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_qs8_vaddc_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_vaddc_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 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 >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
34 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
35 const v128_t va89ABCDEF = wasm_i16x8_load8x8(input_a + 8);
36 input_a += 16;
37
38 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(va01234567), va_multiplier));
39 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va01234567), va_multiplier));
40 v128_t vacc89AB = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(va89ABCDEF), va_multiplier));
41 v128_t vaccCDEF = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va89ABCDEF), va_multiplier));
42
43 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
44 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
45 vacc89AB = wasm_i32x4_shr(vacc89AB, vshift);
46 vaccCDEF = wasm_i32x4_shr(vaccCDEF, vshift);
47
48 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
49 v128_t vout89ABCDEF = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF), voutput_zero_point);
50
51 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
52
53 vout0123456789ABCDEF = wasm_i8x16_max(vout0123456789ABCDEF, voutput_min);
54
55 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
56
57 wasm_v128_store(output, vout0123456789ABCDEF);
58 output += 16;
59 }
60 if XNN_UNLIKELY(n != 0) {
61 do {
62 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
63 input_a += 8;
64
65 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(va01234567), va_multiplier));
66 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va01234567), va_multiplier));
67
68 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
69 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
70
71 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
72
73 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
74 vout0123456701234567 = wasm_i8x16_max(vout0123456701234567, voutput_min);
75 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
76
77 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
78 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
79 output += 8;
80 n -= 8 * sizeof(int8_t);
81 } else {
82 if (n & (4 * sizeof(int8_t))) {
83 *((float*) output) = (float) wasm_f32x4_extract_lane(vout0123456701234567, 0);
84 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
85 output += 4;
86 }
87 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
88 if (n & (2 * sizeof(int8_t))) {
89 *((uint16_t*) output) = (uint16_t) vout0123;
90 vout0123 >>= 16;
91 output += 2;
92 }
93 if (n & (1 * sizeof(int8_t))) {
94 *output = (int8_t) vout0123;
95 }
96 n = 0;
97 }
98 } while (n != 0);
99 }
100 }
101