xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/elementwise_binary/generic/sve2/impl.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2021-2022 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 SRC_CORE_SVE_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H
25 #define SRC_CORE_SVE_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H
26 
27 #include "src/cpu/kernels/elementwise_binary/generic/sve/impl.h"
28 namespace arm_compute
29 {
30 namespace cpu
31 {
32 using namespace arm_compute::wrapper;
33 
load_quantized(const int8_t * ptr,svbool_t pg,const svint32_t & offset,const svfloat32_t & scale)34 inline svfloat32x4_t load_quantized(const int8_t *ptr, svbool_t pg, const svint32_t &offset, const svfloat32_t &scale)
35 {
36     auto x = svld1(pg, ptr);
37 
38     const auto widened = svcreate4(
39                              svmovlb(svmovlb(x)),
40                              svmovlt(svmovlb(x)),
41                              svmovlb(svmovlt(x)),
42                              svmovlt(svmovlt(x)));
43 
44     pg = svptrue_b8();
45 
46     return svcreate4(
47                svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 0), offset)), scale),
48                svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 1), offset)), scale),
49                svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 2), offset)), scale),
50                svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 3), offset)), scale));
51 }
52 
load_quantized(const uint8_t * ptr,svbool_t pg,const svint32_t & offset,const svfloat32_t & scale)53 inline svfloat32x4_t load_quantized(const uint8_t *ptr, svbool_t pg, const svint32_t &offset, const svfloat32_t &scale)
54 {
55     auto x = svld1(pg, ptr);
56 
57     //vprint(x);
58 
59     const auto widened = svcreate4(
60                              svmovlb(svmovlb(x)),
61                              svmovlt(svmovlb(x)),
62                              svmovlb(svmovlt(x)),
63                              svmovlt(svmovlt(x)));
64 
65     pg = svptrue_b8();
66 
67     return svcreate4(
68                svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 0)), offset)), scale),
69                svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 1)), offset)), scale),
70                svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 2)), offset)), scale),
71                svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 3)), offset)), scale));
72 }
73 
store_quantized(uint8_t * ptr,svbool_t pg,svfloat32x4_t data,const svint32_t & offset,const svfloat32_t & inv_scale)74 inline void store_quantized(uint8_t *ptr, svbool_t pg, svfloat32x4_t data, const svint32_t &offset, const svfloat32_t &inv_scale)
75 {
76     const auto quantized = svcreate4(
77                                svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 0), inv_scale))), offset),
78                                svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 1), inv_scale))), offset),
79                                svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 2), inv_scale))), offset),
80                                svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 3), inv_scale))), offset));
81 
82     const auto narrowed_bottom = svqxtunt(svqxtunb(svget4(quantized, 0)), svget4(quantized, 1));
83     const auto narrowed_top    = svqxtunt(svqxtunb(svget4(quantized, 2)), svget4(quantized, 3));
84     const auto narrowed        = svqxtnt(svqxtnb(narrowed_bottom), narrowed_top);
85     svst1(pg, ptr, narrowed);
86 }
87 
store_quantized(int8_t * ptr,svbool_t pg,svfloat32x4_t data,const svint32_t & offset,const svfloat32_t & inv_scale)88 inline void store_quantized(int8_t *ptr, svbool_t pg, svfloat32x4_t data, const svint32_t &offset, const svfloat32_t &inv_scale)
89 {
90     const auto quantized = svcreate4(
91                                svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 0), inv_scale))), offset),
92                                svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 1), inv_scale))), offset),
93                                svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 2), inv_scale))), offset),
94                                svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 3), inv_scale))), offset));
95 
96     const auto narrowed_bottom = svqxtnt(svqxtnb(svget4(quantized, 0)), svget4(quantized, 1));
97     const auto narrowed_top    = svqxtnt(svqxtnb(svget4(quantized, 2)), svget4(quantized, 3));
98     const auto narrowed        = svqxtnt(svqxtnb(narrowed_bottom), narrowed_top);
99 
100     svst1(pg, ptr, narrowed);
101 }
102 
103 template <typename ScalarType>
elementwise_arithmetic_quantized_op(const ITensor * in1,const ITensor * in2,ITensor * out,ArithmeticOperation op,const Window & window)104 void elementwise_arithmetic_quantized_op(const ITensor *in1, const ITensor *in2, ITensor *out, ArithmeticOperation op, const Window &window)
105 {
106     const auto all_true_pg = wrapper::svptrue<ScalarType>();
107 
108     // Create input windows
109     Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
110     Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
111 
112     // Clear X Dimension on execution window as we handle manually
113     Window win = window;
114     win.set(Window::DimX, Window::Dimension(0, 1, 1));
115 
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     const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
119 
120     const auto output_voffset = svdup_n(out->info()->quantization_info().uniform().offset);
121     const auto output_vscale  = svdup_n(1.f / out->info()->quantization_info().uniform().scale);
122 
123     if(is_broadcast_across_x)
124     {
125         const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
126         Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
127         Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
128         const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
129         const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
130 
131         const auto non_broadcast_qinfo = is_broadcast_input_2 ? in1->info()->quantization_info() : in2->info()->quantization_info();
132         const auto broadcast_qinfo     = is_broadcast_input_2 ? in2->info()->quantization_info() : in1->info()->quantization_info();
133 
134         const auto non_broadcast_voffset = svdup_n(non_broadcast_qinfo.uniform().offset);
135         const auto non_broadcast_vscale  = svdup_n(non_broadcast_qinfo.uniform().scale);
136 
137         // Clear X Dimension on execution window as we handle manually
138         non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
139 
140         Iterator broadcast_input(broadcast_tensor, broadcast_win);
141         Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
142         Iterator output(out, win);
143 
144         execute_window_loop(win, [&](const Coordinates &)
145         {
146             auto             output_ptr              = reinterpret_cast<ScalarType *>(output.ptr());
147             const auto       non_broadcast_input_ptr = reinterpret_cast<const ScalarType *>(non_broadcast_input.ptr());
148             const ScalarType broadcast_value         = *reinterpret_cast<const ScalarType *>(broadcast_input.ptr());
149             const float      broadcast_value_f       = Qasymm8QuantizationHelper<ScalarType>::dequantize(broadcast_value, broadcast_qinfo);
150             const auto       in2                     = svcreate4(svdup_n(broadcast_value_f), svdup_n(broadcast_value_f), svdup_n(broadcast_value_f), svdup_n(broadcast_value_f));
151 
152             int x = window_start_x;
153 
154             svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
155             do
156             {
157                 const auto in1 = load_quantized(non_broadcast_input_ptr + x, pg, non_broadcast_voffset, non_broadcast_vscale);
158 
159                 svfloat32x4_t result{};
160 
161                 if(!is_broadcast_input_2)
162                 {
163                     result = svcreate4(
164                                  elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in2, 0), svget4(in1, 0), op),
165                                  elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in2, 1), svget4(in1, 1), op),
166                                  elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in2, 2), svget4(in1, 2), op),
167                                  elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in2, 3), svget4(in1, 3), op));
168                 }
169                 else
170                 {
171                     result = svcreate4(
172                                  elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 0), svget4(in2, 0), op),
173                                  elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 1), svget4(in2, 1), op),
174                                  elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 2), svget4(in2, 2), op),
175                                  elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 3), svget4(in2, 3), op));
176                 }
177 
178                 store_quantized(output_ptr + x, pg, result, output_voffset, output_vscale);
179 
180                 x += wrapper::svcnt<ScalarType>();
181                 pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
182             }
183             while(svptest_any(all_true_pg, pg));
184         },
185         broadcast_input, non_broadcast_input, output);
186     }
187     else
188     {
189         // Clear X Dimension on execution window as we handle manually
190         input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
191         input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
192 
193         Iterator input1(in1, input1_win);
194         Iterator input2(in2, input2_win);
195         Iterator output(out, win);
196 
197         const auto in1_voffset = svdup_n(in1->info()->quantization_info().uniform().offset);
198         const auto in1_vscale  = svdup_n(in1->info()->quantization_info().uniform().scale);
199 
200         const auto in2_voffset = svdup_n(in2->info()->quantization_info().uniform().offset);
201         const auto in2_vscale  = svdup_n(in2->info()->quantization_info().uniform().scale);
202 
203         execute_window_loop(win, [&](const Coordinates &)
204         {
205             auto       output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
206             const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr());
207             const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr());
208 
209             int x = window_start_x;
210 
211             svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
212             do
213             {
214                 const auto in1 = load_quantized(input1_ptr + x, pg, in1_voffset, in1_vscale);
215                 const auto in2 = load_quantized(input2_ptr + x, pg, in2_voffset, in2_vscale);
216 
217                 const auto result = svcreate4(
218                                         elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 0), svget4(in2, 0), op),
219                                         elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 1), svget4(in2, 1), op),
220                                         elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 2), svget4(in2, 2), op),
221                                         elementwise_arithmetic_op<svfloat32_t>(pg, svget4(in1, 3), svget4(in2, 3), op));
222 
223                 store_quantized(output_ptr + x, pg, result, output_voffset, output_vscale);
224 
225                 x += wrapper::svcnt<ScalarType>();
226                 pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
227             }
228             while(svptest_any(all_true_pg, pg));
229         },
230         input1, input2, output);
231     }
232 }
233 
234 template <typename InputScalarType, typename OutputScalarType = uint8_t>
elementwise_comparison_quantized_op(const ITensor * in1,const ITensor * in2,ITensor * out,ComparisonOperation op,const Window & window)235 void elementwise_comparison_quantized_op(const ITensor *in1, const ITensor *in2, ITensor *out, ComparisonOperation op, const Window &window)
236 {
237     static_assert(sizeof(InputScalarType) >= sizeof(OutputScalarType), "input data type's width should be equal to or greater than output data type's width");
238 
239     using OutputVectorType = typename wrapper::traits::sve_vector<OutputScalarType>::type;
240     const auto all_true_pg = wrapper::svptrue<InputScalarType>();
241 
242     // Create input windows
243     Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
244     Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
245 
246     // Clear X Dimension on execution window as we handle manually
247     Window win = window;
248     win.set(Window::DimX, Window::Dimension(0, 1, 1));
249 
250     const auto window_start_x        = static_cast<int>(window.x().start());
251     const auto window_end_x          = static_cast<int>(window.x().end());
252     const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
253 
254     if(is_broadcast_across_x)
255     {
256         const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
257         Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
258         Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
259         const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
260         const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
261 
262         const auto non_broadcast_qinfo = is_broadcast_input_2 ? in1->info()->quantization_info() : in2->info()->quantization_info();
263         const auto broadcast_qinfo     = is_broadcast_input_2 ? in2->info()->quantization_info() : in1->info()->quantization_info();
264 
265         const auto non_broadcast_voffset = svdup_n(non_broadcast_qinfo.uniform().offset);
266         const auto non_broadcast_vscale  = svdup_n(non_broadcast_qinfo.uniform().scale);
267 
268         // Clear X Dimension on execution window as we handle manually
269         non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
270 
271         Iterator broadcast_input(broadcast_tensor, broadcast_win);
272         Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
273         Iterator output(out, win);
274 
275         execute_window_loop(win, [&](const Coordinates &)
276         {
277             auto                  output_ptr              = reinterpret_cast<OutputScalarType *>(output.ptr());
278             const auto            non_broadcast_input_ptr = reinterpret_cast<const InputScalarType *>(non_broadcast_input.ptr());
279             const InputScalarType broadcast_value         = *reinterpret_cast<const InputScalarType *>(broadcast_input.ptr());
280             const float           broadcast_value_f       = Qasymm8QuantizationHelper<InputScalarType>::dequantize(broadcast_value, broadcast_qinfo);
281             const auto            in2                     = svcreate4(svdup_n(broadcast_value_f), svdup_n(broadcast_value_f), svdup_n(broadcast_value_f), svdup_n(broadcast_value_f));
282 
283             int x = window_start_x;
284 
285             svbool_t pg = wrapper::svwhilelt<InputScalarType>(x, window_end_x);
286             do
287             {
288                 const auto in1 = load_quantized(non_broadcast_input_ptr + x, pg, non_broadcast_voffset, non_broadcast_vscale);
289 
290                 svuint8x4_t result{};
291 
292                 if(!is_broadcast_input_2)
293                 {
294                     result = svcreate4(
295                                  elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in2, 0), svget4(in1, 0), op),
296                                  elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in2, 1), svget4(in1, 1), op),
297                                  elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in2, 2), svget4(in1, 2), op),
298                                  elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in2, 3), svget4(in1, 3), op));
299                 }
300                 else
301                 {
302                     result = svcreate4(
303                                  elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 0), svget4(in2, 0), op),
304                                  elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 1), svget4(in2, 1), op),
305                                  elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 2), svget4(in2, 2), op),
306                                  elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 3), svget4(in2, 3), op));
307                 }
308 
309                 const auto zipped_bottom = svzip1(svget4(result, 0), svget4(result, 1));
310                 const auto zipped_top    = svzip1(svget4(result, 2), svget4(result, 3));
311                 const auto zipped        = svzip1(zipped_bottom, zipped_top);
312                 svst1(pg, output_ptr + x, zipped);
313 
314                 x += wrapper::svcnt<InputScalarType>();
315                 pg = wrapper::svwhilelt<InputScalarType>(x, window_end_x);
316             }
317             while(svptest_any(all_true_pg, pg));
318         },
319         broadcast_input, non_broadcast_input, output);
320     }
321     else
322     {
323         // Clear X Dimension on execution window as we handle manually
324         input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
325         input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
326 
327         Iterator input1(in1, input1_win);
328         Iterator input2(in2, input2_win);
329         Iterator output(out, win);
330 
331         const auto in1_voffset = svdup_n(in1->info()->quantization_info().uniform().offset);
332         const auto in1_vscale  = svdup_n(in1->info()->quantization_info().uniform().scale);
333 
334         const auto in2_voffset = svdup_n(in2->info()->quantization_info().uniform().offset);
335         const auto in2_vscale  = svdup_n(in2->info()->quantization_info().uniform().scale);
336 
337         execute_window_loop(win, [&](const Coordinates &)
338         {
339             auto       output_ptr = reinterpret_cast<OutputScalarType *>(output.ptr());
340             const auto input1_ptr = reinterpret_cast<const InputScalarType *>(input1.ptr());
341             const auto input2_ptr = reinterpret_cast<const InputScalarType *>(input2.ptr());
342 
343             int x = window_start_x;
344 
345             svbool_t pg = wrapper::svwhilelt<InputScalarType>(x, window_end_x);
346             do
347             {
348                 const auto in1    = load_quantized(input1_ptr + x, pg, in1_voffset, in1_vscale);
349                 const auto in2    = load_quantized(input2_ptr + x, pg, in2_voffset, in2_vscale);
350                 const auto result = svcreate4(
351                                         elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 0), svget4(in2, 0), op),
352                                         elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 1), svget4(in2, 1), op),
353                                         elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 2), svget4(in2, 2), op),
354                                         elementwise_comparison_op<svfloat32_t, OutputVectorType>(pg, svget4(in1, 3), svget4(in2, 3), op));
355 
356                 const auto zipped_bottom = svzip1(svget4(result, 0), svget4(result, 1));
357                 const auto zipped_top    = svzip1(svget4(result, 2), svget4(result, 3));
358                 const auto zipped        = svzip1(zipped_bottom, zipped_top);
359                 svst1(pg, output_ptr + x, zipped);
360 
361                 x += wrapper::svcnt<InputScalarType>();
362                 pg = wrapper::svwhilelt<InputScalarType>(x, window_end_x);
363             }
364             while(svptest_any(all_true_pg, pg));
365         },
366         input1, input2, output);
367     }
368 }
369 } // namespace cpu
370 } // namespace arm_compute
371 
372 #endif /* SRC_CORE_SVE_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H */