Compute Library
 21.02
NESymm.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2019-2020 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 #ifndef ARM_COMPUTE_NESYMM_H
25 #define ARM_COMPUTE_NESYMM_H
26 
28 #include "src/core/NEON/NEMath.h"
29 #include <arm_neon.h>
30 
31 namespace arm_compute
32 {
33 using qsymm8_t = int8_t; /**< 8 bit quantized symmetric scalar value */
34 using qsymm16_t = int16_t; /**< 16 bit quantized symmetric scalar value */
35 
36 using qsymm16x8_t = int16x8_t; /**< 16 bit quantized symmetric vector with 8 elements */
37 using qsymm16x8x2_t = int16x8x2_t; /**< 16 bit quantized symmetric vector with 16 elements */
38 
39 /** Performs final quantization step on 8 signed 16-bit elements
40  *
41  * @tparam is_bounded_relu Specified if a fused bounded relu should be applied
42  *
43  * @param[in] in_s32 Input to be quantized.
44  * @param[in] result_fixedpoint_multiplier Result multiplier parameter
45  * @param[in] result_shift Result shift parameter
46  * @param[in] min_s16 Relu lower bound
47  * @param[in] max_s16 Relu upper bound
48  *
49  * @return Quantized values
50  */
51 template <bool is_bounded_relu>
52 int16x8_t finalize_quantization_int16(int32x4x2_t &in_s32,
53  int result_fixedpoint_multiplier,
54  int32_t result_shift,
55  int16x8_t min_s16,
56  int16x8_t max_s16)
57 {
58  if(result_shift < 0)
59  {
60  in_s32.val[0] = vmulq_n_s32(in_s32.val[0], (1 << -result_shift));
61  in_s32.val[1] = vmulq_n_s32(in_s32.val[1], (1 << -result_shift));
62 
63  in_s32.val[0] = vqrdmulhq_n_s32(in_s32.val[0], result_fixedpoint_multiplier);
64  in_s32.val[1] = vqrdmulhq_n_s32(in_s32.val[1], result_fixedpoint_multiplier);
65  }
66  else
67  {
68  // Fixed point multiplication with vector saturating rounding doubling multiply high with scalar
69  in_s32.val[0] = vqrdmulhq_n_s32(in_s32.val[0], result_fixedpoint_multiplier);
70  in_s32.val[1] = vqrdmulhq_n_s32(in_s32.val[1], result_fixedpoint_multiplier);
71  // Round to the nearest division by a power-of-two using result_shift_s32
72  in_s32.val[0] = rounding_divide_by_pow2(in_s32.val[0], result_shift);
73  in_s32.val[1] = rounding_divide_by_pow2(in_s32.val[1], result_shift);
74  }
75 
76  // Convert S32 to S16
77  int16x8_t out_s16 = vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1]));
78 
79  if(is_bounded_relu)
80  {
81  out_s16 = vmaxq_s16(out_s16, min_s16);
82  out_s16 = vminq_s16(out_s16, max_s16);
83  }
84 
85  return out_s16;
86 }
87 
88 /** Performs final quantization step on single signed 16-bit element
89  *
90  * @tparam is_bounded_relu Specified if a fused bounded relu should be applied
91  *
92  * @param[in] in_value Input to be quantized.
93  * @param[in] result_fixedpoint_multiplier Result multiplier parameter
94  * @param[in] result_shift Result shift parameter
95  * @param[in] min_s16 Relu lower bound
96  * @param[in] max_s16 Relu upper bound
97  *
98  * @return Quantized values
99  */
100 template <bool is_bounded_relu>
101 inline int16_t finalize_quantization_int16(int32_t in_value, int result_fixedpoint_multiplier,
102  int32_t result_shift, int16_t min_s16, int16_t max_s16)
103 {
104  if(result_shift < 0)
105  {
106  const int64_t in_64 = static_cast<int64_t>(in_value) * (1 << (-result_shift)) * static_cast<int64_t>(result_fixedpoint_multiplier);
107  in_value = static_cast<int32_t>((in_64 + (1 << 30)) >> 31);
108  }
109  else
110  {
111  // Fixed point multiplication with vector saturating rounding doubling multiply high with scalar
112  const int64_t in_64 = static_cast<int64_t>(in_value) * static_cast<int64_t>(result_fixedpoint_multiplier);
113  // Shift value by result_shift_s32
114  in_value = rounding_divide_by_pow2(static_cast<int32_t>((in_64 + (1 << 30)) >> 31), result_shift);
115  }
116 
117  // Bound the result
118  int16_t out_s16 = static_cast<int16_t>(std::max<int32_t>(-32768, std::min<int32_t>(32767, in_value)));
119 
120  if(is_bounded_relu)
121  {
122  out_s16 = static_cast<int16_t>(std::max(min_s16, std::min(max_s16, out_s16)));
123  }
124 
125  return out_s16;
126 }
127 
128 /** Dequantize a neon vector holding 8 16-bit quantized values.
129  *
130  * @param[in] qv Input values to be dequantized.
131  * @param[in] scale Quantization scale
132  *
133  * @return Dequantized values in a neon vector
134  */
135 inline float32x4x2_t vdequantize_int16(const int16x8_t &qv, float scale)
136 {
137  const float32x4_t vscale = vdupq_n_f32(scale);
138  const float32x4x2_t vdequantized_input =
139  {
140  {
141  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(qv))), vscale),
142  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(qv))), vscale)
143  }
144  };
145  return vdequantized_input;
146 }
147 
148 /** Quantize a neon vector holding 8 floating point values.
149  *
150  * @param[in] qv Input values to be quantized.
151  * @param[in] scale Quantization scale
152  *
153  * @return A neon vector holding the quantized values
154  */
155 inline int16x8_t vquantize_int16(const float32x4x2_t &qv, float scale)
156 {
157  const float32x4_t vinvscale = vdupq_n_f32(1.f / scale);
158 
159  const int32x4x2_t rf =
160  {
161  {
162 #ifdef __aarch64__
163  vcvtnq_s32_f32(vmulq_f32(qv.val[0], vinvscale)),
164  vcvtnq_s32_f32(vmulq_f32(qv.val[1], vinvscale))
165 #else //__aarch64__
166  vcvtq_s32_f32(vmulq_f32(qv.val[0], vinvscale)),
167  vcvtq_s32_f32(vmulq_f32(qv.val[1], vinvscale))
168 #endif //__aarch64__
169  }
170  };
171  return vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1]));
172 }
173 
174 /** Dequantize a neon vector holding 16 16-bit quantized values.
175  *
176  * @param[in] qv Input values to be dequantized.
177  * @param[in] qi Quantization information to be used in the computation.
178  *
179  * @return Dequantized values in a neon vector
180  */
181 inline float32x4x4_t vdequantize(const int16x8x2_t &qv, const UniformQuantizationInfo &qi)
182 {
183  const float scale = qi.scale;
184  const float32x4_t vscale = vdupq_n_f32(scale);
185  const float32x4x4_t vdequantized_input =
186  {
187  {
188  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(qv.val[0]))), vscale),
189  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(qv.val[0]))), vscale),
190  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(qv.val[1]))), vscale),
191  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(qv.val[1]))), vscale),
192  }
193  };
194  return vdequantized_input;
195 }
196 
197 /** Quantize a neon vector holding 16 floating point values.
198  *
199  * @param[in] qv Input values to be quantized.
200  * @param[in] qi Quantization information to be used in the computation.
201  *
202  * @return A neon vector holding the quantized values
203  */
204 inline qsymm16x8x2_t vquantize_qsymm16(const float32x4x4_t &qv, const UniformQuantizationInfo &qi)
205 {
206  const float scale = qi.scale;
207  ARM_COMPUTE_ERROR_ON(scale == 0.f);
208  const float32x4_t vinvscale = vdupq_n_f32(1.f / scale);
209  const int32x4x4_t rf =
210  {
211  {
212 #ifdef __aarch64__
213  vcvtnq_s32_f32(vmulq_f32(qv.val[0], vinvscale)),
214  vcvtnq_s32_f32(vmulq_f32(qv.val[1], vinvscale)),
215  vcvtnq_s32_f32(vmulq_f32(qv.val[2], vinvscale)),
216  vcvtnq_s32_f32(vmulq_f32(qv.val[3], vinvscale)),
217 #else //__aarch64__
218  vcvtq_s32_f32(vmulq_f32(qv.val[0], vinvscale)),
219  vcvtq_s32_f32(vmulq_f32(qv.val[1], vinvscale)),
220  vcvtq_s32_f32(vmulq_f32(qv.val[2], vinvscale)),
221  vcvtq_s32_f32(vmulq_f32(qv.val[3], vinvscale)),
222 #endif //__aarch64__
223  }
224  };
225  const qsymm16x8x2_t res =
226  {
227  vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])),
228  vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])),
229  };
230 
231  return res;
232 }
233 
234 /** Multiply a neon vector using quantized multiplier and shift
235  *
236  * @param[in] input Input vector to mutiply values to be quantized.
237  * @param[in] qmul Quantized multipler
238  * @param[in] shift Left bit shift
239  *
240  * @return A neon vector holding the multiplied value
241  */
242 inline int32x4x2_t multiply_by_quantized_multiplier_2row(int32x4x2_t input, int32_t qmul, int32_t shift)
243 {
244  const auto left_shift = shift > 0 ? shift : 0;
245  const auto right_shift = shift > 0 ? 0 : -shift;
246  const auto one_shifted = 1 << left_shift;
247 
248  int32x4x2_t result;
249  result.val[0] = rounding_divide_by_pow2(vqrdmulhq_n_s32(vmulq_n_s32(input.val[0], one_shifted), qmul), right_shift);
250  result.val[1] = rounding_divide_by_pow2(vqrdmulhq_n_s32(vmulq_n_s32(input.val[1], one_shifted), qmul), right_shift);
251 
252  return result;
253 }
254 
255 } // namespace arm_compute
256 #endif // ARM_COMPUTE_NESYMM_H
int8_t qsymm8_t
8 bit quantized symmetric scalar value
Definition: NESymm.h:33
qsymm16x8x2_t vquantize_qsymm16(const float32x4x4_t &qv, const UniformQuantizationInfo &qi)
Quantize a neon vector holding 16 floating point values.
Definition: NESymm.h:204
float32x4x2_t vdequantize(const uint8x8_t &qv, const UniformQuantizationInfo &qi)
Dequantize a neon vector holding 8 quantized values.
Definition: NEAsymm.h:415
float32x4x2_t vdequantize_int16(const int16x8_t &qv, float scale)
Dequantize a neon vector holding 8 16-bit quantized values.
Definition: NESymm.h:135
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
Quantization info when assuming per layer quantization.
int16x8_t vquantize_int16(const float32x4x2_t &qv, float scale)
Quantize a neon vector holding 8 floating point values.
Definition: NESymm.h:155
Copyright (c) 2017-2021 Arm Limited.
int16x8_t qsymm16x8_t
16 bit quantized symmetric vector with 8 elements
Definition: NESymm.h:36
int16x8_t finalize_quantization_int16(int32x4x2_t &in_s32, int result_fixedpoint_multiplier, int32_t result_shift, int16x8_t min_s16, int16x8_t max_s16)
Performs final quantization step on 8 signed 16-bit elements.
Definition: NESymm.h:52
int32x4x2_t multiply_by_quantized_multiplier_2row(int32x4x2_t input, int32_t qmul, int32_t shift)
Multiply a neon vector using quantized multiplier and shift.
Definition: NESymm.h:242
int32x4_t rounding_divide_by_pow2(int32x4_t x, int32x4_t exponent)
Round to the nearest division by a power-of-two using exponent.
Definition: NEMath.inl:299
int16x8x2_t qsymm16x8x2_t
16 bit quantized symmetric vector with 16 elements
Definition: NESymm.h:37
int16_t qsymm16_t
16 bit quantized symmetric scalar value