1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vmul/wasmsimd-mul32-ld64.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/vmul.h>
15
16
xnn_qu8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8(
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_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23
24 {
25 const v128_t va_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.a_zero_point);
26 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.b_zero_point);
27 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
28 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
29 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
30 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
31 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
32
33 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
34 const v128_t va01234567 = wasm_u16x8_load8x8(input_a);
35 const v128_t vb01234567 = wasm_u16x8_load8x8(input_b);
36 input_a += 8;
37 input_b += 8;
38
39 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point);
40 const v128_t vxb01234567 = wasm_i16x8_sub(vb01234567, vb_zero_point);
41
42 v128_t vacc0123 = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa01234567), wasm_i32x4_extend_low_i16x8(vxb01234567));
43 v128_t vacc4567 = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa01234567), wasm_i32x4_extend_high_i16x8(vxb01234567));
44
45 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
46 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
47
48 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
49 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
50
51 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
52 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
53
54 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
55 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
56
57 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
58 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
59
60 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
61
62 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
63
64 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
65
66 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
67 output += 8;
68 }
69 if XNN_UNLIKELY(n != 0) {
70 {
71 const v128_t va01234567 = wasm_u16x8_load8x8(input_a);
72 const v128_t vb01234567 = wasm_u16x8_load8x8(input_b);
73
74 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point);
75 const v128_t vxb01234567 = wasm_i16x8_sub(vb01234567, vb_zero_point);
76
77 v128_t vacc0123 = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa01234567), wasm_i32x4_extend_low_i16x8(vxb01234567));
78 v128_t vacc4567 = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa01234567), wasm_i32x4_extend_high_i16x8(vxb01234567));
79
80 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
81 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
82
83 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
84 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
85
86 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
87 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
88
89 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
90 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
91
92 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
93 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
94
95 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
96 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
97 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
98
99 if (n & (4 * sizeof(uint8_t))) {
100 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
101 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
102 output += 4;
103 }
104 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
105 if (n & (2 * sizeof(uint8_t))) {
106 *((uint16_t*) output) = (uint16_t) vout0123;
107 vout0123 >>= 16;
108 output += 2;
109 }
110 if (n & (1 * sizeof(uint8_t))) {
111 *output = (uint8_t) vout0123;
112 }
113 }
114 }
115 }
116