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/CpuDequantizeKernel.h"
25
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/Utils.h"
29 #include "arm_compute/core/Validate.h"
30 #include "arm_compute/core/Window.h"
31 #include "src/core/CPP/Validate.h"
32 #include "src/core/NEON/NEAsymm.h"
33 #include "src/core/NEON/NESymm.h"
34 #include "src/core/NEON/wrapper/wrapper.h"
35 #include "src/core/helpers/AutoConfiguration.h"
36 #include "src/core/helpers/WindowHelpers.h"
37
38 #include <arm_neon.h>
39
40 namespace arm_compute
41 {
42 namespace cpu
43 {
44 namespace kernels
45 {
46 namespace
47 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst)48 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst)
49 {
50 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
51 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8, DataType::QSYMM16);
52
53 if(dst->tensor_shape().total_size() > 0)
54 {
55 ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(dst);
56 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::F16, DataType::F32);
57 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
58 }
59
60 return Status{};
61 }
62
63 template <typename T>
store_result(T * ptr,const float32x4x4_t & v)64 inline void store_result(T *ptr, const float32x4x4_t &v)
65 {
66 ARM_COMPUTE_UNUSED(ptr, v);
67 }
68
69 template <>
store_result(float * ptr,const float32x4x4_t & v)70 inline void store_result<float>(float *ptr, const float32x4x4_t &v)
71 {
72 wrapper::vstore(ptr, v.val[0]);
73 wrapper::vstore(ptr + 4, v.val[1]);
74 wrapper::vstore(ptr + 8, v.val[2]);
75 wrapper::vstore(ptr + 12, v.val[3]);
76 }
77
78 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
79 template <>
store_result(float16_t * ptr,const float32x4x4_t & v)80 inline void store_result<float16_t>(float16_t *ptr, const float32x4x4_t &v)
81 {
82 wrapper::vstore(ptr, vcombine_f16(vcvt_f16_f32(v.val[0]), vcvt_f16_f32(v.val[1])));
83 wrapper::vstore(ptr + 8, vcombine_f16(vcvt_f16_f32(v.val[2]), vcvt_f16_f32(v.val[3])));
84 }
85 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
86
87 template <typename T>
store_result(T * ptr,const float32x4x2_t & v)88 inline void store_result(T *ptr, const float32x4x2_t &v)
89 {
90 ARM_COMPUTE_UNUSED(ptr, v);
91 }
92
93 template <>
store_result(float * ptr,const float32x4x2_t & v)94 inline void store_result<float>(float *ptr, const float32x4x2_t &v)
95 {
96 wrapper::vstore(ptr, v.val[0]);
97 wrapper::vstore(ptr + 4, v.val[1]);
98 }
99
100 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
101 template <>
store_result(float16_t * ptr,const float32x4x2_t & v)102 inline void store_result<float16_t>(float16_t *ptr, const float32x4x2_t &v)
103 {
104 wrapper::vstore(ptr, vcombine_f16(vcvt_f16_f32(v.val[0]), vcvt_f16_f32(v.val[1])));
105 }
106 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
107
108 template <typename TOut, typename TIn>
run_dequantization_qasymm8(const ITensor * input,ITensor * output,const Window & window)109 void run_dequantization_qasymm8(const ITensor *input, ITensor *output, const Window &window)
110 {
111 const UniformQuantizationInfo &qinfo = input->info()->quantization_info().uniform();
112 const float scale = qinfo.scale;
113 const int32_t offset = qinfo.offset;
114
115 const int window_step_x = 16;
116 const auto window_start_x = static_cast<int>(window.x().start());
117 const auto window_end_x = static_cast<int>(window.x().end());
118
119 // Collapse window and reset first dimension to handle tail calculations manually
120 Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
121 win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
122
123 // Create iterators
124 Iterator in(input, win_collapsed);
125 Iterator out(output, win_collapsed);
126
127 execute_window_loop(win_collapsed, [&](const Coordinates &)
128 {
129 const auto in_ptr = reinterpret_cast<const TIn *>(in.ptr());
130 const auto out_ptr = reinterpret_cast<TOut *>(out.ptr());
131
132 int x = window_start_x;
133 for(; x <= (window_end_x - window_step_x); x += window_step_x)
134 {
135 const auto vin = wrapper::vloadq(in_ptr + x);
136 const auto vdeq = vdequantize(vin, scale, offset);
137
138 store_result(reinterpret_cast<TOut *>(out_ptr + x), vdeq);
139 }
140
141 // Compute left-over elements
142 for(; x < window_end_x; ++x)
143 {
144 auto val = *(in_ptr + x);
145 *(out_ptr + x) = static_cast<TOut>(Qasymm8QuantizationHelper<TIn>::dequantize(val, qinfo));
146 }
147 },
148 in, out);
149 }
150
151 template <typename T>
run_dequantization_qsymm8_per_channel_nchw(const ITensor * input,ITensor * output,const Window & window)152 void run_dequantization_qsymm8_per_channel_nchw(const ITensor *input, ITensor *output, const Window &window)
153 {
154 const auto scale = input->info()->quantization_info().scale();
155
156 const int window_step_x = 16;
157 const auto window_start_x = static_cast<int>(window.x().start());
158 const auto window_end_x = static_cast<int>(window.x().end());
159
160 // Reset first dimension to handle tail calculations manually
161 Window win(window);
162 win.set(Window::DimX, Window::Dimension(0, 1, 1));
163
164 // Create iterators
165 Iterator in(input, win);
166 Iterator out(output, win);
167
168 execute_window_loop(win, [&](const Coordinates & id)
169 {
170 const auto in_ptr = reinterpret_cast<const int8_t *>(in.ptr());
171 const auto out_ptr = reinterpret_cast<T *>(out.ptr());
172
173 int x = window_start_x;
174 for(; x <= (window_end_x - window_step_x); x += window_step_x)
175 {
176 const auto vin = wrapper::vloadq(in_ptr + x);
177 const auto vdeq = vdequantize(vin, scale[id.z()]);
178
179 store_result<T>(reinterpret_cast<T *>(out_ptr + x), vdeq);
180 }
181
182 // Compute left-over elements
183 for(; x < window_end_x; ++x)
184 {
185 int8_t val = *(in_ptr + x);
186 *(out_ptr + x) = static_cast<T>(dequantize(val, scale[id.z()]));
187 }
188 },
189 in, out);
190 }
191
192 template <typename T>
run_dequantization_qsymm8_per_channel_nhwc(const ITensor * input,ITensor * output,const Window & window)193 void run_dequantization_qsymm8_per_channel_nhwc(const ITensor *input, ITensor *output, const Window &window)
194 {
195 const auto scale = input->info()->quantization_info().scale();
196
197 const int window_step_x = 16;
198 const auto window_start_x = static_cast<int>(window.x().start());
199 const auto window_end_x = static_cast<int>(window.x().end());
200
201 // Reset first dimension to handle tail calculations manually
202 Window win(window);
203 win.set(Window::DimX, Window::Dimension(0, 1, 1));
204
205 // Create iterators
206 Iterator in(input, win);
207 Iterator out(output, win);
208
209 execute_window_loop(win, [&](const Coordinates &)
210 {
211 const auto in_ptr = reinterpret_cast<const int8_t *>(in.ptr());
212 const auto out_ptr = reinterpret_cast<T *>(out.ptr());
213
214 int x = window_start_x;
215 for(; x <= (window_end_x - window_step_x); x += window_step_x)
216 {
217 const float32x4x4_t vscale =
218 {
219 {
220 scale[x + 0], scale[x + 1], scale[x + 2], scale[x + 3],
221 scale[x + 4], scale[x + 5], scale[x + 6], scale[x + 7],
222 scale[x + 8], scale[x + 9], scale[x + 10], scale[x + 11],
223 scale[x + 12], scale[x + 13], scale[x + 14], scale[x + 15]
224 }
225 };
226 const auto vin = wrapper::vloadq(in_ptr + x);
227 const auto vdeq = vdequantize(vin, vscale);
228
229 store_result<T>(reinterpret_cast<T *>(out_ptr + x), vdeq);
230 }
231
232 // Compute left-over elements
233 for(; x < window_end_x; ++x)
234 {
235 int8_t val = *(in_ptr + x);
236 *(out_ptr + x) = static_cast<T>(dequantize(val, scale[x]));
237 }
238 },
239 in, out);
240 }
241
242 template <typename T>
run_dequantization_qsymm8(const ITensor * input,ITensor * output,const Window & window)243 void run_dequantization_qsymm8(const ITensor *input, ITensor *output, const Window &window)
244 {
245 const UniformQuantizationInfo &qinfo = input->info()->quantization_info().uniform();
246 const float scale = qinfo.scale;
247
248 const int window_step_x = 16;
249 const auto window_start_x = static_cast<int>(window.x().start());
250 const auto window_end_x = static_cast<int>(window.x().end());
251
252 // Collapse window and reset first dimension to handle tail calculations manually
253 Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
254 win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
255
256 // Create iterators
257 Iterator in(input, win_collapsed);
258 Iterator out(output, win_collapsed);
259
260 execute_window_loop(win_collapsed, [&](const Coordinates &)
261 {
262 const auto in_ptr = reinterpret_cast<const int8_t *>(in.ptr());
263 const auto out_ptr = reinterpret_cast<T *>(out.ptr());
264
265 int x = window_start_x;
266 for(; x <= (window_end_x - window_step_x); x += window_step_x)
267 {
268 const auto vin = wrapper::vloadq(in_ptr + x);
269 const auto vdeq = vdequantize(vin, scale);
270
271 store_result<T>(reinterpret_cast<T *>(out_ptr + x), vdeq);
272 }
273
274 // Compute left-over elements
275 for(; x < window_end_x; ++x)
276 {
277 int8_t val = *(in_ptr + x);
278 *(out_ptr + x) = static_cast<T>(dequantize(val, scale));
279 }
280 },
281 in, out);
282 }
283
284 template <typename T>
run_dequantization_qsymm16(const ITensor * input,ITensor * output,const Window & window)285 void run_dequantization_qsymm16(const ITensor *input, ITensor *output, const Window &window)
286 {
287 const UniformQuantizationInfo &qinfo = input->info()->quantization_info().uniform();
288 const float scale = qinfo.scale;
289
290 const int window_step_x = 8;
291 const auto window_start_x = static_cast<int>(window.x().start());
292 const auto window_end_x = static_cast<int>(window.x().end());
293
294 // Collapse window and reset first dimension to handle tail calculations manually
295 Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
296 win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
297
298 // Create iterators
299 Iterator in(input, win_collapsed);
300 Iterator out(output, win_collapsed);
301
302 execute_window_loop(win_collapsed, [&](const Coordinates &)
303 {
304 const auto in_ptr = reinterpret_cast<const int16_t *>(in.ptr());
305 const auto out_ptr = reinterpret_cast<T *>(out.ptr());
306
307 int x = window_start_x;
308 for(; x <= (window_end_x - window_step_x); x += window_step_x)
309 {
310 const auto vin = wrapper::vloadq(in_ptr + x);
311 const auto vdeq = vdequantize_int16(vin, scale);
312
313 store_result<T>(reinterpret_cast<T *>(out_ptr + x), vdeq);
314 }
315
316 // Compute left-over elements
317 for(; x < window_end_x; ++x)
318 {
319 int16_t val = *(in_ptr + x);
320 *(out_ptr + x) = static_cast<T>(dequantize_qsymm16(val, scale));
321 }
322 },
323 in, out);
324 }
325
326 template <typename T>
run_dequantization_core(const ITensor * input,ITensor * output,const Window & window)327 void run_dequantization_core(const ITensor *input, ITensor *output, const Window &window)
328 {
329 switch(input->info()->data_type())
330 {
331 case DataType::QASYMM8:
332 run_dequantization_qasymm8<T, uint8_t>(input, output, window);
333 break;
334 case DataType::QASYMM8_SIGNED:
335 run_dequantization_qasymm8<T, int8_t>(input, output, window);
336 break;
337 case DataType::QSYMM8_PER_CHANNEL:
338 input->info()->data_layout() == DataLayout::NHWC ? run_dequantization_qsymm8_per_channel_nhwc<T>(input, output, window) : run_dequantization_qsymm8_per_channel_nchw<T>(input, output, window);
339 break;
340 case DataType::QSYMM8:
341 run_dequantization_qsymm8<T>(input, output, window);
342 break;
343 case DataType::QSYMM16:
344 run_dequantization_qsymm16<T>(input, output, window);
345 break;
346 default:
347 ARM_COMPUTE_ERROR("Unsupported data type.");
348 }
349 }
350 } // namespace
351
configure(const ITensorInfo * src,ITensorInfo * dst)352 void CpuDequantizeKernel::configure(const ITensorInfo *src, ITensorInfo *dst)
353 {
354 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst));
355
356 // Configure kernel window
357 Window win = calculate_max_window(*src, Steps());
358
359 // Output tensor auto initialization if not yet initialized
360 auto_init_if_empty(*dst, src->tensor_shape(), 1, DataType::F32);
361
362 ICpuKernel::configure(win);
363 }
364
validate(const ITensorInfo * src,const ITensorInfo * dst)365 Status CpuDequantizeKernel::validate(const ITensorInfo *src, const ITensorInfo *dst)
366 {
367 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst));
368 return Status{};
369 }
370
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)371 void CpuDequantizeKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
372 {
373 ARM_COMPUTE_UNUSED(info);
374 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
375 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
376
377 const auto src = tensors.get_const_tensor(TensorType::ACL_SRC);
378 auto dst = tensors.get_tensor(TensorType::ACL_DST);
379
380 switch(dst->info()->data_type())
381 {
382 case DataType::F32:
383 run_dequantization_core<float>(src, dst, window);
384 break;
385 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
386 case DataType::F16:
387 run_dequantization_core<float16_t>(src, dst, window);
388 break;
389 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
390 default:
391 ARM_COMPUTE_ERROR("Unsupported data type.");
392 }
393 }
name() const394 const char *CpuDequantizeKernel::name() const
395 {
396 return "CpuDequantizeKernel";
397 }
398 } // namespace kernels
399 } // namespace cpu
400 } // namespace arm_compute
401