1 // Auto-generated file. Do not edit!
2 // Template: src/s8-ibilinear/wasmsimd-mul32.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_mul32_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_mul32_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 = wasm_i32x4_extend_low_i16x8(wasm_v128_load16_splat(weights));
38 const v128_t valphav = wasm_i32x4_extend_low_i16x8(wasm_v128_load16_splat(weights + 1));
39 weights += 2;
40
41 const v128_t vrounding = wasm_i32x4_const_splat(0x00200000);
42
43 size_t c = channels;
44 for (; c >= 8 * sizeof(int8_t); c -= 8 * sizeof(int8_t)) {
45 const v128_t vtl01234567 = wasm_i16x8_load8x8(i0);
46 i0 += 8;
47 const v128_t vtr01234567 = wasm_i16x8_load8x8(i1);
48 i1 += 8;
49 const v128_t vbl01234567 = wasm_i16x8_load8x8(i2);
50 i2 += 8;
51 const v128_t vbr01234567 = wasm_i16x8_load8x8(i3);
52 i3 += 8;
53
54 const v128_t vtd01234567 = wasm_i16x8_sub(vtr01234567, vtl01234567);
55 const v128_t vbd01234567 = wasm_i16x8_sub(vbr01234567, vbl01234567);
56 const v128_t vdl01234567 = wasm_i16x8_sub(vbl01234567, vtl01234567);
57 const v128_t vdd01234567 = wasm_i16x8_sub(vbd01234567, vtd01234567);
58
59 const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vtd01234567), valphah));
60 const v128_t vt4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vtd01234567), valphah));
61 const v128_t vd0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vdd01234567), valphah));
62 const v128_t vd4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vdd01234567), valphah));
63
64 v128_t vacc0123 = wasm_i32x4_mul(vd0123, valphav);
65 v128_t vacc4567 = wasm_i32x4_mul(vd4567, valphav);
66
67 vacc0123 = wasm_i32x4_add(wasm_i32x4_shl(vt0123, 11), vacc0123);
68 vacc4567 = wasm_i32x4_add(wasm_i32x4_shl(vt4567, 11), vacc4567);
69
70 vacc0123 = wasm_i32x4_shr(wasm_i16x8_add(vacc0123, vrounding), 22);
71 vacc4567 = wasm_i32x4_shr(wasm_i16x8_add(vacc4567, vrounding), 22);
72
73 const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
74
75 const v128_t vo01234567 = wasm_i8x16_narrow_i16x8(vacc01234567, vacc01234567);
76
77 *((double*) output) = wasm_f64x2_extract_lane(vo01234567, 0);
78 output += 8;
79 }
80 if XNN_UNLIKELY(c != 0) {
81 const v128_t vtl01234567 = wasm_i16x8_load8x8(i0);
82 const v128_t vtr01234567 = wasm_i16x8_load8x8(i1);
83 const v128_t vbl01234567 = wasm_i16x8_load8x8(i2);
84 const v128_t vbr01234567 = wasm_i16x8_load8x8(i3);
85
86 const v128_t vtd01234567 = wasm_i16x8_sub(vtr01234567, vtl01234567);
87 const v128_t vbd01234567 = wasm_i16x8_sub(vbr01234567, vbl01234567);
88 const v128_t vdl01234567 = wasm_i16x8_sub(vbl01234567, vtl01234567);
89 const v128_t vdd01234567 = wasm_i16x8_sub(vbd01234567, vtd01234567);
90
91 const v128_t vt0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vtd01234567), valphah));
92 const v128_t vt4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vtl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vtd01234567), valphah));
93 const v128_t vd0123 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_low_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vdd01234567), valphah));
94 const v128_t vd4567 = wasm_i32x4_add(wasm_i32x4_shl(wasm_i32x4_extend_high_i16x8(vdl01234567), 11), wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vdd01234567), valphah));
95
96 v128_t vacc0123 = wasm_i32x4_mul(vd0123, valphav);
97 v128_t vacc4567 = wasm_i32x4_mul(vd4567, valphav);
98
99 vacc0123 = wasm_i32x4_add(wasm_i32x4_shl(vt0123, 11), vacc0123);
100 vacc4567 = wasm_i32x4_add(wasm_i32x4_shl(vt4567, 11), vacc4567);
101
102 vacc0123 = wasm_i32x4_shr(wasm_i16x8_add(vacc0123, vrounding), 22);
103 vacc4567 = wasm_i32x4_shr(wasm_i16x8_add(vacc4567, vrounding), 22);
104
105 const v128_t vacc01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
106
107 v128_t vo01234567 = wasm_i8x16_narrow_i16x8(vacc01234567, vacc01234567);
108
109 if (c & (4 * sizeof(int8_t))) {
110 *((float*) output) = wasm_f32x4_extract_lane(vo01234567, 0);
111 output += 4;
112 vo01234567 = wasm_u64x2_shr(vo01234567, 32);
113 }
114 uint32_t vo0123 = (uint32_t) wasm_i32x4_extract_lane(vo01234567, 0);
115 if (c & (2 * sizeof(int8_t))) {
116 *((uint16_t*) output) = (uint16_t) vo0123;
117 output += 2;
118 vo0123 >>= 16;
119 }
120 if (c & (1 * sizeof(int8_t))) {
121 *output++ = (uint8_t) vo0123;
122 }
123 }
124
125 output = (int8_t*) ((uintptr_t) output + output_increment);
126 } while (--output_pixels != 0);
127 }
128