xref: /aosp_15_r20/external/XNNPACK/src/qs8-vmul/gen/minmax-fp32-neon-ld64-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_ld64_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_ld64_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   const int8x8_t va_zero_point = vld1_dup_s8(params->fp32_neon.a_zero_point);
25   const int8x8_t vb_zero_point = vld1_dup_s8(params->fp32_neon.b_zero_point);
26   const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
27   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
28   const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
29   const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neon.output_min);
30   const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neon.output_max);
31 
32   for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
33     const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
34     const int8x8_t vb01234567 = vld1_s8(input_b); input_b += 8;
35     const int8x8_t va89ABCDEF = vld1_s8(input_a); input_a += 8;
36     const int8x8_t vb89ABCDEF = vld1_s8(input_b); input_b += 8;
37 
38     const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
39     const int16x8_t vxb01234567 = vsubl_s8(vb01234567, vb_zero_point);
40     const int16x8_t vxa89ABCDEF = vsubl_s8(va89ABCDEF, va_zero_point);
41     const int16x8_t vxb89ABCDEF = vsubl_s8(vb89ABCDEF, vb_zero_point);
42 
43     int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567));
44     int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567));
45     int32x4_t vacc89AB = vmull_s16(vget_low_s16(vxa89ABCDEF), vget_low_s16(vxb89ABCDEF));
46     int32x4_t vaccCDEF = vmull_s16(vget_high_s16(vxa89ABCDEF), vget_high_s16(vxb89ABCDEF));
47 
48     float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
49     float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
50     float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
51     float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
52 
53     vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
54     vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
55     vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
56     vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
57 
58     vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
59     vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
60     vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
61     vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
62 
63     vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
64     vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
65     vacc89AB = vqsubq_s32(vacc89AB, vmagic_bias_less_output_zero_point);
66     vaccCDEF = vqsubq_s32(vaccCDEF, vmagic_bias_less_output_zero_point);
67 
68     #if XNN_ARCH_ARM64
69       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
70       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
71     #else
72       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
73       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
74     #endif
75 
76 
77     #if XNN_ARCH_ARM64
78       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
79     #else
80       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
81     #endif
82 
83     vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
84 
85     vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
86 
87     vst1q_s8(output, vout0123456789ABCDEF); output += 16;
88   }
89   if XNN_UNLIKELY(n != 0) {
90     do {
91       const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
92       const int8x8_t vb01234567 = vld1_s8(input_b); input_b += 8;
93 
94       const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
95       const int16x8_t vxb01234567 = vsubl_s8(vb01234567, vb_zero_point);
96 
97       int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567));
98       int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567));
99 
100       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
101       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
102 
103       vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
104       vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
105 
106       vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
107       vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
108 
109       vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
110       vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
111 
112       #if XNN_ARCH_ARM64
113         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
114       #else
115         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
116       #endif
117 
118 
119       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
120 
121       vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
122       vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
123       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
124         vst1_s8(output, vout01234567); output += 8;
125         n -= 8 * sizeof(int8_t);
126       } else {
127         if (n & (4 * sizeof(int8_t))) {
128           vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
129           vout01234567 = vext_s8(vout01234567, vout01234567, 4);
130         }
131         if (n & (2 * sizeof(int8_t))) {
132           vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
133           vout01234567 = vext_s8(vout01234567, vout01234567, 2);
134         }
135         if (n & (1 * sizeof(int8_t))) {
136           vst1_lane_s8(output, vout01234567, 0);
137         }
138         n = 0;
139       }
140     } while (n != 0);
141   }
142 }
143