1 /*
2 * Copyright (c) 2017-2021 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/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.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/NEAsymm.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::QASYMM8);
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 CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
75 {
76 const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(_result_offset_after_shift);
77 const uint8x16_t min_u8 = vdupq_n_u8(static_cast<uint8_t>(_min));
78 const uint8x16_t max_u8 = vdupq_n_u8(static_cast<uint8_t>(_max));
79
80 ARM_COMPUTE_UNUSED(min_u8);
81 ARM_COMPUTE_UNUSED(max_u8);
82
83 const int window_step_x = 16;
84 const auto window_start_x = static_cast<int>(window.x().start());
85 const auto window_end_x = static_cast<int>(window.x().end());
86
87 Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
88 win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
89
90 Iterator in(src, win_collapsed);
91 Iterator out(dst, win_collapsed);
92 if(bias != nullptr)
93 {
94 Window win_biases;
95 win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
96 win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
97
98 Iterator bias_i(bias, win_biases);
99 execute_window_loop(win_collapsed, [&](const Coordinates &)
100 {
101 // Compute 16 elements per iteration
102 int x = window_start_x;
103 for(; x <= (window_end_x - window_step_x); x += window_step_x)
104 {
105 int32x4x4_t in_s32 =
106 {
107 {
108 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
109 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
110 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
111 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
112 }
113 };
114
115 const int32x4x4_t bias_s32 =
116 {
117 {
118 vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
119 vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4),
120 vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 8),
121 vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 12)
122 }
123 };
124
125 // Add the bias to GEMM's result
126 in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
127 in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
128 in_s32.val[2] = vaddq_s32(in_s32.val[2], bias_s32.val[2]);
129 in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]);
130
131 vst1q_u8(out.ptr() + x, finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_u8, max_u8, is_bounded_relu));
132 }
133
134 // Compute left-over elements
135 for(; x < window_end_x; ++x)
136 {
137 const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x);
138 int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
139
140 // Add bias
141 in_value += bias_value;
142 // Finalize and store the result
143 *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift, static_cast<uint8_t>(_min), static_cast<uint8_t>(_max), is_bounded_relu);
144 }
145 },
146 in, out, bias_i);
147 }
148 else
149 {
150 execute_window_loop(win_collapsed, [&](const Coordinates &)
151 {
152 // Compute 16 elements per iteration
153 int x = window_start_x;
154 for(; x <= (window_end_x - window_step_x); x += window_step_x)
155 {
156 int32x4x4_t in_s32 =
157 {
158 {
159 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
160 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
161 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
162 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
163 }
164 };
165
166 vst1q_u8(out.ptr() + x, finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_u8, max_u8, is_bounded_relu));
167 }
168
169 // Compute left-over elements
170 for(; x < window_end_x; ++x)
171 {
172 const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
173
174 // Finalize and store the result
175 *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift, static_cast<uint8_t>(_min), static_cast<uint8_t>(_max), is_bounded_relu);
176 }
177 },
178 in, out);
179 }
180 }
181
configure(ITensorInfo * src,ITensorInfo * bias,ITensorInfo * dst,int result_fixedpoint_multiplier,int result_shift,int result_offset_after_shift,int min,int max)182 void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift,
183 int result_offset_after_shift, int min, int max)
184 {
185 ARM_COMPUTE_UNUSED(bias);
186 // Perform validate step
187 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
188 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, min, max));
189
190 _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
191 _result_shift = result_shift;
192 _result_offset_after_shift = result_offset_after_shift;
193 _min = min;
194 _max = max;
195
196 // Output auto inizialitation if not yet initialized
197 auto_init_if_empty(*dst, src->clone()->set_data_type(DataType::QASYMM8));
198
199 // Configure kernel window
200 auto win_config = calculate_max_window(*src, Steps());
201 ICpuKernel::configure(win_config);
202
203 // Check if we need to clamp the result using min and max
204 const bool is_bounded_relu = !(min <= 0 && max >= 255);
205 _func = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal<true> :
206 &CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal<false>;
207 }
208
validate(const ITensorInfo * src,const ITensorInfo * bias,const ITensorInfo * dst,int min,int max)209 Status CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
210 {
211 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
212 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, min, max));
213 return Status{};
214 }
215
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)216 void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
217 {
218 ARM_COMPUTE_UNUSED(info);
219 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
220 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
221 ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
222
223 auto src = tensors.get_const_tensor(TensorType::ACL_SRC);
224 auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
225 auto dst = tensors.get_tensor(TensorType::ACL_DST);
226
227 (this->*_func)(src, bias, dst, window);
228 }
229
name() const230 const char *CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::name() const
231 {
232 return "CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel";
233 }
234 } // namespace kernels
235 } // namespace cpu
236 } // namespace arm_compute