1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/wasmsimd.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/vadd.h>
15
16
xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16(
18 size_t n,
19 const int8_t* input_a,
20 const int8_t* input_b,
21 int8_t* output,
22 const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 const v128_t vbias = wasm_v128_load64_splat(params->wasmsimd.bias);
25 const v128_t va_multiplier = wasm_v128_load64_splat(params->wasmsimd.a_multiplier);
26 const v128_t vb_multiplier = wasm_v128_load64_splat(params->wasmsimd.b_multiplier);
27 const uint32_t vshift = params->wasmsimd.shift;
28 const v128_t voutput_zero_point = wasm_v128_load64_splat(params->wasmsimd.output_zero_point);
29 const v128_t voutput_min = wasm_v128_load64_splat(params->wasmsimd.output_min);
30 const v128_t voutput_max = wasm_v128_load64_splat(params->wasmsimd.output_max);
31
32 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
33 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
34 const v128_t vb01234567 = wasm_i16x8_load8x8(input_b);
35 const v128_t va89ABCDEF = wasm_i16x8_load8x8(input_a + 8);
36 const v128_t vb89ABCDEF = wasm_i16x8_load8x8(input_b + 8);
37 input_a += 16;
38 input_b += 16;
39
40 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(va01234567), va_multiplier));
41 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va01234567), va_multiplier));
42 v128_t vacc89AB = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(va89ABCDEF), va_multiplier));
43 v128_t vaccCDEF = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va89ABCDEF), va_multiplier));
44
45 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vb01234567), vb_multiplier));
46 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vb01234567), vb_multiplier));
47 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vb89ABCDEF), vb_multiplier));
48 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vb89ABCDEF), vb_multiplier));
49
50 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
51 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
52 vacc89AB = wasm_i32x4_shr(vacc89AB, vshift);
53 vaccCDEF = wasm_i32x4_shr(vaccCDEF, vshift);
54
55 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
56 v128_t vout89ABCDEF = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF), voutput_zero_point);
57
58 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
59
60 vout0123456789ABCDEF = wasm_i8x16_max(vout0123456789ABCDEF, voutput_min);
61
62 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
63
64 wasm_v128_store(output, vout0123456789ABCDEF);
65 output += 16;
66 }
67 if XNN_UNLIKELY(n != 0) {
68 do {
69 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
70 const v128_t vb01234567 = wasm_i16x8_load8x8(input_b);
71 input_a += 8;
72 input_b += 8;
73
74 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(va01234567), va_multiplier));
75 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va01234567), va_multiplier));
76
77 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vb01234567), vb_multiplier));
78 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vb01234567), vb_multiplier));
79
80 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
81 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
82
83 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
84
85 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
86 vout0123456701234567 = wasm_i8x16_max(vout0123456701234567, voutput_min);
87 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
88
89 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
90 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
91 output += 8;
92 n -= 8 * sizeof(int8_t);
93 } else {
94 if (n & (4 * sizeof(int8_t))) {
95 *((float*) output) = (float) wasm_f32x4_extract_lane(vout0123456701234567, 0);
96 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
97 output += 4;
98 }
99 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
100 if (n & (2 * sizeof(int8_t))) {
101 *((uint16_t*) output) = (uint16_t) vout0123;
102 vout0123 >>= 16;
103 output += 2;
104 }
105 if (n & (1 * sizeof(int8_t))) {
106 *output = (int8_t) vout0123;
107 }
108 n = 0;
109 }
110 } while (n != 0);
111 }
112 }
113