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_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8(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)])17 void xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8(
18 size_t n,
19 const uint8_t* input_a,
20 const uint8_t* input_b,
21 uint8_t* output,
22 const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 const uint8x8_t va_zero_point = vld1_dup_u8(params->rndnu_neon.a_zero_point);
25 const uint8x8_t vb_zero_point = vld1_dup_u8(params->rndnu_neon.b_zero_point);
26 const int32x4_t vleft_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.left_pre_shift);
27 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
28 const int32x4_t vleft_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.left_post_shift);
29 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
30 const uint8x8_t voutput_min = vld1_dup_u8(¶ms->rndnu_neon.output_min);
31 const uint8x8_t voutput_max = vld1_dup_u8(¶ms->rndnu_neon.output_max);
32
33 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
34 const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
35 const uint8x8_t vb01234567 = vld1_u8(input_b); input_b += 8;
36
37 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
38 const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));
39
40 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567));
41 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567));
42
43 vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
44 vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
45
46 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
47 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
48
49 vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
50 vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
51
52 #if XNN_ARCH_ARM64
53 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
54 #else
55 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
56 #endif
57
58 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
59
60 #if XNN_ARCH_ARM64
61 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
62 #else
63 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
64 #endif
65
66 vout01234567 = vmax_u8(vout01234567, voutput_min);
67
68 vout01234567 = vmin_u8(vout01234567, voutput_max);
69
70 vst1_u8(output, vout01234567); output += 8;
71 }
72 if XNN_UNLIKELY(n != 0) {
73 {
74 const uint8x8_t va01234567 = vld1_u8(input_a);
75 const uint8x8_t vb01234567 = vld1_u8(input_b);
76
77 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
78 const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));
79
80 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567));
81 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567));
82
83 vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
84 vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
85
86 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
87 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
88
89 vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
90 vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
91
92 #if XNN_ARCH_ARM64
93 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
94 #else
95 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
96 #endif
97
98 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
99
100 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
101
102 vout01234567 = vmax_u8(vout01234567, voutput_min);
103 vout01234567 = vmin_u8(vout01234567, voutput_max);
104 if (n & (4 * sizeof(uint8_t))) {
105 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
106 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
107 }
108 if (n & (2 * sizeof(uint8_t))) {
109 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
110 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
111 }
112 if (n & (1 * sizeof(uint8_t))) {
113 vst1_lane_u8(output, vout01234567, 0);
114 }
115 }
116 }
117 }
118