xref: /aosp_15_r20/external/XNNPACK/src/qu8-vmulc/gen/minmax-rndnu-neon-ld64-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_ld64_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_ld64_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   const uint8x8_t va_zero_point = vld1_dup_u8(params->rndnu_neon.a_zero_point);
25   const int32x4_t vleft_pre_shift = vld1q_dup_s32(&params->rndnu_neon.left_pre_shift);
26   const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
27   const int32x4_t vleft_post_shift = vld1q_dup_s32(&params->rndnu_neon.left_post_shift);
28   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
29   const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
30   const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
31 
32   const uint8x8_t vb = vld1_dup_u8(input_b);
33   const uint8x8_t vb_zero_point = vld1_dup_u8(params->rndnu_neon.b_zero_point);
34   const int16x8_t vxb = vreinterpretq_s16_u16(vsubl_u8(vb, vb_zero_point));
35   for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
36     const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
37     const uint8x8_t va89ABCDEF = vld1_u8(input_a); input_a += 8;
38 
39     const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
40     const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(va89ABCDEF, va_zero_point));
41 
42     int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
43     int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
44     int32x4_t vacc89AB = vmull_s16(vget_low_s16(vxa89ABCDEF), vget_low_s16(vxb));
45     int32x4_t vaccCDEF = vmull_s16(vget_high_s16(vxa89ABCDEF), vget_high_s16(vxb));
46 
47     vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
48     vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
49     vacc89AB = vqshlq_s32(vacc89AB, vleft_pre_shift);
50     vaccCDEF = vqshlq_s32(vaccCDEF, vleft_pre_shift);
51 
52     vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
53     vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
54     vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
55     vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
56 
57     vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
58     vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
59     vacc89AB = vrshlq_s32(vacc89AB, vleft_post_shift);
60     vaccCDEF = vrshlq_s32(vaccCDEF, vleft_post_shift);
61 
62     #if XNN_ARCH_ARM64
63       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
64       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
65     #else
66       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
67       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
68     #endif
69 
70     vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
71     vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
72 
73     #if XNN_ARCH_ARM64
74       uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
75     #else
76       uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
77     #endif
78 
79     vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
80 
81     vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
82 
83     vst1q_u8(output, vout0123456789ABCDEF); output += 16;
84   }
85   if XNN_UNLIKELY(n != 0) {
86     do {
87       const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
88 
89       const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
90 
91       int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
92       int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
93 
94       vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
95       vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
96 
97       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
98       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
99 
100       vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
101       vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
102 
103       #if XNN_ARCH_ARM64
104         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
105       #else
106         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
107       #endif
108 
109       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
110 
111       uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
112 
113       vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
114       vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
115       if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
116         vst1_u8(output, vout01234567); output += 8;
117         n -= 8 * sizeof(uint8_t);
118       } else {
119         if (n & (4 * sizeof(uint8_t))) {
120           vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
121           vout01234567 = vext_u8(vout01234567, vout01234567, 4);
122         }
123         if (n & (2 * sizeof(uint8_t))) {
124           vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
125           vout01234567 = vext_u8(vout01234567, vout01234567, 2);
126         }
127         if (n & (1 * sizeof(uint8_t))) {
128           vst1_lane_u8(output, vout01234567, 0);
129         }
130         n = 0;
131       }
132     } while (n != 0);
133   }
134 }
135