xref: /aosp_15_r20/external/XNNPACK/src/qu8-vlrelu/gen/vlrelu-wasmrelaxedsimd-x86-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-vlrelu/wasmsimd-x86.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_vlrelu_ukernel__wasmrelaxedsimd_x86_x16(size_t n,const uint8_t * x,uint8_t * y,const union xnn_qu8_lrelu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_vlrelu_ukernel__wasmrelaxedsimd_x86_x16(
19     size_t n,
20     const uint8_t* x,
21     uint8_t* y,
22     const union xnn_qu8_lrelu_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_x86.input_zero_point);
30   const v128_t vmultiplier_diff = wasm_v128_load64_splat(params->wasmsimd_x86.multiplier_diff);
31   const v128_t vmultiplier_base = wasm_v128_load64_splat(params->wasmsimd_x86.multiplier_base);
32   const v128_t voutput_zero_point = wasm_v128_load64_splat(params->wasmsimd_x86.output_zero_point);
33   for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
34     v128_t vacc0 = wasm_u16x8_load8x8(x);
35     v128_t vacc1 = wasm_u16x8_load8x8(x + 8);
36     x += 16;
37 
38     v128_t vmultiplier0 = wasm_i16x8_gt(vacc0, vinput_zero_point);
39     vacc0 = wasm_i16x8_sub(vinput_zero_point, vacc0);
40     v128_t vmultiplier1 = wasm_i16x8_gt(vacc1, vinput_zero_point);
41     vacc1 = wasm_i16x8_sub(vinput_zero_point, vacc1);
42 
43     vmultiplier0 = wasm_v128_and(vmultiplier0, vmultiplier_diff);
44     vacc0 = wasm_i16x8_shl(vacc0, 7);
45     vmultiplier0 = wasm_v128_xor(vmultiplier0, vmultiplier_base);
46     vmultiplier1 = wasm_v128_and(vmultiplier1, vmultiplier_diff);
47     vacc1 = wasm_i16x8_shl(vacc1, 7);
48     vmultiplier1 = wasm_v128_xor(vmultiplier1, vmultiplier_base);
49 
50     vacc0 = __builtin_wasm_relaxed_q15mulr_s_i16x8(vacc0, vmultiplier0);
51     vacc1 = __builtin_wasm_relaxed_q15mulr_s_i16x8(vacc1, vmultiplier1);
52 
53     vacc0 = wasm_i16x8_add_sat(vacc0, voutput_zero_point);
54     vacc1 = wasm_i16x8_add_sat(vacc1, voutput_zero_point);
55 
56     const v128_t vy0 = wasm_u8x16_narrow_i16x8(vacc0, vacc1);
57 
58     wasm_v128_store(y, vy0);
59     y += 16;
60   }
61   for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
62     v128_t vacc = wasm_u16x8_load8x8(x);
63     v128_t vmultiplier = wasm_i16x8_gt(vacc, vinput_zero_point);
64     vacc = wasm_i16x8_sub(vinput_zero_point, vacc);
65     vmultiplier = wasm_v128_and(vmultiplier, vmultiplier_diff);
66     vacc = wasm_i16x8_shl(vacc, 7);
67     vmultiplier = wasm_v128_xor(vmultiplier, vmultiplier_base);
68     vacc = __builtin_wasm_relaxed_q15mulr_s_i16x8(vacc, vmultiplier);
69     vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
70     x += 8;
71 
72     const v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
73     wasm_v128_store64_lane(y, vy, 0);
74     y += 8;
75   }
76   if XNN_UNLIKELY(n != 0) {
77     assert(n >= 1 * sizeof(uint8_t));
78     assert(n <= 7 * sizeof(uint8_t));
79 
80     v128_t vacc = wasm_u16x8_load8x8(x);
81     v128_t vmultiplier = wasm_i16x8_gt(vacc, vinput_zero_point);
82     vacc = wasm_i16x8_sub(vinput_zero_point, vacc);
83     vmultiplier = wasm_v128_and(vmultiplier, vmultiplier_diff);
84     vacc = wasm_i16x8_shl(vacc, 7);
85     vmultiplier = wasm_v128_xor(vmultiplier, vmultiplier_base);
86     vacc = __builtin_wasm_relaxed_q15mulr_s_i16x8(vacc, vmultiplier);
87     vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
88 
89     v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc);
90     if (n & (4 * sizeof(uint8_t))) {
91       wasm_v128_store32_lane(y, vy, 0);
92       vy = wasm_u64x2_shr(vy, 32);
93       y += 4;
94     }
95     if (n & (2 * sizeof(uint8_t))) {
96       wasm_v128_store16_lane(y, vy, 0);
97       vy = wasm_u32x4_shr(vy, 16);
98       y += 2;
99     }
100     if (n & (1 * sizeof(uint8_t))) {
101       wasm_v128_store8_lane(y, vy, 0);
102     }
103   }
104 }
105