• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2017-2020 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/CL/functions/CLGEMMLowpOutputStage.h"
25 
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/Types.h"
28 #include "src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFixedPointKernel.h"
29 #include "src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h"
30 #include "src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleKernel.h"
31 #include "support/MemorySupport.h"
32 
33 #include <algorithm>
34 
35 namespace arm_compute
36 {
configure(const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,int result_fixedpoint_multiplier,int result_shift,int result_offset_after_shift,int min,int max)37 void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
38                                                                     int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift,
39                                                                     int min, int max)
40 {
41     configure(CLKernelLibrary::get().get_compile_context(), input, bias, output, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
42 }
43 
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,int result_fixedpoint_multiplier,int result_shift,int result_offset_after_shift,int min,int max)44 void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
45                                                                     int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift,
46                                                                     int min, int max)
47 {
48     GEMMLowpOutputStageInfo info{};
49     info.gemmlowp_multiplier = result_fixedpoint_multiplier;
50     info.gemmlowp_shift      = result_shift;
51     info.gemmlowp_offset     = result_offset_after_shift;
52     info.gemmlowp_min_bound  = min;
53     info.gemmlowp_max_bound  = max;
54     info.output_data_type    = DataType::QASYMM8;
55     auto k                   = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ScaleByFixedPointKernel>();
56     k->configure(compile_context, input, bias, output, &info);
57     _kernel = std::move(k);
58 }
59 
validate(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,int min,int max)60 Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
61                                                                      int min, int max)
62 {
63     GEMMLowpOutputStageInfo info{};
64     info.gemmlowp_min_bound = min;
65     info.gemmlowp_max_bound = max;
66     info.output_data_type   = DataType::QASYMM8;
67     return CLGEMMLowpQuantizeDownInt32ScaleByFixedPointKernel::validate(input, bias, output, &info);
68 }
69 
configure(const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,int result_fixedpoint_multiplier,int result_shift,int result_offset_after_shift,int min,int max)70 void CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
71                                                                    int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift,
72                                                                    int min, int max)
73 {
74     configure(CLKernelLibrary::get().get_compile_context(), input, bias, output, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
75 }
76 
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,int result_fixedpoint_multiplier,int result_shift,int result_offset_after_shift,int min,int max)77 void CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
78                                                                    int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift,
79                                                                    int min, int max)
80 {
81     GEMMLowpOutputStageInfo info{};
82     info.gemmlowp_multiplier = result_fixedpoint_multiplier;
83     info.gemmlowp_shift      = result_shift;
84     info.gemmlowp_offset     = result_offset_after_shift;
85     info.gemmlowp_min_bound  = min;
86     info.gemmlowp_max_bound  = max;
87     info.output_data_type    = DataType::QASYMM8_SIGNED;
88     auto k                   = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ScaleByFixedPointKernel>();
89     k->configure(compile_context, input, bias, output, &info);
90     _kernel = std::move(k);
91 }
92 
validate(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,int min,int max)93 Status CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
94                                                                     int min, int max)
95 {
96     GEMMLowpOutputStageInfo info{};
97     info.gemmlowp_min_bound = min;
98     info.gemmlowp_max_bound = max;
99     info.output_data_type   = DataType::QASYMM8_SIGNED;
100     return CLGEMMLowpQuantizeDownInt32ScaleByFixedPointKernel::validate(input, bias, output, &info);
101 }
102 
configure(const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,int result_fixedpoint_multiplier,int result_shift,int min,int max)103 void CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
104                                                                     int result_fixedpoint_multiplier, int result_shift,
105                                                                     int min, int max)
106 {
107     configure(CLKernelLibrary::get().get_compile_context(), input, bias, output, result_fixedpoint_multiplier, result_shift, min, max);
108 }
109 
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,int result_fixedpoint_multiplier,int result_shift,int min,int max)110 void CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
111                                                                     int result_fixedpoint_multiplier, int result_shift,
112                                                                     int min, int max)
113 {
114     GEMMLowpOutputStageInfo info{};
115     info.gemmlowp_multiplier = result_fixedpoint_multiplier;
116     info.gemmlowp_shift      = result_shift;
117     info.gemmlowp_min_bound  = min;
118     info.gemmlowp_max_bound  = max;
119     info.output_data_type    = DataType::QSYMM16;
120     auto k                   = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ScaleByFixedPointKernel>();
121     k->configure(compile_context, input, bias, output, &info);
122     _kernel = std::move(k);
123 }
124 
validate(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,int min,int max)125 Status CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
126                                                                      int min, int max)
127 {
128     GEMMLowpOutputStageInfo info{};
129     info.gemmlowp_min_bound = min;
130     info.gemmlowp_max_bound = max;
131     info.output_data_type   = DataType::QSYMM16;
132     return CLGEMMLowpQuantizeDownInt32ScaleByFixedPointKernel::validate(input, bias, output, &info);
133 }
134 
configure(const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,const GEMMLowpOutputStageInfo & info)135 void CLGEMMLowpOutputStage::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const GEMMLowpOutputStageInfo &info)
136 {
137     configure(CLKernelLibrary::get().get_compile_context(), input, bias, output, info);
138 }
139 
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,const GEMMLowpOutputStageInfo & info)140 void CLGEMMLowpOutputStage::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const GEMMLowpOutputStageInfo &info)
141 {
142     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
143 
144     switch(info.type)
145     {
146         case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
147         {
148             auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ScaleByFixedPointKernel>();
149             k->configure(compile_context, input, bias, output, &info);
150             _kernel = std::move(k);
151             break;
152         }
153         case GEMMLowpOutputStageType::QUANTIZE_DOWN:
154         {
155             auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ScaleKernel>();
156             k->configure(compile_context, input, bias, output, &info);
157             _kernel = std::move(k);
158             break;
159         }
160         case GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT:
161         {
162             auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel>();
163             k->configure(compile_context, input, bias, output, &info);
164             _kernel = std::move(k);
165             break;
166         }
167         default:
168             ARM_COMPUTE_ERROR("Unsupported GEMMLowpOutputStage type.");
169     }
170 }
171 
validate(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,const GEMMLowpOutputStageInfo & info)172 Status CLGEMMLowpOutputStage::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo &info)
173 {
174     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
175     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16);
176 
177     switch(info.type)
178     {
179         case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
180             return CLGEMMLowpQuantizeDownInt32ScaleByFixedPointKernel::validate(input, bias, output, &info);
181         case GEMMLowpOutputStageType::QUANTIZE_DOWN:
182             return CLGEMMLowpQuantizeDownInt32ScaleKernel::validate(input, bias, output, &info);
183         case GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT:
184             return CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::validate(input, bias, output, &info);
185         default:
186             return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported GEMMLowpOutputStage type.");
187     }
188 }
189 } // namespace arm_compute
190