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/vmul.h>
15
16
xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8(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_vmulc_minmax_fp32_ukernel__neon_ld64_x8(
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 float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
26 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
27 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
28 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neon.output_min);
29 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neon.output_max);
30
31 const int8x8_t vb = vld1_dup_s8(input_b);
32 const int8x8_t vb_zero_point = vld1_dup_s8(params->fp32_neon.b_zero_point);
33 const int16x8_t vxb = vsubl_s8(vb, vb_zero_point);
34 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
35 const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
36
37 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
38
39 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
40 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
41
42 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
43 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
44
45 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
46 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
47
48 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
49 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
50
51 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
52 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
53
54 #if XNN_ARCH_ARM64
55 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
56 #else
57 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
58 #endif
59
60
61 #if XNN_ARCH_ARM64
62 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
63 #else
64 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
65 #endif
66
67 vout01234567 = vmax_s8(vout01234567, voutput_min);
68
69 vout01234567 = vmin_s8(vout01234567, voutput_max);
70
71 vst1_s8(output, vout01234567); output += 8;
72 }
73 if XNN_UNLIKELY(n != 0) {
74 {
75 const int8x8_t va01234567 = vld1_s8(input_a);
76
77 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
78
79 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
80 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
81
82 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
83 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
84
85 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
86 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
87
88 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
89 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
90
91 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
92 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
93
94 #if XNN_ARCH_ARM64
95 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
96 #else
97 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
98 #endif
99
100
101 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
102
103 vout01234567 = vmax_s8(vout01234567, voutput_min);
104 vout01234567 = vmin_s8(vout01234567, voutput_max);
105 if (n & (4 * sizeof(int8_t))) {
106 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
107 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
108 }
109 if (n & (2 * sizeof(int8_t))) {
110 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
111 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
112 }
113 if (n & (1 * sizeof(int8_t))) {
114 vst1_lane_s8(output, vout01234567, 0);
115 }
116 }
117 }
118 }
119