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_x32(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_x32(
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 >= 32 * sizeof(int8_t); n -= 32 * 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 const v128_t vaGHIJKLMN = wasm_i16x8_load8x8(input_a + 16);
38 const v128_t vbGHIJKLMN = wasm_i16x8_load8x8(input_b + 16);
39 const v128_t vaOPQRSTUV = wasm_i16x8_load8x8(input_a + 24);
40 const v128_t vbOPQRSTUV = wasm_i16x8_load8x8(input_b + 24);
41 input_a += 32;
42 input_b += 32;
43
44 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(va01234567), va_multiplier));
45 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va01234567), va_multiplier));
46 v128_t vacc89AB = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(va89ABCDEF), va_multiplier));
47 v128_t vaccCDEF = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va89ABCDEF), va_multiplier));
48 v128_t vaccGHIJ = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vaGHIJKLMN), va_multiplier));
49 v128_t vaccKLMN = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vaGHIJKLMN), va_multiplier));
50 v128_t vaccOPQR = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vaOPQRSTUV), va_multiplier));
51 v128_t vaccSTUV = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vaOPQRSTUV), va_multiplier));
52
53 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vb01234567), vb_multiplier));
54 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vb01234567), vb_multiplier));
55 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vb89ABCDEF), vb_multiplier));
56 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vb89ABCDEF), vb_multiplier));
57 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vbGHIJKLMN), vb_multiplier));
58 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vbGHIJKLMN), vb_multiplier));
59 vaccOPQR = wasm_i32x4_add(vaccOPQR, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vbOPQRSTUV), vb_multiplier));
60 vaccSTUV = wasm_i32x4_add(vaccSTUV, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vbOPQRSTUV), vb_multiplier));
61
62 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
63 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
64 vacc89AB = wasm_i32x4_shr(vacc89AB, vshift);
65 vaccCDEF = wasm_i32x4_shr(vaccCDEF, vshift);
66 vaccGHIJ = wasm_i32x4_shr(vaccGHIJ, vshift);
67 vaccKLMN = wasm_i32x4_shr(vaccKLMN, vshift);
68 vaccOPQR = wasm_i32x4_shr(vaccOPQR, vshift);
69 vaccSTUV = wasm_i32x4_shr(vaccSTUV, vshift);
70
71 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
72 v128_t vout89ABCDEF = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF), voutput_zero_point);
73 v128_t voutGHIJKLMN = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN), voutput_zero_point);
74 v128_t voutOPQRSTUV = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vaccOPQR, vaccSTUV), voutput_zero_point);
75
76 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
77 v128_t voutGHIJKLMNOPQRSTUV = wasm_i8x16_narrow_i16x8(voutGHIJKLMN, voutOPQRSTUV);
78
79 vout0123456789ABCDEF = wasm_i8x16_max(vout0123456789ABCDEF, voutput_min);
80 voutGHIJKLMNOPQRSTUV = wasm_i8x16_max(voutGHIJKLMNOPQRSTUV, voutput_min);
81
82 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
83 voutGHIJKLMNOPQRSTUV = wasm_i8x16_min(voutGHIJKLMNOPQRSTUV, voutput_max);
84
85 wasm_v128_store(output, vout0123456789ABCDEF);
86 wasm_v128_store(output + 16, voutGHIJKLMNOPQRSTUV);
87 output += 32;
88 }
89 if XNN_UNLIKELY(n != 0) {
90 do {
91 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
92 const v128_t vb01234567 = wasm_i16x8_load8x8(input_b);
93 input_a += 8;
94 input_b += 8;
95
96 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(va01234567), va_multiplier));
97 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va01234567), va_multiplier));
98
99 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vb01234567), vb_multiplier));
100 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vb01234567), vb_multiplier));
101
102 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
103 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
104
105 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
106
107 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
108 vout0123456701234567 = wasm_i8x16_max(vout0123456701234567, voutput_min);
109 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
110
111 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
112 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
113 output += 8;
114 n -= 8 * sizeof(int8_t);
115 } else {
116 if (n & (4 * sizeof(int8_t))) {
117 *((float*) output) = (float) wasm_f32x4_extract_lane(vout0123456701234567, 0);
118 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
119 output += 4;
120 }
121 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
122 if (n & (2 * sizeof(int8_t))) {
123 *((uint16_t*) output) = (uint16_t) vout0123;
124 vout0123 >>= 16;
125 output += 2;
126 }
127 if (n & (1 * sizeof(int8_t))) {
128 *output = (int8_t) vout0123;
129 }
130 n = 0;
131 }
132 } while (n != 0);
133 }
134 }
135