1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vmulc/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_vmulc_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_vmulc_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 #else
28 const uint8x8_t va_zero_point = vld1_dup_u8(params->fp32_neonv8.a_zero_point);
29 #endif
30 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
31 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
32 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->fp32_neonv8.output_min);
33 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->fp32_neonv8.output_max);
34
35 const uint8x8_t vb = vld1_dup_u8(input_b);
36 const uint8x8_t vb_zero_point = vld1_dup_u8(params->fp32_neonv8.b_zero_point);
37 const int16x8_t vxb = vreinterpretq_s16_u16(vsubl_u8(vb, vb_zero_point));
38 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
39 const uint8x16_t va0123456789ABCDEF = vld1q_u8(input_a); input_a += 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 #else // !XNN_ARCH_ARM64
45 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_zero_point));
46 const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(va0123456789ABCDEF), va_zero_point));
47 #endif // XNN_ARCH_ARM64
48
49 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
50 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
51 int32x4_t vacc89AB = vmull_s16(vget_low_s16(vxa89ABCDEF), vget_low_s16(vxb));
52 int32x4_t vaccCDEF = vmull_s16(vget_high_s16(vxa89ABCDEF), vget_high_s16(vxb));
53
54 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
55 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
56 float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
57 float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
58
59 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
60 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
61 vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
62 vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
63
64 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
65 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
66 vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
67 vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
68
69 #if XNN_ARCH_ARM64
70 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
71 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
72 #else
73 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
74 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
75 #endif
76
77 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
78 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
79
80 #if XNN_ARCH_ARM64
81 uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
82 #else
83 uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
84 #endif
85
86 vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
87
88 vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
89
90 vst1q_u8(output, vout0123456789ABCDEF); output += 16;
91 }
92 if XNN_UNLIKELY(n != 0) {
93 do {
94 const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
95
96 #if XNN_ARCH_ARM64
97 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point)));
98 #else // !XNN_ARCH_ARM64
99 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
100 #endif // XNN_ARCH_ARM64
101
102 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
103 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
104
105 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
106 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
107
108 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
109 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
110
111 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
112 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
113
114 #if XNN_ARCH_ARM64
115 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
116 #else
117 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
118 #endif
119
120 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
121
122 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
123
124 vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
125 vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
126 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
127 vst1_u8(output, vout01234567); output += 8;
128 n -= 8 * sizeof(uint8_t);
129 } else {
130 if (n & (4 * sizeof(uint8_t))) {
131 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
132 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
133 }
134 if (n & (2 * sizeof(uint8_t))) {
135 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
136 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
137 }
138 if (n & (1 * sizeof(uint8_t))) {
139 vst1_lane_u8(output, vout01234567, 0);
140 }
141 n = 0;
142 }
143 } while (n != 0);
144 }
145 }
146