1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/neon-ld64.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_x,const int8_t * input_y,int8_t * output,const union xnn_qs8_add_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_x,
20 const int8_t* input_y,
21 int8_t* output,
22 const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24 const int8x8_t vx_zero_point = vld1_dup_s8(¶ms->neon.x_zero_point);
25 const int8x8_t vy_zero_point = vld1_dup_s8(¶ms->neon.y_zero_point);
26 const int32x4_t vx_multiplier = vld1q_dup_s32(¶ms->neon.x_multiplier);
27 const int32x4_t vy_multiplier = vld1q_dup_s32(¶ms->neon.y_multiplier);
28 const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift);
29 const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
30 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point);
31 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min);
32 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max);
33
34 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
35 const int8x8_t vx01234567 = vld1_s8(input_x); input_x += 8;
36 const int8x8_t vy01234567 = vld1_s8(input_y); input_y += 8;
37
38 const int16x8_t vex01234567 = vsubl_s8(vx01234567, vx_zero_point);
39 const int16x8_t vey01234567 = vsubl_s8(vy01234567, vy_zero_point);
40
41 int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vex01234567)), vx_multiplier);
42 int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vex01234567)), vx_multiplier);
43
44 vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vey01234567)), vy_multiplier);
45 vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vey01234567)), vy_multiplier);
46
47 vacc0123 = vsraq_n_s32(vacc0123, vbicq_s32(vacc0123, vzero_shift_mask), 31);
48 vacc4567 = vsraq_n_s32(vacc4567, vbicq_s32(vacc4567, vzero_shift_mask), 31);
49
50 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
51 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
52
53 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
54
55 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
56
57 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
58
59 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
60
61 vst1_s8(output, vout01234567); output += 8;
62 }
63 if XNN_UNLIKELY(n != 0) {
64 {
65 const int8x8_t vx01234567 = vld1_s8(input_x);
66 const int8x8_t vy01234567 = vld1_s8(input_y);
67
68 const int16x8_t vex01234567 = vsubl_s8(vx01234567, vx_zero_point);
69 const int16x8_t vey01234567 = vsubl_s8(vy01234567, vy_zero_point);
70
71 int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vex01234567)), vx_multiplier);
72 int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vex01234567)), vx_multiplier);
73
74 vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vey01234567)), vy_multiplier);
75 vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vey01234567)), vy_multiplier);
76
77 vacc0123 = vsraq_n_s32(vacc0123, vbicq_s32(vacc0123, vzero_shift_mask), 31);
78 vacc4567 = vsraq_n_s32(vacc4567, vbicq_s32(vacc4567, vzero_shift_mask), 31);
79
80 vacc0123 = vrshlq_s32(vacc0123, vright_shift);
81 vacc4567 = vrshlq_s32(vacc4567, vright_shift);
82
83 const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
84
85 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
86 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
87 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
88
89 if (n & (4 * sizeof(int8_t))) {
90 vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_s8(vout01234567), 0); output += 4;
91 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
92 }
93 if (n & (2 * sizeof(int8_t))) {
94 vst1_lane_u16(__builtin_assume_aligned(output, 1), vreinterpret_u16_s8(vout01234567), 0); output += 2;
95 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
96 }
97 if (n & (1 * sizeof(int8_t))) {
98 vst1_lane_s8(output, vout01234567, 0);
99 }
100 }
101 }
102 }
103