xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3) 
1 /*
2  * Copyright (c) 2020-2021 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/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.h"
25 
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/ITensor.h"
29 #include "arm_compute/core/Types.h"
30 #include "arm_compute/core/Utils.h"
31 #include "arm_compute/core/Validate.h"
32 #include "arm_compute/core/Window.h"
33 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
34 #include "src/core/AccessWindowStatic.h"
35 #include "src/core/NEON/wrapper/wrapper.h"
36 #include "src/core/helpers/AutoConfiguration.h"
37 #include "src/core/helpers/WindowHelpers.h"
38 
39 #include <arm_neon.h>
40 
41 namespace arm_compute
42 {
43 namespace cpu
44 {
45 namespace kernels
46 {
47 namespace
48 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * bias,const ITensorInfo * dst,const GEMMLowpOutputStageInfo * output_stage)49 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage)
50 {
51     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
52     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::S32);
53 
54     ARM_COMPUTE_RETURN_ERROR_ON(output_stage->gemmlowp_max_bound > std::get<1>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type)));
55     ARM_COMPUTE_RETURN_ERROR_ON(output_stage->gemmlowp_min_bound < std::get<0>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))
56                                 || output_stage->gemmlowp_min_bound > output_stage->gemmlowp_max_bound);
57 
58     // Check biases if exist
59     if(bias != nullptr)
60     {
61         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias);
62         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
63         ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
64     }
65 
66     if(dst->total_size() != 0)
67     {
68         if(dst->data_type() != output_stage->output_data_type && (output_stage->output_data_type == DataType::QASYMM8 || output_stage->output_data_type == DataType::QASYMM8_SIGNED))
69         {
70             ARM_COMPUTE_RETURN_ERROR_MSG("Mismatching data types");
71         }
72 
73         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
74     }
75 
76     return Status{};
77 }
78 
scale_input(int32x4x4_t & in_s32,int32x4_t result_offset_s32,int32_t result_mult_int)79 inline void scale_input(int32x4x4_t &in_s32, int32x4_t result_offset_s32, int32_t result_mult_int)
80 {
81     // Add the offset terms to GEMM's result
82     in_s32.val[0] = vaddq_s32(in_s32.val[0], result_offset_s32);
83     in_s32.val[1] = vaddq_s32(in_s32.val[1], result_offset_s32);
84     in_s32.val[2] = vaddq_s32(in_s32.val[2], result_offset_s32);
85     in_s32.val[3] = vaddq_s32(in_s32.val[3], result_offset_s32);
86 
87     // Multiply by result_mult_int
88     in_s32.val[0] = vmulq_n_s32(in_s32.val[0], result_mult_int);
89     in_s32.val[1] = vmulq_n_s32(in_s32.val[1], result_mult_int);
90     in_s32.val[2] = vmulq_n_s32(in_s32.val[2], result_mult_int);
91     in_s32.val[3] = vmulq_n_s32(in_s32.val[3], result_mult_int);
92 }
93 
94 template <typename T>
95 inline typename std::enable_if<std::is_same<T, uint8_t>::value,
96        typename wrapper::traits::neon_vector<T, 16>::type>::type
convert_to_8bit(const int16x8x2_t in_s16)97        convert_to_8bit(const int16x8x2_t in_s16)
98 {
99     return wrapper::vcombine(wrapper::vqmovun(in_s16.val[0]), wrapper::vqmovun(in_s16.val[1]));
100 }
101 
102 template <typename T>
103 inline typename std::enable_if<std::is_same<T, int8_t>::value,
104        typename wrapper::traits::neon_vector<T, 16>::type>::type
convert_to_8bit(const int16x8x2_t in_s16)105        convert_to_8bit(const int16x8x2_t in_s16)
106 {
107     return wrapper::vcombine(wrapper::vqmovn(in_s16.val[0]), wrapper::vqmovn(in_s16.val[1]));
108 }
109 
110 template <typename T>
finalize_quantization(int32x4x4_t & in_s32,int32x4_t result_shift_s32,typename wrapper::traits::neon_vector<T,16>::type min,typename wrapper::traits::neon_vector<T,16>::type max)111 inline typename wrapper::traits::neon_vector<T, 16>::type finalize_quantization(int32x4x4_t &in_s32, int32x4_t result_shift_s32, typename wrapper::traits::neon_vector<T, 16>::type min,
112                                                                                 typename wrapper::traits::neon_vector<T, 16>::type max)
113 {
114     // Shift final result (negative value shift right)
115     in_s32.val[0] = vshlq_s32(in_s32.val[0], result_shift_s32);
116     in_s32.val[1] = vshlq_s32(in_s32.val[1], result_shift_s32);
117     in_s32.val[2] = vshlq_s32(in_s32.val[2], result_shift_s32);
118     in_s32.val[3] = vshlq_s32(in_s32.val[3], result_shift_s32);
119 
120     // Convert S32 to S16
121     const int16x8x2_t in_s16 =
122     {
123         {
124             vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),
125             vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3]))
126         }
127     };
128 
129     // Convert S16 to S8 or U8
130     typename wrapper::traits::neon_vector<T, 16>::type out = convert_to_8bit<T>(in_s16);
131 
132     out = wrapper::vmax(out, min);
133     out = wrapper::vmin(out, max);
134 
135     return out;
136 }
137 } // namespace
138 
139 template <typename T>
run_internal(const ITensor * src,const ITensor * bias,ITensor * dst,const Window & window)140 void CpuGemmLowpQuantizeDownInt32ScaleKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
141 {
142     using VectorType = typename wrapper::traits::neon_vector<T, 16>::type;
143 
144     const int32x4_t result_offset_s32 = vdupq_n_s32(_output_stage->gemmlowp_offset);
145     const int32x4_t result_shift_s32  = vdupq_n_s32(-_output_stage->gemmlowp_shift);
146     const int       window_step_x     = 16;
147     const auto      window_start_x    = static_cast<int>(window.x().start());
148     const auto      window_end_x      = static_cast<int>(window.x().end());
149 
150     const int clamp_min = (_is_bounded_relu) ? _output_stage->gemmlowp_min_bound : std::numeric_limits<T>::lowest();
151     const int clamp_max = (_is_bounded_relu) ? _output_stage->gemmlowp_max_bound : std::numeric_limits<T>::max();
152 
153     VectorType min = wrapper::vdup_n(static_cast<T>(clamp_min), wrapper::traits::vector_128_tag{});
154     VectorType max = wrapper::vdup_n(static_cast<T>(clamp_max), wrapper::traits::vector_128_tag{});
155 
156     Window win(window);
157     win.set(Window::DimX, Window::Dimension(0, 1, 1));
158 
159     Iterator in(src, win);
160     Iterator out(dst, win);
161 
162     if(bias != nullptr)
163     {
164         Window win_biases;
165         win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
166         win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
167 
168         Iterator bias_i(bias, win_biases);
169         execute_window_loop(win, [&](const Coordinates &)
170         {
171             // Compute 16 elements per iteration
172             int x = window_start_x;
173             for(; x <= (window_end_x - window_step_x); x += window_step_x)
174             {
175                 int32x4x4_t in_s32 =
176                 {
177                     {
178                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
179                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
180                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
181                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
182                     }
183                 };
184 
185                 const int32x4x4_t bias_s32 =
186                 {
187                     {
188                         vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
189                         vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4),
190                         vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 8),
191                         vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 12)
192                     }
193                 };
194 
195                 // Add the bias to GEMM's result
196                 in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
197                 in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
198                 in_s32.val[2] = vaddq_s32(in_s32.val[2], bias_s32.val[2]);
199                 in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]);
200 
201                 // Add the offset terms to GEMM's result and multiply by result_mult_int
202                 scale_input(in_s32, result_offset_s32, _output_stage->gemmlowp_multiplier);
203 
204                 wrapper::vstore(reinterpret_cast<T *>(out.ptr() + x), finalize_quantization<T>(in_s32, result_shift_s32, min, max));
205             }
206 
207             // Compute left-over elements
208             for(; x < window_end_x; ++x)
209             {
210                 const int bias_value = *(reinterpret_cast<const int *>(bias_i.ptr()) + x);
211                 int       in_value   = *(reinterpret_cast<const int *>(in.ptr()) + x);
212 
213                 // Quantize
214                 in_value = ((in_value + bias_value + _output_stage->gemmlowp_offset) * _output_stage->gemmlowp_multiplier) >> _output_stage->gemmlowp_shift;
215 
216                 // Store the result
217                 *(out.ptr() + x) = static_cast<T>(utility::clamp<int>(in_value, clamp_min, clamp_max));
218             }
219         },
220         in, bias_i, out);
221     }
222     else
223     {
224         execute_window_loop(win, [&](const Coordinates &)
225         {
226             // Compute 16 elements per iteration
227             int x = window_start_x;
228             for(; x <= (window_end_x - window_step_x); x += window_step_x)
229             {
230                 int32x4x4_t in_s32 =
231                 {
232                     {
233                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
234                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
235                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
236                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
237                     }
238                 };
239 
240                 // Add the offset terms to GEMM's result and multiply by result_mult_int
241                 scale_input(in_s32, result_offset_s32, _output_stage->gemmlowp_multiplier);
242 
243                 wrapper::vstore(reinterpret_cast<T *>(out.ptr() + x), finalize_quantization<T>(in_s32, result_shift_s32, min, max));
244             }
245 
246             // Compute left-over elements
247             for(; x < window_end_x; ++x)
248             {
249                 int in_value = *(reinterpret_cast<const int *>(in.ptr()) + x);
250 
251                 // Quantize
252                 in_value = ((in_value + _output_stage->gemmlowp_offset) * _output_stage->gemmlowp_multiplier) >> _output_stage->gemmlowp_shift;
253 
254                 // Store the result
255                 *(out.ptr() + x) = static_cast<T>(utility::clamp<int>(in_value, clamp_min, clamp_max));
256             }
257         },
258         in, out);
259     }
260 }
261 
configure(ITensorInfo * src,ITensorInfo * bias,ITensorInfo * dst,const GEMMLowpOutputStageInfo * output_stage)262 void CpuGemmLowpQuantizeDownInt32ScaleKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage)
263 {
264     ARM_COMPUTE_UNUSED(bias);
265     // Perform validate step
266     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst, output_stage);
267 
268     // Output auto inizialitation if not yet initialized
269     auto_init_if_empty(*dst, src->clone()->set_data_type(output_stage->output_data_type));
270 
271     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src,
272                                                   bias,
273                                                   dst,
274                                                   output_stage));
275 
276     _output_stage = output_stage;
277 
278     // Configure kernel window
279     Window win = calculate_max_window(*src, Steps());
280 
281     ICpuKernel::configure(win);
282 
283     // Check if we need to clamp the result using min and max
284     _is_bounded_relu = ((_output_stage->gemmlowp_min_bound != _output_stage->gemmlowp_max_bound)
285                         && !(_output_stage->gemmlowp_min_bound == std::get<0>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))
286                              && _output_stage->gemmlowp_max_bound == std::get<1>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))));
287     if(_output_stage->output_data_type == DataType::QASYMM8)
288     {
289         _func = &CpuGemmLowpQuantizeDownInt32ScaleKernel::run_internal<uint8_t>;
290     }
291     else if(_output_stage->output_data_type == DataType::QASYMM8_SIGNED)
292     {
293         _func = &CpuGemmLowpQuantizeDownInt32ScaleKernel::run_internal<int8_t>;
294     }
295     else
296     {
297         ARM_COMPUTE_ERROR("Data type not supported");
298     }
299 }
300 
validate(const ITensorInfo * src,const ITensorInfo * bias,const ITensorInfo * dst,const GEMMLowpOutputStageInfo * output_stage)301 Status CpuGemmLowpQuantizeDownInt32ScaleKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage)
302 {
303     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, output_stage));
304     return Status{};
305 }
306 
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)307 void CpuGemmLowpQuantizeDownInt32ScaleKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
308 {
309     ARM_COMPUTE_UNUSED(info);
310     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
311     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
312     ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
313 
314     auto src  = tensors.get_const_tensor(TensorType::ACL_SRC);
315     auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
316     auto dst  = tensors.get_tensor(TensorType::ACL_DST);
317     (this->*_func)(src, bias, dst, window);
318 }
319 
name() const320 const char *CpuGemmLowpQuantizeDownInt32ScaleKernel::name() const
321 {
322     return "CpuGemmLowpQuantizeDownInt32ScaleKernel";
323 }
324 } // namespace kernels
325 } // namespace cpu
326 } // namespace arm_compute