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_qs8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x16(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_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_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 >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
34 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
35 const v128_t vb01234567 = wasm_i16x8_load8x8(input_b);
36 const v128_t va89ABCDEF = wasm_i16x8_load8x8(input_a + 8);
37 const v128_t vb89ABCDEF = wasm_i16x8_load8x8(input_b + 8);
38 input_a += 16;
39 input_b += 16;
40
41 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point);
42 const v128_t vxb01234567 = wasm_i16x8_sub(vb01234567, vb_zero_point);
43 const v128_t vxa89ABCDEF = wasm_i16x8_sub(va89ABCDEF, va_zero_point);
44 const v128_t vxb89ABCDEF = wasm_i16x8_sub(vb89ABCDEF, vb_zero_point);
45
46 v128_t vacc0123 = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa01234567), wasm_i32x4_extend_low_i16x8(vxb01234567));
47 v128_t vacc4567 = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa01234567), wasm_i32x4_extend_high_i16x8(vxb01234567));
48 v128_t vacc89AB = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa89ABCDEF), wasm_i32x4_extend_low_i16x8(vxb89ABCDEF));
49 v128_t vaccCDEF = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa89ABCDEF), wasm_i32x4_extend_high_i16x8(vxb89ABCDEF));
50
51 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
52 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
53 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
54 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
55
56 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
57 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
58 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
59 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
60
61 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
62 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
63 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
64 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
65
66 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
67 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
68 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
69 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
70
71 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
72 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
73 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
74 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
75
76 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
77 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
78
79 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
80
81 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
82
83 wasm_v128_store(output, vout0123456789ABCDEF);
84 output += 16;
85 }
86 if XNN_UNLIKELY(n != 0) {
87 do {
88 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
89 const v128_t vb01234567 = wasm_i16x8_load8x8(input_b);
90 input_a += 8;
91 input_b += 8;
92
93 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point);
94 const v128_t vxb01234567 = wasm_i16x8_sub(vb01234567, vb_zero_point);
95
96 v128_t vacc0123 = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa01234567), wasm_i32x4_extend_low_i16x8(vxb01234567));
97 v128_t vacc4567 = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa01234567), wasm_i32x4_extend_high_i16x8(vxb01234567));
98
99 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
100 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
101
102 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
103 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
104
105 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
106 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
107
108 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
109 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
110
111 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
112 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
113
114 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
115 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
116 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
117
118 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
119 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
120 output += 8;
121 n -= 8 * sizeof(int8_t);
122 } else {
123 if (n & (4 * sizeof(int8_t))) {
124 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
125 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
126 output += 4;
127 }
128 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
129 if (n & (2 * sizeof(int8_t))) {
130 *((uint16_t*) output) = (uint16_t) vout0123;
131 vout0123 >>= 16;
132 output += 2;
133 }
134 if (n & (1 * sizeof(int8_t))) {
135 *output = (int8_t) vout0123;
136 }
137 n = 0;
138 }
139 } while (n != 0);
140 }
141 }
142