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_s8_ibilinear_ukernel__wasmsimd_dot16x2_c8(size_t output_pixels,size_t channels,const int8_t ** restrict input,size_t input_offset,const int16_t * restrict weights,int8_t * restrict output,size_t output_increment)18 void xnn_s8_ibilinear_ukernel__wasmsimd_dot16x2_c8(
19 size_t output_pixels,
20 size_t channels,
21 const int8_t**restrict input,
22 size_t input_offset,
23 const int16_t*restrict weights,
24 int8_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 int8_t* i0 = (const int8_t*) ((uintptr_t) input[0] + input_offset);
32 const int8_t* i1 = (const int8_t*) ((uintptr_t) input[1] + input_offset);
33 const int8_t* i2 = (const int8_t*) ((uintptr_t) input[2] + input_offset);
34 const int8_t* i3 = (const int8_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 >= 8 * sizeof(int8_t); c -= 8 * sizeof(int8_t)) {
50 const v128_t vtl01234567 = wasm_i16x8_load8x8(i0);
51 i0 += 8;
52 const v128_t vtr01234567 = wasm_i16x8_load8x8(i1);
53 i1 += 8;
54 const v128_t vbl01234567 = wasm_i16x8_load8x8(i2);
55 i2 += 8;
56 const v128_t vbr01234567 = wasm_i16x8_load8x8(i3);
57 i3 += 8;
58
59 const v128_t vdr01234567 = wasm_i16x8_sub(vbr01234567, vtr01234567);
60 const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
61 const v128_t vdl01234567 = wasm_i16x8_sub(vbl01234567, vtl01234567);
62 const v128_t vt4567 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
63
64 const v128_t vd0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr01234567, vdl01234567, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
65 const v128_t vd4567 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr01234567, vdl01234567, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
66
67 v128_t vacc0123 = wasm_i32x4_mul(vd0123, valphav);
68 v128_t vacc4567 = wasm_i32x4_mul(vd4567, valphav);
69
70 vacc0123 = wasm_i32x4_add(wasm_i32x4_shl(vt0123, 11), vacc0123);
71 vacc4567 = wasm_i32x4_add(wasm_i32x4_shl(vt4567, 11), vacc4567);
72
73 vacc0123 = wasm_i32x4_shr(wasm_i16x8_add(vacc0123, vrounding), 22);
74 vacc4567 = wasm_i32x4_shr(wasm_i16x8_add(vacc4567, vrounding), 22);
75
76 const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
77
78 const v128_t vo01234567 = wasm_i8x16_narrow_i16x8(vacc01234567, vacc01234567);
79
80 *((double*) output) = wasm_f64x2_extract_lane(vo01234567, 0);
81 output += 8;
82 }
83 if XNN_UNLIKELY(c != 0) {
84 const v128_t vtl01234567 = wasm_i16x8_load8x8(i0);
85 const v128_t vtr01234567 = wasm_i16x8_load8x8(i1);
86 const v128_t vbl01234567 = wasm_i16x8_load8x8(i2);
87 const v128_t vbr01234567 = wasm_i16x8_load8x8(i3);
88
89 const v128_t vdr01234567 = wasm_i16x8_sub(vbr01234567, vtr01234567);
90 const v128_t vt0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
91 const v128_t vdl01234567 = wasm_i16x8_sub(vbl01234567, vtl01234567);
92 const v128_t vt4567 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vtr01234567, vtl01234567, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
93
94 const v128_t vd0123 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr01234567, vdl01234567, 0, 8, 1, 9, 2, 10, 3, 11), valphah);
95 const v128_t vd4567 = wasm_i32x4_dot_i16x8(wasm_v16x8_shuffle(vdr01234567, vdl01234567, 4, 12, 5, 13, 6, 14, 7, 15), valphah);
96
97 v128_t vacc0123 = wasm_i32x4_mul(vd0123, valphav);
98 v128_t vacc4567 = wasm_i32x4_mul(vd4567, valphav);
99
100 vacc0123 = wasm_i32x4_add(wasm_i32x4_shl(vt0123, 11), vacc0123);
101 vacc4567 = wasm_i32x4_add(wasm_i32x4_shl(vt4567, 11), vacc4567);
102
103 vacc0123 = wasm_i32x4_shr(wasm_i16x8_add(vacc0123, vrounding), 22);
104 vacc4567 = wasm_i32x4_shr(wasm_i16x8_add(vacc4567, vrounding), 22);
105
106 const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
107
108 v128_t vo01234567 = wasm_i8x16_narrow_i16x8(vacc01234567, vacc01234567);
109
110 if (c & (4 * sizeof(int8_t))) {
111 *((float*) output) = wasm_f32x4_extract_lane(vo01234567, 0);
112 output += 4;
113 vo01234567 = wasm_u64x2_shr(vo01234567, 32);
114 }
115 uint32_t vo0123 = (uint32_t) wasm_i32x4_extract_lane(vo01234567, 0);
116 if (c & (2 * sizeof(int8_t))) {
117 *((uint16_t*) output) = (uint16_t) vo0123;
118 output += 2;
119 vo0123 >>= 16;
120 }
121 if (c & (1 * sizeof(int8_t))) {
122 *output++ = (uint8_t) vo0123;
123 }
124 }
125
126 output = (int8_t*) ((uintptr_t) output + output_increment);
127 } while (--output_pixels != 0);
128 }
129