1 // Auto-generated file. Do not edit!
2 // Template: src/s8-ibilinear/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/common.h>
15 #include <xnnpack/ibilinear.h>
16
17
xnn_u8_ibilinear_ukernel__neon_c8(size_t output_pixels,size_t channels,const uint8_t ** restrict input,size_t input_offset,const int16_t * restrict weights,uint8_t * restrict output,size_t output_increment)18 void xnn_u8_ibilinear_ukernel__neon_c8(
19 size_t output_pixels,
20 size_t channels,
21 const uint8_t**restrict input,
22 size_t input_offset,
23 const int16_t*restrict weights,
24 uint8_t*restrict output,
25 size_t output_increment) XNN_OOB_READS
26 {
27 assert(output_pixels != 0);
28 assert(channels != 0);
29
30 do {
31 const uint8_t* i0 = (const uint8_t*) ((uintptr_t) input[0] + input_offset);
32 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) input[1] + input_offset);
33 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) input[2] + input_offset);
34 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) input[3] + input_offset);
35 input += 4;
36
37 #if XNN_ARCH_ARM64
38 const int16x8_t valphah = vld1q_dup_s16(weights); weights += 1;
39 #else
40 const int16x4_t valphah = vld1_dup_s16(weights); weights += 1;
41 #endif
42 const int32x4_t valphav = vmovl_s16(vld1_dup_s16(weights)); weights += 1;
43
44 size_t c = channels;
45 for (; c >= 8 * sizeof(uint8_t); c -= 8 * sizeof(uint8_t)) {
46 const uint8x8_t vtl01234567 = vld1_u8(i0); i0 += 8;
47 const uint8x8_t vtr01234567 = vld1_u8(i1); i1 += 8;
48 const uint8x8_t vbl01234567 = vld1_u8(i2); i2 += 8;
49 const uint8x8_t vbr01234567 = vld1_u8(i3); i3 += 8;
50
51 const int16x8_t vtd01234567 = vreinterpretq_s16_u16(vsubl_u8(vtr01234567, vtl01234567));
52 const int16x8_t vbd01234567 = vreinterpretq_s16_u16(vsubl_u8(vbr01234567, vbl01234567));
53 const int16x8_t vdl01234567 = vreinterpretq_s16_u16(vsubl_u8(vbl01234567, vtl01234567));
54 const int16x8_t vxtl01234567 = vreinterpretq_s16_u16(vmovl_u8(vtl01234567));
55
56 const int16x8_t vdd01234567 = vsubq_s16(vbd01234567, vtd01234567);
57
58 #if XNN_ARCH_ARM64
59 const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01234567), vget_low_s16(valphah));
60 const int32x4_t vt4567 = vmlal_high_s16(vshll_n_s16(vget_high_s16(vxtl01234567), 11), vtd01234567, valphah);
61
62 const int32x4_t vd0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vdl01234567), 11), vget_low_s16(vdd01234567), vget_low_s16(valphah));
63 const int32x4_t vd4567 = vmlal_high_s16(vshll_n_s16(vget_high_s16(vdl01234567), 11), vdd01234567, valphah);
64 #else // !XNN_ARCH_ARM64
65 const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01234567), valphah);
66 const int32x4_t vt4567 = vmlal_s16(vshll_n_s16(vget_high_s16(vxtl01234567), 11), vget_high_s16(vtd01234567), valphah);
67
68 const int32x4_t vd0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vdl01234567), 11), vget_low_s16(vdd01234567), valphah);
69 const int32x4_t vd4567 = vmlal_s16(vshll_n_s16(vget_high_s16(vdl01234567), 11), vget_high_s16(vdd01234567), valphah);
70 #endif // !XNN_ARCH_ARM64
71
72 const int32x4_t vacc0123 = vmlaq_s32(vshlq_n_s32(vt0123, 11), vd0123, valphav);
73 const int32x4_t vacc4567 = vmlaq_s32(vshlq_n_s32(vt4567, 11), vd4567, valphav);
74
75 #if XNN_ARCH_ARM64
76 const int16x8_t vacc01234567 = vuzp2q_s16(vreinterpretq_s16_s32(vacc0123), vreinterpretq_s16_s32(vacc4567));
77 #else // !XNN_ARCH_ARM64
78 const int16x8_t vacc01234567 = vcombine_s16(vshrn_n_s32(vacc0123, 16), vshrn_n_s32(vacc4567, 16));
79 #endif // !XNN_ARCH_ARM64
80
81 const uint8x8_t vo01234567 = vrshrn_n_u16(vreinterpretq_u16_s16(vacc01234567), 6);
82
83 vst1_u8(output, vo01234567); output += 8;
84 }
85 if XNN_UNLIKELY(c != 0) {
86 const uint8x8_t vtl01234567 = vld1_u8(i0);
87 const uint8x8_t vtr01234567 = vld1_u8(i1);
88 const uint8x8_t vbl01234567 = vld1_u8(i2);
89 const uint8x8_t vbr01234567 = vld1_u8(i3);
90
91 const int16x8_t vtd01234567 = vreinterpretq_s16_u16(vsubl_u8(vtr01234567, vtl01234567));
92 const int16x8_t vbd01234567 = vreinterpretq_s16_u16(vsubl_u8(vbr01234567, vbl01234567));
93 const int16x8_t vdl01234567 = vreinterpretq_s16_u16(vsubl_u8(vbl01234567, vtl01234567));
94 const int16x8_t vxtl01234567 = vreinterpretq_s16_u16(vmovl_u8(vtl01234567));
95
96 const int16x8_t vdd01234567 = vsubq_s16(vbd01234567, vtd01234567);
97
98 #if XNN_ARCH_ARM64
99 const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01234567), vget_low_s16(valphah));
100 const int32x4_t vt4567 = vmlal_high_s16(vshll_n_s16(vget_high_s16(vxtl01234567), 11), vtd01234567, valphah);
101
102 const int32x4_t vd0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vdl01234567), 11), vget_low_s16(vdd01234567), vget_low_s16(valphah));
103 const int32x4_t vd4567 = vmlal_high_s16(vshll_n_s16(vget_high_s16(vdl01234567), 11), vdd01234567, valphah);
104 #else // !XNN_ARCH_ARM64
105 const int32x4_t vt0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vxtl01234567), 11), vget_low_s16(vtd01234567), valphah);
106 const int32x4_t vt4567 = vmlal_s16(vshll_n_s16(vget_high_s16(vxtl01234567), 11), vget_high_s16(vtd01234567), valphah);
107
108 const int32x4_t vd0123 = vmlal_s16(vshll_n_s16(vget_low_s16(vdl01234567), 11), vget_low_s16(vdd01234567), valphah);
109 const int32x4_t vd4567 = vmlal_s16(vshll_n_s16(vget_high_s16(vdl01234567), 11), vget_high_s16(vdd01234567), valphah);
110 #endif // !XNN_ARCH_ARM64
111
112 const int32x4_t vacc0123 = vmlaq_s32(vshlq_n_s32(vt0123, 11), vd0123, valphav);
113 const int32x4_t vacc4567 = vmlaq_s32(vshlq_n_s32(vt4567, 11), vd4567, valphav);
114
115 #if XNN_ARCH_ARM64
116 const int16x8_t vacc01234567 = vuzp2q_s16(vreinterpretq_s16_s32(vacc0123), vreinterpretq_s16_s32(vacc4567));
117 #else // !XNN_ARCH_ARM64
118 const int16x8_t vacc01234567 = vcombine_s16(vshrn_n_s32(vacc0123, 16), vshrn_n_s32(vacc4567, 16));
119 #endif // !XNN_ARCH_ARM64
120
121 uint8x8_t vo01234567 = vrshrn_n_u16(vreinterpretq_u16_s16(vacc01234567), 6);
122
123 if (c & (4 * sizeof(uint8_t))) {
124 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vo01234567), 0); output += 4;
125 vo01234567 = vext_u8(vo01234567, vo01234567, 4);
126 }
127 if (c & (2 * sizeof(uint8_t))) {
128 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vo01234567), 0); output += 2;
129 vo01234567 = vext_u8(vo01234567, vo01234567, 2);
130 }
131 if (c & (1 * sizeof(uint8_t))) {
132 vst1_lane_u8(output, vo01234567, 0); output += 1;
133 }
134 }
135
136 output = (uint8_t*) ((uintptr_t) output + output_increment);
137 } while (--output_pixels != 0);
138 }
139