1 /*
2 * Copyright (c) 2020-2021 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/core/NEON/kernels/NELogicalKernel.h"
25
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/Validate.h"
28 #include "src/common/utils/Validate.h"
29 #include "src/core/helpers/AutoConfiguration.h"
30 #include "src/core/helpers/WindowHelpers.h"
31
32 #include <arm_neon.h>
33
34 namespace arm_compute
35 {
36 namespace kernels
37 {
38 namespace
39 {
40 static const uint8x8_t c0_x8 = vdup_n_u8(0);
41 static const uint8x16_t c0_x16 = vdupq_n_u8(0);
42 static const uint8x8_t c1_x8 = vdup_n_u8(1);
43 static const uint8x16_t c1_x16 = vdupq_n_u8(1);
44 static const uint32_t step = 16;
45 static const uint32_t half_step = step / 2;
46
neon_logical_and(const uint8_t * src0,const uint8_t * src1,uint8_t * dst,uint32_t len)47 void neon_logical_and(const uint8_t *src0, const uint8_t *src1, uint8_t *dst, uint32_t len)
48 {
49 ARM_COMPUTE_ASSERT_NOT_NULLPTR(src0);
50 ARM_COMPUTE_ASSERT_NOT_NULLPTR(src1);
51 ARM_COMPUTE_ASSERT_NOT_NULLPTR(dst);
52
53 for(; len >= step; len -= step)
54 {
55 vst1q_u8(dst, vandq_u8(vminq_u8(vld1q_u8(src0), c1_x16), vminq_u8(vld1q_u8(src1), c1_x16)));
56 src0 += step;
57 src1 += step;
58 dst += step;
59 }
60
61 for(; len >= half_step; len -= half_step)
62 {
63 vst1_u8(dst, vand_u8(vmin_u8(vld1_u8(src0), c1_x8), vmin_u8(vld1_u8(src1), c1_x8)));
64 src0 += half_step;
65 src1 += half_step;
66 dst += half_step;
67 }
68
69 for(; len > 0; --len)
70 {
71 *dst = (*src0) && (*src1);
72 ++src0;
73 ++src1;
74 ++dst;
75 }
76 }
77
neon_logical_and_broadcast(const uint8_t * src,uint8_t broadcast_val,uint8_t * dst,uint32_t len)78 void neon_logical_and_broadcast(const uint8_t *src, uint8_t broadcast_val, uint8_t *dst, uint32_t len)
79 {
80 ARM_COMPUTE_ASSERT_NOT_NULLPTR(src);
81 ARM_COMPUTE_ASSERT_NOT_NULLPTR(dst);
82
83 const auto broadcast_val_clamped_s = std::min<uint8_t>(broadcast_val, 1);
84 const auto broadcast_val_clamped_x16 = vdupq_n_u8(broadcast_val_clamped_s);
85 const auto broadcast_val_clamped_x8 = vdup_n_u8(broadcast_val_clamped_s);
86
87 for(; len >= step; len -= step)
88 {
89 vst1q_u8(dst, vandq_u8(vminq_u8(vld1q_u8(src), c1_x16), broadcast_val_clamped_x16));
90 src += step;
91 dst += step;
92 }
93
94 for(; len >= half_step; len -= half_step)
95 {
96 vst1_u8(dst, vand_u8(vmin_u8(vld1_u8(src), c1_x8), broadcast_val_clamped_x8));
97 src += half_step;
98 dst += half_step;
99 }
100
101 for(; len > 0; --len)
102 {
103 *dst = (*src) && broadcast_val_clamped_s;
104 ++src;
105 ++dst;
106 }
107 }
108
neon_logical_or(const uint8_t * src0,const uint8_t * src1,uint8_t * dst,uint32_t len)109 void neon_logical_or(const uint8_t *src0, const uint8_t *src1, uint8_t *dst, uint32_t len)
110 {
111 ARM_COMPUTE_ASSERT_NOT_NULLPTR(src0);
112 ARM_COMPUTE_ASSERT_NOT_NULLPTR(src1);
113 ARM_COMPUTE_ASSERT_NOT_NULLPTR(dst);
114
115 for(; len >= step; len -= step)
116 {
117 vst1q_u8(dst, vorrq_u8(vminq_u8(vld1q_u8(src0), c1_x16), vminq_u8(vld1q_u8(src1), c1_x16)));
118 src0 += step;
119 src1 += step;
120 dst += step;
121 }
122
123 for(; len >= half_step; len -= half_step)
124 {
125 vst1_u8(dst, vorr_u8(vmin_u8(vld1_u8(src0), c1_x8), vmin_u8(vld1_u8(src1), c1_x8)));
126 src0 += half_step;
127 src1 += half_step;
128 dst += half_step;
129 }
130
131 for(; len > 0; --len)
132 {
133 *dst = (*src0) || (*src1);
134 ++src0;
135 ++src1;
136 ++dst;
137 }
138 }
139
neon_logical_or_broadcast(const uint8_t * src,uint8_t broadcast_val,uint8_t * dst,uint32_t len)140 void neon_logical_or_broadcast(const uint8_t *src, uint8_t broadcast_val, uint8_t *dst, uint32_t len)
141 {
142 ARM_COMPUTE_ASSERT_NOT_NULLPTR(src);
143 ARM_COMPUTE_ASSERT_NOT_NULLPTR(dst);
144
145 const auto broadcast_val_clamped_s = std::min<uint8_t>(broadcast_val, 1);
146 const auto broadcast_val_clamped_x16 = vdupq_n_u8(broadcast_val_clamped_s);
147 const auto broadcast_val_clamped_x8 = vdup_n_u8(broadcast_val_clamped_s);
148
149 for(; len >= step; len -= step)
150 {
151 vst1q_u8(dst, vorrq_u8(vminq_u8(vld1q_u8(src), c1_x16), broadcast_val_clamped_x16));
152 src += step;
153 dst += step;
154 }
155
156 for(; len >= half_step; len -= half_step)
157 {
158 vst1_u8(dst, vorr_u8(vmin_u8(vld1_u8(src), c1_x8), broadcast_val_clamped_x8));
159 src += half_step;
160 dst += half_step;
161 }
162
163 for(; len > 0; --len)
164 {
165 *dst = (*src) || broadcast_val_clamped_s;
166 ++src;
167 ++dst;
168 }
169 }
170
neon_logical_not(const uint8_t * src,uint8_t * dst,uint32_t len)171 void neon_logical_not(const uint8_t *src, uint8_t *dst, uint32_t len)
172 {
173 ARM_COMPUTE_ASSERT_NOT_NULLPTR(src);
174 ARM_COMPUTE_ASSERT_NOT_NULLPTR(dst);
175
176 for(; len >= step; len -= step)
177 {
178 vst1q_u8(dst, vbslq_u8(vceqq_u8(vld1q_u8(src), c0_x16), c1_x16, c0_x16));
179 src += step;
180 dst += step;
181 }
182
183 for(; len >= half_step; len -= half_step)
184 {
185 vst1_u8(dst, vbsl_u8(vceq_u8(vld1_u8(src), c0_x8), c1_x8, c0_x8));
186 src += half_step;
187 dst += half_step;
188 }
189
190 for(; len > 0; --len)
191 {
192 *dst = !(*src);
193 ++src;
194 ++dst;
195 }
196 }
197
run_unary(const Window & window,const ITensor * src,ITensor * dst)198 void run_unary(const Window &window, const ITensor *src, ITensor *dst)
199 {
200 Window win{ window };
201 win.set(Window::DimX, Window::Dimension(0, 1, 1));
202 const auto len = window.x().end() - window.x().start();
203
204 Iterator in(src, win);
205 Iterator out(dst, win);
206
207 execute_window_loop(win, [&](const Coordinates &)
208 {
209 neon_logical_not(in.ptr(), out.ptr(), len);
210 },
211 in, out);
212 }
213
run_binary(const Window & window,const ITensor * src0,const ITensor * src1,ITensor * dst,LogicalOperation op)214 void run_binary(const Window &window, const ITensor *src0, const ITensor *src1, ITensor *dst, LogicalOperation op)
215 {
216 Window src0_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
217 Window src1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
218
219 Window win{ window };
220 win.set(Window::DimX, Window::Dimension(0, 1, 1));
221
222 const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
223 const auto len = window.x().end() - window.x().start();
224
225 if(is_broadcast_across_x)
226 {
227 using LogicalBroadcastUKernelPtr = std::add_pointer<void(const uint8_t *, uint8_t, uint8_t *, uint32_t)>::type;
228 LogicalBroadcastUKernelPtr logical_func = op == LogicalOperation::Or ? &neon_logical_or_broadcast : &neon_logical_and_broadcast;
229
230 const bool is_broadcast_input_1 = src1_win.x().step() == 0;
231 Window broadcast_win = is_broadcast_input_1 ? src1_win : src0_win;
232 Window non_broadcast_win = !is_broadcast_input_1 ? src1_win : src0_win;
233 const ITensor *broadcast_tensor = is_broadcast_input_1 ? src1 : src0;
234 const ITensor *non_broadcast_tensor = !is_broadcast_input_1 ? src1 : src0;
235 non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
236
237 Iterator broadcast_in(broadcast_tensor, broadcast_win);
238 Iterator non_broadcast_in(non_broadcast_tensor, non_broadcast_win);
239 Iterator out(dst, win);
240
241 execute_window_loop(win, [&](const Coordinates &)
242 {
243 const uint8_t broadcast_value = *broadcast_in.ptr();
244 logical_func(non_broadcast_in.ptr(), broadcast_value, out.ptr(), len);
245
246 },
247 broadcast_in, non_broadcast_in, out);
248 }
249 else
250 {
251 using LogicalUKernelPtr = std::add_pointer<void(const uint8_t *, const uint8_t *, uint8_t *, uint32_t)>::type;
252 LogicalUKernelPtr logical_func = op == LogicalOperation::Or ? &neon_logical_or : &neon_logical_and;
253
254 src0_win.set(Window::DimX, Window::Dimension(0, 1, 1));
255 src1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
256
257 Iterator in0(src0, src0_win);
258 Iterator in1(src1, src1_win);
259 Iterator out(dst, win);
260 execute_window_loop(win, [&](const Coordinates &)
261 {
262 logical_func(in0.ptr(), in1.ptr(), out.ptr(), len);
263 },
264 in0, in1, out);
265 }
266 }
267 } // namespace
name() const268 const char *NELogicalKernel::name() const
269 {
270 return "NELogicalKernel";
271 }
272
configure(const ITensorInfo * input1,const ITensorInfo * input2,ITensorInfo * output,LogicalOperation op)273 void NELogicalKernel::configure(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output, LogicalOperation op)
274 {
275 ARM_COMPUTE_ERROR_ON_NULLPTR(input1, output);
276 ARM_COMPUTE_ERROR_THROW_ON(validate(input1, input2, output, op));
277
278 _op = op;
279
280 Window win = calculate_max_window(*input1, Steps());
281 TensorShape out_shape = input1->tensor_shape();
282 if(op != LogicalOperation::Not)
283 {
284 ARM_COMPUTE_ERROR_ON_NULLPTR(input2);
285 out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape());
286 win = calculate_max_window(out_shape, Steps());
287 }
288 ICPPKernel::configure(win);
289
290 // Auto initialize if empty
291 set_shape_if_empty(*output, out_shape);
292 set_data_type_if_unknown(*output, input1->data_type());
293 }
294
validate(const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * output,LogicalOperation op)295 Status NELogicalKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, LogicalOperation op)
296 {
297 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8);
298 ARM_COMPUTE_RETURN_ERROR_ON(op == LogicalOperation::Unknown);
299
300 TensorShape out_shape = input1->tensor_shape();
301 if(op != LogicalOperation::Not)
302 {
303 out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape());
304 ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible");
305 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2);
306 }
307
308 // Checks performed when output is configured
309 if((output != nullptr) && (output->total_size() != 0))
310 {
311 ARM_COMPUTE_RETURN_ERROR_ON(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0));
312 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, output);
313 }
314
315 return Status{};
316 }
317
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)318 void NELogicalKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
319 {
320 ARM_COMPUTE_UNUSED(info);
321 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
322 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
323 ARM_COMPUTE_ERROR_ON(tensors.empty());
324
325 const ITensor *src0 = tensors.get_const_tensor(TensorType::ACL_SRC_0);
326 const ITensor *src1 = tensors.get_const_tensor(TensorType::ACL_SRC_1);
327 ITensor *dst = tensors.get_tensor(TensorType::ACL_DST);
328
329 if(_op == LogicalOperation::Not)
330 {
331 run_unary(window, src0, dst);
332 }
333 else
334 {
335 run_binary(window, src0, src1, dst, _op);
336 }
337 }
338 } // namespace kernels
339 } // namespace arm_compute
340