Debian Codegraph @

src/qu8-vadd/gen/qu8-vadd-minmax-wasmsimd-u16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/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/common.h"
15 #include "xnnpack/vbinary.h"
16
17
18 void xnn_qu8_vadd_minmax_ukernel__wasmsimd_u16(
19 size_t batch,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const struct xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(uint8_t) == 0);
27 assert(input_a != NULL);
28 assert(input_b != NULL);
29 assert(output != NULL);
30
31 const v128_t vbias = wasm_v128_load32_splat(&params->scalar.bias);
32 const v128_t va_multiplier = wasm_v128_load32_splat(&params->scalar.a_multiplier);
33 const v128_t vb_multiplier = wasm_v128_load32_splat(&params->scalar.b_multiplier);
34 const uint32_t vshift = params->scalar.shift;
35 const v128_t voutput_zero_point = wasm_v128_load16_splat(&params->scalar.output_zero_point);
36 const v128_t voutput_min = wasm_v128_load8_splat(&params->scalar.output_min);
37 const v128_t voutput_max = wasm_v128_load8_splat(&params->scalar.output_max);
38
39 XNN_FORCE_REALIZATION(vbias);
40 XNN_FORCE_REALIZATION(va_multiplier);
41 XNN_FORCE_REALIZATION(vb_multiplier);
42 XNN_FORCE_REALIZATION(voutput_zero_point);
43 XNN_FORCE_REALIZATION(voutput_min);
44 XNN_FORCE_REALIZATION(voutput_max);
45
46 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
47 const v128_t va01234567 = wasm_u16x8_load8x8(input_a);
48 const v128_t vb01234567 = wasm_u16x8_load8x8(input_b);
49 const v128_t va89ABCDEF = wasm_u16x8_load8x8(input_a + 8);
50 const v128_t vb89ABCDEF = wasm_u16x8_load8x8(input_b + 8);
51 input_a += 16;
52 input_b += 16;
53
54 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier));
55 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier));
56 v128_t vacc89AB = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va89ABCDEF), va_multiplier));
57 v128_t vaccCDEF = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va89ABCDEF), va_multiplier));
58
59 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vb01234567), vb_multiplier));
60 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vb01234567), vb_multiplier));
61 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vb89ABCDEF), vb_multiplier));
62 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vb89ABCDEF), vb_multiplier));
63
64 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
65 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
66 vacc89AB = wasm_i32x4_shr(vacc89AB, vshift);
67 vaccCDEF = wasm_i32x4_shr(vaccCDEF, vshift);
68
69 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
70 v128_t vout89ABCDEF = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF), voutput_zero_point);
71
72 v128_t vout0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
73
74 vout0123456789ABCDEF = wasm_u8x16_max(vout0123456789ABCDEF, voutput_min);
75
76 vout0123456789ABCDEF = wasm_u8x16_min(vout0123456789ABCDEF, voutput_max);
77
78 wasm_v128_store(output, vout0123456789ABCDEF);
79 output += 16;
80 }
81 if XNN_UNLIKELY(batch != 0) {
82 do {
83 const v128_t va01234567 = wasm_u16x8_load8x8(input_a);
84 const v128_t vb01234567 = wasm_u16x8_load8x8(input_b);
85 input_a += 8;
86 input_b += 8;
87
88 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier));
89 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier));
90
91 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vb01234567), vb_multiplier));
92 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vb01234567), vb_multiplier));
93
94 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
95 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
96
97 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
98
99 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
100 vout0123456701234567 = wasm_u8x16_max(vout0123456701234567, voutput_min);
101 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
102
103 if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) {
104 wasm_v128_store64_lane(output, vout0123456701234567, 0);
105 output += 8;
106 batch -= 8 * sizeof(uint8_t);
107 } else {
108 if (batch & (4 * sizeof(uint8_t))) {
109 wasm_v128_store32_lane(output, vout0123456701234567, 0);
110 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
111 output += 4;
112 }
113 if (batch & (2 * sizeof(uint8_t))) {
114 wasm_v128_store16_lane(output, vout0123456701234567, 0);
115 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
116 output += 2;
117 }
118 if (batch & (1 * sizeof(uint8_t))) {
119 wasm_v128_store8_lane(output, vout0123456701234567, 0);
120 }
121 batch = 0;
122 }
123 } while (batch != 0);
124 }
125 }