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_x8(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_x8(
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 >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
34 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
35 input_a += 8;
36
37 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(va01234567), va_multiplier));
38 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va01234567), va_multiplier));
39
40 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
41 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
42
43 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
44
45 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
46
47 vout0123456701234567 = wasm_i8x16_max(vout0123456701234567, voutput_min);
48
49 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
50
51 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
52 output += 8;
53 }
54 if XNN_UNLIKELY(n != 0) {
55 {
56 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
57
58 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(va01234567), va_multiplier));
59 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va01234567), va_multiplier));
60
61 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
62 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
63
64 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
65
66 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
67 vout0123456701234567 = wasm_i8x16_max(vout0123456701234567, voutput_min);
68 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
69
70 if (n & (4 * sizeof(int8_t))) {
71 *((float*) output) = (float) wasm_f32x4_extract_lane(vout0123456701234567, 0);
72 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
73 output += 4;
74 }
75 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
76 if (n & (2 * sizeof(int8_t))) {
77 *((uint16_t*) output) = (uint16_t) vout0123;
78 vout0123 >>= 16;
79 output += 2;
80 }
81 if (n & (1 * sizeof(int8_t))) {
82 *output = (int8_t) vout0123;
83 }
84 }
85 }
86 }
87