Debian Codegraph @

src/qu8-vaddc/gen/qu8-vaddc-minmax-wasmsimd-u32.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_u32(
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 >= 32 * sizeof(uint8_t); batch -= 32 * 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 const v128_t vaGHIJKLMN = wasm_u16x8_load8x8(input_a + 16);
48 const v128_t vaOPQRSTUV = wasm_u16x8_load8x8(input_a + 24);
49 input_a += 32;
50
51 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier));
52 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier));
53 v128_t vacc89AB = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va89ABCDEF), va_multiplier));
54 v128_t vaccCDEF = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va89ABCDEF), va_multiplier));
55 v128_t vaccGHIJ = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vaGHIJKLMN), va_multiplier));
56 v128_t vaccKLMN = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vaGHIJKLMN), va_multiplier));
57 v128_t vaccOPQR = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vaOPQRSTUV), va_multiplier));
58 v128_t vaccSTUV = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vaOPQRSTUV), va_multiplier));
59
60 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
61 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
62 vacc89AB = wasm_i32x4_shr(vacc89AB, vshift);
63 vaccCDEF = wasm_i32x4_shr(vaccCDEF, vshift);
64 vaccGHIJ = wasm_i32x4_shr(vaccGHIJ, vshift);
65 vaccKLMN = wasm_i32x4_shr(vaccKLMN, vshift);
66 vaccOPQR = wasm_i32x4_shr(vaccOPQR, vshift);
67 vaccSTUV = wasm_i32x4_shr(vaccSTUV, 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 v128_t voutGHIJKLMN = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN), voutput_zero_point);
72 v128_t voutOPQRSTUV = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vaccOPQR, vaccSTUV), voutput_zero_point);
73
74 v128_t vout0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
75 v128_t voutGHIJKLMNOPQRSTUV = wasm_u8x16_narrow_i16x8(voutGHIJKLMN, voutOPQRSTUV);
76
77 vout0123456789ABCDEF = wasm_u8x16_max(vout0123456789ABCDEF, voutput_min);
78 voutGHIJKLMNOPQRSTUV = wasm_u8x16_max(voutGHIJKLMNOPQRSTUV, voutput_min);
79
80 vout0123456789ABCDEF = wasm_u8x16_min(vout0123456789ABCDEF, voutput_max);
81 voutGHIJKLMNOPQRSTUV = wasm_u8x16_min(voutGHIJKLMNOPQRSTUV, voutput_max);
82
83 wasm_v128_store(output, vout0123456789ABCDEF);
84 wasm_v128_store(output + 16, voutGHIJKLMNOPQRSTUV);
85 output += 32;
86 }
87 if XNN_UNLIKELY(batch != 0) {
88 do {
89 const v128_t va01234567 = wasm_u16x8_load8x8(input_a);
90 input_a += 8;
91
92 v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier));
93 v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier));
94
95 vacc0123 = wasm_i32x4_shr(vacc0123, vshift);
96 vacc4567 = wasm_i32x4_shr(vacc4567, vshift);
97
98 v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point);
99
100 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
101 vout0123456701234567 = wasm_u8x16_max(vout0123456701234567, voutput_min);
102 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
103
104 if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) {
105 wasm_v128_store64_lane(output, vout0123456701234567, 0);
106 output += 8;
107 batch -= 8 * sizeof(uint8_t);
108 } else {
109 if (batch & (4 * sizeof(uint8_t))) {
110 wasm_v128_store32_lane(output, vout0123456701234567, 0);
111 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
112 output += 4;
113 }
114 if (batch & (2 * sizeof(uint8_t))) {
115 wasm_v128_store16_lane(output, vout0123456701234567, 0);
116 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
117 output += 2;
118 }
119 if (batch & (1 * sizeof(uint8_t))) {
120 wasm_v128_store8_lane(output, vout0123456701234567, 0);
121 }
122 batch = 0;
123 }
124 } while (batch != 0);
125 }
126 }