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_OCL_MANY_INPUTS_SUM_HPP |
18 | #define GPU_OCL_MANY_INPUTS_SUM_HPP |
19 | |
20 | #include "common/c_types_map.hpp" |
21 | #include "common/primitive.hpp" |
22 | #include "gpu/compute/compute.hpp" |
23 | #include "gpu/gpu_primitive.hpp" |
24 | #include "gpu/gpu_resource.hpp" |
25 | #include "gpu/gpu_sum_pd.hpp" |
26 | #include "gpu/ocl/ocl_stream.hpp" |
27 | #include "gpu/ocl/ocl_utils.hpp" |
28 | #include "gpu/primitive_conf.hpp" |
29 | |
30 | namespace dnnl { |
31 | namespace impl { |
32 | namespace gpu { |
33 | namespace ocl { |
34 | |
35 | struct many_inputs_sum_t : public gpu_primitive_t { |
36 | using gpu_primitive_t::gpu_primitive_t; |
37 | struct pd_t : public gpu_sum_pd_t { |
38 | using gpu_sum_pd_t::gpu_sum_pd_t; |
39 | |
40 | DECLARE_SUM_PD_T("ocl:many_inputs" , many_inputs_sum_t); |
41 | |
42 | status_t init(engine_t *engine) { |
43 | const int n = n_inputs(); |
44 | |
45 | bool ok = gpu_sum_pd_t::init(engine) == status::success; |
46 | |
47 | if (!ok) return status::unimplemented; |
48 | |
49 | const memory_desc_wrapper o_d(dst_md()); |
50 | |
51 | for (int i = 0; i < n; ++i) { |
52 | const memory_desc_wrapper i_d(src_md(i)); |
53 | if (i_d != o_d) return status::unimplemented; |
54 | } |
55 | |
56 | if (scales()[0] != 1.0f) return status::unimplemented; |
57 | return status::success; |
58 | } |
59 | }; |
60 | |
61 | status_t init(engine_t *engine) override { |
62 | compute::kernel_ctx_t kernel_ctx; |
63 | |
64 | const memory_desc_wrapper data_d(pd()->dst_md()); |
65 | const memory_desc_wrapper data_s(pd()->src_md()); |
66 | |
67 | kernel_ctx.set_data_type(data_s.data_type()); |
68 | |
69 | kernel_ctx.define_int("N_ELEMS" , data_d.nelems(true)); |
70 | |
71 | const int num_arrs = pd()->n_inputs() - 1; |
72 | int N_INPUTS = (num_arrs) % max_num_arrs; |
73 | if (N_INPUTS == 0) { N_INPUTS = max_num_arrs; }; |
74 | kernel_ctx.define_int("N_INPUTS" , N_INPUTS); |
75 | kernel_ctx.define_int("MAX_N_INPUTS" , max_num_arrs); |
76 | |
77 | def_memory_desc_info( |
78 | kernel_ctx, memory_desc_info_t::create(data_d), "SRC" ); |
79 | def_memory_desc_info( |
80 | kernel_ctx, memory_desc_info_t::create(data_s), "DST" ); |
81 | |
82 | std::vector<compute::kernel_t> kernels; |
83 | std::vector<const char *> kernel_names; |
84 | kernel_names.push_back("many_inputs_sum" ); |
85 | kernel_names.push_back("many_inputs_sum_batched" ); |
86 | CHECK(create_kernels(engine, &kernels, kernel_names, kernel_ctx)); |
87 | kernel_ = kernels[0]; |
88 | batched_kernel_ = kernels[1]; |
89 | if (!kernel_ || !batched_kernel_) return status::runtime_error; |
90 | return status::success; |
91 | } |
92 | |
93 | status_t init_res_storage( |
94 | engine_t *engine, gpu_resource_t *r) const override { |
95 | const dim_t count = pd()->n_inputs(); |
96 | const float *s_data = pd()->scales(); |
97 | |
98 | const size_t size = count * sizeof(float); |
99 | std::unique_ptr<memory_storage_t> scales; |
100 | memory_storage_t *scale = nullptr; |
101 | auto s = engine->create_memory_storage(&scale, size); |
102 | if (s != status::success) return s; |
103 | float *mapped_mem_storage = nullptr; |
104 | s = scale->map_data((void **)&mapped_mem_storage, nullptr, size); |
105 | if (s != status::success) return s; |
106 | utils::array_copy(mapped_mem_storage, s_data, count); |
107 | s = scale->unmap_data((void *)mapped_mem_storage, nullptr); |
108 | if (s != status::success) return s; |
109 | scales.reset(scale); |
110 | r->add_memory_storage(SCALES_, std::move(scales)); |
111 | return status::success; |
112 | } |
113 | |
114 | status_t execute(const exec_ctx_t &ctx) const override; |
115 | |
116 | private: |
117 | enum { max_num_arrs = 94 }; |
118 | enum { SCALES_ = 0 }; |
119 | const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); } |
120 | compute::kernel_t kernel_; |
121 | compute::kernel_t batched_kernel_; |
122 | }; |
123 | |
124 | } // namespace ocl |
125 | } // namespace gpu |
126 | } // namespace impl |
127 | } // namespace dnnl |
128 | |
129 | #endif |
130 | |
131 | // vim: et ts=4 sw=4 cindent cino+=l0,\:4,N-s |
132 | |