Compute Library
 21.02
PixelWiseMultiplication.cpp
Go to the documentation of this file.
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  */
25 
27 
28 namespace arm_compute
29 {
30 namespace test
31 {
32 namespace validation
33 {
34 namespace reference
35 {
36 template <class T>
37 struct is_floating_point
38  : std::integral_constant < bool,
39  std::is_same<float, typename std::remove_cv<T>::type>::value || std::is_same<half_float::half, typename std::remove_cv<T>::type>::value
40  || std::is_same<double, typename std::remove_cv<T>::type>::value || std::is_same<long double, typename std::remove_cv<T>::type>::value >
41 {
42 };
43 
44 namespace
45 {
46 constexpr float scale1_constant = 1.f;
47 
48 /** Compute the result of `src1 * src2 * scale`. The result type always matches the type of @p src2.
49  *
50  * @param[in] src1 An input value. Data types supported: U8/S16/F16/F32.
51  * @param[in] src2 An input value. Data types supported: same as @p src1.
52  * @param[in] scale Scale to apply after multiplication.
53  * Scale must be positive and its value must be either 1/255 or 1/2^n where n is between 0 and 15.
54  * @param[in] convert_policy Overflow policy. Supported overflow policies: Wrap, Saturate
55  * @param[in] rounding_policy Rounding policy. Supported rounding modes: to zero, to nearest even.
56  */
57 template <typename T1, typename T2, typename T3>
58 T3 mul(const T1 src1, const T2 src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy)
59 {
60  using intermediate_type = typename common_promoted_signed_type<T1, T2, T3>::intermediate_type;
61 
62  const double val = static_cast<intermediate_type>(src1) * static_cast<intermediate_type>(src2) * static_cast<double>(scale);
63 
64  if(is_floating_point<T3>::value)
65  {
66  const auto result = static_cast<T3>(val);
67 
68  return result;
69  }
70  else
71  {
72  double rounded_val = 0;
73  switch(rounding_policy)
74  {
76  rounded_val = support::cpp11::trunc(val);
77  break;
79  rounded_val = round_half_up(val);
80  break;
82  rounded_val = round_half_even(val);
83  break;
84  default:
85  ARM_COMPUTE_ERROR("Unsupported rounding policy");
86  }
87 
88  const auto result = static_cast<T3>((convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T3>(rounded_val) : rounded_val);
89 
90  return result;
91  }
92 }
93 
94 template <>
95 int32_t mul(const int32_t src1, const int32_t src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy)
96 {
97  const int64_t intermediate_val = static_cast<int64_t>(src1) * static_cast<int64_t>(src2);
98 
99  if(std::abs(scale - scale1_constant) < 0.00001f)
100  {
101  // Use bit-accurate integer arithmetic for scale == 1
102  // Apply conversion
103  if(convert_policy == ConvertPolicy::SATURATE)
104  {
105  return saturate_cast<int32_t>(intermediate_val);
106  }
107  else
108  {
109  // Correct wrapping behaviour for int32_t
110  const auto i32_hi = static_cast<int64_t>(std::numeric_limits<int32_t>::max());
111  const auto i32_lo = static_cast<int64_t>(std::numeric_limits<int32_t>::lowest());
112  const auto i32_wi = static_cast<int64_t>(1) << 32;
113  int64_t wrapped_rounded_val = intermediate_val - i32_wi * static_cast<int64_t>(support::cpp11::trunc(static_cast<double>(intermediate_val) / i32_wi));
114  if(wrapped_rounded_val <= i32_hi)
115  {
116  return static_cast<int32_t>(wrapped_rounded_val);
117  }
118  else
119  {
120  // Values beyond i32_hi wrap around to negatives
121  return static_cast<int32_t>((wrapped_rounded_val - i32_hi) + i32_lo - 1);
122  }
123  }
124  }
125  else
126  {
127  // Use double arithmetic for scale != 1; may not be bit-accurate
128  // Apply scaling
129  // scale == 1 / 2^scale_exponent
130  int scale_exponent = 0;
131  std::frexp(scale, &scale_exponent);
132  // Store the positive exponent. We know that we compute 1/2^n
133  // Additionally we need to subtract 1 to compensate that frexp used a mantissa of 0.5
134  scale_exponent = std::abs(scale_exponent - 1);
135  const double scale_inv = static_cast<int64_t>(1) << scale_exponent;
136  const double val = intermediate_val / scale_inv;
137  // Apply rounding
138  double rounded_val = 0;
139  switch(rounding_policy)
140  {
142  rounded_val = support::cpp11::trunc(val);
143  break;
145  rounded_val = round_half_up(val);
146  break;
148  rounded_val = round_half_even(val);
149  break;
150  default:
151  ARM_COMPUTE_ERROR("Unsupported rounding policy");
152  }
153  // Apply conversion
154  if(convert_policy == ConvertPolicy::SATURATE)
155  {
156  return saturate_cast<int32_t>(rounded_val);
157  }
158  else
159  {
160  // Correct wrapping behaviour for int32_t
161  const auto i32_hi = static_cast<double>(std::numeric_limits<int32_t>::max());
162  const auto i32_lo = static_cast<double>(std::numeric_limits<int32_t>::lowest());
163  const auto i32_wi = static_cast<double>(static_cast<int64_t>(1) << 32);
164  double wrapped_rounded_val = rounded_val - i32_wi * std::floor(rounded_val / i32_wi);
165  if(wrapped_rounded_val <= i32_hi)
166  {
167  return static_cast<int32_t>(wrapped_rounded_val);
168  }
169  else
170  {
171  // Values beyond i32_hi wrap around to negatives
172  return static_cast<int32_t>((wrapped_rounded_val - i32_hi) + i32_lo - 1);
173  }
174  }
175  }
176 }
177 
178 template <size_t dim>
179 struct BroadcastUnroll
180 {
181  template <typename T1, typename T2, typename T3>
182  static void unroll(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, SimpleTensor<T3> &dst,
183  float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy,
184  Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst)
185  {
186  const bool src1_is_broadcast = (src1.shape()[dim - 1] != dst.shape()[dim - 1]);
187  const bool src2_is_broadcast = (src2.shape()[dim - 1] != dst.shape()[dim - 1]);
188 
189  id_src1.set(dim - 1, 0);
190  id_src2.set(dim - 1, 0);
191  id_dst.set(dim - 1, 0);
192 
193  for(size_t i = 0; i < dst.shape()[dim - 1]; ++i, ++id_dst[dim - 1])
194  {
195  BroadcastUnroll < dim - 1 >::unroll(src1, src2, dst, scale, convert_policy, rounding_policy, id_src1, id_src2, id_dst);
196 
197  id_src1[dim - 1] += !src1_is_broadcast;
198  id_src2[dim - 1] += !src2_is_broadcast;
199  }
200  }
201 };
202 
203 template <>
204 struct BroadcastUnroll<0>
205 {
206  template <typename T1, typename T2, typename T3>
207  static void unroll(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, SimpleTensor<T3> &dst,
208  float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy,
209  Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst)
210  {
211  dst[coord2index(dst.shape(), id_dst)] = mul<T1, T2, T3>(src1[coord2index(src1.shape(), id_src1)], src2[coord2index(src2.shape(), id_src2)], scale, convert_policy, rounding_policy);
212  }
213 };
214 } // namespace
215 
216 template <typename T1, typename T2, typename T3>
217 SimpleTensor<T3> pixel_wise_multiplication(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy,
218  DataType dt_out, const QuantizationInfo &qout)
219 {
220  ARM_COMPUTE_UNUSED(qout);
221 
223 
224  if(scale < 0)
225  {
226  ARM_COMPUTE_ERROR("Scale of pixel-wise multiplication must be non-negative");
227  }
228 
229  Coordinates id_src1{};
230  Coordinates id_src2{};
231  Coordinates id_dst{};
232 
233  BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, scale, convert_policy, rounding_policy, id_src1, id_src2, id_dst);
234 
235  return dst;
236 }
237 
238 template <>
240  DataType dt_out, const QuantizationInfo &qout)
241 {
242  SimpleTensor<uint8_t> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), dt_out, 1, qout);
243 
244  if(src1.data_type() == DataType::QASYMM8 && src2.data_type() == DataType::QASYMM8)
245  {
248  SimpleTensor<float> dst_tmp = pixel_wise_multiplication<float, float, float>(src1_tmp, src2_tmp, scale, convert_policy, rounding_policy, DataType::F32, qout);
249  dst = convert_to_asymmetric<uint8_t>(dst_tmp, qout);
250  }
251  else
252  {
253  if(scale < 0)
254  {
255  ARM_COMPUTE_ERROR("Scale of pixel-wise multiplication must be non-negative");
256  }
257 
258  Coordinates id_src1{};
259  Coordinates id_src2{};
260  Coordinates id_dst{};
261  BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, scale, convert_policy, rounding_policy, id_src1, id_src2, id_dst);
262  }
263  return dst;
264 }
265 
266 template <>
268  DataType dt_out, const QuantizationInfo &qout)
269 {
270  SimpleTensor<int16_t> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), dt_out, 1, qout);
271 
272  if(src1.data_type() == DataType::QASYMM8 && src2.data_type() == DataType::QASYMM8)
273  {
276  SimpleTensor<float> dst_tmp = pixel_wise_multiplication<float, float, float>(src1_tmp, src2_tmp, scale, convert_policy, rounding_policy, DataType::F32, qout);
277  dst = convert_to_symmetric<int16_t>(dst_tmp, qout);
278  }
279  else
280  {
281  if(scale < 0)
282  {
283  ARM_COMPUTE_ERROR("Scale of pixel-wise multiplication must be non-negative");
284  }
285 
286  Coordinates id_src1{};
287  Coordinates id_src2{};
288  Coordinates id_dst{};
289  BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, scale, convert_policy, rounding_policy, id_src1, id_src2, id_dst);
290  }
291  return dst;
292 }
293 
294 template <>
295 SimpleTensor<int8_t> pixel_wise_multiplication(const SimpleTensor<int8_t> &src1, const SimpleTensor<int8_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy,
296  DataType dt_out, const QuantizationInfo &qout)
297 {
298  SimpleTensor<int8_t> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), dt_out, 1, qout);
299 
301  {
304  SimpleTensor<float> dst_tmp = pixel_wise_multiplication<float, float, float>(src1_tmp, src2_tmp, scale, convert_policy, rounding_policy, DataType::F32, qout);
305  dst = convert_to_asymmetric<int8_t>(dst_tmp, qout);
306  }
307  else
308  {
309  if(scale < 0)
310  {
311  ARM_COMPUTE_ERROR("Scale of pixel-wise multiplication must be non-negative");
312  }
313 
314  Coordinates id_src1{};
315  Coordinates id_src2{};
316  Coordinates id_dst{};
317  BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, scale, convert_policy, rounding_policy, id_src1, id_src2, id_dst);
318  }
319  return dst;
320 }
321 
322 template <>
324  DataType dt_out, const QuantizationInfo &qout)
325 {
326  SimpleTensor<int16_t> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), dt_out, 1, qout);
327 
328  if(src1.data_type() == DataType::QSYMM16 && src2.data_type() == DataType::QSYMM16)
329  {
330  SimpleTensor<float> src1_tmp = convert_from_symmetric<int16_t>(src1);
331  SimpleTensor<float> src2_tmp = convert_from_symmetric<int16_t>(src2);
332  SimpleTensor<float> dst_tmp = pixel_wise_multiplication<float, float, float>(src1_tmp, src2_tmp, scale, convert_policy, rounding_policy, DataType::F32, qout);
333  dst = convert_to_symmetric<int16_t>(dst_tmp, qout);
334  }
335  else
336  {
337  if(scale < 0)
338  {
339  ARM_COMPUTE_ERROR("Scale of pixel-wise multiplication must be non-negative");
340  }
341 
342  Coordinates id_src1{};
343  Coordinates id_src2{};
344  Coordinates id_dst{};
345  BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, scale, convert_policy, rounding_policy, id_src1, id_src2, id_dst);
346  }
347  return dst;
348 }
349 // *INDENT-OFF*
350 // clang-format off
351 template SimpleTensor<int16_t> pixel_wise_multiplication(const SimpleTensor<uint8_t> &src1, const SimpleTensor<int16_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout);
352 template SimpleTensor<int32_t> pixel_wise_multiplication(const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout);
353 template SimpleTensor<int32_t> pixel_wise_multiplication(const SimpleTensor<int32_t> &src1, const SimpleTensor<int32_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout);
354 template SimpleTensor<float> pixel_wise_multiplication(const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout);
355 template SimpleTensor<half_float::half> pixel_wise_multiplication(const SimpleTensor<half_float::half> &src1, const SimpleTensor<half_float::half> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout);
356 // clang-format on
357 // *INDENT-ON*
358 } // namespace reference
359 } // namespace validation
360 } // namespace test
361 } // namespace arm_compute
T trunc(T value)
Truncate floating-point value.
quantized, symmetric fixed-point 16-bit number
Rounds to nearest value; half rounds away from zero.
#define ARM_COMPUTE_ERROR(msg)
Print the given message then throw an std::runtime_error.
Definition: Error.h:352
SimpleTensor< T3 > pixel_wise_multiplication(const SimpleTensor< T1 > &src1, const SimpleTensor< T2 > &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, DataType dt_out, const QuantizationInfo &qout)
1 channel, 1 F32 per channel
DataType data_type() const override
Data type of the tensor.
Definition: SimpleTensor.h:357
static TensorShape broadcast_shape(const Shapes &... shapes)
If shapes are broadcast compatible, return the broadcasted shape.
Definition: TensorShape.h:211
SimpleTensor< float > convert_from_asymmetric(const SimpleTensor< uint8_t > &src)
Definition: Helpers.cpp:112
TensorShape shape() const override
Shape of the tensor.
Definition: SimpleTensor.h:320
Copyright (c) 2017-2021 Arm Limited.
int coord2index(const TensorShape &shape, const Coordinates &coord)
Linearise the given coordinate.
Definition: Utils.h:489
T round_half_even(T value, T epsilon=std::numeric_limits< T >::epsilon())
Round floating-point value with half value rounding to nearest even.
Definition: Utils.h:88
Quantization information.
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
quantized, asymmetric fixed-point 8-bit number unsigned
Coordinates of an item.
Definition: Coordinates.h:37
RoundingPolicy
Rounding method.
Definition: Rounding.h:30
Simple tensor object that stores elements in a consecutive chunk of memory.
Definition: SimpleTensor.h:58
Rounds to nearest value; half rounds to nearest even.
typename traits::make_signed_conditional_t< promoted_type >::type intermediate_type
Intermediate type.
Definition: Utils.h:437
T saturate_cast(T val)
Saturate a value of type T against the numeric limits of type U.
Definition: Utils.h:414
quantized, asymmetric fixed-point 8-bit number signed
T round_half_up(T value)
Round floating-point value with half value rounding to positive infinity.
Definition: Utils.h:75
SimpleTensor< T > scale(const SimpleTensor< T > &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, T constant_border_value, SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners)
Definition: Scale.cpp:185
DataType
Available data types.
Definition: Types.h:77
Truncates the least significant values that are lost in operations.
ConvertPolicy
Policy to handle overflow.
Definition: Types.h:385