1 // Copyright 2021 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <wasm_simd128.h>
9
10 #include <xnnpack/vunary.h>
11
12
xnn_s8_vclamp_ukernel__wasmsimd_x64(size_t n,const int8_t * x,int8_t * y,const union xnn_s8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])13 void xnn_s8_vclamp_ukernel__wasmsimd_x64(
14 size_t n,
15 const int8_t* x,
16 int8_t* y,
17 const union xnn_s8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
18 {
19 assert(n != 0);
20
21 const v128_t voutput_max = wasm_v128_load64_splat(params->wasmsimd.max);
22 const v128_t voutput_min = wasm_v128_load64_splat(params->wasmsimd.min);
23 for (; n >= 64; n -= 64) {
24 v128_t vacc0 = wasm_v128_load(x);
25 v128_t vacc1 = wasm_v128_load(x + 16);
26 v128_t vacc2 = wasm_v128_load(x + 32);
27 v128_t vacc3 = wasm_v128_load(x + 48);
28 x += 64;
29
30 vacc0 = wasm_i8x16_max(vacc0, voutput_min);
31 vacc1 = wasm_i8x16_max(vacc1, voutput_min);
32 vacc2 = wasm_i8x16_max(vacc2, voutput_min);
33 vacc3 = wasm_i8x16_max(vacc3, voutput_min);
34
35 vacc0 = wasm_i8x16_min(vacc0, voutput_max);
36 vacc1 = wasm_i8x16_min(vacc1, voutput_max);
37 vacc2 = wasm_i8x16_min(vacc2, voutput_max);
38 vacc3 = wasm_i8x16_min(vacc3, voutput_max);
39
40 wasm_v128_store(y, vacc0);
41 wasm_v128_store(y + 16, vacc1);
42 wasm_v128_store(y + 32, vacc2);
43 wasm_v128_store(y + 48, vacc3);
44 y += 64;
45 }
46 for (; n >= 16; n -= 16) {
47 v128_t vacc = wasm_v128_load(x);
48 x += 16;
49
50 vacc = wasm_i8x16_min(vacc, voutput_max);
51 vacc = wasm_i8x16_max(vacc, voutput_min);
52
53 wasm_v128_store(y, vacc);
54 y += 16;
55 }
56 if XNN_UNLIKELY(n != 0) {
57 v128_t vacc = wasm_v128_load(x);
58
59 vacc = wasm_i8x16_min(vacc, voutput_max);
60 vacc = wasm_i8x16_max(vacc, voutput_min);
61
62 if (n & 8) {
63 *((double*) y) = wasm_f64x2_extract_lane(vacc, 0);
64 y += 8;
65 vacc = wasm_v64x2_shuffle(vacc, vacc, 1, 1);
66 }
67 if (n & 4) {
68 *((float*) y) = wasm_f32x4_extract_lane(vacc, 0);
69 y += 4;
70 vacc = wasm_u64x2_shr(vacc, 32);
71 }
72 uint32_t vacc_lo = wasm_i32x4_extract_lane(vacc, 0);
73 if (n & 2) {
74 *((uint16_t*) y) = (uint16_t) vacc_lo;
75 vacc_lo >>= 16;
76 y += 2;
77 }
78 if (n & 1) {
79 *y = (uint8_t) vacc_lo;
80 }
81 }
82 }
83