24 #ifndef SRC_CORE_NEON_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H 25 #define SRC_CORE_NEON_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H 36 const float32x4x4_t out =
39 vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(x))))),
offset)),
scale),
40 vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(x))))),
offset)),
scale),
41 vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(x))))),
offset)),
scale),
42 vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(x))))),
offset)),
scale),
51 const float32x4x4_t out =
54 vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(x)))),
offset)),
scale),
55 vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(x)))),
offset)),
scale),
56 vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(x)))),
offset)),
scale),
57 vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(x)))),
offset)),
scale),
65 const uint8x8_t pa = vqmovn_u16(vcombine_u16(vqmovn_u32(out.val[0]), vqmovn_u32(out.val[1])));
66 const uint8x8_t pb = vqmovn_u16(vcombine_u16(vqmovn_u32(out.val[2]), vqmovn_u32(out.val[3])));
67 vst1q_u8(output_ptr, vcombine_u8(pa, pb));
72 const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(out.val[0]), vqmovn_s32(out.val[1])));
73 const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(out.val[2]), vqmovn_s32(out.val[3])));
74 vst1q_u8(output_ptr, vcombine_u8(pa, pb));
77 void store_quantized(uint8_t *output_ptr,
const float32x4x4_t &rf,
const float32x4_t &
offset,
const float32x4_t &invscale)
82 vcvtq_s32_f32(vmlaq_f32(
offset, rf.val[0], invscale)),
83 vcvtq_s32_f32(vmlaq_f32(
offset, rf.val[1], invscale)),
84 vcvtq_s32_f32(vmlaq_f32(
offset, rf.val[2], invscale)),
85 vcvtq_s32_f32(vmlaq_f32(
offset, rf.val[3], invscale)),
93 const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(out.val[0]), vqmovn_s32(out.val[1])));
94 const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(out.val[2]), vqmovn_s32(out.val[3])));
95 vst1q_s8(output_ptr, vcombine_s8(pa, pb));
103 vcvtq_s32_f32(vmlaq_f32(
offset, rf.val[0], invscale)),
104 vcvtq_s32_f32(vmlaq_f32(
offset, rf.val[1], invscale)),
105 vcvtq_s32_f32(vmlaq_f32(
offset, rf.val[2], invscale)),
106 vcvtq_s32_f32(vmlaq_f32(
offset, rf.val[3], invscale)),
112 template <ArithmeticOperation op>
118 template <ArithmeticOperation op>
124 template <ArithmeticOperation op>
131 elementwise_arithm_op<op, neon_vector_float>(a.val[0],
b.val[0]),
132 elementwise_arithm_op<op, neon_vector_float>(a.val[1],
b.val[1]),
133 elementwise_arithm_op<op, neon_vector_float>(a.val[2],
b.val[2]),
134 elementwise_arithm_op<op, neon_vector_float>(a.val[3],
b.val[3]),
140 template <ComparisonOperation op>
144 return elementwise_comp_op_scalar<op>(a,
b);
147 template <ComparisonOperation op>
153 elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[0],
b.val[0]),
154 elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[1],
b.val[1]),
155 elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[2],
b.val[2]),
156 elementwise_comp_op<op, float32x4_t, uint32x4_t>(a.val[3],
b.val[3])
162 template <ArithmeticOperation op>
164 const uint8_t *input1_ptr,
const uint8_t *input2_ptr, uint8_t *output_ptr,
165 int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2,
166 float32x4_t voffseto, float32x4_t invvscaleo)
168 int x = window_start_x;
169 for(; x <= (window_end_x - window_step_x); x += window_step_x)
172 const float32x4x4_t af =
load_quantized(input1_ptr + x, voffset1, vscale1);
173 const float32x4x4_t bf =
load_quantized(input2_ptr + x, voffset2, vscale2);
174 const float32x4x4_t rf = elementwise_arithm_op<op>(af, bf);
180 template <ArithmeticOperation op>
182 const int8_t *input1_ptr,
const int8_t *input2_ptr, int8_t *output_ptr,
183 int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2,
184 float32x4_t voffseto, float32x4_t invvscaleo)
186 int x = window_start_x;
187 for(; x <= (window_end_x - window_step_x); x += window_step_x)
192 const float32x4x4_t rf = elementwise_arithm_op<op>(af, bf);
198 template <ArithmeticOperation op>
200 const uint8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr,
201 int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast,
202 float32x4_t voffseto, float32x4_t invvscaleo,
bool reorder)
204 int x = window_start_x;
205 for(; x <= (window_end_x - window_step_x); x += window_step_x)
207 const float32x4x4_t af =
load_quantized(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast);
208 const float32x4x4_t rf = elementwise_arithm_op<op>(
reorder ? broadcast_vector : af,
reorder ? af : broadcast_vector);
213 template <ArithmeticOperation op>
215 const int8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, int8_t *output_ptr,
216 int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast,
217 float32x4_t voffseto, float32x4_t invvscaleo,
bool reorder)
219 int x = window_start_x;
220 for(; x <= (window_end_x - window_step_x); x += window_step_x)
222 const float32x4x4_t af =
load_quantized_signed(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast);
223 const float32x4x4_t rf = elementwise_arithm_op<op>(
reorder ? broadcast_vector : af,
reorder ? af : broadcast_vector);
229 template <ComparisonOperation op>
231 const uint8_t *input1_ptr,
const uint8_t *input2_ptr, uint8_t *output_ptr,
232 int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2,
233 float32x4_t voffseto, float32x4_t invvscaleo)
236 int x = window_start_x;
237 for(; x <= (window_end_x - window_step_x); x += window_step_x)
239 const float32x4x4_t af =
load_quantized(input1_ptr + x, voffset1, vscale1);
240 const float32x4x4_t bf =
load_quantized(input2_ptr + x, voffset2, vscale2);
241 const uint32x4x4_t rf = elementwise_comp_op<op>(af, bf);
247 template <ComparisonOperation op>
249 const int8_t *input1_ptr,
const int8_t *input2_ptr, uint8_t *output_ptr,
250 int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2,
251 float32x4_t voffseto, float32x4_t invvscaleo)
254 int x = window_start_x;
255 for(; x <= (window_end_x - window_step_x); x += window_step_x)
259 const uint32x4x4_t rf = elementwise_comp_op<op>(af, bf);
265 template <ComparisonOperation op>
267 const uint8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr,
268 int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast,
269 float32x4_t voffseto, float32x4_t invvscaleo,
bool reorder)
272 int x = window_start_x;
273 for(; x <= (window_end_x - window_step_x); x += window_step_x)
275 const float32x4x4_t af =
load_quantized(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast);
276 const uint32x4x4_t rf = elementwise_comp_op<op>(
reorder ? broadcast_vector : af,
reorder ? af : broadcast_vector);
282 template <ComparisonOperation op>
284 const int8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr,
285 int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast,
286 float32x4_t voffseto, float32x4_t invvscaleo,
bool reorder)
289 int x = window_start_x;
290 for(; x <= (window_end_x - window_step_x); x += window_step_x)
292 const float32x4x4_t af =
load_quantized_signed(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast);
293 const uint32x4x4_t rf = elementwise_comp_op<op>(
reorder ? broadcast_vector : af,
reorder ? af : broadcast_vector);
301 int (*broadcast_func)(
int,
int,
int,
const uint8_t *, float32x4x4_t, uint8_t *, int32x4_t, float32x4_t,
302 float32x4_t, float32x4_t,
const bool),
303 int (*neon_func)(
int,
int,
int,
const uint8_t *,
const uint8_t *, uint8_t *,
304 int32x4_t, int32x4_t, float32x4_t, float32x4_t,
305 float32x4_t, float32x4_t))
315 const int window_step_x = 16;
316 const auto window_start_x = static_cast<int>(window.
x().
start());
317 const auto window_end_x = static_cast<int>(window.
x().
end());
323 const float32x4_t voffseto = vdupq_n_f32(output_qinfo.
offset + 0.5f);
324 const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.
scale);
326 if(is_broadcast_across_x)
329 const bool is_broadcast_input_2 = input2_win.
x().
step() == 0;
330 Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win;
331 Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win;
332 const ITensor *broadcast_tensor = is_broadcast_input_2 ? in2 : in1;
333 const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
338 const int32x4_t voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.
offset);
339 const float32x4_t vscale_non_broadcast = vdupq_n_f32(non_broadcast_qinfo.
scale);
344 Iterator broadcast_input(broadcast_tensor, broadcast_win);
345 Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
350 const auto non_broadcast_input_ptr = reinterpret_cast<const uint8_t *>(non_broadcast_input.
ptr());
351 const auto output_ptr = reinterpret_cast<uint8_t *>(output.
ptr());
353 const uint8_t broadcast_value = *reinterpret_cast<const uint8_t *>(broadcast_input.
ptr());
354 const float32x4x4_t broadcast_vector =
vdequantize(vdupq_n_u8(broadcast_value), broadcast_qinfo);
356 int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr,
357 voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2);
358 for(; x < window_end_x; ++x)
360 const float afs =
dequantize_qasymm8(*(non_broadcast_input_ptr + x), non_broadcast_qinfo);
362 *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ?
bfs : afs, !is_broadcast_input_2 ? afs :
bfs, output_qinfo);
365 broadcast_input, non_broadcast_input, output);
373 const int32x4_t voffset1 = vdupq_n_s32(input1_qinfo.
offset);
374 const float32x4_t vscale1 = vdupq_n_f32(input1_qinfo.
scale);
377 const int32x4_t voffset2 = vdupq_n_s32(input2_qinfo.
offset);
378 const float32x4_t vscale2 = vdupq_n_f32(input2_qinfo.
scale);
390 const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.
ptr());
391 const auto input2_ptr = reinterpret_cast<const uint8_t *>(input2.
ptr());
392 const auto output_ptr = reinterpret_cast<uint8_t *>(output.
ptr());
394 int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr, voffset1, voffset2,
395 vscale1, vscale2, voffseto, invvscaleo);
396 for(; x < window_end_x; ++x)
400 *(output_ptr + x) = (*scalar_func)(afs,
bfs, output_qinfo);
403 input1, input2, output);
409 int (*broadcast_func)(
int,
int,
int,
const int8_t *, float32x4x4_t, uint8_t *, int32x4_t, float32x4_t,
410 float32x4_t, float32x4_t,
const bool),
411 int (*neon_func)(
int,
int,
int,
const int8_t *,
const int8_t *, uint8_t *,
412 int32x4_t, int32x4_t, float32x4_t, float32x4_t,
413 float32x4_t, float32x4_t))
423 const int window_step_x = 16;
424 const auto window_start_x = static_cast<int>(window.
x().
start());
425 const auto window_end_x = static_cast<int>(window.
x().
end());
430 const float32x4_t voffseto = vdupq_n_f32(output_qinfo.
offset);
431 const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.
scale);
433 if(is_broadcast_across_x)
436 const bool is_broadcast_input_2 = input2_win.
x().
step() == 0;
437 Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win;
438 Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win;
439 const ITensor *broadcast_tensor = is_broadcast_input_2 ? in2 : in1;
440 const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
445 const int32x4_t voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.
offset);
446 const float32x4_t vscale_non_broadcast = vdupq_n_f32(non_broadcast_qinfo.
scale);
451 Iterator broadcast_input(broadcast_tensor, broadcast_win);
452 Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
457 const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.
ptr());
458 const auto output_ptr = reinterpret_cast<uint8_t *>(output.
ptr());
460 const int8_t broadcast_value = *reinterpret_cast<const int8_t *>(broadcast_input.
ptr());
461 const float32x4x4_t broadcast_vector =
vdequantize(vdupq_n_s8(broadcast_value), broadcast_qinfo);
463 int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr,
464 voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2);
465 for(; x < window_end_x; ++x)
469 *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ?
bfs : afs, !is_broadcast_input_2 ? afs :
bfs, output_qinfo);
472 broadcast_input, non_broadcast_input, output);
480 const int32x4_t voffset1 = vdupq_n_s32(input1_qinfo.
offset);
481 const float32x4_t vscale1 = vdupq_n_f32(input1_qinfo.
scale);
484 const int32x4_t voffset2 = vdupq_n_s32(input2_qinfo.
offset);
485 const float32x4_t vscale2 = vdupq_n_f32(input2_qinfo.
scale);
497 const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.
ptr());
498 const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.
ptr());
499 const auto output_ptr = reinterpret_cast<uint8_t *>(output.
ptr());
501 int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr, voffset1, voffset2,
502 vscale1, vscale2, voffseto, invvscaleo);
503 for(; x < window_end_x; ++x)
507 *(output_ptr + x) = (*scalar_func)(afs,
bfs, output_qinfo);
510 input1, input2, output);
516 int (*broadcast_func)(
int,
int,
int,
const int8_t *, float32x4x4_t, int8_t *, int32x4_t, float32x4_t,
517 float32x4_t, float32x4_t,
const bool),
518 int (*neon_func)(
int,
int,
int,
const int8_t *,
const int8_t *, int8_t *,
519 int32x4_t, int32x4_t, float32x4_t, float32x4_t,
520 float32x4_t, float32x4_t))
530 const int window_step_x = 16;
531 const auto window_start_x = static_cast<int>(window.
x().
start());
532 const auto window_end_x = static_cast<int>(window.
x().
end());
537 const float32x4_t voffseto = vdupq_n_f32(output_qinfo.
offset);
538 const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.
scale);
540 if(is_broadcast_across_x)
543 const bool is_broadcast_input_2 = input2_win.
x().
step() == 0;
544 Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win;
545 Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win;
546 const ITensor *broadcast_tensor = is_broadcast_input_2 ? in2 : in1;
547 const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
552 const int32x4_t voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.
offset);
553 const float32x4_t vscale_non_broadcast = vdupq_n_f32(non_broadcast_qinfo.
scale);
558 Iterator broadcast_input(broadcast_tensor, broadcast_win);
559 Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
564 const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.
ptr());
565 const auto output_ptr = reinterpret_cast<int8_t *>(output.
ptr());
567 const int8_t broadcast_value = *reinterpret_cast<const int8_t *>(broadcast_input.
ptr());
568 const float32x4x4_t broadcast_vector =
vdequantize(vdupq_n_s8(broadcast_value), broadcast_qinfo);
570 int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr,
571 voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2);
572 for(; x < window_end_x; ++x)
576 *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ?
bfs : afs, !is_broadcast_input_2 ? afs :
bfs, output_qinfo);
579 broadcast_input, non_broadcast_input, output);
587 const int32x4_t voffset1 = vdupq_n_s32(input1_qinfo.
offset);
588 const float32x4_t vscale1 = vdupq_n_f32(input1_qinfo.
scale);
591 const int32x4_t voffset2 = vdupq_n_s32(input2_qinfo.
offset);
592 const float32x4_t vscale2 = vdupq_n_f32(input2_qinfo.
scale);
604 const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.
ptr());
605 const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.
ptr());
606 const auto output_ptr = reinterpret_cast<int8_t *>(output.
ptr());
608 int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr, voffset1, voffset2,
609 vscale1, vscale2, voffseto, invvscaleo);
610 for(; x < window_end_x; ++x)
614 *(output_ptr + x) = (*scalar_func)(afs,
bfs, output_qinfo);
617 input1, input2, output);
621 template <ArithmeticOperation op>
625 &elementwise_arithm_op_quantized_broadcast_loop<op>,
626 &elementwise_arithm_op_quantized_loop<op>);
628 template <ArithmeticOperation op>
632 &elementwise_arithm_op_quantized_signed_broadcast_loop<op>,
633 &elementwise_arithm_op_quantized_singed_loop<op>);
636 template <ComparisonOperation op>
640 &elementwise_comp_op_quantized_broadcast_loop<op>,
641 &elementwise_comp_op_quantized_loop<op>);
644 template <ComparisonOperation op>
648 &elementwise_comp_op_quantized_signed_broadcast_loop<op>,
649 &elementwise_comp_op_quantized_signed_loop<op>);
__global uchar * offset(const Image *img, int x, int y)
Get the pointer position of a Image.
int elementwise_comp_op_quantized_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, const uint8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr, int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast, float32x4_t voffseto, float32x4_t invvscaleo, bool reorder)
int elementwise_arithm_op_quantized_singed_loop(int window_start_x, int window_end_x, int window_step_x, const int8_t *input1_ptr, const int8_t *input2_ptr, int8_t *output_ptr, int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2, float32x4_t voffseto, float32x4_t invvscaleo)
float32x4x4_t load_quantized(const uint8_t *input1_ptr, const int32x4_t &offset, const float32x4_t &scale)
float dequantize_qasymm8(uint8_t value, const INFO_TYPE &qinfo)
Dequantize a value given an unsigned 8-bit asymmetric quantization scheme.
int elementwise_arithm_op_quantized_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, const uint8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr, int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast, float32x4_t voffseto, float32x4_t invvscaleo, bool reorder)
uint8_t quantize_qasymm8(float value, const INFO_TYPE &qinfo, RoundingPolicy rounding_policy=RoundingPolicy::TO_NEAREST_UP)
Quantize a value given an unsigned 8-bit asymmetric quantization scheme.
constexpr int step() const
Return the step of the dimension.
float32x4x2_t vdequantize(const uint8x8_t &qv, const UniformQuantizationInfo &qi)
Dequantize a neon vector holding 8 quantized values.
int elementwise_comp_op_quantized_loop(int window_start_x, int window_end_x, int window_step_x, const uint8_t *input1_ptr, const uint8_t *input2_ptr, uint8_t *output_ptr, int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2, float32x4_t voffseto, float32x4_t invvscaleo)
Describe one of the image's dimensions with a start, end and step.
Interface for CPU tensor.
Copyright (c) 2017-2021 Arm Limited.
int8_t elementwise_arithm_op_quantized_signed_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo)
T x() const
Alias to access the size of the first dimension.
VectorType::type elementwise_arithm_op(const typename VectorType::type &a, const typename VectorType::type &b)
uint8_t elementwise_arithm_op_quantized_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo)
void elementwise_arithm_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Create the appropriate SIMD vector given its type and size in terms of elements.
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
int8_t quantize_qasymm8_signed(float value, const INFO_TYPE &qinfo, RoundingPolicy rounding_policy=RoundingPolicy::TO_NEAREST_UP)
Quantize a value given a signed 8-bit asymmetric quantization scheme.
UniformQuantizationInfo uniform() const
Return per layer quantization info.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
constexpr uint8_t * ptr() const
Return a pointer to the current pixel.
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
int elementwise_arithm_op_quantized_signed_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, const int8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, int8_t *output_ptr, int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast, float32x4_t voffseto, float32x4_t invvscaleo, bool reorder)
Window broadcast_if_dimension_le_one(const TensorShape &shape) const
Don't advance in the dimension where shape is less equal to 1.
OutputVectorType elementwise_comp_op(const InputVectorType &a, const InputVectorType &b)
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
void elementwise_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, int8_t(*scalar_func)(const float &, const float &, UniformQuantizationInfo), int(*broadcast_func)(int, int, int, const int8_t *, float32x4x4_t, int8_t *, int32x4_t, float32x4_t, float32x4_t, float32x4_t, const bool), int(*neon_func)(int, int, int, const int8_t *, const int8_t *, int8_t *, int32x4_t, int32x4_t, float32x4_t, float32x4_t, float32x4_t, float32x4_t))
void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, uint8_t(*scalar_func)(const float &, const float &, UniformQuantizationInfo), int(*broadcast_func)(int, int, int, const uint8_t *, float32x4x4_t, uint8_t *, int32x4_t, float32x4_t, float32x4_t, float32x4_t, const bool), int(*neon_func)(int, int, int, const uint8_t *, const uint8_t *, uint8_t *, int32x4_t, int32x4_t, float32x4_t, float32x4_t, float32x4_t, float32x4_t))
int elementwise_comp_op_quantized_signed_loop(int window_start_x, int window_end_x, int window_step_x, const int8_t *input1_ptr, const int8_t *input2_ptr, uint8_t *output_ptr, int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2, float32x4_t voffseto, float32x4_t invvscaleo)
int elementwise_arithm_op_quantized_loop(int window_start_x, int window_end_x, int window_step_x, const uint8_t *input1_ptr, const uint8_t *input2_ptr, uint8_t *output_ptr, int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2, float32x4_t voffseto, float32x4_t invvscaleo)
void elementwise_comp_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
float32x4x4_t load_quantized_signed(const int8_t *input1_ptr, const int32x4_t &offset, const float32x4_t &scale)
const QuantizationInfo qinfo
float dequantize_qasymm8_signed(int8_t value, const INFO_TYPE &qinfo)
Dequantize a value given a signed 8-bit asymmetric quantization scheme.
void store_quantized_signed(int8_t *output_ptr, const int32x4x4_t &out)
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...
void elementwise_comp_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
std::vector< NodeID > bfs(Graph &g)
Breadth first search traversal.
int8x16_t qasymm8x16_signed_t
8 bit quantized signed asymmetric vector with 16 elements
void elementwise_comp_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, uint8_t(*scalar_func)(const float &, const float &, UniformQuantizationInfo), int(*broadcast_func)(int, int, int, const int8_t *, float32x4x4_t, uint8_t *, int32x4_t, float32x4_t, float32x4_t, float32x4_t, const bool), int(*neon_func)(int, int, int, const int8_t *, const int8_t *, uint8_t *, int32x4_t, int32x4_t, float32x4_t, float32x4_t, float32x4_t, float32x4_t))
uint8_t elementwise_comp_op_quantized_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo)
constexpr int end() const
Return the end of the dimension.
Iterator updated by execute_window_loop for each window element.
constexpr int start() const
Return the start of the dimension.
Describe a multidimensional execution window.
void store_quantized(uint8_t *output_ptr, const uint32x4x4_t &out)
void elementwise_arithm_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
Sets the macro arm_any if compiling for Aarch32 or Aarch64.
constexpr const Dimension & x() const
Alias to access the first dimension of the window.
uint8x16_t qasymm8x16_t
8 bit quantized asymmetric vector with 16 elements
int elementwise_comp_op_quantized_signed_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, const int8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr, int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast, float32x4_t voffseto, float32x4_t invvscaleo, bool reorder)