xref: /aosp_15_r20/external/XNNPACK/src/qu8-vadd/gen/minmax-neon-ld64-x32.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_qu8_vadd_minmax_ukernel__neon_ld64_x32(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_vadd_minmax_ukernel__neon_ld64_x32(
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 uint8x8_t vb_zero_point = vld1_dup_u8(&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 uint8x16_t voutput_min = vld1q_dup_u8(&params->neon.output_min);
31   const uint8x16_t voutput_max = vld1q_dup_u8(&params->neon.output_max);
32 
33   for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
34     const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
35     const uint8x8_t vb01234567 = vld1_u8(input_b); input_b += 8;
36     const uint8x8_t va89ABCDEF = vld1_u8(input_a); input_a += 8;
37     const uint8x8_t vb89ABCDEF = vld1_u8(input_b); input_b += 8;
38     const uint8x8_t vaGHIJKLMN = vld1_u8(input_a); input_a += 8;
39     const uint8x8_t vbGHIJKLMN = vld1_u8(input_b); input_b += 8;
40     const uint8x8_t vaOPQRSTUV = vld1_u8(input_a); input_a += 8;
41     const uint8x8_t vbOPQRSTUV = vld1_u8(input_b); input_b += 8;
42 
43     const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
44     const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));
45     const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(va89ABCDEF, va_zero_point));
46     const int16x8_t vxb89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEF, vb_zero_point));
47     const int16x8_t vxaGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vaGHIJKLMN, va_zero_point));
48     const int16x8_t vxbGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vbGHIJKLMN, vb_zero_point));
49     const int16x8_t vxaOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vaOPQRSTUV, va_zero_point));
50     const int16x8_t vxbOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vbOPQRSTUV, vb_zero_point));
51 
52     int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
53     int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
54     int32x4_t vacc89AB = vmulq_s32(vmovl_s16(vget_low_s16(vxa89ABCDEF)), va_multiplier);
55     int32x4_t vaccCDEF = vmulq_s32(vmovl_s16(vget_high_s16(vxa89ABCDEF)), va_multiplier);
56     int32x4_t vaccGHIJ = vmulq_s32(vmovl_s16(vget_low_s16(vxaGHIJKLMN)), va_multiplier);
57     int32x4_t vaccKLMN = vmulq_s32(vmovl_s16(vget_high_s16(vxaGHIJKLMN)), va_multiplier);
58     int32x4_t vaccOPQR = vmulq_s32(vmovl_s16(vget_low_s16(vxaOPQRSTUV)), va_multiplier);
59     int32x4_t vaccSTUV = vmulq_s32(vmovl_s16(vget_high_s16(vxaOPQRSTUV)), va_multiplier);
60 
61     vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier);
62     vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier);
63     vacc89AB = vmlaq_s32(vacc89AB, vmovl_s16(vget_low_s16(vxb89ABCDEF)), vb_multiplier);
64     vaccCDEF = vmlaq_s32(vaccCDEF, vmovl_s16(vget_high_s16(vxb89ABCDEF)), vb_multiplier);
65     vaccGHIJ = vmlaq_s32(vaccGHIJ, vmovl_s16(vget_low_s16(vxbGHIJKLMN)), vb_multiplier);
66     vaccKLMN = vmlaq_s32(vaccKLMN, vmovl_s16(vget_high_s16(vxbGHIJKLMN)), vb_multiplier);
67     vaccOPQR = vmlaq_s32(vaccOPQR, vmovl_s16(vget_low_s16(vxbOPQRSTUV)), vb_multiplier);
68     vaccSTUV = vmlaq_s32(vaccSTUV, vmovl_s16(vget_high_s16(vxbOPQRSTUV)), vb_multiplier);
69 
70     vacc0123 = vrshlq_s32(vacc0123, vright_shift);
71     vacc4567 = vrshlq_s32(vacc4567, vright_shift);
72     vacc89AB = vrshlq_s32(vacc89AB, vright_shift);
73     vaccCDEF = vrshlq_s32(vaccCDEF, vright_shift);
74     vaccGHIJ = vrshlq_s32(vaccGHIJ, vright_shift);
75     vaccKLMN = vrshlq_s32(vaccKLMN, vright_shift);
76     vaccOPQR = vrshlq_s32(vaccOPQR, vright_shift);
77     vaccSTUV = vrshlq_s32(vaccSTUV, vright_shift);
78 
79     const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
80     const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point);
81     const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)), voutput_zero_point);
82     const int16x8_t vaccOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV)), voutput_zero_point);
83 
84     uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
85     uint8x16_t voutGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vaccGHIJKLMN), vqmovun_s16(vaccOPQRSTUV));
86 
87     vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
88     voutGHIJKLMNOPQRSTUV = vmaxq_u8(voutGHIJKLMNOPQRSTUV, voutput_min);
89 
90     vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
91     voutGHIJKLMNOPQRSTUV = vminq_u8(voutGHIJKLMNOPQRSTUV, voutput_max);
92 
93     vst1q_u8(output, vout0123456789ABCDEF); output += 16;
94     vst1q_u8(output, voutGHIJKLMNOPQRSTUV); output += 16;
95   }
96   if XNN_UNLIKELY(n != 0) {
97     do {
98       const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
99       const uint8x8_t vb01234567 = vld1_u8(input_b); input_b += 8;
100 
101       const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
102       const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));
103 
104       int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
105       int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
106 
107       vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier);
108       vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier);
109 
110       vacc0123 = vrshlq_s32(vacc0123, vright_shift);
111       vacc4567 = vrshlq_s32(vacc4567, vright_shift);
112 
113       const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
114 
115       uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
116       vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
117       vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
118 
119       if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
120         vst1_u8(output, vout01234567); output += 8;
121         n -= 8 * sizeof(uint8_t);
122       } else {
123         if (n & (4 * sizeof(uint8_t))) {
124           vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
125           vout01234567 = vext_u8(vout01234567, vout01234567, 4);
126         }
127         if (n & (2 * sizeof(uint8_t))) {
128           vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
129           vout01234567 = vext_u8(vout01234567, vout01234567, 2);
130         }
131         if (n & (1 * sizeof(uint8_t))) {
132           vst1_lane_u8(output, vout01234567, 0);
133         }
134         n = 0;
135       }
136     } while (n != 0);
137   }
138 }
139