1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vmulc/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_vmulc_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_vmulc_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 vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
27 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
28 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
29 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
30 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
31
32 const v128_t vxb = wasm_i16x8_sub(
33 wasm_i16x8_splat((int16_t) *input_b), wasm_v128_load64_splat(params->fp32_wasmsimd.b_zero_point));
34 const v128_t vxblo = wasm_i32x4_extend_low_i16x8(vxb);
35 const v128_t vxbhi = wasm_i32x4_extend_high_i16x8(vxb);
36 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
37 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
38 const v128_t va89ABCDEF = wasm_i16x8_load8x8(input_a + 8);
39 input_a += 16;
40
41 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point);
42 const v128_t vxa89ABCDEF = wasm_i16x8_sub(va89ABCDEF, va_zero_point);
43
44 v128_t vacc0123 = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa01234567), vxblo);
45 v128_t vacc4567 = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa01234567), vxbhi);
46 v128_t vacc89AB = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa89ABCDEF), vxblo);
47 v128_t vaccCDEF = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa89ABCDEF), vxbhi);
48
49 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
50 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
51 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
52 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
53
54 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
55 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
56 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
57 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
58
59 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
60 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
61 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
62 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
63
64 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
65 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
66 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
67 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
68
69 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
70 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
71 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
72 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
73
74 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
75 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
76
77 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
78
79 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
80
81 wasm_v128_store(output, vout0123456789ABCDEF);
82 output += 16;
83 }
84 if XNN_UNLIKELY(n != 0) {
85 do {
86 const v128_t va01234567 = wasm_i16x8_load8x8(input_a);
87 input_a += 8;
88
89 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point);
90
91 v128_t vacc0123 = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa01234567), vxblo);
92 v128_t vacc4567 = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa01234567), vxbhi);
93
94 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
95 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
96
97 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
98 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
99
100 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
101 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
102
103 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
104 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
105
106 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
107 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
108
109 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
110 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
111 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
112
113 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
114 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
115 output += 8;
116 n -= 8 * sizeof(int8_t);
117 } else {
118 if (n & (4 * sizeof(int8_t))) {
119 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
120 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
121 output += 4;
122 }
123 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
124 if (n & (2 * sizeof(int8_t))) {
125 *((uint16_t*) output) = (uint16_t) vout0123;
126 vout0123 >>= 16;
127 output += 2;
128 }
129 if (n & (1 * sizeof(int8_t))) {
130 *output = (int8_t) vout0123;
131 }
132 n = 0;
133 }
134 } while (n != 0);
135 }
136 }
137