1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vcvt/wasmsimd.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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/vcvt.h>
16
17
xnn_qu8_vcvt_ukernel__wasmrelaxedsimd_x32(size_t n,const uint8_t * x,uint8_t * y,const union xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_vcvt_ukernel__wasmrelaxedsimd_x32(
19 size_t n,
20 const uint8_t* x,
21 uint8_t* y,
22 const union xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n != 0);
25 assert(n % sizeof(uint8_t) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const v128_t vinput_zero_point = wasm_v128_load64_splat(params->wasmsimd.input_zero_point);
30 const v128_t vmultiplier = wasm_v128_load64_splat(params->wasmsimd.multiplier);
31 const v128_t voutput_zero_point = wasm_v128_load64_splat(params->wasmsimd.output_zero_point);
32 for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
33 v128_t vacc0 = wasm_u16x8_load8x8(x);
34 v128_t vacc1 = wasm_u16x8_load8x8(x + 8);
35 v128_t vacc2 = wasm_u16x8_load8x8(x + 16);
36 v128_t vacc3 = wasm_u16x8_load8x8(x + 24);
37 x += 32;
38
39 vacc0 = wasm_i16x8_sub(vinput_zero_point, vacc0);
40 vacc1 = wasm_i16x8_sub(vinput_zero_point, vacc1);
41 vacc2 = wasm_i16x8_sub(vinput_zero_point, vacc2);
42 vacc3 = wasm_i16x8_sub(vinput_zero_point, vacc3);
43
44 vacc0 = wasm_i16x8_shl(vacc0, 7);
45 vacc1 = wasm_i16x8_shl(vacc1, 7);
46 vacc2 = wasm_i16x8_shl(vacc2, 7);
47 vacc3 = wasm_i16x8_shl(vacc3, 7);
48
49 vacc0 = __builtin_wasm_relaxed_q15mulr_s_i16x8(vacc0, vmultiplier);
50 vacc1 = __builtin_wasm_relaxed_q15mulr_s_i16x8(vacc1, vmultiplier);
51 vacc2 = __builtin_wasm_relaxed_q15mulr_s_i16x8(vacc2, vmultiplier);
52 vacc3 = __builtin_wasm_relaxed_q15mulr_s_i16x8(vacc3, vmultiplier);
53
54 vacc0 = wasm_i16x8_add_sat(vacc0, voutput_zero_point);
55 vacc1 = wasm_i16x8_add_sat(vacc1, voutput_zero_point);
56 vacc2 = wasm_i16x8_add_sat(vacc2, voutput_zero_point);
57 vacc3 = wasm_i16x8_add_sat(vacc3, voutput_zero_point);
58
59 const v128_t vy0 = wasm_u8x16_narrow_i16x8(vacc0, vacc1);
60 const v128_t vy1 = wasm_u8x16_narrow_i16x8(vacc2, vacc3);
61
62 wasm_v128_store(y, vy0);
63 wasm_v128_store((y + 16), vy1);
64 y += 32;
65 }
66 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
67 v128_t vacc = wasm_u16x8_load8x8(x);
68 vacc = wasm_i16x8_sub(vinput_zero_point, vacc);
69 vacc = wasm_i16x8_shl(vacc, 7);
70 vacc = __builtin_wasm_relaxed_q15mulr_s_i16x8(vacc, vmultiplier);
71 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
72 x += 8;
73
74 const v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
75 wasm_v128_store64_lane(y, vy, 0);
76 y += 8;
77 }
78 if XNN_UNLIKELY(n != 0) {
79 assert(n >= 1 * sizeof(uint8_t));
80 assert(n <= 7 * sizeof(uint8_t));
81
82 v128_t vacc = wasm_u16x8_load8x8(x);
83 vacc = wasm_i16x8_sub(vinput_zero_point, vacc);
84 vacc = wasm_i16x8_shl(vacc, 7);
85 vacc = __builtin_wasm_relaxed_q15mulr_s_i16x8(vacc, vmultiplier);
86 vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
87
88 v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
89 if (n & (4 * sizeof(uint8_t))) {
90 wasm_v128_store32_lane(y, vy, 0);
91 vy = wasm_u64x2_shr(vy, 32);
92 y += 4;
93 }
94 if (n & (2 * sizeof(uint8_t))) {
95 wasm_v128_store16_lane(y, vy, 0);
96 vy = wasm_u32x4_shr(vy, 16);
97 y += 2;
98 }
99 if (n & (1 * sizeof(uint8_t))) {
100 wasm_v128_store8_lane(y, vy, 0);
101 }
102 }
103 }
104