1 /*
2  * Copyright (c) 2019-2021 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/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.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/TensorInfo.h"
30 #include "arm_compute/core/Types.h"
31 #include "arm_compute/core/Utils.h"
32 #include "arm_compute/core/Validate.h"
33 #include "arm_compute/core/Window.h"
34 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
35 #include "src/core/NEON/NESymm.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,int min,int max)49 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
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     ARM_COMPUTE_RETURN_ERROR_ON(min > max);
54 
55     // Check biases if exist
56     if(bias != nullptr)
57     {
58         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias);
59         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
60         ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
61     }
62 
63     if(dst->total_size() != 0)
64     {
65         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QSYMM16);
66         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, src);
67     }
68 
69     return Status{};
70 }
71 } // namespace
72 
73 template <bool is_bounded_relu>
run_internal(const ITensor * src,const ITensor * bias,ITensor * dst,const Window & window)74 void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
75 {
76     const int16x8_t min_s16 = vdupq_n_s16(static_cast<int16_t>(_min));
77     const int16x8_t max_s16 = vdupq_n_s16(static_cast<int16_t>(_max));
78 
79     ARM_COMPUTE_UNUSED(min_s16);
80     ARM_COMPUTE_UNUSED(max_s16);
81 
82     const int  window_step_x  = 8;
83     const auto window_start_x = static_cast<int>(window.x().start());
84     const auto window_end_x   = static_cast<int>(window.x().end());
85 
86     Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
87     win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
88 
89     Iterator in(src, win_collapsed);
90     Iterator out(dst, win_collapsed);
91     if(bias != nullptr)
92     {
93         Window win_biases;
94         win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
95         win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
96 
97         Iterator bias_i(bias, win_biases);
98         execute_window_loop(win_collapsed, [&](const Coordinates &)
99         {
100             // Compute 16 elements per iteration
101             int x = window_start_x;
102             for(; x <= (window_end_x - window_step_x); x += window_step_x)
103             {
104                 int32x4x2_t in_s32 =
105                 {
106                     {
107                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
108                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
109                     }
110                 };
111 
112                 const int32x4x2_t bias_s32 =
113                 {
114                     {
115                         vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
116                         vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4)
117                     }
118                 };
119 
120                 // Add the bias to GEMM's result
121                 in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
122                 in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
123 
124                 vst1q_s16(reinterpret_cast<int16_t *>(out.ptr()) + x, finalize_quantization_int16<is_bounded_relu>(in_s32, _result_fixedpoint_multiplier, _result_shift, min_s16, max_s16));
125             }
126 
127             // Compute left-over elements
128             for(; x < window_end_x; ++x)
129             {
130                 const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x);
131                 int32_t       in_value   = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
132 
133                 // Add bias
134                 in_value += bias_value;
135                 // Finalize and store the result
136                 *(reinterpret_cast<int16_t *>(out.ptr()) + x) = finalize_quantization_int16<is_bounded_relu>(in_value, _result_fixedpoint_multiplier, _result_shift, static_cast<int16_t>(_min),
137                                                                                                              static_cast<int16_t>(_max));
138             }
139         },
140         in, out, bias_i);
141     }
142     else
143     {
144         execute_window_loop(win_collapsed, [&](const Coordinates &)
145         {
146             // Compute 16 elements per iteration
147             int x = window_start_x;
148             for(; x <= (window_end_x - window_step_x); x += window_step_x)
149             {
150                 int32x4x2_t in_s32 =
151                 {
152                     {
153                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
154                         vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
155                     }
156                 };
157 
158                 vst1q_s16(reinterpret_cast<int16_t *>(out.ptr()) + x, finalize_quantization_int16<is_bounded_relu>(in_s32, _result_fixedpoint_multiplier, _result_shift, min_s16, max_s16));
159             }
160 
161             // Compute left-over elements
162             for(; x < window_end_x; ++x)
163             {
164                 const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
165                 ARM_COMPUTE_UNUSED(in_value);
166                 // Finalize and store the result
167                 *(reinterpret_cast<int16_t *>(out.ptr()) + x) = finalize_quantization_int16<is_bounded_relu>(in_value, _result_fixedpoint_multiplier, _result_shift, static_cast<int16_t>(_min),
168                                                                                                              static_cast<int16_t>(_max));
169             }
170         },
171         in, out);
172     }
173 }
174 
configure(ITensorInfo * src,ITensorInfo * bias,ITensorInfo * dst,int result_fixedpoint_multiplier,int result_shift,int min,int max)175 void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift,
176                                                                            int min, int max)
177 {
178     // Perform validate step
179     ARM_COMPUTE_UNUSED(bias, dst);
180     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
181     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, min, max));
182 
183     _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
184     _result_shift                 = result_shift;
185     _min                          = min;
186     _max                          = max;
187 
188     // Output auto inizialitation if not yet initialized
189     auto_init_if_empty(*src, src->clone()->set_data_type(DataType::QSYMM16));
190     // Configure kernel window
191     Window win_config = calculate_max_window(*src, Steps());
192     ICpuKernel::configure(win_config);
193 
194     // Check if we need to clamp the result using min and max
195     const bool is_bounded_relu = !(min <= -32768 && max >= 32767);
196     _func                      = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal<true> :
197                                  &CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal<false>;
198 }
199 
validate(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,int min,int max)200 Status CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
201 {
202     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
203     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
204     return Status{};
205 }
206 
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)207 void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
208 {
209     ARM_COMPUTE_UNUSED(info);
210     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
211     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
212     ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
213 
214     auto src  = tensors.get_const_tensor(TensorType::ACL_SRC);
215     auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
216     auto dst  = tensors.get_tensor(TensorType::ACL_DST);
217 
218     (this->*_func)(src, bias, dst, window);
219 }
220 
name() const221 const char *CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::name() const
222 {
223     return "CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel";
224 }
225 } // namespace kernels
226 } // namespace cpu
227 } // namespace arm_compute
228