xref: /aosp_15_r20/external/XNNPACK/src/qs8-vadd/gen/minmax-neon-ld64-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-vadd/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_qs8_vadd_minmax_ukernel__neon_ld64_x8(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8(
18     size_t n,
19     const int8_t* input_a,
20     const int8_t* input_b,
21     int8_t* output,
22     const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24   const int8x8_t va_zero_point = vld1_dup_s8(&params->neon.a_zero_point);
25   const int8x8_t vb_zero_point = vld1_dup_s8(&params->neon.b_zero_point);
26   const int32x4_t va_multiplier = vld1q_dup_s32(&params->neon.a_multiplier);
27   const int32x4_t vb_multiplier = vld1q_dup_s32(&params->neon.b_multiplier);
28   const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
29   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
30   const int8x8_t voutput_min = vld1_dup_s8(&params->neon.output_min);
31   const int8x8_t voutput_max = vld1_dup_s8(&params->neon.output_max);
32 
33   for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
34     const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
35     const int8x8_t vb01234567 = vld1_s8(input_b); input_b += 8;
36 
37     const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
38     const int16x8_t vxb01234567 = vsubl_s8(vb01234567, vb_zero_point);
39 
40     int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
41     int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
42 
43     vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier);
44     vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier);
45 
46     vacc0123 = vrshlq_s32(vacc0123, vright_shift);
47     vacc4567 = vrshlq_s32(vacc4567, vright_shift);
48 
49     const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
50 
51     int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
52 
53     vout01234567 = vmax_s8(vout01234567, voutput_min);
54 
55     vout01234567 = vmin_s8(vout01234567, voutput_max);
56 
57     vst1_s8(output, vout01234567); output += 8;
58   }
59   if XNN_UNLIKELY(n != 0) {
60     {
61       const int8x8_t va01234567 = vld1_s8(input_a);
62       const int8x8_t vb01234567 = vld1_s8(input_b);
63 
64       const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
65       const int16x8_t vxb01234567 = vsubl_s8(vb01234567, vb_zero_point);
66 
67       int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
68       int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
69 
70       vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier);
71       vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier);
72 
73       vacc0123 = vrshlq_s32(vacc0123, vright_shift);
74       vacc4567 = vrshlq_s32(vacc4567, vright_shift);
75 
76       const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
77 
78       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
79       vout01234567 = vmax_s8(vout01234567, voutput_min);
80       vout01234567 = vmin_s8(vout01234567, voutput_max);
81 
82       if (n & (4 * sizeof(int8_t))) {
83         vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
84         vout01234567 = vext_s8(vout01234567, vout01234567, 4);
85       }
86       if (n & (2 * sizeof(int8_t))) {
87         vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
88         vout01234567 = vext_s8(vout01234567, vout01234567, 2);
89       }
90       if (n & (1 * sizeof(int8_t))) {
91         vst1_lane_s8(output, vout01234567, 0);
92       }
93     }
94   }
95 }
96