Debian Codegraph @

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

raw
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/common.h"
15 #include "xnnpack/vbinary.h"
16
17
18 void xnn_qu8_vaddc_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_i32x4_splat((int32_t) *input_b * params->scalar.b_multiplier + params->scalar.bias);
32 const v128_t va_multiplier = wasm_v128_load32_splat(&params->scalar.a_multiplier);
33 const uint32_t vshift = params->scalar.shift;
34 const v128_t voutput_zero_point = wasm_v128_load16_splat(&params->scalar.output_zero_point);
35 const v128_t voutput_min = wasm_v128_load8_splat(&params->scalar.output_min);
36 const v128_t voutput_max = wasm_v128_load8_splat(&params->scalar.output_max);
37
38 XNN_FORCE_REALIZATION(vbias);
39 XNN_FORCE_REALIZATION(va_multiplier);
40 XNN_FORCE_REALIZATION(voutput_zero_point);
41 XNN_FORCE_REALIZATION(voutput_min);
42 XNN_FORCE_REALIZATION(voutput_max);
43
44 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
45 const v128_t va01234567 = wasm_u16x8_load8x8(input_a);
46 const v128_t va89ABCDEF = wasm_u16x8_load8x8(input_a + 8);
47 input_a += 16;
48
49 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier));
50 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier));
51 v128_t vacc89AB = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va89ABCDEF), va_multiplier));
52 v128_t vaccCDEF = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va89ABCDEF), va_multiplier));
53
54 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
55 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
56 vacc89AB = wasm_i32x4_shr(vacc89AB, vshift);
57 vaccCDEF = wasm_i32x4_shr(vaccCDEF, vshift);
58
59 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
60 v128_t vout89ABCDEF = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF), voutput_zero_point);
61
62 v128_t vout0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
63
64 vout0123456789ABCDEF = wasm_u8x16_max(vout0123456789ABCDEF, voutput_min);
65
66 vout0123456789ABCDEF = wasm_u8x16_min(vout0123456789ABCDEF, voutput_max);
67
68 wasm_v128_store(output, vout0123456789ABCDEF);
69 output += 16;
70 }
71 if XNN_UNLIKELY(batch != 0) {
72 do {
73 const v128_t va01234567 = wasm_u16x8_load8x8(input_a);
74 input_a += 8;
75
76 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier));
77 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier));
78
79 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
80 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
81
82 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
83
84 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
85 vout0123456701234567 = wasm_u8x16_max(vout0123456701234567, voutput_min);
86 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
87
88 if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) {
89 wasm_v128_store64_lane(output, vout0123456701234567, 0);
90 output += 8;
91 batch -= 8 * sizeof(uint8_t);
92 } else {
93 if (batch & (4 * sizeof(uint8_t))) {
94 wasm_v128_store32_lane(output, vout0123456701234567, 0);
95 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
96 output += 4;
97 }
98 if (batch & (2 * sizeof(uint8_t))) {
99 wasm_v128_store16_lane(output, vout0123456701234567, 0);
100 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
101 output += 2;
102 }
103 if (batch & (1 * sizeof(uint8_t))) {
104 wasm_v128_store8_lane(output, vout0123456701234567, 0);
105 }
106 batch = 0;
107 }
108 } while (batch != 0);
109 }
110 }