1 | /******************************************************************************* |
2 | * Copyright 2022 Intel Corporation |
3 | * |
4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
5 | * you may not use this file except in compliance with the License. |
6 | * You may obtain a copy of the License at |
7 | * |
8 | * http://www.apache.org/licenses/LICENSE-2.0 |
9 | * |
10 | * Unless required by applicable law or agreed to in writing, software |
11 | * distributed under the License is distributed on an "AS IS" BASIS, |
12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
13 | * See the License for the specific language governing permissions and |
14 | * limitations under the License. |
15 | *******************************************************************************/ |
16 | |
17 | #ifndef GPU_JIT_IR_REORDER_HPP |
18 | #define GPU_JIT_IR_REORDER_HPP |
19 | |
20 | #include "gpu/jit/ir/ir.hpp" |
21 | #include "gpu/jit/ir/tensor.hpp" |
22 | |
23 | namespace dnnl { |
24 | namespace impl { |
25 | namespace gpu { |
26 | namespace jit { |
27 | |
28 | // Implements reorder between GRF buffers in given layouts. Conversion between |
29 | // data types is supported. |
30 | class reorder_t : public func_impl_t { |
31 | public: |
32 | IR_DECL_DERIVED_TYPE_ID(reorder_t, func_impl_t) |
33 | |
34 | static func_t make(const layout_t &src_layout, const layout_t &dst_layout) { |
35 | return func_t(new reorder_t(src_layout, dst_layout)); |
36 | } |
37 | |
38 | bool is_equal(const object_impl_t &obj) const override { |
39 | if (!obj.is<self_type>()) return false; |
40 | auto &other = obj.as<self_type>(); |
41 | |
42 | return (src_layout == other.src_layout) |
43 | && (dst_layout == other.dst_layout); |
44 | } |
45 | |
46 | size_t get_hash() const override { |
47 | return ir_utils::get_hash(src_layout, dst_layout); |
48 | } |
49 | |
50 | std::string str() const override { |
51 | std::ostringstream oss; |
52 | oss << "reorder[" << src_layout << ", " << dst_layout << "]" ; |
53 | return oss.str(); |
54 | } |
55 | |
56 | IR_DEFINE_ARG_GET(dst_buf, 0) |
57 | IR_DEFINE_ARG_GET(src_buf, 1) |
58 | |
59 | layout_t src_layout; |
60 | layout_t dst_layout; |
61 | |
62 | private: |
63 | reorder_t(const layout_t &src_layout, const layout_t &dst_layout) |
64 | : func_impl_t(_type_info()) |
65 | , src_layout(src_layout) |
66 | , dst_layout(dst_layout) {} |
67 | }; |
68 | |
69 | inline stmt_t create_reorder_stmt(const layout_t &src, const layout_t &dst, |
70 | const expr_t &src_buf, const expr_t &dst_buf) { |
71 | ir_assert(src.ndims() == dst.ndims()) << "Layouts are incompatible." ; |
72 | ir_assert(src.elems() == dst.elems()) << "Layouts are incompatible." ; |
73 | auto func = reorder_t::make(src, dst); |
74 | return func.call({dst_buf, src_buf}); |
75 | } |
76 | |
77 | } // namespace jit |
78 | } // namespace gpu |
79 | } // namespace impl |
80 | } // namespace dnnl |
81 | |
82 | #endif |
83 | |