xref: /aosp_15_r20/external/XNNPACK/src/qu8-vmulc/gen/minmax-rndnu-neon-ld128-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-vmulc/neon.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 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/vmul.h>
15 
16 
xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16(
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_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24   #if XNN_ARCH_ARM64
25     const uint8x16_t va_zero_point = vld1q_dup_u8(params->rndnu_neon.a_zero_point);
26   #else
27     const uint8x8_t va_zero_point = vld1_dup_u8(params->rndnu_neon.a_zero_point);
28   #endif
29   const int32x4_t vleft_pre_shift = vld1q_dup_s32(&params->rndnu_neon.left_pre_shift);
30   const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
31   const int32x4_t vleft_post_shift = vld1q_dup_s32(&params->rndnu_neon.left_post_shift);
32   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
33   const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
34   const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
35 
36   const uint8x8_t vb = vld1_dup_u8(input_b);
37   const uint8x8_t vb_zero_point = vld1_dup_u8(params->rndnu_neon.b_zero_point);
38   const int16x8_t vxb = vreinterpretq_s16_u16(vsubl_u8(vb, vb_zero_point));
39   for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
40     const uint8x16_t va0123456789ABCDEF = vld1q_u8(input_a); input_a += 16;
41 
42     #if XNN_ARCH_ARM64
43       const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), vget_low_u8(va_zero_point)));
44       const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_high_u8(va0123456789ABCDEF, va_zero_point));
45     #else  // !XNN_ARCH_ARM64
46       const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_zero_point));
47       const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(va0123456789ABCDEF), va_zero_point));
48     #endif  // XNN_ARCH_ARM64
49 
50     int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
51     int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
52     int32x4_t vacc89AB = vmull_s16(vget_low_s16(vxa89ABCDEF), vget_low_s16(vxb));
53     int32x4_t vaccCDEF = vmull_s16(vget_high_s16(vxa89ABCDEF), vget_high_s16(vxb));
54 
55     vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
56     vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
57     vacc89AB = vqshlq_s32(vacc89AB, vleft_pre_shift);
58     vaccCDEF = vqshlq_s32(vaccCDEF, vleft_pre_shift);
59 
60     vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
61     vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
62     vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
63     vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
64 
65     vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
66     vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
67     vacc89AB = vrshlq_s32(vacc89AB, vleft_post_shift);
68     vaccCDEF = vrshlq_s32(vaccCDEF, vleft_post_shift);
69 
70     #if XNN_ARCH_ARM64
71       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
72       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
73     #else
74       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
75       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
76     #endif
77 
78     vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
79     vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
80 
81     #if XNN_ARCH_ARM64
82       uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
83     #else
84       uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
85     #endif
86 
87     vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
88 
89     vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
90 
91     vst1q_u8(output, vout0123456789ABCDEF); output += 16;
92   }
93   if XNN_UNLIKELY(n != 0) {
94     do {
95       const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
96 
97       #if XNN_ARCH_ARM64
98         const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point)));
99       #else  // !XNN_ARCH_ARM64
100         const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
101       #endif  // XNN_ARCH_ARM64
102 
103       int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
104       int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
105 
106       vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
107       vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
108 
109       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
110       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
111 
112       vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
113       vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
114 
115       #if XNN_ARCH_ARM64
116         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
117       #else
118         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
119       #endif
120 
121       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
122 
123       uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
124 
125       vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
126       vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
127       if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
128         vst1_u8(output, vout01234567); output += 8;
129         n -= 8 * sizeof(uint8_t);
130       } else {
131         if (n & (4 * sizeof(uint8_t))) {
132           vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
133           vout01234567 = vext_u8(vout01234567, vout01234567, 4);
134         }
135         if (n & (2 * sizeof(uint8_t))) {
136           vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
137           vout01234567 = vext_u8(vout01234567, vout01234567, 2);
138         }
139         if (n & (1 * sizeof(uint8_t))) {
140           vst1_lane_u8(output, vout01234567, 0);
141         }
142         n = 0;
143       }
144     } while (n != 0);
145   }
146 }
147