Debian Codegraph @

src/qs8-vmulc/gen/qs8-vmulc-minmax-fp32-neon-ld64-u8.c

raw
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/vbinary.h"
15
16
17 void xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_u8(
18 size_t batch,
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 assert(batch != 0);
25 assert(batch % sizeof(int8_t) == 0);
26 assert(input_a != NULL);
27 assert(input_b != NULL);
28 assert(output != NULL);
29
30 const int8x8_t va_zero_point = vld1_dup_s8(&params->scalar.a_zero_point);
31 const float32x4_t vscale = vld1q_dup_f32(&params->scalar.scale);
32 const float32x4_t vmagic_bias = vdupq_n_f32(12582912.0f);
33 const int32x4_t vmagic_bias_less_output_zero_point = vdupq_n_s32(INT32_C(0x4B400000) - (int32_t) params->scalar.output_zero_point);
34 const int8x8_t voutput_min = vld1_dup_s8(&params->scalar.output_min);
35 const int8x8_t voutput_max = vld1_dup_s8(&params->scalar.output_max);
36
37 const int8x8_t vb = vld1_dup_s8(input_b);
38 const int8x8_t vb_zero_point = vld1_dup_s8(&params->scalar.b_zero_point);
39 const int16x8_t vxb = vsubl_s8(vb, vb_zero_point);
40 for (; batch >= 8 * sizeof(int8_t); batch -= 8 * sizeof(int8_t)) {
41 const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
42
43 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
44
45 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
46 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
47
48 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
49 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
50
51 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
52 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
53
54 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
55 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
56
57 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
58 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
59
60 #if XNN_ARCH_ARM64
61 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
62 #else
63 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
64 #endif
65
66
67 #if XNN_ARCH_ARM64
68 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
69 #else
70 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
71 #endif
72
73 vout01234567 = vmax_s8(vout01234567, voutput_min);
74
75 vout01234567 = vmin_s8(vout01234567, voutput_max);
76
77 vst1_s8(output, vout01234567); output += 8;
78 }
79 if XNN_UNLIKELY(batch != 0) {
80 {
81 const int8x8_t va01234567 = vld1_s8(input_a);
82
83 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
84
85 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
86 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
87
88 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
89 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
90
91 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
92 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
93
94 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
95 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
96
97 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
98 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
99
100 #if XNN_ARCH_ARM64
101 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
102 #else
103 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
104 #endif
105
106
107 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
108
109 vout01234567 = vmax_s8(vout01234567, voutput_min);
110 vout01234567 = vmin_s8(vout01234567, voutput_max);
111 if (batch & (4 * sizeof(int8_t))) {
112 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
113 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
114 }
115 if (batch & (2 * sizeof(int8_t))) {
116 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
117 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
118 }
119 if (batch & (1 * sizeof(int8_t))) {
120 vst1_lane_s8(output, vout01234567, 0);
121 }
122 }
123 }
124 }