1 /*
2 * Copyright (c) 2021 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 "src/cpu/operators/CpuGemmLowpOutputStage.h"
25
26 #include "arm_compute/core/ITensor.h"
27 #include "arm_compute/core/Validate.h"
28 #include "arm_compute/runtime/NEON/NEScheduler.h"
29 #include "src/common/utils/Log.h"
30 #include "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.h"
31 #include "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
32 #include "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h"
33 #include "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
34
35 namespace arm_compute
36 {
37 namespace cpu
38 {
configure(ITensorInfo * src,ITensorInfo * bias,ITensorInfo * dst,const GEMMLowpOutputStageInfo & info)39 void CpuGemmLowpOutputStage::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const GEMMLowpOutputStageInfo &info)
40 {
41 // Perform validate step
42 ARM_COMPUTE_ERROR_THROW_ON(CpuGemmLowpOutputStage::validate(src, bias, dst, info));
43 ARM_COMPUTE_LOG_PARAMS(src, bias, dst, info);
44
45 switch(info.type)
46 {
47 case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
48 {
49 switch(info.output_data_type)
50 {
51 case DataType::QASYMM8:
52 {
53 auto k = std::make_unique<kernels::CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel>();
54 k->configure(src, bias, dst, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
55 _kernel = std::move(k);
56 break;
57 }
58 case DataType::QASYMM8_SIGNED:
59 {
60 auto k = std::make_unique<kernels::CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel>();
61 k->configure(src, bias, dst, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
62 _kernel = std::move(k);
63 break;
64 }
65 case DataType::QSYMM16:
66 {
67 auto k = std::make_unique<kernels::CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel>();
68 k->configure(src, bias, dst, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
69 _kernel = std::move(k);
70 break;
71 }
72 default:
73 {
74 ARM_COMPUTE_ERROR("Unsupported output data type.");
75 break;
76 }
77 }
78 break;
79 }
80 case GEMMLowpOutputStageType::QUANTIZE_DOWN:
81 {
82 switch(info.output_data_type)
83 {
84 case DataType::QASYMM8:
85 case DataType::QASYMM8_SIGNED:
86 {
87 auto k = std::make_unique<kernels::CpuGemmLowpQuantizeDownInt32ScaleKernel>();
88 k->configure(src, bias, dst, &info);
89 _kernel = std::move(k);
90 break;
91 }
92 default:
93 {
94 ARM_COMPUTE_ERROR("Unsupported output data type.");
95 break;
96 }
97 }
98 break;
99 }
100 default:
101 ARM_COMPUTE_ERROR("Unsupported GEMMLowpOutputStage type.");
102 }
103 }
104
validate(const ITensorInfo * src,const ITensorInfo * bias,const ITensorInfo * dst,const GEMMLowpOutputStageInfo & info)105 Status CpuGemmLowpOutputStage::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo &info)
106 {
107 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
108 ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst->data_type() == DataType::UNKNOWN, "CpuGemmLowpOutputStage cannot be used with UNKNOWN output data type.");
109 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16);
110 ARM_COMPUTE_RETURN_ERROR_ON((info.type != GEMMLowpOutputStageType::QUANTIZE_DOWN) && (info.type != GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT));
111
112 switch(info.type)
113 {
114 case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
115 {
116 switch(dst->data_type())
117 {
118 case DataType::QASYMM8:
119 return kernels::CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(src, bias, dst, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
120 case DataType::QASYMM8_SIGNED:
121 return kernels::CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(src, bias, dst, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
122 case DataType::QSYMM16:
123 return kernels::CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(src, bias, dst, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
124 default:
125 return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
126 }
127 }
128 case GEMMLowpOutputStageType::QUANTIZE_DOWN:
129 {
130 switch(dst->data_type())
131 {
132 case DataType::QASYMM8:
133 case DataType::QASYMM8_SIGNED:
134 return kernels::CpuGemmLowpQuantizeDownInt32ScaleKernel::validate(src, bias, dst, &info);
135 default:
136 return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
137 }
138 }
139 default:
140 return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported GEMMLowpOutputStage type.");
141 }
142 }
143
run(ITensorPack & tensors)144 void CpuGemmLowpOutputStage::run(ITensorPack &tensors)
145 {
146 NEScheduler::get().schedule_op(_kernel.get(), Window::DimY, _kernel->window(), tensors);
147 }
148 } // namespace cpu
149 } // namespace arm_compute