xref: /aosp_15_r20/external/XNNPACK/src/qu8-vaddc/gen/minmax-neon-ld64-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-vaddc/neon.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/vadd.h>
15 
16 
xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8(
18     size_t n,
19     const uint8_t* input_a,
20     const uint8_t* input_b,
21     uint8_t* output,
22     const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24   const uint8x8_t va_zero_point = vld1_dup_u8(&params->neon.a_zero_point);
25   const int32x4_t va_multiplier = vld1q_dup_s32(&params->neon.a_multiplier);
26   const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
27   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
28   const uint8x8_t voutput_min = vld1_dup_u8(&params->neon.output_min);
29   const uint8x8_t voutput_max = vld1_dup_u8(&params->neon.output_max);
30 
31   const int32_t vxb = (int32_t) *input_b - (int32_t) params->neon.b_zero_point;
32   const int32_t vb = params->neon.b_multiplier;
33   const int32x4_t vbias = vdupq_n_s32(vxb * vb);
34 
35   for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
36     const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
37 
38     const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
39 
40     int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
41     int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
42 
43     vacc0123 = vrshlq_s32(vacc0123, vright_shift);
44     vacc4567 = vrshlq_s32(vacc4567, vright_shift);
45 
46     const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
47 
48     uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
49 
50     vout01234567 = vmax_u8(vout01234567, voutput_min);
51 
52     vout01234567 = vmin_u8(vout01234567, voutput_max);
53 
54     vst1_u8(output, vout01234567); output += 8;
55   }
56   if XNN_UNLIKELY(n != 0) {
57     {
58       const uint8x8_t va01234567 = vld1_u8(input_a);
59 
60       const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
61 
62       int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
63       int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
64 
65       vacc0123 = vrshlq_s32(vacc0123, vright_shift);
66       vacc4567 = vrshlq_s32(vacc4567, vright_shift);
67 
68       const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
69 
70       uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
71       vout01234567 = vmax_u8(vout01234567, voutput_min);
72       vout01234567 = vmin_u8(vout01234567, voutput_max);
73 
74       if (n & (4 * sizeof(uint8_t))) {
75         vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
76         vout01234567 = vext_u8(vout01234567, vout01234567, 4);
77       }
78       if (n & (2 * sizeof(uint8_t))) {
79         vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
80         vout01234567 = vext_u8(vout01234567, vout01234567, 2);
81       }
82       if (n & (1 * sizeof(uint8_t))) {
83         vst1_lane_u8(output, vout01234567, 0);
84       }
85     }
86   }
87 }
88