Compute Library
 22.05
Data Structures | Public Member Functions | Static Public Member Functions
CpuCastKernel Class Reference

Casts a given tensor to a new type. More...

#include <CpuCastKernel.h>

Collaboration diagram for CpuCastKernel:
[legend]

Data Structures

struct  CastKernel
 

Public Member Functions

 CpuCastKernel ()=default
 
 ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE (CpuCastKernel)
 
void configure (const ITensorInfo *src, ITensorInfo *dst, ConvertPolicy policy)
 Set the src and dst of the kernel. More...
 
void run_op (ITensorPack &tensors, const Window &window, const ThreadInfo &info) override
 Execute the kernel on the passed window. More...
 
const char * name () const override
 Name of the kernel. More...
 
- Public Member Functions inherited from ICPPKernel
virtual ~ICPPKernel ()=default
 Default destructor. More...
 
virtual void run (const Window &window, const ThreadInfo &info)
 Execute the kernel on the passed window. More...
 
virtual void run_nd (const Window &window, const ThreadInfo &info, const Window &thread_locator)
 legacy compatibility layer for implemantions which do not support thread_locator In these cases we simply narrow the interface down the legacy version More...
 
virtual size_t get_mws (const CPUInfo &platform, size_t thread_count) const
 Return minimum workload size of the relevant kernel. More...
 
- Public Member Functions inherited from IKernel
 IKernel ()
 Constructor. More...
 
virtual ~IKernel ()=default
 Destructor. More...
 
virtual bool is_parallelisable () const
 Indicates whether or not the kernel is parallelisable. More...
 
virtual BorderSize border_size () const
 The size of the border for that kernel. More...
 
const Window & window () const
 The maximum window the kernel can be executed on. More...
 
bool is_window_configured () const
 Function to check if the embedded window of this kernel has been configured. More...
 

Static Public Member Functions

static Status validate (const ITensorInfo *src, const ITensorInfo *dst, ConvertPolicy policy)
 Static function to check if given info will lead to a valid configuration. More...
 
static const std::vector< CastKernel > & get_available_kernels ()
 
- Static Public Member Functions inherited from ICpuKernel< CpuCastKernel >
static const auto * get_implementation (const SelectorType &selector, KernelSelectionType selection_type=KernelSelectionType::Supported)
 Micro-kernel selector. More...
 

Additional Inherited Members

- Static Public Attributes inherited from ICPPKernel
static constexpr size_t default_mws = 1
 

Detailed Description

Casts a given tensor to a new type.

Note
When casting between quantized types the scale and zeroPoint are ignored

Definition at line 40 of file CpuCastKernel.h.

Constructor & Destructor Documentation

◆ CpuCastKernel()

CpuCastKernel ( )
default

Member Function Documentation

◆ ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE()

ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE ( CpuCastKernel  )

◆ configure()

void configure ( const ITensorInfo *  src,
ITensorInfo *  dst,
ConvertPolicy  policy 
)

Set the src and dst of the kernel.

Valid conversions src -> dst :

  • QASYMM8_SIGNED -> S16, S32, F32, F16
  • QASYMM8 -> U16, S16, S32, F32, F16
  • U8 -> U16, S16, S32, F32, F16
  • U16 -> U8, U32
  • S16 -> QASYMM8_SIGNED, U8, S32
  • BFLOAT16 -> F32
  • F16 -> QASYMM8_SIGNED, QASYMM8, F32, S32, U8
  • S32 -> QASYMM8_SIGNED, QASYMM8, F16, F32, U8
  • F32 -> QASYMM8_SIGNED, QASYMM8, BFLOAT16, F16, S32, U8
Parameters
[in]srcThe src tensor to convert. Data types supported: QASYMM8_SIGNED/QASYMM8/U8/U16/S16/BFLOAT16/F16/F32.
[out]dstThe dst tensor. Data types supported: QASYMM8_SIGNED/QASYMM8/U8/U16/S16/U32/S32/BFLOAT16/F16/F32.
[in]policyConversion policy.

Definition at line 155 of file CpuCastKernel.cpp.

References ARM_COMPUTE_ERROR_ON_NULLPTR, ARM_COMPUTE_ERROR_THROW_ON, arm_compute::calculate_max_window(), arm_compute::set_shape_if_empty(), ITensorInfo::tensor_shape(), and arm_compute::cpu::kernels::validate_arguments().

156 {
157  ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
158 
159  // Auto initialize dst shape if not initialized (We can only auto-configure the shape, datatype must be given)
160  set_shape_if_empty(*dst, src->tensor_shape());
161 
162  _policy = policy;
163 
164  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, policy));
165 
166  // Configure kernel window
167  Window win = calculate_max_window(*src, Steps());
168 
169  ICPPKernel::configure(win);
170 }
arm_compute::calculate_max_window
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
Definition: WindowHelpers.cpp:28
ARM_COMPUTE_ERROR_THROW_ON
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
arm_compute::cpu::kernels::validate_arguments
Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const PadStrideInfo &conv_info)
Definition: CpuDirectConv2dKernel.cpp:60
arm_compute::test::validation::src
SimpleTensor< float > src
Definition: DFT.cpp:155
arm_compute::set_shape_if_empty
bool set_shape_if_empty(ITensorInfo &info, const TensorShape &shape)
Set the shape to the specified value if the current assignment is empty.
Definition: AutoConfiguration.h:88
arm_compute::test::validation::dst
auto dst
Definition: DFT.cpp:170
ARM_COMPUTE_ERROR_ON_NULLPTR
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:157

◆ get_available_kernels()

const std::vector< CpuCastKernel::CastKernel > & get_available_kernels ( )
static

Definition at line 1110 of file CpuCastKernel.cpp.

1111 {
1112  return available_kernels;
1113 }

◆ name()

const char * name ( ) const
overridevirtual

Name of the kernel.

Returns
Kernel name

Implements ICPPKernel.

Definition at line 1105 of file CpuCastKernel.cpp.

1106 {
1107  return "CpuCastKernel.cpp";
1108 }

◆ run_op()

void run_op ( ITensorPack &  tensors,
const Window &  window,
const ThreadInfo &  info 
)
overridevirtual

Execute the kernel on the passed window.

Warning
If is_parallelisable() returns false then the passed window must be equal to window()
Note
The window has to be a region within the window returned by the window() method
The width of the window has to be a multiple of num_elems_processed_per_iteration().
Parameters
[in]tensorsA vector containing the tensors to operate on.
[in]windowRegion on which to execute the kernel. (Must be a region of the window returned by window())
[in]infoInfo about executing thread and CPU.

Reimplemented from ICPPKernel.

Definition at line 178 of file CpuCastKernel.cpp.

References arm_compute::ACL_DST, arm_compute::ACL_SRC, ARM_COMPUTE_ERROR, ARM_COMPUTE_ERROR_ON, ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW, ARM_COMPUTE_ERROR_ON_NULLPTR, ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL, ARM_COMPUTE_UNUSED, arm_compute::BFLOAT16, ITensorInfo::data_type(), Window::DimX, arm_compute::test::validation::dst, Window::Dimension::end(), arm_compute::execute_window_loop(), arm_compute::F16, arm_compute::F32, CPUInfo::get(), ITensorPack::get_const_tensor(), ICpuKernel< CpuCastKernel >::get_implementation(), CPUInfo::get_isa(), ITensorPack::get_tensor(), ITensor::info(), Iterator::ptr(), arm_compute::QASYMM8, arm_compute::QASYMM8_SIGNED, arm_compute::S16, arm_compute::S32, arm_compute::SATURATE, Window::set(), arm_compute::test::validation::src, Window::Dimension::start(), arm_compute::U16, arm_compute::U32, arm_compute::U8, IKernel::window(), and Window::x().

179 {
180  ARM_COMPUTE_UNUSED(info);
181  ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
182  ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
183 
184  const auto window_start_x = static_cast<int>(window.x().start());
185  const auto window_end_x = static_cast<int>(window.x().end());
186  const int window_step_x = 16;
187 
188  const ITensor *_src = tensors.get_const_tensor(TensorType::ACL_SRC);
189  ITensor *_dst = tensors.get_tensor(TensorType::ACL_DST);
190  ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _dst);
191  ARM_COMPUTE_ERROR_ON(_src == _dst);
192 
193  ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _dst);
194 
195  Window win{ window };
196  win.set(Window::DimX, Window::Dimension(0, 1, 1));
197 
198  Iterator src(_src, win);
199  Iterator dst(_dst, win);
200 
201  /*ukernel runs only when using fp16/bfloat16, so we validate it isn't a nullptr only before using it */
202  const auto *uk = CpuCastKernel::get_implementation(CastDataTypeISASelectorData{ _src->info()->data_type(), _dst->info()->data_type(), CPUInfo::get().get_isa() });
203 
204  switch(_src->info()->data_type())
205  {
206  case DataType::QASYMM8_SIGNED:
207  {
208  switch(_dst->info()->data_type())
209  {
210  case DataType::S16:
211  {
212  /* Up-conversion QASYMM8_SIGNED -> S16 */
213  execute_window_loop(win, [&](const Coordinates &)
214  {
215  const auto src_ptr = reinterpret_cast<const int8_t *>(src.ptr());
216  const auto dst_ptr = reinterpret_cast<int16_t *>(dst.ptr());
217  int x = window_start_x;
218 
219  for(; x <= (window_end_x - window_step_x); x += window_step_x)
220  {
221  const int8x16_t texels_s8 = vld1q_s8(src_ptr + x);
222 
223  const int16x8x2_t texels =
224  {
225  {
226  vmovl_s8(vget_low_s8(texels_s8)),
227  vmovl_s8(vget_high_s8(texels_s8))
228  }
229  };
230 
231  vst1q_s16(dst_ptr + x, texels.val[0]);
232  vst1q_s16(dst_ptr + x + 8, texels.val[1]);
233  }
234 
235  // Compute left-over elements
236  for(; x < window_end_x; ++x)
237  {
238  *(dst_ptr + x) = static_cast<int16_t>(*(src_ptr + x));
239  }
240  },
241  src, dst);
242  break;
243  }
244  case DataType::S32:
245  {
246  /* Up-conversion QASYMM8_SIGNED -> S32 */
247  execute_window_loop(win, [&](const Coordinates &)
248  {
249  const auto src_ptr = reinterpret_cast<const int8_t *>(src.ptr());
250  const auto dst_ptr = reinterpret_cast<int32_t *>(dst.ptr());
251  int x = window_start_x;
252 
253  for(; x <= (window_end_x - window_step_x); x += window_step_x)
254  {
255  const int8x16_t texels_s8 = vld1q_s8(src_ptr + x);
256 
257  const int16x8x2_t texels =
258  {
259  {
260  vmovl_s8(vget_low_s8(texels_s8)),
261  vmovl_s8(vget_high_s8(texels_s8))
262  }
263  };
264 
265  vst1q_s32(dst_ptr + x, vmovl_s16(vget_low_s16(texels.val[0])));
266  vst1q_s32(dst_ptr + x + 4, vmovl_s16(vget_high_s16(texels.val[0])));
267  vst1q_s32(dst_ptr + x + 8, vmovl_s16(vget_low_s16(texels.val[1])));
268  vst1q_s32(dst_ptr + x + 12, vmovl_s16(vget_high_s16(texels.val[1])));
269  }
270 
271  // Compute left-over elements
272  for(; x < window_end_x; ++x)
273  {
274  *(dst_ptr + x) = static_cast<int32_t>(*(src_ptr + x));
275  }
276  },
277  src, dst);
278  break;
279  }
280  case DataType::F32:
281  {
282  /* Up-conversion QASYMM8_SIGNED -> F32 */
283  execute_window_loop(win, [&](const Coordinates &)
284  {
285  const auto src_ptr = reinterpret_cast<const int8_t *>(src.ptr());
286  const auto dst_ptr = reinterpret_cast<float *>(dst.ptr());
287 
288  int x = window_start_x;
289  for(; x <= (window_end_x - window_step_x); x += window_step_x)
290  {
291  const int8x16_t texels_s8 = vld1q_s8(reinterpret_cast<int8_t *>(src.ptr()));
292 
293  const int16x8x2_t texels =
294  {
295  {
296  vmovl_s8(vget_low_s8(texels_s8)),
297  vmovl_s8(vget_high_s8(texels_s8))
298  }
299  };
300  vst1q_f32(dst_ptr + x, vcvtq_f32_s32(vmovl_s16(vget_low_s16(texels.val[0]))));
301  vst1q_f32(dst_ptr + x + 4, vcvtq_f32_s32(vmovl_s16(vget_high_s16(texels.val[0]))));
302  vst1q_f32(dst_ptr + x + 8, vcvtq_f32_s32(vmovl_s16(vget_low_s16(texels.val[1]))));
303  vst1q_f32(dst_ptr + x + 12, vcvtq_f32_s32(vmovl_s16(vget_high_s16(texels.val[1]))));
304  }
305 
306  // Compute left-over elements
307  for(; x < window_end_x; ++x)
308  {
309  *(dst_ptr + x) = static_cast<float>(*(src_ptr + x));
310  }
311  },
312  src, dst);
313  break;
314  }
315  case DataType::F16:
316  {
317  /* Up-conversion QASYMM8_SIGNED -> F16 */
318  ARM_COMPUTE_ERROR_ON(uk->ukernel == nullptr);
319  uk->ukernel(_src, _dst, info, _policy, window);
320  break;
321  }
322  default:
323  ARM_COMPUTE_ERROR("dst data type not supported");
324  }
325  break;
326  }
327 
328  case DataType::QASYMM8:
329  case DataType::U8:
330  {
331  switch(_dst->info()->data_type())
332  {
333  case DataType::S16:
334  {
335  /* Up-conversion U8 -> S16 */
336  execute_window_loop(win, [&](const Coordinates &)
337  {
338  const auto src_ptr = reinterpret_cast<const uint8_t *>(src.ptr());
339  const auto dst_ptr = reinterpret_cast<int16_t *>(dst.ptr());
340 
341  int x = window_start_x;
342  for(; x <= (window_end_x - window_step_x); x += window_step_x)
343  {
344  const uint8x16_t texels_u8 = vld1q_u8(src_ptr + x);
345 
346  const int16x8x2_t texels =
347  {
348  {
349  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(texels_u8))),
350  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(texels_u8)))
351  }
352  };
353 
354  vst1q_s16(dst_ptr + x, texels.val[0]);
355  vst1q_s16(dst_ptr + x + 8, texels.val[1]);
356  }
357 
358  // Compute left-over elements
359  for(; x < window_end_x; ++x)
360  {
361  *(dst_ptr + x) = static_cast<int32_t>(*(src_ptr + x));
362  }
363  },
364  src, dst);
365  break;
366  }
367  case DataType::S32:
368  {
369  /* Up-conversion U8 -> S32 */
370  execute_window_loop(win, [&](const Coordinates &)
371  {
372  const auto src_ptr = reinterpret_cast<const uint8_t *>(src.ptr());
373  const auto dst_ptr = reinterpret_cast<int32_t *>(dst.ptr());
374 
375  int x = window_start_x;
376  for(; x <= (window_end_x - window_step_x); x += window_step_x)
377  {
378  const uint8x16_t texels_u8 = vld1q_u8(src_ptr + x);
379 
380  const int16x8x2_t texels =
381  {
382  {
383  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(texels_u8))),
384  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(texels_u8)))
385  }
386  };
387 
388  vst1q_s32(dst_ptr + x, vmovl_s16(vget_low_s16(texels.val[0])));
389  vst1q_s32(dst_ptr + x + 4, vmovl_s16(vget_high_s16(texels.val[0])));
390  vst1q_s32(dst_ptr + x + 8, vmovl_s16(vget_low_s16(texels.val[1])));
391  vst1q_s32(dst_ptr + x + 12, vmovl_s16(vget_high_s16(texels.val[1])));
392  }
393 
394  // Compute left-over elements
395  for(; x < window_end_x; ++x)
396  {
397  *(dst_ptr + x) = static_cast<uint32_t>(*(src_ptr + x));
398  }
399  },
400  src, dst);
401  break;
402  }
403  case DataType::F32:
404  {
405  /* Up-conversion U8 -> F32 */
406  execute_window_loop(win, [&](const Coordinates &)
407  {
408  const auto src_ptr = reinterpret_cast<const uint8_t *>(src.ptr());
409  const auto dst_ptr = reinterpret_cast<float *>(dst.ptr());
410 
411  int x = window_start_x;
412  for(; x <= (window_end_x - window_step_x); x += window_step_x)
413  {
414  const uint8x16_t texels_u8 = vld1q_u8(src_ptr + x);
415 
416  const int16x8x2_t texels =
417  {
418  {
419  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(texels_u8))),
420  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(texels_u8)))
421  }
422  };
423  vst1q_f32(dst_ptr + x, vcvtq_f32_s32(vmovl_s16(vget_low_s16(texels.val[0]))));
424  vst1q_f32(dst_ptr + x + 4, vcvtq_f32_s32(vmovl_s16(vget_high_s16(texels.val[0]))));
425  vst1q_f32(dst_ptr + x + 8, vcvtq_f32_s32(vmovl_s16(vget_low_s16(texels.val[1]))));
426  vst1q_f32(dst_ptr + x + 12, vcvtq_f32_s32(vmovl_s16(vget_high_s16(texels.val[1]))));
427  }
428 
429  // Compute left-over elements
430  for(; x < window_end_x; ++x)
431  {
432  *(dst_ptr + x) = static_cast<uint32_t>(*(src_ptr + x));
433  }
434  },
435  src, dst);
436  break;
437  }
438  case DataType::F16:
439  {
440  /* Up-conversion U8 -> FP16 */
441  ARM_COMPUTE_ERROR_ON(uk->ukernel == nullptr);
442  uk->ukernel(_src, _dst, info, _policy, window);
443  break;
444  }
445  case DataType::U16:
446  {
447  /* Up-conversion U8 -> U16 */
448  execute_window_loop(win, [&](const Coordinates &)
449  {
450  const auto src_ptr = reinterpret_cast<const uint8_t *>(src.ptr());
451  const auto dst_ptr = reinterpret_cast<uint16_t *>(dst.ptr());
452 
453  int x = window_start_x;
454  for(; x <= (window_end_x - window_step_x); x += window_step_x)
455  {
456  const uint8x16_t texels_u8 = vld1q_u8(src_ptr + x);
457 
458  const uint16x8x2_t texels =
459  {
460  {
461  vmovl_u8(vget_low_u8(texels_u8)),
462  vmovl_u8(vget_high_u8(texels_u8))
463  }
464  };
465 
466  vst1q_u16(dst_ptr + x, texels.val[0]);
467  vst1q_u16(dst_ptr + x + 8, texels.val[1]);
468  }
469 
470  // Compute left-over elements
471  for(; x < window_end_x; ++x)
472  {
473  *(dst_ptr + x) = static_cast<uint16_t>(*(src_ptr + x));
474  }
475  },
476  src, dst);
477  break;
478  }
479  default:
480  ARM_COMPUTE_ERROR("dst data type not supported");
481  }
482  break;
483  }
484  case DataType::S16:
485  {
486  switch(_dst->info()->data_type())
487  {
488  case DataType::QASYMM8_SIGNED:
489  {
490  /* Down-conversion S16 -> QASYMM8_SIGNED */
491  if(ConvertPolicy::SATURATE == _policy)
492  {
493  execute_window_loop(win, [&](const Coordinates &)
494  {
495  const auto src_ptr = reinterpret_cast<const int16_t *>(src.ptr());
496  const auto dst_ptr = reinterpret_cast<int8_t *>(dst.ptr());
497 
498  int x = window_start_x;
499  for(; x <= (window_end_x - window_step_x); x += window_step_x)
500  {
501  const int16x8x2_t texels =
502  {
503  {
504  vld1q_s16(src_ptr + x),
505  vld1q_s16(src_ptr + x + 8)
506  }
507  };
508 
509  vst1q_s8(dst_ptr + x, vcombine_s8(vqmovn_s16(texels.val[0]), vqmovn_s16(texels.val[1])));
510  }
511 
512  // Compute left-over elements
513  for(; x < window_end_x; ++x)
514  {
515  *(dst_ptr + x) = utils::cast::saturate_cast<int8_t>(*(src_ptr + x));
516  }
517  },
518  src, dst);
519  }
520  else
521  {
522  execute_window_loop(win, [&](const Coordinates &)
523  {
524  const auto src_ptr = reinterpret_cast<const int16_t *>(src.ptr());
525  const auto dst_ptr = reinterpret_cast<int8_t *>(dst.ptr());
526 
527  int x = window_start_x;
528  for(; x <= (window_end_x - window_step_x); x += window_step_x)
529  {
530  const int16x8x2_t texels =
531  {
532  {
533  vld1q_s16(src_ptr + x),
534  vld1q_s16(src_ptr + x + 8)
535  }
536  };
537 
538  vst1q_s8(dst_ptr + x, vcombine_s8(vmovn_s16(texels.val[0]), vmovn_s16(texels.val[1])));
539  }
540 
541  // Compute left-over elements
542  for(; x < window_end_x; ++x)
543  {
544  *(dst_ptr + x) = static_cast<int8_t>(*(src_ptr + x));
545  }
546  },
547  src, dst);
548  }
549  break;
550  }
551  case DataType::U8:
552  {
553  /* Down-conversion S16 -> U8 */
554  if(ConvertPolicy::SATURATE == _policy)
555  {
556  execute_window_loop(win, [&](const Coordinates &)
557  {
558  const auto src_ptr = reinterpret_cast<const int16_t *>(src.ptr());
559  const auto dst_ptr = reinterpret_cast<uint8_t *>(dst.ptr());
560 
561  int x = window_start_x;
562  for(; x <= (window_end_x - window_step_x); x += window_step_x)
563  {
564  const int16x8x2_t texels =
565  {
566  {
567  vld1q_s16(src_ptr + x),
568  vld1q_s16(src_ptr + x + 8)
569  }
570  };
571 
572  vst1q_u8(dst_ptr + x, vcombine_u8(vqmovun_s16(texels.val[0]), vqmovun_s16(texels.val[1])));
573  }
574 
575  // Compute left-over elements
576  for(; x < window_end_x; ++x)
577  {
578  *(dst_ptr + x) = utils::cast::saturate_cast<uint8_t>(*(src_ptr + x));
579  }
580  },
581  src, dst);
582  }
583  else
584  {
585  execute_window_loop(win, [&](const Coordinates &)
586  {
587  const auto src_ptr = reinterpret_cast<const int16_t *>(src.ptr());
588  const auto dst_ptr = reinterpret_cast<uint8_t *>(dst.ptr());
589 
590  int x = window_start_x;
591  for(; x <= (window_end_x - window_step_x); x += window_step_x)
592  {
593  const int16x8x2_t texels =
594  {
595  {
596  vld1q_s16(src_ptr + x),
597  vld1q_s16(src_ptr + x + 8)
598  }
599  };
600 
601  vst1q_u8(dst_ptr + x, vcombine_u8(vmovn_u16(vreinterpretq_u16_s16(texels.val[0])),
602  vmovn_u16(vreinterpretq_u16_s16(texels.val[1]))));
603  }
604 
605  // Compute left-over elements
606  for(; x < window_end_x; ++x)
607  {
608  *(dst_ptr + x) = static_cast<uint8_t>(*(src_ptr + x));
609  }
610  },
611  src, dst);
612  }
613  break;
614  }
615  case DataType::S32:
616  {
617  /* Up-conversion S16 -> S32 */
618  execute_window_loop(win, [&](const Coordinates &)
619  {
620  const auto src_ptr = reinterpret_cast<const int16_t *>(src.ptr());
621  const auto dst_ptr = reinterpret_cast<int32_t *>(dst.ptr());
622 
623  int x = window_start_x;
624  for(; x <= (window_end_x - window_step_x); x += window_step_x)
625  {
626  const int16x8x2_t texels =
627  {
628  {
629  vld1q_s16(src_ptr + x),
630  vld1q_s16(src_ptr + x + 8)
631  }
632  };
633 
634  const int32x4x4_t texels_s32 =
635  {
636  {
637  vmovl_s16(vget_low_s16(texels.val[0])),
638  vmovl_s16(vget_high_s16(texels.val[0])),
639  vmovl_s16(vget_low_s16(texels.val[1])),
640  vmovl_s16(vget_high_s16(texels.val[1]))
641  }
642  };
643 
644  vst1q_s32(dst_ptr + x, texels_s32.val[0]);
645  vst1q_s32(dst_ptr + x + 4, texels_s32.val[1]);
646  vst1q_s32(dst_ptr + x + 8, texels_s32.val[2]);
647  vst1q_s32(dst_ptr + x + 12, texels_s32.val[3]);
648  }
649 
650  // Compute left-over elements
651  for(; x < window_end_x; ++x)
652  {
653  *(dst_ptr + x) = static_cast<int32_t>(*(src_ptr + x));
654  }
655  },
656  src, dst);
657  break;
658  }
659  default:
660  ARM_COMPUTE_ERROR("dst data type not supported");
661  }
662  break;
663  }
664 
665  case DataType::U16:
666  {
667  switch(_dst->info()->data_type())
668  {
669  case DataType::U8:
670  {
671  /* Down-conversion U16 -> U8 */
672  if(ConvertPolicy::SATURATE == _policy)
673  {
674  execute_window_loop(win, [&](const Coordinates &)
675  {
676  const auto src_ptr = reinterpret_cast<const uint16_t *>(src.ptr());
677  const auto dst_ptr = reinterpret_cast<uint8_t *>(dst.ptr());
678 
679  int x = window_start_x;
680  for(; x <= (window_end_x - window_step_x); x += window_step_x)
681  {
682  const uint16x8x2_t texels =
683  {
684  {
685  vld1q_u16(src_ptr + x),
686  vld1q_u16(src_ptr + x + 8)
687  }
688  };
689 
690  vst1q_u8(dst_ptr + x, vcombine_u8(vqmovn_u16(texels.val[0]), vqmovn_u16(texels.val[1])));
691  }
692 
693  // Compute left-over elements
694  for(; x < window_end_x; ++x)
695  {
696  *(dst_ptr + x) = utils::cast::saturate_cast<uint8_t>(*(src_ptr + x));
697  }
698  },
699  src, dst);
700  }
701  else
702  {
703  execute_window_loop(win, [&](const Coordinates &)
704  {
705  const auto src_ptr = reinterpret_cast<const uint16_t *>(src.ptr());
706  const auto dst_ptr = reinterpret_cast<uint8_t *>(dst.ptr());
707 
708  int x = window_start_x;
709  for(; x <= (window_end_x - window_step_x); x += window_step_x)
710  {
711  const uint16x8x2_t texels =
712  {
713  {
714  vld1q_u16(src_ptr + x),
715  vld1q_u16(src_ptr + x + 8)
716  }
717  };
718 
719  vst1q_u8(dst_ptr + x, vcombine_u8(vmovn_u16(texels.val[0]), vmovn_u16(texels.val[1])));
720  }
721 
722  // Compute left-over elements
723  for(; x < window_end_x; ++x)
724  {
725  *(dst_ptr + x) = static_cast<uint8_t>(*(src_ptr + x));
726  }
727 
728  },
729  src, dst);
730  }
731  break;
732  }
733  case DataType::U32:
734  {
735  /* Up-conversion U16 -> U32 */
736  execute_window_loop(win, [&](const Coordinates &)
737  {
738  const auto src_ptr = reinterpret_cast<const uint16_t *>(src.ptr());
739  const auto dst_ptr = reinterpret_cast<uint32_t *>(dst.ptr());
740 
741  int x = window_start_x;
742  for(; x <= (window_end_x - window_step_x); x += window_step_x)
743  {
744  const uint16x8x2_t texels =
745  {
746  {
747  vld1q_u16(src_ptr + x),
748  vld1q_u16(src_ptr + x + 8)
749  }
750  };
751 
752  vst1q_u32(dst_ptr + x, vmovl_u16(vget_low_u16(texels.val[0])));
753  vst1q_u32(dst_ptr + x + 4, vmovl_u16(vget_high_u16(texels.val[0])));
754  vst1q_u32(dst_ptr + x + 8, vmovl_u16(vget_low_u16(texels.val[1])));
755  vst1q_u32(dst_ptr + x + 12, vmovl_u16(vget_high_u16(texels.val[1])));
756  }
757  // Compute left-over elements
758  for(; x < window_end_x; ++x)
759  {
760  *(dst_ptr + x) = static_cast<uint32_t>(*(src_ptr + x));
761  }
762 
763  },
764  src, dst);
765  break;
766  }
767  default:
768  ARM_COMPUTE_ERROR("dst data type not supported");
769  }
770  break;
771  }
772  case DataType::BFLOAT16:
773  {
774  /* Up-conversion BFLOAT16 -> F32 */
775  ARM_COMPUTE_ERROR_ON(uk->ukernel == nullptr);
776  uk->ukernel(_src, _dst, info, _policy, window);
777  break;
778  }
779  case DataType::F16:
780  {
781  /* conversion F16 -> any data type */
782  ARM_COMPUTE_ERROR_ON(uk->ukernel == nullptr);
783  uk->ukernel(_src, _dst, info, _policy, window);
784  break;
785  }
786  case DataType::F32:
787  switch(_dst->info()->data_type())
788  {
789  case DataType::F16:
790  {
791  /* Down-conversion F32 -> F16 */
792  ARM_COMPUTE_ERROR_ON(uk->ukernel == nullptr);
793  uk->ukernel(_src, _dst, info, _policy, window);
794  break;
795  }
796  case DataType::BFLOAT16:
797  {
798  /* Down-conversion F32 -> BFLOAT16 */
799  ARM_COMPUTE_ERROR_ON(uk->ukernel == nullptr);
800  uk->ukernel(_src, _dst, info, _policy, window);
801  break;
802  }
803  case DataType::S32:
804  {
805  /* Conversion F32 -> S32 */
806  execute_window_loop(win, [&](const Coordinates &)
807  {
808  const auto src_ptr = reinterpret_cast<const float *>(src.ptr());
809  const auto dst_ptr = reinterpret_cast<int32_t *>(dst.ptr());
810 
811  int x = window_start_x;
812  for(; x <= (window_end_x - window_step_x); x += window_step_x)
813  {
814  const float32x4x4_t texels =
815  {
816  {
817  vld1q_f32(src_ptr + x),
818  vld1q_f32(src_ptr + x + 4),
819  vld1q_f32(src_ptr + x + 8),
820  vld1q_f32(src_ptr + x + 12),
821  }
822  };
823 
824  vst1q_s32(dst_ptr + x, vcvtq_s32_f32(texels.val[0]));
825  vst1q_s32(dst_ptr + x + 4, vcvtq_s32_f32(texels.val[1]));
826  vst1q_s32(dst_ptr + x + 8, vcvtq_s32_f32(texels.val[2]));
827  vst1q_s32(dst_ptr + x + 12, vcvtq_s32_f32(texels.val[3]));
828  }
829 
830  // Compute left-over elements
831  for(; x < window_end_x; ++x)
832  {
833  *(dst_ptr + x) = static_cast<int32_t>(*(src_ptr + x));
834  }
835  },
836  src, dst);
837  break;
838  }
839  case DataType::QASYMM8:
840  case DataType::U8:
841  {
842  /* Down-conversion F32 -> U8 */
843  execute_window_loop(win, [&](const Coordinates &)
844  {
845  const auto src_ptr = reinterpret_cast<const float *>(src.ptr());
846  const auto dst_ptr = reinterpret_cast<uint8_t *>(dst.ptr());
847 
848  int x = window_start_x;
849  for(; x <= (window_end_x - window_step_x); x += window_step_x)
850  {
851  const float32x4x4_t texels =
852  {
853  {
854  vld1q_f32(src_ptr + x),
855  vld1q_f32(src_ptr + x + 4),
856  vld1q_f32(src_ptr + x + 8),
857  vld1q_f32(src_ptr + x + 12),
858  }
859  };
860 
861  vst1_u8(dst_ptr + x, vqmovn_u16(vcombine_u16(vqmovun_s32(vcvtq_s32_f32(texels.val[0])), vqmovun_s32(vcvtq_s32_f32(texels.val[1])))));
862  vst1_u8(dst_ptr + x + 8, vqmovn_u16(vcombine_u16(vqmovun_s32(vcvtq_s32_f32(texels.val[2])), vqmovun_s32(vcvtq_s32_f32(texels.val[3])))));
863  }
864 
865  // Compute left-over elements
866  for(; x < window_end_x; ++x)
867  {
868  *(dst_ptr + x) = utils::cast::saturate_cast<uint8_t>(*(src_ptr + x));
869  }
870  },
871  src, dst);
872  break;
873  }
874  case DataType::QASYMM8_SIGNED:
875  {
876  /* Down-conversion F32 -> QASYMM8_SIGNED */
877  execute_window_loop(win, [&](const Coordinates &)
878  {
879  const auto src_ptr = reinterpret_cast<const float *>(src.ptr());
880  const auto dst_ptr = reinterpret_cast<int8_t *>(dst.ptr());
881 
882  int x = window_start_x;
883  for(; x <= (window_end_x - window_step_x); x += window_step_x)
884  {
885  const float32x4x4_t texels =
886  {
887  {
888  vld1q_f32(src_ptr + x),
889  vld1q_f32(src_ptr + x + 4),
890  vld1q_f32(src_ptr + x + 8),
891  vld1q_f32(src_ptr + x + 12),
892  }
893  };
894 
895  vst1_s8(dst_ptr + x, vqmovn_s16(vcombine_s16(vqmovn_s32(vcvtq_s32_f32(texels.val[0])), vqmovn_s32(vcvtq_s32_f32(texels.val[1])))));
896  vst1_s8(dst_ptr + x + 8, vqmovn_s16(vcombine_s16(vqmovn_s32(vcvtq_s32_f32(texels.val[2])), vqmovn_s32(vcvtq_s32_f32(texels.val[3])))));
897  }
898  // Compute left-over elements
899  for(; x < window_end_x; ++x)
900  {
901  *(dst_ptr + x) = utils::cast::saturate_cast<int8_t>(*(src_ptr + x));
902  }
903  },
904  src, dst);
905  break;
906  }
907 
908  default:
909  ARM_COMPUTE_ERROR("dst data type not supported");
910  }
911  break;
912 
913  case DataType::S32:
914  switch(_dst->info()->data_type())
915  {
916  case DataType::F16:
917  {
918  /* Down-conversion S32 -> F16 */
919  ARM_COMPUTE_ERROR_ON(uk->ukernel == nullptr);
920  uk->ukernel(_src, _dst, info, _policy, window);
921  break;
922  }
923  case DataType::F32:
924  {
925  /* Conversion S32 -> F32 */
926  execute_window_loop(win, [&](const Coordinates &)
927  {
928  const auto src_ptr = reinterpret_cast<const int32_t *>(src.ptr());
929  const auto dst_ptr = reinterpret_cast<float *>(dst.ptr());
930 
931  int x = window_start_x;
932  for(; x <= (window_end_x - window_step_x); x += window_step_x)
933  {
934  const int32x4x4_t texels =
935  {
936  {
937  vld1q_s32(src_ptr + x),
938  vld1q_s32(src_ptr + x + 4),
939  vld1q_s32(src_ptr + x + 8),
940  vld1q_s32(src_ptr + x + 12),
941  }
942  };
943 
944  vst1q_f32(dst_ptr + x, vcvtq_f32_s32(texels.val[0]));
945  vst1q_f32(dst_ptr + x + 4, vcvtq_f32_s32(texels.val[1]));
946  vst1q_f32(dst_ptr + x + 8, vcvtq_f32_s32(texels.val[2]));
947  vst1q_f32(dst_ptr + x + 12, vcvtq_f32_s32(texels.val[3]));
948  }
949 
950  // Compute left-over elements
951  for(; x < window_end_x; ++x)
952  {
953  *(dst_ptr + x) = static_cast<float>(*(src_ptr + x));
954  }
955  },
956  src, dst);
957  break;
958  }
959  case DataType::QASYMM8_SIGNED:
960  {
961  /* Down-conversion S32 -> QASYMM8_SIGNED */
962  if(ConvertPolicy::SATURATE == _policy)
963  {
964  execute_window_loop(win, [&](const Coordinates &)
965  {
966  const auto src_ptr = reinterpret_cast<const int32_t *>(src.ptr());
967  const auto dst_ptr = reinterpret_cast<int8_t *>(dst.ptr());
968 
969  int x = window_start_x;
970  for(; x <= (window_end_x - window_step_x); x += window_step_x)
971  {
972  const int32x4x4_t texels =
973  {
974  {
975  vld1q_s32(src_ptr + x),
976  vld1q_s32(src_ptr + x + 4),
977  vld1q_s32(src_ptr + x + 8),
978  vld1q_s32(src_ptr + x + 12),
979  }
980  };
981  vst1_s8(dst_ptr + x, vqmovn_s16(vcombine_s16(vqmovn_s32(texels.val[0]), vqmovn_s32(texels.val[1]))));
982  vst1_s8(dst_ptr + x + 8, vqmovn_s16(vcombine_s16(vqmovn_s32(texels.val[2]), vqmovn_s32(texels.val[3]))));
983  }
984 
985  // Compute left-over elements
986  for(; x < window_end_x; ++x)
987  {
988  *(dst_ptr + x) = utils::cast::saturate_cast<int8_t>(*(src_ptr + x));
989  }
990  },
991  src, dst);
992  }
993  else
994  {
995  execute_window_loop(win, [&](const Coordinates &)
996  {
997  const auto src_ptr = reinterpret_cast<const int32_t *>(src.ptr());
998  const auto dst_ptr = reinterpret_cast<int8_t *>(dst.ptr());
999 
1000  int x = window_start_x;
1001  for(; x <= (window_end_x - window_step_x); x += window_step_x)
1002  {
1003  const int32x4x4_t texels =
1004  {
1005  {
1006  vld1q_s32(src_ptr + x),
1007  vld1q_s32(src_ptr + x + 4),
1008  vld1q_s32(src_ptr + x + 8),
1009  vld1q_s32(src_ptr + x + 12)
1010  }
1011  };
1012 
1013  vst1_s8(dst_ptr + x, vmovn_s16(vcombine_s16(vmovn_s32(texels.val[0]), vmovn_s32(texels.val[1]))));
1014  vst1_s8(dst_ptr + x + 8, vmovn_s16(vcombine_s16(vmovn_s32(texels.val[2]), vmovn_s32(texels.val[3]))));
1015  }
1016 
1017  // Compute left-over elements
1018  for(; x < window_end_x; ++x)
1019  {
1020  *(dst_ptr + x) = static_cast<int8_t>(*(src_ptr + x));
1021  }
1022  },
1023  src, dst);
1024  }
1025  break;
1026  }
1027  case DataType::QASYMM8:
1028  case DataType::U8:
1029  {
1030  /* Down-conversion S32 -> U8 */
1031  if(ConvertPolicy::SATURATE == _policy)
1032  {
1033  execute_window_loop(win, [&](const Coordinates &)
1034  {
1035  const auto src_ptr = reinterpret_cast<const int32_t *>(src.ptr());
1036  const auto dst_ptr = reinterpret_cast<uint8_t *>(dst.ptr());
1037 
1038  int x = window_start_x;
1039  for(; x <= (window_end_x - window_step_x); x += window_step_x)
1040  {
1041  const int32x4x4_t texels =
1042  {
1043  {
1044  vld1q_s32(src_ptr + x),
1045  vld1q_s32(src_ptr + x + 4),
1046  vld1q_s32(src_ptr + x + 8),
1047  vld1q_s32(src_ptr + x + 12)
1048  }
1049  };
1050  vst1_u8(dst_ptr + x, vqmovn_u16(vcombine_u16(vqmovun_s32(texels.val[0]), vqmovun_s32(texels.val[1]))));
1051  vst1_u8(dst_ptr + x + 8, vqmovn_u16(vcombine_u16(vqmovun_s32(texels.val[2]), vqmovun_s32(texels.val[3]))));
1052  }
1053 
1054  // Compute left-over elements
1055  for(; x < window_end_x; ++x)
1056  {
1057  *(dst_ptr + x) = utils::cast::saturate_cast<uint8_t>(*(src_ptr + x));
1058  }
1059  },
1060  src, dst);
1061  }
1062  else
1063  {
1064  execute_window_loop(win, [&](const Coordinates &)
1065  {
1066  const auto src_ptr = reinterpret_cast<const int32_t *>(src.ptr());
1067  const auto dst_ptr = reinterpret_cast<uint8_t *>(dst.ptr());
1068 
1069  int x = window_start_x;
1070  for(; x <= (window_end_x - window_step_x); x += window_step_x)
1071  {
1072  const int32x4x4_t texels =
1073  {
1074  {
1075  vld1q_s32(src_ptr + x),
1076  vld1q_s32(src_ptr + x + 4),
1077  vld1q_s32(src_ptr + x + 8),
1078  vld1q_s32(src_ptr + x + 12)
1079  }
1080  };
1081 
1082  vst1_u8(dst_ptr + x, vmovn_u16(vcombine_u16(vmovn_u32(vreinterpretq_u32_s32(texels.val[0])), vmovn_u32(vreinterpretq_u32_s32(texels.val[1])))));
1083  vst1_u8(dst_ptr + x + 8, vmovn_u16(vcombine_u16(vmovn_u32(vreinterpretq_u32_s32(texels.val[2])), vmovn_u32(vreinterpretq_u32_s32(texels.val[3])))));
1084  }
1085 
1086  // Compute left-over elements
1087  for(; x < window_end_x; ++x)
1088  {
1089  *(dst_ptr + x) = static_cast<uint8_t>(*(src_ptr + x));
1090  }
1091  },
1092  src, dst);
1093  }
1094  break;
1095  }
1096  default:
1097  ARM_COMPUTE_ERROR("dst data type not supported");
1098  }
1099  break;
1100  default:
1101  ARM_COMPUTE_ERROR("Not supported");
1102  }
1103 }
arm_compute::IKernel::window
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
arm_compute::cpu::ICpuKernel< CpuCastKernel >::get_implementation
static const auto * get_implementation(const SelectorType &selector, KernelSelectionType selection_type=KernelSelectionType::Supported)
Micro-kernel selector.
Definition: ICpuKernel.h:53
arm_compute::ACL_DST
Definition: Types.h:55
ARM_COMPUTE_ERROR
#define ARM_COMPUTE_ERROR(msg)
Print the given message then throw an std::runtime_error.
Definition: Error.h:352
arm_compute::Format::U8
1 channel, 1 U8 per channel
arm_compute::Format::F32
1 channel, 1 F32 per channel
arm_compute::ACL_SRC
Definition: Types.h:44
ARM_COMPUTE_ERROR_ON
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
arm_compute::Format::U16
1 channel, 1 U16 per channel
arm_compute::test::validation::src
SimpleTensor< float > src
Definition: DFT.cpp:155
arm_compute::Format::F16
1 channel, 1 F16 per channel
arm_compute::Format::S32
1 channel, 1 S32 per channel
arm_compute::Format::BFLOAT16
16-bit brain floating-point number
arm_compute::Window::DimX
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
ARM_COMPUTE_UNUSED
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
arm_compute::Format::U32
1 channel, 1 U32 per channel
arm_compute::DataType::QASYMM8
quantized, asymmetric fixed-point 8-bit number unsigned
arm_compute::test::validation::dst
auto dst
Definition: DFT.cpp:170
arm_compute::Window::set
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:915
arm_compute::Format::S16
1 channel, 1 S16 per channel
arm_compute::test::validation::info
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
ARM_COMPUTE_ERROR_ON_NULLPTR
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:157
arm_compute::execute_window_loop
void execute_window_loop(const Window &w, L &&lambda_function, Ts &&... iterators)
Iterate through the passed window, automatically adjusting the iterators and calling the lambda_funct...
Definition: Helpers.inl:77
arm_compute::DataType::QASYMM8_SIGNED
quantized, asymmetric fixed-point 8-bit number signed
arm_compute::CPUInfo::get
static CPUInfo & get()
Access the KernelLibrary singleton.
Definition: CPPTypes.cpp:40
arm_compute::Window::Dimension::end
constexpr int end() const
Return the end of the dimension.
Definition: Window.h:101
arm_compute::ConvertPolicy::SATURATE
Saturate.
arm_compute::Window::Dimension::start
constexpr int start() const
Return the start of the dimension.
Definition: Window.h:96
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:201
arm_compute::CPUInfo::get_isa
cpuinfo::CpuIsaInfo get_isa() const
Gets the current cpu&#39;s ISA information.
Definition: CPPTypes.cpp:114
arm_compute::Window::x
constexpr const Dimension & x() const
Alias to access the first dimension of the window.
Definition: Window.h:158

◆ validate()

Status validate ( const ITensorInfo *  src,
const ITensorInfo *  dst,
ConvertPolicy  policy 
)
static

Static function to check if given info will lead to a valid configuration.

Similar to CpuCastKernel::configure()

Returns
a status

Definition at line 172 of file CpuCastKernel.cpp.

References ARM_COMPUTE_RETURN_ON_ERROR, and arm_compute::cpu::kernels::validate_arguments().

Referenced by CpuCast::validate().

173 {
174  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, policy));
175  return Status{};
176 }
ARM_COMPUTE_RETURN_ON_ERROR
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
arm_compute::cpu::kernels::validate_arguments
Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const PadStrideInfo &conv_info)
Definition: CpuDirectConv2dKernel.cpp:60
arm_compute::test::validation::src
SimpleTensor< float > src
Definition: DFT.cpp:155
arm_compute::test::validation::dst
auto dst
Definition: DFT.cpp:170
The documentation for this class was generated from the following files: