xref: /aosp_15_r20/external/XNNPACK/src/qs8-vadd/gen/minmax-neon-ld128-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_qs8_vadd_minmax_ukernel__neon_ld128_x32(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_ld128_x32(
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   #if XNN_ARCH_ARM64
25     const int8x16_t va_zero_point = vld1q_dup_s8(&params->neon.a_zero_point);
26     const int8x16_t vb_zero_point = vld1q_dup_s8(&params->neon.b_zero_point);
27   #else
28     const int8x8_t va_zero_point = vld1_dup_s8(&params->neon.a_zero_point);
29     const int8x8_t vb_zero_point = vld1_dup_s8(&params->neon.b_zero_point);
30   #endif
31   const int32x4_t va_multiplier = vld1q_dup_s32(&params->neon.a_multiplier);
32   const int32x4_t vb_multiplier = vld1q_dup_s32(&params->neon.b_multiplier);
33   const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
34   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
35   const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
36   const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
37 
38   for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
39     const int8x16_t va0123456789ABCDEF = vld1q_s8(input_a); input_a += 16;
40     const int8x16_t vb0123456789ABCDEF = vld1q_s8(input_b); input_b += 16;
41     const int8x16_t vaGHIJKLMNOPQRSTUV = vld1q_s8(input_a); input_a += 16;
42     const int8x16_t vbGHIJKLMNOPQRSTUV = vld1q_s8(input_b); input_b += 16;
43 
44     #if XNN_ARCH_ARM64
45       const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point));
46       const int16x8_t vxa89ABCDEF = vsubl_high_s8(va0123456789ABCDEF, va_zero_point);
47       const int16x8_t vxb01234567 = vsubl_s8(vget_low_s8(vb0123456789ABCDEF), vget_low_s8(vb_zero_point));
48       const int16x8_t vxb89ABCDEF = vsubl_high_s8(vb0123456789ABCDEF, vb_zero_point);
49       const int16x8_t vxaGHIJKLMN = vsubl_s8(vget_low_s8(vaGHIJKLMNOPQRSTUV), vget_low_s8(va_zero_point));
50       const int16x8_t vxaOPQRSTUV = vsubl_high_s8(vaGHIJKLMNOPQRSTUV, va_zero_point);
51       const int16x8_t vxbGHIJKLMN = vsubl_s8(vget_low_s8(vbGHIJKLMNOPQRSTUV), vget_low_s8(vb_zero_point));
52       const int16x8_t vxbOPQRSTUV = vsubl_high_s8(vbGHIJKLMNOPQRSTUV, vb_zero_point);
53     #else  // !XNN_ARCH_ARM64
54       const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point);
55       const int16x8_t vxa89ABCDEF = vsubl_s8(vget_high_s8(va0123456789ABCDEF), va_zero_point);
56       const int16x8_t vxb01234567 = vsubl_s8(vget_low_s8(vb0123456789ABCDEF), vb_zero_point);
57       const int16x8_t vxb89ABCDEF = vsubl_s8(vget_high_s8(vb0123456789ABCDEF), vb_zero_point);
58       const int16x8_t vxaGHIJKLMN = vsubl_s8(vget_low_s8(vaGHIJKLMNOPQRSTUV), va_zero_point);
59       const int16x8_t vxaOPQRSTUV = vsubl_s8(vget_high_s8(vaGHIJKLMNOPQRSTUV), va_zero_point);
60       const int16x8_t vxbGHIJKLMN = vsubl_s8(vget_low_s8(vbGHIJKLMNOPQRSTUV), vb_zero_point);
61       const int16x8_t vxbOPQRSTUV = vsubl_s8(vget_high_s8(vbGHIJKLMNOPQRSTUV), vb_zero_point);
62     #endif  // XNN_ARCH_ARM64
63 
64     int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
65     int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
66     int32x4_t vacc89AB = vmulq_s32(vmovl_s16(vget_low_s16(vxa89ABCDEF)), va_multiplier);
67     int32x4_t vaccCDEF = vmulq_s32(vmovl_s16(vget_high_s16(vxa89ABCDEF)), va_multiplier);
68     int32x4_t vaccGHIJ = vmulq_s32(vmovl_s16(vget_low_s16(vxaGHIJKLMN)), va_multiplier);
69     int32x4_t vaccKLMN = vmulq_s32(vmovl_s16(vget_high_s16(vxaGHIJKLMN)), va_multiplier);
70     int32x4_t vaccOPQR = vmulq_s32(vmovl_s16(vget_low_s16(vxaOPQRSTUV)), va_multiplier);
71     int32x4_t vaccSTUV = vmulq_s32(vmovl_s16(vget_high_s16(vxaOPQRSTUV)), va_multiplier);
72 
73     vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier);
74     vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier);
75     vacc89AB = vmlaq_s32(vacc89AB, vmovl_s16(vget_low_s16(vxb89ABCDEF)), vb_multiplier);
76     vaccCDEF = vmlaq_s32(vaccCDEF, vmovl_s16(vget_high_s16(vxb89ABCDEF)), vb_multiplier);
77     vaccGHIJ = vmlaq_s32(vaccGHIJ, vmovl_s16(vget_low_s16(vxbGHIJKLMN)), vb_multiplier);
78     vaccKLMN = vmlaq_s32(vaccKLMN, vmovl_s16(vget_high_s16(vxbGHIJKLMN)), vb_multiplier);
79     vaccOPQR = vmlaq_s32(vaccOPQR, vmovl_s16(vget_low_s16(vxbOPQRSTUV)), vb_multiplier);
80     vaccSTUV = vmlaq_s32(vaccSTUV, vmovl_s16(vget_high_s16(vxbOPQRSTUV)), vb_multiplier);
81 
82     vacc0123 = vrshlq_s32(vacc0123, vright_shift);
83     vacc4567 = vrshlq_s32(vacc4567, vright_shift);
84     vacc89AB = vrshlq_s32(vacc89AB, vright_shift);
85     vaccCDEF = vrshlq_s32(vaccCDEF, vright_shift);
86     vaccGHIJ = vrshlq_s32(vaccGHIJ, vright_shift);
87     vaccKLMN = vrshlq_s32(vaccKLMN, vright_shift);
88     vaccOPQR = vrshlq_s32(vaccOPQR, vright_shift);
89     vaccSTUV = vrshlq_s32(vaccSTUV, vright_shift);
90 
91     const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
92     const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point);
93     const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)), voutput_zero_point);
94     const int16x8_t vaccOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV)), voutput_zero_point);
95 
96     int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
97     int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV));
98 
99     vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
100     voutGHIJKLMNOPQRSTUV = vmaxq_s8(voutGHIJKLMNOPQRSTUV, voutput_min);
101 
102     vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
103     voutGHIJKLMNOPQRSTUV = vminq_s8(voutGHIJKLMNOPQRSTUV, voutput_max);
104 
105     vst1q_s8(output, vout0123456789ABCDEF); output += 16;
106     vst1q_s8(output, voutGHIJKLMNOPQRSTUV); output += 16;
107   }
108   if XNN_UNLIKELY(n != 0) {
109     do {
110       const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
111       const int8x8_t vb01234567 = vld1_s8(input_b); input_b += 8;
112 
113       #if XNN_ARCH_ARM64
114         const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point));
115         const int16x8_t vxb01234567 = vsubl_s8(vb01234567, vget_low_s8(vb_zero_point));
116       #else  // !XNN_ARCH_ARM64
117         const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
118         const int16x8_t vxb01234567 = vsubl_s8(vb01234567, vb_zero_point);
119       #endif  // XNN_ARCH_ARM64
120 
121       int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
122       int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);
123 
124       vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier);
125       vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier);
126 
127       vacc0123 = vrshlq_s32(vacc0123, vright_shift);
128       vacc4567 = vrshlq_s32(vacc4567, vright_shift);
129 
130       const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
131 
132       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
133       vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
134       vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
135 
136       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
137         vst1_s8(output, vout01234567); output += 8;
138         n -= 8 * sizeof(int8_t);
139       } else {
140         if (n & (4 * sizeof(int8_t))) {
141           vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
142           vout01234567 = vext_s8(vout01234567, vout01234567, 4);
143         }
144         if (n & (2 * sizeof(int8_t))) {
145           vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
146           vout01234567 = vext_s8(vout01234567, vout01234567, 2);
147         }
148         if (n & (1 * sizeof(int8_t))) {
149           vst1_lane_s8(output, vout01234567, 0);
150         }
151         n = 0;
152       }
153     } while (n != 0);
154   }
155 }
156