Home
last modified time | relevance | path

Searched refs:vbicq_s32 (Results 1 – 25 of 121) sorted by relevance

12345

/external/XNNPACK/src/q8-vadd/
Dneon.c68 vacc0_lo = vsraq_n_s32(vacc0_lo, vbicq_s32(vacc0_lo, vzero_shift_mask), 31); in xnn_q8_vadd_ukernel__neon()
69 vacc1_lo = vsraq_n_s32(vacc1_lo, vbicq_s32(vacc1_lo, vzero_shift_mask), 31); in xnn_q8_vadd_ukernel__neon()
70 vacc2_lo = vsraq_n_s32(vacc2_lo, vbicq_s32(vacc2_lo, vzero_shift_mask), 31); in xnn_q8_vadd_ukernel__neon()
71 vacc3_lo = vsraq_n_s32(vacc3_lo, vbicq_s32(vacc3_lo, vzero_shift_mask), 31); in xnn_q8_vadd_ukernel__neon()
72 vacc0_hi = vsraq_n_s32(vacc0_hi, vbicq_s32(vacc0_hi, vzero_shift_mask), 31); in xnn_q8_vadd_ukernel__neon()
73 vacc1_hi = vsraq_n_s32(vacc1_hi, vbicq_s32(vacc1_hi, vzero_shift_mask), 31); in xnn_q8_vadd_ukernel__neon()
74 vacc2_hi = vsraq_n_s32(vacc2_hi, vbicq_s32(vacc2_hi, vzero_shift_mask), 31); in xnn_q8_vadd_ukernel__neon()
75 vacc3_hi = vsraq_n_s32(vacc3_hi, vbicq_s32(vacc3_hi, vzero_shift_mask), 31); in xnn_q8_vadd_ukernel__neon()
129 vacc0_lo = vsraq_n_s32(vacc0_lo, vbicq_s32(vacc0_lo, vzero_shift_mask), 31); in xnn_q8_vadd_ukernel__neon()
130 vacc1_lo = vsraq_n_s32(vacc1_lo, vbicq_s32(vacc1_lo, vzero_shift_mask), 31); in xnn_q8_vadd_ukernel__neon()
[all …]
/external/XNNPACK/src/requantization/
Dq31-neon.c71 const int32x4_t x_adjusted_product = vsraq_n_s32(x_product, vbicq_s32(x, vshift_eq_0_mask), 31); in xnn_requantize_q31__neon()
72 const int32x4_t y_adjusted_product = vsraq_n_s32(y_product, vbicq_s32(y, vshift_eq_0_mask), 31); in xnn_requantize_q31__neon()
73 const int32x4_t z_adjusted_product = vsraq_n_s32(z_product, vbicq_s32(z, vshift_eq_0_mask), 31); in xnn_requantize_q31__neon()
74 const int32x4_t w_adjusted_product = vsraq_n_s32(w_product, vbicq_s32(w, vshift_eq_0_mask), 31); in xnn_requantize_q31__neon()
/external/XNNPACK/src/f32-sigmoid/gen/
Dneonfma-rr1-lut64-p2-div-x24.c86 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
87 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
88 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
89 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
90 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
91 …const int32x4_t veKLMN = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnKLMN), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
275 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
354 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
Dneonfma-rr1-lut2048-p1-div-x24.c86 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
87 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
88 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
89 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
90 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
91 …const int32x4_t veKLMN = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnKLMN), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
267 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x7FF)))… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
344 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x7FF)))… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
Dneonfma-rr1-lut2048-p1-div-x20.c83 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
84 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
85 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
86 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
87 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
244 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x7FF)))… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
321 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x7FF)))… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
Dneonfma-rr1-lut2048-p1-div-x16.c80 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
81 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
82 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
83 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
221 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x7FF)))… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
298 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x7FF)))… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
Dneonfma-rr1-lut64-p2-div-x16.c80 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
81 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
82 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
83 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
227 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
306 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
Dneonfma-rr1-lut64-p2-div-x20.c83 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
84 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
85 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
86 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
87 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
251 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
330 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
Dneonfma-rr1-lut64-p2-nr1recps1fma-x24.c86 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
87 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
88 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
89 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
90 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
91 …const int32x4_t veKLMN = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnKLMN), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
299 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
387 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
Dneonfma-rr1-lut2048-p1-nr2fma-x24.c86 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
87 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
88 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
89 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
90 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
91 …const int32x4_t veKLMN = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnKLMN), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
291 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x7FF)))… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
377 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x7FF)))… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
Dneonfma-rr1-lut64-p2-nr2fma-x24.c86 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
87 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
88 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
89 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
90 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
91 …const int32x4_t veKLMN = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnKLMN), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
299 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
387 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
Dneonfma-rr1-lut2048-p1-nr2recps-x24.c86 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
87 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
88 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
89 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
90 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
91 …const int32x4_t veKLMN = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnKLMN), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
291 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x7FF)))… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
377 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x7FF)))… in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Dneonfma-lut64-p2-x20.c81 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
82 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
83 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
84 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
85 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
233 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
309 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
Dneonfma-lut64-p2-x20-acc5.c85 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
86 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
87 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
88 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
89 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
242 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
318 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
Dneonfma-lut64-p2-x20-acc2.c82 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
83 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
84 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
85 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
86 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
236 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
312 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
Dneon-lut64-p2-x20-acc2.c83 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
84 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
85 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
86 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
87 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
237 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
313 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
Dneon-lut64-p2-x20.c82 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
83 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
84 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
85 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
86 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
234 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
310 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
Dneon-lut64-p2-x20-acc5.c86 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
87 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
88 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
89 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
90 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
243 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
319 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
Dneonfma-lut64-p2-x16-acc4.c81 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
82 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
83 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
84 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
218 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
294 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
Dneon-lut64-p2-x16-acc2.c80 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
81 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
82 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
83 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
215 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
291 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
Dneon-lut64-p2-x16.c79 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
80 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
81 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
82 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
212 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
288 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
Dneonfma-lut64-p2-x16-acc2.c79 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
80 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
81 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
82 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
214 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
290 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
Dneon-lut64-p2-x16-acc4.c82 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
83 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
84 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
85 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
219 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
295 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
Dneonfma-lut64-p2-x16.c78 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
79 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
80 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
81 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
211 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
287 …const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))),… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
/external/XNNPACK/src/q8-dwconv/
Dup8x9-neon.c187 vacc0_lo = vsraq_n_s32(vacc0_lo, vbicq_s32(vacc0_lo, vzero_shift_mask), 31); in xnn_q8_dwconv_ukernel_up8x9__neon()
188 vacc0_hi = vsraq_n_s32(vacc0_hi, vbicq_s32(vacc0_hi, vzero_shift_mask), 31); in xnn_q8_dwconv_ukernel_up8x9__neon()
189 vacc1_lo = vsraq_n_s32(vacc1_lo, vbicq_s32(vacc1_lo, vzero_shift_mask), 31); in xnn_q8_dwconv_ukernel_up8x9__neon()
190 vacc1_hi = vsraq_n_s32(vacc1_hi, vbicq_s32(vacc1_hi, vzero_shift_mask), 31); in xnn_q8_dwconv_ukernel_up8x9__neon()
191 vacc2_lo = vsraq_n_s32(vacc2_lo, vbicq_s32(vacc2_lo, vzero_shift_mask), 31); in xnn_q8_dwconv_ukernel_up8x9__neon()
192 vacc2_hi = vsraq_n_s32(vacc2_hi, vbicq_s32(vacc2_hi, vzero_shift_mask), 31); in xnn_q8_dwconv_ukernel_up8x9__neon()
345 vacc0_lo = vsraq_n_s32(vacc0_lo, vbicq_s32(vacc0_lo, vzero_shift_mask), 31); in xnn_q8_dwconv_ukernel_up8x9__neon()
346 vacc0_hi = vsraq_n_s32(vacc0_hi, vbicq_s32(vacc0_hi, vzero_shift_mask), 31); in xnn_q8_dwconv_ukernel_up8x9__neon()
347 vacc1_lo = vsraq_n_s32(vacc1_lo, vbicq_s32(vacc1_lo, vzero_shift_mask), 31); in xnn_q8_dwconv_ukernel_up8x9__neon()
348 vacc1_hi = vsraq_n_s32(vacc1_hi, vbicq_s32(vacc1_hi, vzero_shift_mask), 31); in xnn_q8_dwconv_ukernel_up8x9__neon()
[all …]

12345