xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/CpuTransposeKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 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/CpuTransposeKernel.h"
25 
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/ITensor.h"
29 #include "arm_compute/core/TensorInfo.h"
30 #include "arm_compute/core/Types.h"
31 #include "arm_compute/core/Validate.h"
32 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
33 #include "src/core/helpers/AutoConfiguration.h"
34 #include "src/core/helpers/WindowHelpers.h"
35 
36 #include <arm_neon.h>
37 
38 namespace arm_compute
39 {
40 namespace cpu
41 {
42 namespace kernels
43 {
44 namespace
45 {
num_elems_processed(size_t element_size)46 unsigned int num_elems_processed(size_t element_size)
47 {
48     switch(element_size)
49     {
50         case 1:
51             return 8;
52         case 2:
53         case 4:
54             return 4;
55         default:
56             break;
57     }
58 
59     ARM_COMPUTE_ERROR("Element size not supported");
60 }
61 
transpose_8bit_elements(const ITensor * in,ITensor * out,const Window & window)62 void transpose_8bit_elements(const ITensor *in, ITensor *out, const Window &window)
63 {
64     const int    window_step_x            = 8;
65     const int    window_step_y            = 8;
66     const int    window_start_x           = window.x().start();
67     const int    window_end_x             = window.x().end();
68     const int    window_start_y           = window.y().start();
69     const int    window_end_y             = std::min(window.y().end(), static_cast<int>(in->info()->dimension(1)));
70     const int    window_end_y_multiple_of = ((window_end_y - window_start_y) / window_step_y) * window_step_y;
71     const size_t input_stride_in_bytes    = in->info()->strides_in_bytes()[1];
72     const size_t output_stride_in_bytes   = out->info()->strides_in_bytes()[1];
73 
74     // Check if we need a left-over loop for the y dimension
75     bool left_over_loop_y = (((window_end_y - window_start_y) % window_step_y) != 0);
76 
77     Window window_in(window);
78     window_in.set(Window::DimX, Window::Dimension(0, 1, 1));
79     if(left_over_loop_y)
80     {
81         // Check if window_end_y_multiple_of is greater than window_start_y
82         if(window_end_y_multiple_of > window_start_y)
83         {
84             window_in.set(Window::DimY, Window::Dimension(window_start_y, window_end_y_multiple_of, window_step_y));
85         }
86         else
87         {
88             window_in.set(Window::DimY, Window::Dimension(0, 0, 1));
89         }
90     }
91 
92     Window window_out(window);
93     window_out.set(Window::DimX, Window::Dimension(0, 0, 0));
94     window_out.set(Window::DimY, Window::Dimension(0, 0, 0));
95 
96     Iterator output(out, window_out);
97 
98     // Run the SIMD path if and only if the input is not a row-vector
99     if(in->info()->dimension(1) != 1)
100     {
101         Iterator input(in, window_in);
102         execute_window_loop(window_in, [&](const Coordinates & id)
103         {
104             // Compute 8x8 elements per iteration
105             int x = window_start_x;
106             for(; x <= (window_end_x - window_step_x); x += window_step_x)
107             {
108                 const uint8x8_t row0 = vld1_u8(reinterpret_cast<const uint8_t *>(input.ptr() + x + 0 * input_stride_in_bytes));
109                 const uint8x8_t row1 = vld1_u8(reinterpret_cast<const uint8_t *>(input.ptr() + x + 1 * input_stride_in_bytes));
110                 const uint8x8_t row2 = vld1_u8(reinterpret_cast<const uint8_t *>(input.ptr() + x + 2 * input_stride_in_bytes));
111                 const uint8x8_t row3 = vld1_u8(reinterpret_cast<const uint8_t *>(input.ptr() + x + 3 * input_stride_in_bytes));
112                 const uint8x8_t row4 = vld1_u8(reinterpret_cast<const uint8_t *>(input.ptr() + x + 4 * input_stride_in_bytes));
113                 const uint8x8_t row5 = vld1_u8(reinterpret_cast<const uint8_t *>(input.ptr() + x + 5 * input_stride_in_bytes));
114                 const uint8x8_t row6 = vld1_u8(reinterpret_cast<const uint8_t *>(input.ptr() + x + 6 * input_stride_in_bytes));
115                 const uint8x8_t row7 = vld1_u8(reinterpret_cast<const uint8_t *>(input.ptr() + x + 7 * input_stride_in_bytes));
116 
117                 // Transpose 2x2
118                 const uint8x8x2_t k0_u8 = vtrn_u8(row0, row1);
119                 const uint8x8x2_t k1_u8 = vtrn_u8(row2, row3);
120                 const uint8x8x2_t k2_u8 = vtrn_u8(row4, row5);
121                 const uint8x8x2_t k3_u8 = vtrn_u8(row6, row7);
122 
123                 // Transpose 4x4
124                 const uint16x4x2_t k0_u16 = vtrn_u16(vreinterpret_u16_u8(k0_u8.val[0]), vreinterpret_u16_u8(k1_u8.val[0]));
125                 const uint16x4x2_t k1_u16 = vtrn_u16(vreinterpret_u16_u8(k0_u8.val[1]), vreinterpret_u16_u8(k1_u8.val[1]));
126                 const uint16x4x2_t k2_u16 = vtrn_u16(vreinterpret_u16_u8(k2_u8.val[0]), vreinterpret_u16_u8(k3_u8.val[0]));
127                 const uint16x4x2_t k3_u16 = vtrn_u16(vreinterpret_u16_u8(k2_u8.val[1]), vreinterpret_u16_u8(k3_u8.val[1]));
128 
129                 // Transpose 8x8
130                 const uint32x2x2_t k0_u32 = vtrn_u32(vreinterpret_u32_u16(k0_u16.val[0]), vreinterpret_u32_u16(k2_u16.val[0]));
131                 const uint32x2x2_t k1_u32 = vtrn_u32(vreinterpret_u32_u16(k0_u16.val[1]), vreinterpret_u32_u16(k2_u16.val[1]));
132                 const uint32x2x2_t k2_u32 = vtrn_u32(vreinterpret_u32_u16(k1_u16.val[0]), vreinterpret_u32_u16(k3_u16.val[0]));
133                 const uint32x2x2_t k3_u32 = vtrn_u32(vreinterpret_u32_u16(k1_u16.val[1]), vreinterpret_u32_u16(k3_u16.val[1]));
134 
135                 // Compute destination address
136                 const size_t dst_offset_in_bytes = id.y() * sizeof(uint8_t) + x * output_stride_in_bytes;
137 
138                 vst1_u8(reinterpret_cast<uint8_t *>(output.ptr() + dst_offset_in_bytes + 0 * output_stride_in_bytes), vreinterpret_u8_u16(vreinterpret_u16_u32(k0_u32.val[0])));
139                 vst1_u8(reinterpret_cast<uint8_t *>(output.ptr() + dst_offset_in_bytes + 1 * output_stride_in_bytes), vreinterpret_u8_u16(vreinterpret_u16_u32(k2_u32.val[0])));
140                 vst1_u8(reinterpret_cast<uint8_t *>(output.ptr() + dst_offset_in_bytes + 2 * output_stride_in_bytes), vreinterpret_u8_u16(vreinterpret_u16_u32(k1_u32.val[0])));
141                 vst1_u8(reinterpret_cast<uint8_t *>(output.ptr() + dst_offset_in_bytes + 3 * output_stride_in_bytes), vreinterpret_u8_u16(vreinterpret_u16_u32(k3_u32.val[0])));
142                 vst1_u8(reinterpret_cast<uint8_t *>(output.ptr() + dst_offset_in_bytes + 4 * output_stride_in_bytes), vreinterpret_u8_u16(vreinterpret_u16_u32(k0_u32.val[1])));
143                 vst1_u8(reinterpret_cast<uint8_t *>(output.ptr() + dst_offset_in_bytes + 5 * output_stride_in_bytes), vreinterpret_u8_u16(vreinterpret_u16_u32(k2_u32.val[1])));
144                 vst1_u8(reinterpret_cast<uint8_t *>(output.ptr() + dst_offset_in_bytes + 6 * output_stride_in_bytes), vreinterpret_u8_u16(vreinterpret_u16_u32(k1_u32.val[1])));
145                 vst1_u8(reinterpret_cast<uint8_t *>(output.ptr() + dst_offset_in_bytes + 7 * output_stride_in_bytes), vreinterpret_u8_u16(vreinterpret_u16_u32(k3_u32.val[1])));
146             }
147 
148             // Compute left-over elements along the x dimension (1x8)
149             for(; x < window_end_x; ++x)
150             {
151                 const uint8_t val0 = *(input.ptr() + x + 0 * input_stride_in_bytes);
152                 const uint8_t val1 = *(input.ptr() + x + 1 * input_stride_in_bytes);
153                 const uint8_t val2 = *(input.ptr() + x + 2 * input_stride_in_bytes);
154                 const uint8_t val3 = *(input.ptr() + x + 3 * input_stride_in_bytes);
155                 const uint8_t val4 = *(input.ptr() + x + 4 * input_stride_in_bytes);
156                 const uint8_t val5 = *(input.ptr() + x + 5 * input_stride_in_bytes);
157                 const uint8_t val6 = *(input.ptr() + x + 6 * input_stride_in_bytes);
158                 const uint8_t val7 = *(input.ptr() + x + 7 * input_stride_in_bytes);
159 
160                 uint8x8_t result = vdup_n_u8(0);
161                 result           = vset_lane_u8(val0, result, 0);
162                 result           = vset_lane_u8(val1, result, 1);
163                 result           = vset_lane_u8(val2, result, 2);
164                 result           = vset_lane_u8(val3, result, 3);
165                 result           = vset_lane_u8(val4, result, 4);
166                 result           = vset_lane_u8(val5, result, 5);
167                 result           = vset_lane_u8(val6, result, 6);
168                 result           = vset_lane_u8(val7, result, 7);
169 
170                 // Compute destination address
171                 const size_t dst_offset_in_bytes = id.y() * sizeof(uint8_t) + x * output_stride_in_bytes;
172 
173                 vst1_u8(output.ptr() + dst_offset_in_bytes, result);
174             }
175         },
176         input, output);
177     }
178 
179     if(left_over_loop_y)
180     {
181         window_in.set(Window::DimX, Window::Dimension(window.x().start(), window.x().end(), 1));
182         window_in.set(Window::DimY, Window::Dimension(window_end_y_multiple_of, window_end_y, 1));
183 
184         Iterator input(in, window_in);
185         Iterator output(out, window_out);
186 
187         // Compute left-over elements along the y dimension (1x1)
188         execute_window_loop(window_in, [&](const Coordinates & id)
189         {
190             const uint8_t val0 = *input.ptr();
191 
192             // Compute destination address
193             const size_t dst_offset_in_bytes = id.y() * sizeof(uint8_t) + id.x() * output_stride_in_bytes;
194 
195             *(output.ptr() + dst_offset_in_bytes) = val0;
196         },
197         input, output);
198     }
199 }
200 
transpose_16bit_elements(const ITensor * in,ITensor * out,const Window & window)201 void transpose_16bit_elements(const ITensor *in, ITensor *out, const Window &window)
202 {
203     const int    window_step_x            = 4;
204     const int    window_step_y            = 4;
205     const int    window_start_x           = window.x().start();
206     const int    window_end_x             = window.x().end();
207     const int    window_start_y           = window.y().start();
208     const int    window_end_y             = std::min(window.y().end(), static_cast<int>(in->info()->dimension(1)));
209     const int    window_end_y_multiple_of = ((window_end_y - window_start_y) / window_step_y) * window_step_y;
210     const size_t input_stride_in_bytes    = in->info()->strides_in_bytes()[1];
211     const size_t output_stride_in_bytes   = out->info()->strides_in_bytes()[1];
212 
213     // Check if we need a left-over loop for the y dimension
214     bool left_over_loop_y = (((window_end_y - window_start_y) % window_step_y) != 0);
215 
216     Window window_in(window);
217     window_in.set(Window::DimX, Window::Dimension(0, 1, 1));
218     if(left_over_loop_y)
219     {
220         // Check if window_end_y_multiple_of is greater than window_start_y
221         if(window_end_y_multiple_of > window_start_y)
222         {
223             window_in.set(Window::DimY, Window::Dimension(window_start_y, window_end_y_multiple_of, window_step_y));
224         }
225         else
226         {
227             window_in.set(Window::DimY, Window::Dimension(0, 0, 1));
228         }
229     }
230 
231     Window window_out(window);
232     window_out.set(Window::DimX, Window::Dimension(0, 0, 0));
233     window_out.set(Window::DimY, Window::Dimension(0, 0, 0));
234 
235     Iterator output(out, window_out);
236 
237     // Run the SIMD path if and only if the input is not a row-vector
238     if(in->info()->dimension(1) != 1)
239     {
240         Iterator input(in, window_in);
241         execute_window_loop(window_in, [&](const Coordinates & id)
242         {
243             // Compute 4x4 elements per iteration
244             int x = window_start_x;
245             for(; x <= (window_end_x - window_step_x); x += window_step_x)
246             {
247                 const uint16x4_t row0 = vld1_u16(reinterpret_cast<const uint16_t *>(input.ptr() + 0 * input_stride_in_bytes) + x);
248                 const uint16x4_t row1 = vld1_u16(reinterpret_cast<const uint16_t *>(input.ptr() + 1 * input_stride_in_bytes) + x);
249                 const uint16x4_t row2 = vld1_u16(reinterpret_cast<const uint16_t *>(input.ptr() + 2 * input_stride_in_bytes) + x);
250                 const uint16x4_t row3 = vld1_u16(reinterpret_cast<const uint16_t *>(input.ptr() + 3 * input_stride_in_bytes) + x);
251 
252                 // Transpose 2x2
253                 const uint16x4x2_t k0_u16 = vtrn_u16(row0, row1);
254                 const uint16x4x2_t k1_u16 = vtrn_u16(row2, row3);
255 
256                 // Transpose 4x4
257                 const uint32x2x2_t k0_u32 = vtrn_u32(vreinterpret_u32_u16(k0_u16.val[0]), vreinterpret_u32_u16(k1_u16.val[0]));
258                 const uint32x2x2_t k1_u32 = vtrn_u32(vreinterpret_u32_u16(k0_u16.val[1]), vreinterpret_u32_u16(k1_u16.val[1]));
259 
260                 // Compute destination address
261                 const size_t dst_offset_in_bytes = id.y() * sizeof(uint16_t) + x * output_stride_in_bytes;
262 
263                 vst1_u16(reinterpret_cast<uint16_t *>(output.ptr() + dst_offset_in_bytes + 0 * output_stride_in_bytes), vreinterpret_u16_u32(k0_u32.val[0]));
264                 vst1_u16(reinterpret_cast<uint16_t *>(output.ptr() + dst_offset_in_bytes + 1 * output_stride_in_bytes), vreinterpret_u16_u32(k1_u32.val[0]));
265                 vst1_u16(reinterpret_cast<uint16_t *>(output.ptr() + dst_offset_in_bytes + 2 * output_stride_in_bytes), vreinterpret_u16_u32(k0_u32.val[1]));
266                 vst1_u16(reinterpret_cast<uint16_t *>(output.ptr() + dst_offset_in_bytes + 3 * output_stride_in_bytes), vreinterpret_u16_u32(k1_u32.val[1]));
267             }
268 
269             // Compute left-over elements (1x4)
270             for(; x < window_end_x; ++x)
271             {
272                 const uint16_t val0 = *(reinterpret_cast<uint16_t *>(input.ptr() + 0 * input_stride_in_bytes) + x);
273                 const uint16_t val1 = *(reinterpret_cast<uint16_t *>(input.ptr() + 1 * input_stride_in_bytes) + x);
274                 const uint16_t val2 = *(reinterpret_cast<uint16_t *>(input.ptr() + 2 * input_stride_in_bytes) + x);
275                 const uint16_t val3 = *(reinterpret_cast<uint16_t *>(input.ptr() + 3 * input_stride_in_bytes) + x);
276 
277                 uint16x4_t result = vdup_n_u16(0);
278                 result            = vset_lane_u16(val0, result, 0);
279                 result            = vset_lane_u16(val1, result, 1);
280                 result            = vset_lane_u16(val2, result, 2);
281                 result            = vset_lane_u16(val3, result, 3);
282 
283                 // Compute destination address
284                 const size_t dst_offset_in_bytes = id.y() * sizeof(uint16_t) + x * output_stride_in_bytes;
285 
286                 vst1_u16(reinterpret_cast<uint16_t *>(output.ptr() + dst_offset_in_bytes), result);
287             }
288         },
289         input, output);
290     }
291 
292     if(left_over_loop_y)
293     {
294         window_in.set(Window::DimX, Window::Dimension(window.x().start(), window.x().end(), 1));
295         window_in.set(Window::DimY, Window::Dimension(window_end_y_multiple_of, window_end_y, 1));
296 
297         Iterator input(in, window_in);
298         Iterator output(out, window_out);
299 
300         // Compute left-over elements along the y dimension (1x1)
301         execute_window_loop(window_in, [&](const Coordinates & id)
302         {
303             const uint16_t val0 = *(reinterpret_cast<uint16_t *>(input.ptr()));
304 
305             // Compute destination address
306             const size_t dst_offset_in_bytes = id.y() * sizeof(uint16_t) + id.x() * output_stride_in_bytes;
307 
308             *(reinterpret_cast<uint16_t *>(output.ptr() + dst_offset_in_bytes)) = val0;
309         },
310         input, output);
311     }
312 }
313 
transpose_32bit_elements(const ITensor * in,ITensor * out,const Window & window)314 void transpose_32bit_elements(const ITensor *in, ITensor *out, const Window &window)
315 {
316     const int    window_step_x            = 4;
317     const int    window_step_y            = 4;
318     const int    window_start_x           = window.x().start();
319     const int    window_end_x             = window.x().end();
320     const int    window_start_y           = window.y().start();
321     const int    window_end_y             = std::min(window.y().end(), static_cast<int>(in->info()->dimension(1)));
322     const int    window_end_y_multiple_of = ((window_end_y - window_start_y) / window_step_y) * window_step_y;
323     const size_t input_stride_in_bytes    = in->info()->strides_in_bytes()[1];
324     const size_t output_stride_in_bytes   = out->info()->strides_in_bytes()[1];
325 
326     // Check if we need a left-over loop for the y dimension
327     bool left_over_loop_y = (((window_end_y - window_start_y) % window_step_y) != 0);
328 
329     Window window_in(window);
330     window_in.set(Window::DimX, Window::Dimension(0, 1, 1));
331     if(left_over_loop_y)
332     {
333         // Check if window_end_y_multiple_of is greater than window_start_y
334         if(window_end_y_multiple_of > window_start_y)
335         {
336             window_in.set(Window::DimY, Window::Dimension(window_start_y, window_end_y_multiple_of, window_step_y));
337         }
338         else
339         {
340             window_in.set(Window::DimY, Window::Dimension(0, 0, 1));
341         }
342     }
343 
344     Window window_out(window);
345     window_out.set(Window::DimX, Window::Dimension(0, 0, 0));
346     window_out.set(Window::DimY, Window::Dimension(0, 0, 0));
347 
348     Iterator output(out, window_out);
349 
350     // Run the SIMD path if and only if the input is not a row-vector
351     if(in->info()->dimension(1) != 1)
352     {
353         Iterator input(in, window_in);
354         execute_window_loop(window_in, [&](const Coordinates & id)
355         {
356             // Compute 4x4 elements per iteration
357             int x = window_start_x;
358             for(; x <= (window_end_x - window_step_x); x += window_step_x)
359             {
360                 const uint32x4_t row0 = vld1q_u32(reinterpret_cast<const uint32_t *>(input.ptr() + 0 * input_stride_in_bytes) + x);
361                 const uint32x4_t row1 = vld1q_u32(reinterpret_cast<const uint32_t *>(input.ptr() + 1 * input_stride_in_bytes) + x);
362                 const uint32x4_t row2 = vld1q_u32(reinterpret_cast<const uint32_t *>(input.ptr() + 2 * input_stride_in_bytes) + x);
363                 const uint32x4_t row3 = vld1q_u32(reinterpret_cast<const uint32_t *>(input.ptr() + 3 * input_stride_in_bytes) + x);
364 
365                 // Transpose 2x2
366                 const uint32x2x2_t k0_u32 = vtrn_u32(vget_low_u32(row0), vget_low_u32(row1));
367                 const uint32x2x2_t k1_u32 = vtrn_u32(vget_high_u32(row2), vget_high_u32(row3));
368                 const uint32x2x2_t k2_u32 = vtrn_u32(vget_high_u32(row0), vget_high_u32(row1));
369                 const uint32x2x2_t k3_u32 = vtrn_u32(vget_low_u32(row2), vget_low_u32(row3));
370 
371                 // Compute destination address
372                 const size_t dst_offset_in_bytes = id.y() * sizeof(uint32_t) + x * output_stride_in_bytes;
373 
374                 // Swap block 01 with block 10 and store
375                 vst1q_u32(reinterpret_cast<uint32_t *>(output.ptr() + dst_offset_in_bytes + 0 * output_stride_in_bytes), vcombine_u32(k0_u32.val[0], k3_u32.val[0]));
376                 vst1q_u32(reinterpret_cast<uint32_t *>(output.ptr() + dst_offset_in_bytes + 1 * output_stride_in_bytes), vcombine_u32(k0_u32.val[1], k3_u32.val[1]));
377                 vst1q_u32(reinterpret_cast<uint32_t *>(output.ptr() + dst_offset_in_bytes + 2 * output_stride_in_bytes), vcombine_u32(k2_u32.val[0], k1_u32.val[0]));
378                 vst1q_u32(reinterpret_cast<uint32_t *>(output.ptr() + dst_offset_in_bytes + 3 * output_stride_in_bytes), vcombine_u32(k2_u32.val[1], k1_u32.val[1]));
379             }
380 
381             // Compute left-over elements (1x4)
382             for(; x < window_end_x; ++x)
383             {
384                 const uint32_t val0 = *(reinterpret_cast<uint32_t *>(input.ptr() + 0 * input_stride_in_bytes) + x);
385                 const uint32_t val1 = *(reinterpret_cast<uint32_t *>(input.ptr() + 1 * input_stride_in_bytes) + x);
386                 const uint32_t val2 = *(reinterpret_cast<uint32_t *>(input.ptr() + 2 * input_stride_in_bytes) + x);
387                 const uint32_t val3 = *(reinterpret_cast<uint32_t *>(input.ptr() + 3 * input_stride_in_bytes) + x);
388 
389                 uint32x4_t result = vdupq_n_u32(0);
390                 result            = vsetq_lane_u32(val0, result, 0);
391                 result            = vsetq_lane_u32(val1, result, 1);
392                 result            = vsetq_lane_u32(val2, result, 2);
393                 result            = vsetq_lane_u32(val3, result, 3);
394 
395                 // Compute destination address
396                 const size_t dst_offset_in_bytes = id.y() * sizeof(uint32_t) + x * output_stride_in_bytes;
397 
398                 vst1q_u32(reinterpret_cast<uint32_t *>(output.ptr() + dst_offset_in_bytes), result);
399             }
400         },
401         input, output);
402     }
403 
404     if(left_over_loop_y)
405     {
406         window_in.set(Window::DimX, Window::Dimension(window.x().start(), window.x().end(), 1));
407         window_in.set(Window::DimY, Window::Dimension(window_end_y_multiple_of, window_end_y, 1));
408 
409         Iterator input(in, window_in);
410         Iterator output(out, window_out);
411 
412         // Compute left-over elements along the y dimension (1x1)
413         execute_window_loop(window_in, [&](const Coordinates & id)
414         {
415             const uint32_t val0 = *(reinterpret_cast<uint32_t *>(input.ptr()));
416 
417             // Compute destination address
418             const size_t dst_offset_in_bytes = id.y() * sizeof(uint32_t) + id.x() * output_stride_in_bytes;
419 
420             *(reinterpret_cast<uint32_t *>(output.ptr() + dst_offset_in_bytes)) = val0;
421         },
422         input, output);
423     }
424 }
425 } // namespace
426 
configure(const ITensorInfo * src,ITensorInfo * dst)427 void CpuTransposeKernel::configure(const ITensorInfo *src, ITensorInfo *dst)
428 {
429     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
430 
431     // Destination auto inizialitation if not yet initialized
432     const TensorShape dst_shape = misc::shape_calculator::compute_transposed_shape(*src);
433     auto_init_if_empty(*dst, src->clone()->set_tensor_shape(dst_shape));
434 
435     // Perform validation step
436     ARM_COMPUTE_ERROR_THROW_ON(validate(src, dst));
437 
438     // Note: This kernel performs 16 elements per iteration.
439     // However, since we use a left-over for loop on both dimensions (X and Y), we cannot have any read or write out of memory
440     // For this reason num_elems_processed_per_iteration_x is set to 1
441     const unsigned int num_elems_processed_per_iteration_x = 1;
442     const unsigned int num_elems_processed_per_iteration_y = num_elems_processed(src->element_size());
443 
444     // Configure kernel window
445     Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
446 
447     // The CpuTranspose doesn't need padding so update_window_and_padding() can be skipped
448     Coordinates coord;
449     coord.set_num_dimensions(dst->num_dimensions());
450     dst->set_valid_region(ValidRegion(coord, dst->tensor_shape()));
451 
452     ICpuKernel::configure(win);
453 }
454 
validate(const ITensorInfo * src,const ITensorInfo * dst)455 Status CpuTransposeKernel::validate(const ITensorInfo *src, const ITensorInfo *dst)
456 {
457     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src);
458     //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use CPU FP16 instructions.
459     ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN);
460 
461     // Error if input is not 8 bit, 16bit or 32bit
462     ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->element_size() != 1 && src->element_size() != 2 && src->element_size() != 4,
463                                     "Element size not supported");
464 
465     // Validate configured destination
466     if(dst->total_size() != 0)
467     {
468         const TensorShape dst_shape = misc::shape_calculator::compute_transposed_shape(*src);
469 
470         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), dst_shape);
471         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst);
472         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
473     }
474 
475     return Status{};
476 }
477 
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)478 void CpuTransposeKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
479 {
480     ARM_COMPUTE_UNUSED(info);
481     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
482     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
483 
484     const auto src = tensors.get_const_tensor(TensorType::ACL_SRC);
485     auto       dst = tensors.get_tensor(TensorType::ACL_DST);
486 
487     switch(src->info()->element_size())
488     {
489         case 1:
490             transpose_8bit_elements(src, dst, window);
491             break;
492         case 2:
493             transpose_16bit_elements(src, dst, window);
494             break;
495         case 4:
496             transpose_32bit_elements(src, dst, window);
497             break;
498         default:
499             ARM_COMPUTE_ERROR("Element size not supported");
500             break;
501     }
502 }
503 
name() const504 const char *CpuTransposeKernel::name() const
505 {
506     return "CpuTransposeKernel";
507 }
508 } // namespace kernels
509 } // namespace cpu
510 } // namespace arm_compute
511