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_qs8_vmul_minmax_fp32_ukernel__neonv8_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)])18 void xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x16(
19 size_t n,
20 const int8_t* input_a,
21 const int8_t* input_b,
22 int8_t* output,
23 const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 const int8x8_t va_zero_point = vld1_dup_s8(params->fp32_neonv8.a_zero_point);
26 const int8x8_t vb_zero_point = vld1_dup_s8(params->fp32_neonv8.b_zero_point);
27 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
28 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
29 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min);
30 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neonv8.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 = vcvtnq_s32_f32(vfpacc0123);
59 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
60 vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
61 vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
62
63 #if XNN_ARCH_ARM64
64 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
65 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
66 #else
67 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
68 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
69 #endif
70
71 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
72 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
73
74 #if XNN_ARCH_ARM64
75 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
76 #else
77 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
78 #endif
79
80 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
81
82 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
83
84 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
85 }
86 if XNN_UNLIKELY(n != 0) {
87 do {
88 const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
89 const int8x8_t vb01234567 = vld1_s8(input_b); input_b += 8;
90
91 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
92 const int16x8_t vxb01234567 = vsubl_s8(vb01234567, vb_zero_point);
93
94 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567));
95 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567));
96
97 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
98 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
99
100 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
101 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
102
103 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
104 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
105
106 #if XNN_ARCH_ARM64
107 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
108 #else
109 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
110 #endif
111
112 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
113
114 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
115
116 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
117 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
118 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
119 vst1_s8(output, vout01234567); output += 8;
120 n -= 8 * sizeof(int8_t);
121 } else {
122 if (n & (4 * sizeof(int8_t))) {
123 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
124 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
125 }
126 if (n & (2 * sizeof(int8_t))) {
127 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
128 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
129 }
130 if (n & (1 * sizeof(int8_t))) {
131 vst1_lane_s8(output, vout01234567, 0);
132 }
133 n = 0;
134 }
135 } while (n != 0);
136 }
137 }
138