xref: /aosp_15_r20/external/XNNPACK/src/qs8-vmul/gen/minmax-fp32-neon-ld128-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-vmul/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_qs8_vmul_minmax_fp32_ukernel__neon_ld128_x16(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld128_x16(
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_mul_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->fp32_neon.a_zero_point);
26     const int8x16_t vb_zero_point = vld1q_dup_s8(params->fp32_neon.b_zero_point);
27   #else
28     const int8x8_t va_zero_point = vld1_dup_s8(params->fp32_neon.a_zero_point);
29     const int8x8_t vb_zero_point = vld1_dup_s8(params->fp32_neon.b_zero_point);
30   #endif
31   const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
32   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
33   const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
34   const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neon.output_min);
35   const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.output_max);
36 
37   for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
38     const int8x16_t va0123456789ABCDEF = vld1q_s8(input_a); input_a += 16;
39     const int8x16_t vb0123456789ABCDEF = vld1q_s8(input_b); input_b += 16;
40 
41     #if XNN_ARCH_ARM64
42       const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point));
43       const int16x8_t vxa89ABCDEF = vsubl_high_s8(va0123456789ABCDEF, va_zero_point);
44       const int16x8_t vxb01234567 = vsubl_s8(vget_low_s8(vb0123456789ABCDEF), vget_low_s8(vb_zero_point));
45       const int16x8_t vxb89ABCDEF = vsubl_high_s8(vb0123456789ABCDEF, vb_zero_point);
46     #else  // !XNN_ARCH_ARM64
47       const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point);
48       const int16x8_t vxa89ABCDEF = vsubl_s8(vget_high_s8(va0123456789ABCDEF), va_zero_point);
49       const int16x8_t vxb01234567 = vsubl_s8(vget_low_s8(vb0123456789ABCDEF), vb_zero_point);
50       const int16x8_t vxb89ABCDEF = vsubl_s8(vget_high_s8(vb0123456789ABCDEF), vb_zero_point);
51     #endif  // XNN_ARCH_ARM64
52 
53     int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567));
54     int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567));
55     int32x4_t vacc89AB = vmull_s16(vget_low_s16(vxa89ABCDEF), vget_low_s16(vxb89ABCDEF));
56     int32x4_t vaccCDEF = vmull_s16(vget_high_s16(vxa89ABCDEF), vget_high_s16(vxb89ABCDEF));
57 
58     float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
59     float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
60     float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
61     float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
62 
63     vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
64     vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
65     vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
66     vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
67 
68     vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
69     vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
70     vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
71     vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
72 
73     vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
74     vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
75     vacc89AB = vqsubq_s32(vacc89AB, vmagic_bias_less_output_zero_point);
76     vaccCDEF = vqsubq_s32(vaccCDEF, vmagic_bias_less_output_zero_point);
77 
78     #if XNN_ARCH_ARM64
79       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
80       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
81     #else
82       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
83       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
84     #endif
85 
86 
87     #if XNN_ARCH_ARM64
88       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
89     #else
90       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
91     #endif
92 
93     vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
94 
95     vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
96 
97     vst1q_s8(output, vout0123456789ABCDEF); output += 16;
98   }
99   if XNN_UNLIKELY(n != 0) {
100     do {
101       const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
102       const int8x8_t vb01234567 = vld1_s8(input_b); input_b += 8;
103 
104       #if XNN_ARCH_ARM64
105         const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point));
106         const int16x8_t vxb01234567 = vsubl_s8(vb01234567, vget_low_s8(vb_zero_point));
107       #else  // !XNN_ARCH_ARM64
108         const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
109         const int16x8_t vxb01234567 = vsubl_s8(vb01234567, vb_zero_point);
110       #endif  // XNN_ARCH_ARM64
111 
112       int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567));
113       int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567));
114 
115       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
116       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
117 
118       vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
119       vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
120 
121       vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
122       vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
123 
124       vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
125       vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
126 
127       #if XNN_ARCH_ARM64
128         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
129       #else
130         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
131       #endif
132 
133 
134       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
135 
136       vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
137       vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
138       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
139         vst1_s8(output, vout01234567); output += 8;
140         n -= 8 * sizeof(int8_t);
141       } else {
142         if (n & (4 * sizeof(int8_t))) {
143           vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
144           vout01234567 = vext_s8(vout01234567, vout01234567, 4);
145         }
146         if (n & (2 * sizeof(int8_t))) {
147           vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
148           vout01234567 = vext_s8(vout01234567, vout01234567, 2);
149         }
150         if (n & (1 * sizeof(int8_t))) {
151           vst1_lane_s8(output, vout01234567, 0);
152         }
153         n = 0;
154       }
155     } while (n != 0);
156   }
157 }
158