Debian Codegraph @

src/qs8-vmulc/gen/qs8-vmulc-minmax-fp32-neonv8-ld64-u16.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/intrinsics-polyfill.h"
15 #include "xnnpack/vbinary.h"
16
17
18 void xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_u16(
19 size_t batch,
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 assert(batch != 0);
26 assert(batch % sizeof(int8_t) == 0);
27 assert(input_a != NULL);
28 assert(input_b != NULL);
29 assert(output != NULL);
30
31 const int8x8_t va_zero_point = vld1_dup_s8(&params->scalar.a_zero_point);
32 const float32x4_t vscale = vld1q_dup_f32(&params->scalar.scale);
33 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->scalar.output_zero_point);
34 const int8x16_t voutput_min = vld1q_dup_s8(&params->scalar.output_min);
35 const int8x16_t voutput_max = vld1q_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 >= 16 * sizeof(int8_t); batch -= 16 * sizeof(int8_t)) {
41 const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
42 const int8x8_t va89ABCDEF = vld1_s8(input_a); input_a += 8;
43
44 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
45 const int16x8_t vxa89ABCDEF = vsubl_s8(va89ABCDEF, va_zero_point);
46
47 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
48 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
49 int32x4_t vacc89AB = vmull_s16(vget_low_s16(vxa89ABCDEF), vget_low_s16(vxb));
50 int32x4_t vaccCDEF = vmull_s16(vget_high_s16(vxa89ABCDEF), vget_high_s16(vxb));
51
52 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
53 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
54 float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
55 float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
56
57 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
58 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
59 vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
60 vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
61
62 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
63 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
64 vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
65 vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
66
67 #if XNN_ARCH_ARM64
68 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
69 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
70 #else
71 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
72 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
73 #endif
74
75 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
76 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
77
78 #if XNN_ARCH_ARM64
79 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
80 #else
81 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
82 #endif
83
84 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
85
86 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
87
88 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
89 }
90 if XNN_UNLIKELY(batch != 0) {
91 do {
92 const int8x8_t va01234567 = vld1_s8(input_a); input_a += 8;
93
94 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point);
95
96 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
97 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
98
99 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
100 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
101
102 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
103 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
104
105 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
106 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
107
108 #if XNN_ARCH_ARM64
109 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
110 #else
111 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
112 #endif
113
114 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
115
116 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
117
118 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
119 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
120 if XNN_LIKELY(batch >= (8 * sizeof(int8_t))) {
121 vst1_s8(output, vout01234567); output += 8;
122 batch -= 8 * sizeof(int8_t);
123 } else {
124 if (batch & (4 * sizeof(int8_t))) {
125 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
126 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
127 }
128 if (batch & (2 * sizeof(int8_t))) {
129 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
130 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
131 }
132 if (batch & (1 * sizeof(int8_t))) {
133 vst1_lane_s8(output, vout01234567, 0);
134 }
135 batch = 0;
136 }
137 } while (batch != 0);
138 }
139 }