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/intrinsics-polyfill.h>
15 #include <xnnpack/vmul.h>
16
17
xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld128_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)])18 void xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16(
19 size_t n,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 #if XNN_ARCH_ARM64
26 const uint8x16_t va_zero_point = vld1q_dup_u8(params->fp32_neonv8.a_zero_point);
27 const uint8x16_t vb_zero_point = vld1q_dup_u8(params->fp32_neonv8.b_zero_point);
28 #else
29 const uint8x8_t va_zero_point = vld1_dup_u8(params->fp32_neonv8.a_zero_point);
30 const uint8x8_t vb_zero_point = vld1_dup_u8(params->fp32_neonv8.b_zero_point);
31 #endif
32 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
33 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
34 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->fp32_neonv8.output_min);
35 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->fp32_neonv8.output_max);
36
37 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
38 const uint8x16_t va0123456789ABCDEF = vld1q_u8(input_a); input_a += 16;
39 const uint8x16_t vb0123456789ABCDEF = vld1q_u8(input_b); input_b += 16;
40
41 #if XNN_ARCH_ARM64
42 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), vget_low_u8(va_zero_point)));
43 const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_high_u8(va0123456789ABCDEF, va_zero_point));
44 const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(vb0123456789ABCDEF), vget_low_u8(vb_zero_point)));
45 const int16x8_t vxb89ABCDEF = vreinterpretq_s16_u16(vsubl_high_u8(vb0123456789ABCDEF, vb_zero_point));
46 #else // !XNN_ARCH_ARM64
47 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_zero_point));
48 const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(va0123456789ABCDEF), va_zero_point));
49 const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(vb0123456789ABCDEF), vb_zero_point));
50 const int16x8_t vxb89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(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 = vcvtnq_s32_f32(vfpacc0123);
69 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
70 vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
71 vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
72
73 #if XNN_ARCH_ARM64
74 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
75 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
76 #else
77 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
78 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
79 #endif
80
81 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
82 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
83
84 #if XNN_ARCH_ARM64
85 uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
86 #else
87 uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
88 #endif
89
90 vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
91
92 vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
93
94 vst1q_u8(output, vout0123456789ABCDEF); output += 16;
95 }
96 if XNN_UNLIKELY(n != 0) {
97 do {
98 const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
99 const uint8x8_t vb01234567 = vld1_u8(input_b); input_b += 8;
100
101 #if XNN_ARCH_ARM64
102 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point)));
103 const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vget_low_u8(vb_zero_point)));
104 #else // !XNN_ARCH_ARM64
105 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
106 const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));
107 #endif // XNN_ARCH_ARM64
108
109 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567));
110 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567));
111
112 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
113 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
114
115 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
116 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
117
118 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
119 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
120
121 #if XNN_ARCH_ARM64
122 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
123 #else
124 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
125 #endif
126
127 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
128
129 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
130
131 vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
132 vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
133 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
134 vst1_u8(output, vout01234567); output += 8;
135 n -= 8 * sizeof(uint8_t);
136 } else {
137 if (n & (4 * sizeof(uint8_t))) {
138 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
139 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
140 }
141 if (n & (2 * sizeof(uint8_t))) {
142 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
143 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
144 }
145 if (n & (1 * sizeof(uint8_t))) {
146 vst1_lane_u8(output, vout01234567, 0);
147 }
148 n = 0;
149 }
150 } while (n != 0);
151 }
152 }
153