xref: /aosp_15_r20/external/XNNPACK/src/qs8-vaddc/gen/minmax-neon-ld64-x24.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-vaddc/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_vaddc_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_vaddc_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 int32x4_t va_multiplier = vld1q_dup_s32(&params->neon.a_multiplier);
26   const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
27   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
28   const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
29   const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
30 
31   const int32_t vxb = (int32_t) *input_b - (int32_t) params->neon.b_zero_point;
32   const int32_t vb = params->neon.b_multiplier;
33   const int32x4_t vbias = vdupq_n_s32(vxb * vb);
34 
35   for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
36     const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
37     const int8x8_t va89ABCDEF = vld1_s8(input_a); input_a += 8;
38     const int8x8_t vaGHIJKLMN = vld1_s8(input_a); input_a += 8;
39 
40     const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
41     const int16x8_t vxa89ABCDEF = vsubl_s8(va89ABCDEF, va_zero_point);
42     const int16x8_t vxaGHIJKLMN = vsubl_s8(vaGHIJKLMN, va_zero_point);
43 
44     int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
45     int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
46     int32x4_t vacc89AB = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa89ABCDEF)), va_multiplier);
47     int32x4_t vaccCDEF = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa89ABCDEF)), va_multiplier);
48     int32x4_t vaccGHIJ = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxaGHIJKLMN)), va_multiplier);
49     int32x4_t vaccKLMN = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxaGHIJKLMN)), va_multiplier);
50 
51     vacc0123 = vrshlq_s32(vacc0123, vright_shift);
52     vacc4567 = vrshlq_s32(vacc4567, vright_shift);
53     vacc89AB = vrshlq_s32(vacc89AB, vright_shift);
54     vaccCDEF = vrshlq_s32(vaccCDEF, vright_shift);
55     vaccGHIJ = vrshlq_s32(vaccGHIJ, vright_shift);
56     vaccKLMN = vrshlq_s32(vaccKLMN, vright_shift);
57 
58     const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
59     const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point);
60     const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)), voutput_zero_point);
61 
62     int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
63     int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN);
64 
65     vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
66     voutGHIJKLMN = vmax_s8(voutGHIJKLMN, vget_low_s8(voutput_min));
67 
68     vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
69     voutGHIJKLMN = vmin_s8(voutGHIJKLMN, vget_low_s8(voutput_max));
70 
71     vst1q_s8(output, vout0123456789ABCDEF); output += 16;
72     vst1_s8(output, voutGHIJKLMN); output += 8;
73   }
74   if XNN_UNLIKELY(n != 0) {
75     do {
76       const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
77 
78       const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
79 
80       int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
81       int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
82 
83       vacc0123 = vrshlq_s32(vacc0123, vright_shift);
84       vacc4567 = vrshlq_s32(vacc4567, vright_shift);
85 
86       const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
87 
88       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
89       vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
90       vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
91 
92       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
93         vst1_s8(output, vout01234567); output += 8;
94         n -= 8 * sizeof(int8_t);
95       } else {
96         if (n & (4 * sizeof(int8_t))) {
97           vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
98           vout01234567 = vext_s8(vout01234567, vout01234567, 4);
99         }
100         if (n & (2 * sizeof(int8_t))) {
101           vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
102           vout01234567 = vext_s8(vout01234567, vout01234567, 2);
103         }
104         if (n & (1 * sizeof(int8_t))) {
105           vst1_lane_s8(output, vout01234567, 0);
106         }
107         n = 0;
108       }
109     } while (n != 0);
110   }
111 }
112