xref: /aosp_15_r20/external/executorch/kernels/optimized/cpu/op_linear.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 <executorch/kernels/optimized/blas/CPUBlas.h>
10 #include <executorch/kernels/portable/cpu/util/matmul_ops_util.h>
11 #include <executorch/runtime/kernel/kernel_includes.h>
12 
13 #include <array>
14 
15 namespace torch {
16 namespace executor {
17 namespace native {
18 
19 using Tensor = exec_aten::Tensor;
20 
opt_linear_out(RuntimeContext & ctx,const Tensor & in,const Tensor & mat2,const optional<Tensor> & bias,Tensor & out)21 Tensor& opt_linear_out(
22     RuntimeContext& ctx,
23     const Tensor& in,
24     const Tensor& mat2,
25     const optional<Tensor>& bias,
26     Tensor& out) {
27   ET_KERNEL_CHECK_MSG(
28       ctx,
29       !bias.has_value(),
30       InvalidArgument,
31       out,
32       "bias not supported yet in linear");
33   ET_KERNEL_CHECK(ctx, check_linear_args(in, mat2, out), InvalidArgument, out);
34 
35   size_t output_ndim = 0;
36   std::array<exec_aten::SizesType, kTensorDimensionLimit> output_sizes;
37   get_linear_out_target_size(in, mat2, output_sizes.data(), &output_ndim);
38   ET_KERNEL_CHECK(
39       ctx,
40       resize_tensor(out, {output_sizes.data(), output_ndim}) == Error::Ok,
41       InvalidArgument,
42       out);
43 
44   // gemm on some platforms doesn't tolerate empty input.
45   if (out.numel() == 0) {
46     return out;
47   }
48 
49   int flattened_input_dim = 1;
50   for (int ii = 0; ii < in.dim() - 1; ++ii) {
51     flattened_input_dim *= in.sizes()[ii];
52   }
53   ET_SWITCH_REAL_TYPES_AND2(
54       Half, BFloat16, in.scalar_type(), ctx, "mm.out", CTYPE, [&]() {
55         size_t n = flattened_input_dim;
56         size_t k = in.sizes()[in.dim() - 1];
57         size_t m = mat2.size(0);
58 
59         executorch::cpublas::gemm(
60             executorch::cpublas::TransposeType::Transpose,
61             executorch::cpublas::TransposeType::NoTranspose,
62             m,
63             n,
64             k,
65             static_cast<CTYPE>(1),
66             mat2.const_data_ptr<CTYPE>(),
67             k,
68             in.const_data_ptr<CTYPE>(),
69             k,
70             static_cast<CTYPE>(0),
71             out.mutable_data_ptr<CTYPE>(),
72             m);
73       });
74 
75   return out;
76 }
77 
78 } // namespace native
79 } // namespace executor
80 } // namespace torch
81