xref: /aosp_15_r20/external/XNNPACK/src/qs8-vadd/gen/minmax-neon-ld64-x24.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_x24(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_x24(
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 int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
31   const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
32 
33   for (; n >= 24 * sizeof(int8_t); n -= 24 * 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     const int8x8_t va89ABCDEF = vld1_s8(input_a); input_a += 8;
37     const int8x8_t vb89ABCDEF = vld1_s8(input_b); input_b += 8;
38     const int8x8_t vaGHIJKLMN = vld1_s8(input_a); input_a += 8;
39     const int8x8_t vbGHIJKLMN = vld1_s8(input_b); input_b += 8;
40 
41     const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
42     const int16x8_t vxb01234567 = vsubl_s8(vb01234567, vb_zero_point);
43     const int16x8_t vxa89ABCDEF = vsubl_s8(va89ABCDEF, va_zero_point);
44     const int16x8_t vxb89ABCDEF = vsubl_s8(vb89ABCDEF, vb_zero_point);
45     const int16x8_t vxaGHIJKLMN = vsubl_s8(vaGHIJKLMN, va_zero_point);
46     const int16x8_t vxbGHIJKLMN = vsubl_s8(vbGHIJKLMN, vb_zero_point);
47 
48     int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
49     int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
50     int32x4_t vacc89AB = vmulq_s32(vmovl_s16(vget_low_s16(vxa89ABCDEF)), va_multiplier);
51     int32x4_t vaccCDEF = vmulq_s32(vmovl_s16(vget_high_s16(vxa89ABCDEF)), va_multiplier);
52     int32x4_t vaccGHIJ = vmulq_s32(vmovl_s16(vget_low_s16(vxaGHIJKLMN)), va_multiplier);
53     int32x4_t vaccKLMN = vmulq_s32(vmovl_s16(vget_high_s16(vxaGHIJKLMN)), va_multiplier);
54 
55     vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier);
56     vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier);
57     vacc89AB = vmlaq_s32(vacc89AB, vmovl_s16(vget_low_s16(vxb89ABCDEF)), vb_multiplier);
58     vaccCDEF = vmlaq_s32(vaccCDEF, vmovl_s16(vget_high_s16(vxb89ABCDEF)), vb_multiplier);
59     vaccGHIJ = vmlaq_s32(vaccGHIJ, vmovl_s16(vget_low_s16(vxbGHIJKLMN)), vb_multiplier);
60     vaccKLMN = vmlaq_s32(vaccKLMN, vmovl_s16(vget_high_s16(vxbGHIJKLMN)), vb_multiplier);
61 
62     vacc0123 = vrshlq_s32(vacc0123, vright_shift);
63     vacc4567 = vrshlq_s32(vacc4567, vright_shift);
64     vacc89AB = vrshlq_s32(vacc89AB, vright_shift);
65     vaccCDEF = vrshlq_s32(vaccCDEF, vright_shift);
66     vaccGHIJ = vrshlq_s32(vaccGHIJ, vright_shift);
67     vaccKLMN = vrshlq_s32(vaccKLMN, vright_shift);
68 
69     const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
70     const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point);
71     const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)), voutput_zero_point);
72 
73     int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
74     int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN);
75 
76     vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
77     voutGHIJKLMN = vmax_s8(voutGHIJKLMN, vget_low_s8(voutput_min));
78 
79     vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
80     voutGHIJKLMN = vmin_s8(voutGHIJKLMN, vget_low_s8(voutput_max));
81 
82     vst1q_s8(output, vout0123456789ABCDEF); output += 16;
83     vst1_s8(output, voutGHIJKLMN); output += 8;
84   }
85   if XNN_UNLIKELY(n != 0) {
86     do {
87       const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
88       const int8x8_t vb01234567 = vld1_s8(input_b); input_b += 8;
89 
90       const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
91       const int16x8_t vxb01234567 = vsubl_s8(vb01234567, vb_zero_point);
92 
93       int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
94       int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
95 
96       vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier);
97       vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier);
98 
99       vacc0123 = vrshlq_s32(vacc0123, vright_shift);
100       vacc4567 = vrshlq_s32(vacc4567, vright_shift);
101 
102       const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
103 
104       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
105       vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
106       vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
107 
108       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
109         vst1_s8(output, vout01234567); output += 8;
110         n -= 8 * sizeof(int8_t);
111       } else {
112         if (n & (4 * sizeof(int8_t))) {
113           vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
114           vout01234567 = vext_s8(vout01234567, vout01234567, 4);
115         }
116         if (n & (2 * sizeof(int8_t))) {
117           vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
118           vout01234567 = vext_s8(vout01234567, vout01234567, 2);
119         }
120         if (n & (1 * sizeof(int8_t))) {
121           vst1_lane_s8(output, vout01234567, 0);
122         }
123         n = 0;
124       }
125     } while (n != 0);
126   }
127 }
128