1 /*
2 * Copyright (c) 2017-2022 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #include "arm_compute/runtime/NEON/functions/NEGEMM.h"
25
26 #include "arm_compute/core/ITensorPack.h"
27 #include "arm_compute/core/TensorInfo.h"
28 #include "arm_compute/core/Types.h"
29 #include "arm_compute/runtime/MemoryGroup.h"
30 #include "arm_compute/runtime/Tensor.h"
31 #include "src/core/CPP/Validate.h"
32 #include "src/core/helpers/MemoryHelpers.h"
33 #include "src/cpu/operators/CpuGemm.h"
34
35 using namespace arm_compute::experimental;
36
37 namespace arm_compute
38 {
39 struct NEGEMM::Impl
40 {
41 MemoryGroup memory_group{};
42 IWeightsManager *weights_manager{ nullptr };
43
44 std::unique_ptr<cpu::CpuGemm> op{ nullptr };
45
46 const ITensor *original_b{ nullptr };
47 bool is_prepared{ false };
48
49 ITensorPack run_pack{};
50 ITensorPack prep_pack{};
51 WorkspaceData<Tensor> workspace{};
52 experimental::MemoryRequirements aux_mem_req{};
53 };
54
NEGEMM(std::shared_ptr<IMemoryManager> memory_manager,IWeightsManager * weights_manager)55 NEGEMM::NEGEMM(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
56 : _impl(std::make_unique<Impl>())
57 {
58 _impl->memory_group = MemoryGroup(std::move(memory_manager));
59 _impl->weights_manager = weights_manager;
60 }
61
62 NEGEMM::~NEGEMM() = default;
63
configure(const ITensor * a,const ITensor * b,const ITensor * c,ITensor * d,float alpha,float beta,const GEMMInfo & gemm_info)64 void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta, const GEMMInfo &gemm_info)
65 {
66 ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, d);
67 ARM_COMPUTE_ERROR_THROW_ON(cpu::CpuGemm::validate(a->info(), b->info(), (c != nullptr) ? c->info() : nullptr, d->info(), alpha, beta, gemm_info));
68
69 // Check if we need to reshape the matrix B only on the first run
70 _impl->is_prepared = false;
71 _impl->original_b = b;
72 _impl->op = std::make_unique<cpu::CpuGemm>();
73
74 _impl->op->configure(a->info(), b->info(), (c != nullptr) ? c->info() : nullptr, d->info(), alpha, beta, gemm_info);
75
76 _impl->aux_mem_req = _impl->op->workspace();
77 _impl->run_pack = { { ACL_SRC_0, a }, { ACL_SRC_1, b }, { ACL_SRC_2, c }, { ACL_DST, d } };
78 _impl->prep_pack = { { ACL_SRC_1, b }, { ACL_SRC_2, c } };
79 _impl->workspace = manage_workspace<Tensor>(_impl->aux_mem_req, _impl->memory_group, _impl->run_pack, _impl->prep_pack);
80 }
81
validate(const ITensorInfo * a,const ITensorInfo * b,const ITensorInfo * c,const ITensorInfo * output,float alpha,float beta,const GEMMInfo & gemm_info)82 Status NEGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info)
83 {
84 return cpu::CpuGemm::validate(a, b, c, output, alpha, beta, gemm_info);
85 }
86
has_opt_impl(arm_compute::WeightFormat & expected_weight_format,const ITensorInfo * a,const ITensorInfo * b,const ITensorInfo * c,const ITensorInfo * output,float alpha,float beta,const GEMMInfo & gemm_info)87 Status NEGEMM::has_opt_impl(arm_compute::WeightFormat &expected_weight_format, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output,
88 float alpha, float beta, const GEMMInfo &gemm_info)
89 {
90 ARM_COMPUTE_UNUSED(alpha, beta);
91 return cpu::CpuGemm::has_opt_impl(expected_weight_format, a, b, c, output, gemm_info);
92 }
93
run()94 void NEGEMM::run()
95 {
96 prepare();
97
98 MemoryGroupResourceScope scope_mg(_impl->memory_group);
99 _impl->op->run(_impl->run_pack);
100 }
101
prepare()102 void NEGEMM::prepare()
103 {
104 if(!_impl->is_prepared)
105 {
106 _impl->op->prepare(_impl->prep_pack);
107
108 auto has_reshape = std::find_if(_impl->aux_mem_req.begin(),
109 _impl->aux_mem_req.end(),
110 [](const MemoryInfo & m) -> bool { return m.lifetime == MemoryLifetime::Persistent; });
111
112 if(has_reshape != std::end(_impl->aux_mem_req))
113 {
114 _impl->original_b->mark_as_unused();
115 }
116 else
117 {
118 _impl->run_pack.add_const_tensor(ACL_SRC_1, _impl->original_b);
119 }
120
121 // Release temporary tensors that are only used in prepare stage
122 release_temporaries<Tensor>(_impl->aux_mem_req, _impl->workspace);
123 _impl->is_prepared = true;
124 }
125 }
126 } // namespace arm_compute
127