1 // Auto-generated file. Do not edit!
2 // Template: src/s8-ibilinear/wasmsimd-dot16x2.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 <wasm_simd128.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/ibilinear.h>
16
17
xnn_u8_ibilinear_ukernel__wasmsimd_dot16x2_c16(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__wasmsimd_dot16x2_c16(
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 const v128_t valphah =
38 wasm_i16x8_add(
39 wasm_v128_xor(
40 wasm_v128_load16_splat(weights),
41 wasm_i32x4_const_splat(0xFFFF0000)),
42 wasm_i32x4_const_splat(0x08010000));
43 const v128_t valphav = wasm_i32x4_extend_low_i16x8(wasm_v128_load16_splat(weights + 1));
44 weights += 2;
45
46 const v128_t vrounding = wasm_i32x4_const_splat(0x00200000);
47
48 size_t c = channels;
49 for (; c >= 16 * sizeof(uint8_t); c -= 16 * sizeof(uint8_t)) {
50 const v128_t vtl01234567 = wasm_u16x8_load8x8(i0);
51 const v128_t vtr01234567 = wasm_u16x8_load8x8(i1);
52 const v128_t vbl01234567 = wasm_u16x8_load8x8(i2);
53 const v128_t vbr01234567 = wasm_u16x8_load8x8(i3);
54 const v128_t vtl89ABCDEF = wasm_u16x8_load8x8(i0 + 8);
55 const v128_t vtr89ABCDEF = wasm_u16x8_load8x8(i1 + 8);
56 const v128_t vbl89ABCDEF = wasm_u16x8_load8x8(i2 + 8);
57 const v128_t vbr89ABCDEF = wasm_u16x8_load8x8(i3 + 8);
58 i0 += 16;
59 i1 += 16;
60 i2 += 16;
61 i3 += 16;
62
63 const v128_t vdr01234567 = wasm_i16x8_sub(vbr01234567, vtr01234567);
64 const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
65 const v128_t vdl01234567 = wasm_i16x8_sub(vbl01234567, vtl01234567);
66 const v128_t vt4567 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
67 const v128_t vdr89ABCDEF = wasm_i16x8_sub(vbr89ABCDEF, vtr89ABCDEF);
68 const v128_t vt89AB = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr89ABCDEF, vtl89ABCDEF, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
69 const v128_t vdl89ABCDEF = wasm_i16x8_sub(vbl89ABCDEF, vtl89ABCDEF);
70 const v128_t vtCDEF = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr89ABCDEF, vtl89ABCDEF, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
71
72 const v128_t vd0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr01234567, vdl01234567, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
73 const v128_t vd4567 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr01234567, vdl01234567, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
74 const v128_t vd89AB = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr89ABCDEF, vdl89ABCDEF, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
75 const v128_t vdCDEF = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr89ABCDEF, vdl89ABCDEF, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
76
77 v128_t vacc0123 = wasm_i32x4_mul(vd0123, valphav);
78 v128_t vacc4567 = wasm_i32x4_mul(vd4567, valphav);
79 v128_t vacc89AB = wasm_i32x4_mul(vd89AB, valphav);
80 v128_t vaccCDEF = wasm_i32x4_mul(vdCDEF, valphav);
81
82 vacc0123 = wasm_i32x4_add(wasm_i32x4_shl(vt0123, 11), vacc0123);
83 vacc4567 = wasm_i32x4_add(wasm_i32x4_shl(vt4567, 11), vacc4567);
84 vacc89AB = wasm_i32x4_add(wasm_i32x4_shl(vt89AB, 11), vacc89AB);
85 vaccCDEF = wasm_i32x4_add(wasm_i32x4_shl(vtCDEF, 11), vaccCDEF);
86
87 vacc0123 = wasm_u32x4_shr(wasm_i16x8_add(vacc0123, vrounding), 22);
88 vacc4567 = wasm_u32x4_shr(wasm_i16x8_add(vacc4567, vrounding), 22);
89 vacc89AB = wasm_u32x4_shr(wasm_i16x8_add(vacc89AB, vrounding), 22);
90 vaccCDEF = wasm_u32x4_shr(wasm_i16x8_add(vaccCDEF, vrounding), 22);
91
92 const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
93 const v128_t vacc89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
94
95 const v128_t vo0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vacc01234567, vacc89ABCDEF);
96
97 wasm_v128_store(output, vo0123456789ABCDEF);
98 output += 16;
99 }
100 for (; c >= 8 * sizeof(uint8_t); c -= 8 * sizeof(uint8_t)) {
101 const v128_t vtl01234567 = wasm_u16x8_load8x8(i0);
102 i0 += 8;
103 const v128_t vtr01234567 = wasm_u16x8_load8x8(i1);
104 i1 += 8;
105 const v128_t vbl01234567 = wasm_u16x8_load8x8(i2);
106 i2 += 8;
107 const v128_t vbr01234567 = wasm_u16x8_load8x8(i3);
108 i3 += 8;
109
110 const v128_t vdr01234567 = wasm_i16x8_sub(vbr01234567, vtr01234567);
111 const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
112 const v128_t vdl01234567 = wasm_i16x8_sub(vbl01234567, vtl01234567);
113 const v128_t vt4567 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
114
115 const v128_t vd0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr01234567, vdl01234567, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
116 const v128_t vd4567 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr01234567, vdl01234567, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
117
118 v128_t vacc0123 = wasm_i32x4_mul(vd0123, valphav);
119 v128_t vacc4567 = wasm_i32x4_mul(vd4567, valphav);
120
121 vacc0123 = wasm_i32x4_add(wasm_i32x4_shl(vt0123, 11), vacc0123);
122 vacc4567 = wasm_i32x4_add(wasm_i32x4_shl(vt4567, 11), vacc4567);
123
124 vacc0123 = wasm_u32x4_shr(wasm_i16x8_add(vacc0123, vrounding), 22);
125 vacc4567 = wasm_u32x4_shr(wasm_i16x8_add(vacc4567, vrounding), 22);
126
127 const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
128
129 const v128_t vo01234567 = wasm_u8x16_narrow_i16x8(vacc01234567, vacc01234567);
130
131 *((double*) output) = wasm_f64x2_extract_lane(vo01234567, 0);
132 output += 8;
133 }
134 if XNN_UNLIKELY(c != 0) {
135 const v128_t vtl01234567 = wasm_u16x8_load8x8(i0);
136 const v128_t vtr01234567 = wasm_u16x8_load8x8(i1);
137 const v128_t vbl01234567 = wasm_u16x8_load8x8(i2);
138 const v128_t vbr01234567 = wasm_u16x8_load8x8(i3);
139
140 const v128_t vdr01234567 = wasm_i16x8_sub(vbr01234567, vtr01234567);
141 const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
142 const v128_t vdl01234567 = wasm_i16x8_sub(vbl01234567, vtl01234567);
143 const v128_t vt4567 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
144
145 const v128_t vd0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr01234567, vdl01234567, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
146 const v128_t vd4567 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr01234567, vdl01234567, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
147
148 v128_t vacc0123 = wasm_i32x4_mul(vd0123, valphav);
149 v128_t vacc4567 = wasm_i32x4_mul(vd4567, valphav);
150
151 vacc0123 = wasm_i32x4_add(wasm_i32x4_shl(vt0123, 11), vacc0123);
152 vacc4567 = wasm_i32x4_add(wasm_i32x4_shl(vt4567, 11), vacc4567);
153
154 vacc0123 = wasm_u32x4_shr(wasm_i16x8_add(vacc0123, vrounding), 22);
155 vacc4567 = wasm_u32x4_shr(wasm_i16x8_add(vacc4567, vrounding), 22);
156
157 const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
158
159 v128_t vo01234567 = wasm_u8x16_narrow_i16x8(vacc01234567, vacc01234567);
160
161 if (c & (4 * sizeof(uint8_t))) {
162 *((float*) output) = wasm_f32x4_extract_lane(vo01234567, 0);
163 output += 4;
164 vo01234567 = wasm_u64x2_shr(vo01234567, 32);
165 }
166 uint32_t vo0123 = (uint32_t) wasm_i32x4_extract_lane(vo01234567, 0);
167 if (c & (2 * sizeof(uint8_t))) {
168 *((uint16_t*) output) = (uint16_t) vo0123;
169 output += 2;
170 vo0123 >>= 16;
171 }
172 if (c & (1 * sizeof(uint8_t))) {
173 *output++ = (uint8_t) vo0123;
174 }
175 }
176
177 output = (uint8_t*) ((uintptr_t) output + output_increment);
178 } while (--output_pixels != 0);
179 }
180