1 | /******************************************************************************* |
2 | * Copyright 2020-2022 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_OCL_SHUFFLE_BY_REORDER_HPP |
18 | #define GPU_OCL_SHUFFLE_BY_REORDER_HPP |
19 | |
20 | #include "common/c_types_map.hpp" |
21 | #include "common/primitive.hpp" |
22 | #include "common/reorder.hpp" |
23 | #include "common/reorder_pd.hpp" |
24 | #include "gpu/compute/compute.hpp" |
25 | #include "gpu/gpu_primitive.hpp" |
26 | #include "gpu/gpu_resource.hpp" |
27 | #include "gpu/gpu_shuffle_pd.hpp" |
28 | #include "gpu/ocl/ocl_engine.hpp" |
29 | #include "gpu/ocl/ocl_stream.hpp" |
30 | #include "gpu/primitive_conf.hpp" |
31 | namespace dnnl { |
32 | namespace impl { |
33 | namespace gpu { |
34 | namespace ocl { |
35 | |
36 | // Implements shuffle using reorder kernel. |
37 | // Pretends that instead of the one dimension to be shuffled there are two |
38 | // smaller dimensions, then reorders the tensor to swap those two. |
39 | // Reorder kernel is used more often so is expected to be better optimized. |
40 | struct shuffle_by_reorder_t : public gpu_primitive_t { |
41 | using gpu_primitive_t::gpu_primitive_t; |
42 | struct pd_t : public gpu_shuffle_pd_t { |
43 | using gpu_shuffle_pd_t::gpu_shuffle_pd_t; |
44 | |
45 | DECLARE_COMMON_PD_T("ocl:reorder:any" , shuffle_by_reorder_t); |
46 | |
47 | status_t init(engine_t *engine) { |
48 | const auto &md_src = is_fwd() ? src_md() : diff_src_md(); |
49 | const auto &md_dst = is_fwd() ? dst_md() : diff_dst_md(); |
50 | const memory_desc_wrapper src_d(md_src); |
51 | const memory_desc_wrapper dst_d(md_dst); |
52 | |
53 | bool ok = src_d.data_type() == dst_d.data_type() |
54 | && md_src->format_kind == format_kind::blocked |
55 | && attr()->has_default_values() |
56 | && set_default_formats_common() && src_d == dst_d |
57 | && src_d.is_dense(); |
58 | if (!ok) return status::unimplemented; |
59 | |
60 | // Abort if there's blocking on the dimension that's going to be |
61 | // shuffled; such shuffle cannot be reduced to simple reorder. |
62 | // TODO: if both group_size and groups are multiples of blocking it |
63 | // still could be possible to use reorder. |
64 | for (int i = 0; i < md_src->format_desc.blocking.inner_nblks; i++) { |
65 | if (md_src->format_desc.blocking.inner_idxs[i] == axis()) { |
66 | return status::unimplemented; |
67 | } |
68 | } |
69 | |
70 | auto tensor_size |
71 | = utils::array_product(md_src->dims, md_src->ndims); |
72 | // groups, group_size() are sizes of the two fake dimensions |
73 | // groups * group_size() == size of the original single dimension |
74 | auto groups = md_src->dims[axis()] / group_size(); |
75 | // prepare 2 dimensions to be reordered |
76 | auto tr_rows = is_fwd() ? group_size() : groups; |
77 | auto tr_cols = is_fwd() ? groups : group_size(); |
78 | // combine all dimensions below axis() together with all blocks |
79 | // into a single dimension that's not going to be reordered |
80 | auto stride_of_axis = md_src->format_desc.blocking.strides[axis()]; |
81 | // combine all dimensions above axis into a single dimension |
82 | // that's not going to be reordered |
83 | auto remaining = tensor_size |
84 | / md_src->format_desc.blocking.strides[axis()] / tr_cols |
85 | / tr_rows; |
86 | |
87 | memory_desc_t fake_src; |
88 | memory_desc_t fake_dst; |
89 | |
90 | dims_t d = {remaining, tr_cols, tr_rows, stride_of_axis}; |
91 | dims_t strides_src = {d[3] * d[2] * d[1], d[3] * d[2], d[3], 1}; |
92 | dims_t strides_dst = {d[3] * d[2] * d[1], d[3], d[1] * d[3], 1}; |
93 | |
94 | CHECK(memory_desc_init_by_strides( |
95 | fake_src, 4, d, md_src->data_type, strides_src)); |
96 | CHECK(memory_desc_init_by_strides( |
97 | fake_dst, 4, d, md_src->data_type, strides_dst)); |
98 | |
99 | CHECK(reorder_primitive_desc_create( |
100 | reorder_pd_, engine, &fake_src, &fake_dst)); |
101 | return status::success; |
102 | } |
103 | |
104 | std::shared_ptr<primitive_desc_t> reorder_pd_; |
105 | }; |
106 | |
107 | status_t init(engine_t *engine) override { |
108 | return create_nested_primitive(reorder_, pd()->reorder_pd_, engine); |
109 | } |
110 | |
111 | status_t execute(const exec_ctx_t &ctx) const override { |
112 | using namespace memory_tracking::names; |
113 | exec_args_t r_args; |
114 | |
115 | auto src = pd()->is_fwd() ? DNNL_ARG_SRC : DNNL_ARG_DIFF_DST; |
116 | auto dst = pd()->is_fwd() ? DNNL_ARG_DST : DNNL_ARG_DIFF_SRC; |
117 | |
118 | r_args[DNNL_ARG_SRC] = ctx.args().at(src); |
119 | r_args[DNNL_ARG_DST] = ctx.args().at(dst); |
120 | exec_ctx_t r_ctx(ctx, std::move(r_args)); |
121 | |
122 | return reorder_->execute(r_ctx); |
123 | } |
124 | |
125 | private: |
126 | const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); } |
127 | std::shared_ptr<primitive_t> reorder_; |
128 | }; |
129 | } // namespace ocl |
130 | } // namespace gpu |
131 | } // namespace impl |
132 | } // namespace dnnl |
133 | |
134 | #endif |
135 | |