out_offset,
39 int64_t n_out) -> int64_t {
40 if (n_out > 0) {
41 auto* rows_start = row_index_write_ptr + out_offset;
42 auto* cols_start = col_index_write_ptr + out_offset;
43 auto* vals_start = values_write_ptr + out_offset;
44 const int64_t first_col = std::max<int64_t>(diag_offset, 0);
45 const int64_t first_row = first_col - diag_offset;
46 auto* data_read = (diagonals_ptr +
47 diagonals_index_stride * diag_index +
48 first_col * diagonals_read_stride);
49 for (int64_t i = 0; i < n_out; ++i) {
50 rows_start[i] = first_row + i;
51 cols_start[i] = first_col + i;
52 vals_start[i] = data_read[i * diagonals_read_stride];
53 }
54 }
55 // dummy return
56 return 0;
57 });
58 });
59 }
60
61 } // namespace
62
63 REGISTER_DISPATCH(spdiags_kernel_stub, &_spdiags_kernel_cpu);
64
65 } // namespace at::native
66