xref: /aosp_15_r20/external/executorch/kernels/portable/cpu/op_embedding.cpp (revision 523fa7a60841cd1ecfb9cc4201f1ca8b03ed023a)
1 /*
2  * Copyright (c) Meta Platforms, Inc. and affiliates.
3  * All rights reserved.
4  *
5  * This source code is licensed under the BSD-style license found in the
6  * LICENSE file in the root directory of this source tree.
7  */
8 
9 #include <cstring>
10 
11 #include <executorch/kernels/portable/cpu/util/kernel_ops_util.h>
12 #include <executorch/runtime/kernel/kernel_includes.h>
13 
14 // A simple lookup table that looks up embeddings in a fixed dictionary and
15 // size.
16 
17 // This module is often used to retrieve word embeddings using indices. The
18 // input to the module is a list of indices, and the embedding matrix, and the
19 // output is the corresponding word embeddings.
20 namespace torch {
21 namespace executor {
22 namespace native {
23 
24 using Tensor = exec_aten::Tensor;
25 using ScalarType = exec_aten::ScalarType;
26 
27 namespace {
28 
29 template <typename CTYPE>
embedding_kernel(KernelRuntimeContext & ctx,const Tensor & weight,const Tensor & indices,Tensor & out)30 void embedding_kernel(
31     KernelRuntimeContext& ctx,
32     const Tensor& weight,
33     const Tensor& indices,
34     Tensor& out) {
35   int64_t nbytes_per_entry = weight.size(1) * weight.element_size();
36   const char* w_data = weight.const_data_ptr<char>();
37   char* out_data = out.mutable_data_ptr<char>();
38   const CTYPE* indices_ptr = indices.const_data_ptr<CTYPE>();
39   ssize_t weight_height = weight.size(0);
40   const auto indices_numel = indices.numel();
41   for (int i = 0; i < indices_numel; i++) {
42     // Ensure index is larger than 0 and smaller than weight.size(0)
43     ET_KERNEL_CHECK_MSG(
44         ctx,
45         indices_ptr[i] < weight_height,
46         InvalidArgument,
47         ,
48         "indices_ptr[%d] %ld >= weight.size(0) %zd",
49         i,
50         static_cast<long>(indices_ptr[i]),
51         weight_height);
52     ET_KERNEL_CHECK_MSG(
53         ctx,
54         indices_ptr[i] >= 0,
55         InvalidArgument,
56         ,
57         "indices_ptr[%d] %ld < 0",
58         i,
59         static_cast<long>(indices_ptr[i]));
60     if (w_data != nullptr) {
61       memcpy(
62           out_data,
63           w_data + nbytes_per_entry * indices_ptr[i],
64           nbytes_per_entry);
65     }
66     out_data += nbytes_per_entry;
67   }
68 }
69 } // namespace
70 
71 // embedding.out(Tensor weight, Tensor indices, int padding_idx=-1, bool
72 // scale_grad_by_freq=False, bool sparse=False, *, Tensor(a!) out) -> Tensor(a!)
embedding_out(KernelRuntimeContext & ctx,const Tensor & weight,const Tensor & indices,int64_t padding_idx,bool scale_grad_by_freq,bool sparse,Tensor & out)73 Tensor& embedding_out(
74     KernelRuntimeContext& ctx,
75     const Tensor& weight,
76     const Tensor& indices,
77     int64_t padding_idx,
78     bool scale_grad_by_freq,
79     bool sparse,
80     Tensor& out) {
81   (void)ctx;
82   (void)padding_idx;
83   (void)scale_grad_by_freq;
84   (void)sparse;
85 
86   ET_KERNEL_CHECK(
87       ctx, check_embedding_args(weight, indices, out), InvalidArgument, out);
88 
89   ET_KERNEL_CHECK(
90       ctx,
91       resize_embedding_output(weight, indices, out) == Error::Ok,
92       InvalidArgument,
93       out);
94 
95   ET_KERNEL_CHECK_MSG(
96       ctx,
97       out.size(out.dim() - 1) == weight.size(1),
98       InvalidArgument,
99       out,
100       "out.size(%zd) %zd != weight.size(1) %zd",
101       out.dim() - 1,
102       out.size(1),
103       weight.size(1));
104 
105   ET_KERNEL_CHECK(
106       ctx,
107       tensors_have_same_dim_order(weight, indices, out),
108       InvalidArgument,
109       out);
110 
111   ET_KERNEL_CHECK(
112       ctx, tensor_is_default_dim_order(weight), InvalidArgument, out);
113 
114   ScalarType ix_type = indices.scalar_type();
115   ET_CHECK_MSG(
116       ix_type == ScalarType::Long || ix_type == ScalarType::Int,
117       "Expected indices tensor to have Long or Int scalar types");
118 
119   ET_SWITCH_TWO_TYPES(
120       Long, Int, ix_type, ctx, "op_embedding.out", CTYPE, [&]() {
121         embedding_kernel<CTYPE>(ctx, weight, indices, out);
122       });
123 
124   return out;
125 }
126 
127 } // namespace native
128 } // namespace executor
129 } // namespace torch
130