1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vaddc/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_qu8_vaddc_minmax_ukernel__wasmsimd_x32(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_vaddc_minmax_ukernel__wasmsimd_x32(
18 size_t n,
19 const uint8_t* input_a,
20 const uint8_t* input_b,
21 uint8_t* output,
22 const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 const v128_t va_multiplier = wasm_v128_load64_splat(params->wasmsimd.a_multiplier);
25 const uint32_t vshift = params->wasmsimd.shift;
26 const v128_t voutput_zero_point = wasm_v128_load64_splat(params->wasmsimd.output_zero_point);
27 const v128_t voutput_min = wasm_v128_load64_splat(params->wasmsimd.output_min);
28 const v128_t voutput_max = wasm_v128_load64_splat(params->wasmsimd.output_max);
29
30 v128_t vbias = wasm_i32x4_splat((int32_t) *input_b * params->wasmsimd.b_multiplier[0]);
31 vbias = wasm_i32x4_add(vbias, wasm_v128_load64_splat(params->wasmsimd.bias));
32
33 for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
34 const v128_t va01234567 = wasm_u16x8_load8x8(input_a);
35 const v128_t va89ABCDEF = wasm_u16x8_load8x8(input_a + 8);
36 const v128_t vaGHIJKLMN = wasm_u16x8_load8x8(input_a + 16);
37 const v128_t vaOPQRSTUV = wasm_u16x8_load8x8(input_a + 24);
38 input_a += 32;
39
40 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier));
41 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier));
42 v128_t vacc89AB = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va89ABCDEF), va_multiplier));
43 v128_t vaccCDEF = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va89ABCDEF), va_multiplier));
44 v128_t vaccGHIJ = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vaGHIJKLMN), va_multiplier));
45 v128_t vaccKLMN = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vaGHIJKLMN), va_multiplier));
46 v128_t vaccOPQR = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vaOPQRSTUV), va_multiplier));
47 v128_t vaccSTUV = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vaOPQRSTUV), va_multiplier));
48
49 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
50 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
51 vacc89AB = wasm_i32x4_shr(vacc89AB, vshift);
52 vaccCDEF = wasm_i32x4_shr(vaccCDEF, vshift);
53 vaccGHIJ = wasm_i32x4_shr(vaccGHIJ, vshift);
54 vaccKLMN = wasm_i32x4_shr(vaccKLMN, vshift);
55 vaccOPQR = wasm_i32x4_shr(vaccOPQR, vshift);
56 vaccSTUV = wasm_i32x4_shr(vaccSTUV, vshift);
57
58 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
59 v128_t vout89ABCDEF = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF), voutput_zero_point);
60 v128_t voutGHIJKLMN = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN), voutput_zero_point);
61 v128_t voutOPQRSTUV = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vaccOPQR, vaccSTUV), voutput_zero_point);
62
63 v128_t vout0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
64 v128_t voutGHIJKLMNOPQRSTUV = wasm_u8x16_narrow_i16x8(voutGHIJKLMN, voutOPQRSTUV);
65
66 vout0123456789ABCDEF = wasm_u8x16_max(vout0123456789ABCDEF, voutput_min);
67 voutGHIJKLMNOPQRSTUV = wasm_u8x16_max(voutGHIJKLMNOPQRSTUV, voutput_min);
68
69 vout0123456789ABCDEF = wasm_u8x16_min(vout0123456789ABCDEF, voutput_max);
70 voutGHIJKLMNOPQRSTUV = wasm_u8x16_min(voutGHIJKLMNOPQRSTUV, voutput_max);
71
72 wasm_v128_store(output, vout0123456789ABCDEF);
73 wasm_v128_store(output + 16, voutGHIJKLMNOPQRSTUV);
74 output += 32;
75 }
76 if XNN_UNLIKELY(n != 0) {
77 do {
78 const v128_t va01234567 = wasm_u16x8_load8x8(input_a);
79 input_a += 8;
80
81 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier));
82 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier));
83
84 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
85 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
86
87 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
88
89 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
90 vout0123456701234567 = wasm_u8x16_max(vout0123456701234567, voutput_min);
91 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
92
93 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
94 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
95 output += 8;
96 n -= 8 * sizeof(uint8_t);
97 } else {
98 if (n & (4 * sizeof(uint8_t))) {
99 *((float*) output) = (float) wasm_f32x4_extract_lane(vout0123456701234567, 0);
100 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
101 output += 4;
102 }
103 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
104 if (n & (2 * sizeof(uint8_t))) {
105 *((uint16_t*) output) = (uint16_t) vout0123;
106 vout0123 >>= 16;
107 output += 2;
108 }
109 if (n & (1 * sizeof(uint8_t))) {
110 *output = (uint8_t) vout0123;
111 }
112 n = 0;
113 }
114 } while (n != 0);
115 }
116 }
117