1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-f32-vcvt/wasmsimd.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/intrinsics-polyfill.h>
16 #include <xnnpack/vcvt.h>
17
18
xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16(size_t n,const uint8_t * x,float * y,const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16(
20 size_t n,
21 const uint8_t* x,
22 float* y,
23 const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(uint8_t) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const v128_t vminus_zero_point = wasm_v128_load64_splat(params->wasmsimd.minus_zero_point);
31 const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd.scale);
32 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
33 v128_t vx01234567 = wasm_u16x8_load8x8(x);
34 v128_t vx89ABCDEF = wasm_u16x8_load8x8(x + 8);
35 x += 16;
36
37 vx01234567 = wasm_i16x8_add(vx01234567, vminus_zero_point);
38 vx89ABCDEF = wasm_i16x8_add(vx89ABCDEF, vminus_zero_point);
39
40 v128_t vy0123 = wasm_i32x4_extend_low_i16x8(vx01234567);
41 v128_t vy4567 = wasm_i32x4_extend_high_i16x8(vx01234567);
42 v128_t vy89AB = wasm_i32x4_extend_low_i16x8(vx89ABCDEF);
43 v128_t vyCDEF = wasm_i32x4_extend_high_i16x8(vx89ABCDEF);
44
45 vy0123 = wasm_f32x4_convert_i32x4(vy0123);
46 vy4567 = wasm_f32x4_convert_i32x4(vy4567);
47 vy89AB = wasm_f32x4_convert_i32x4(vy89AB);
48 vyCDEF = wasm_f32x4_convert_i32x4(vyCDEF);
49
50 vy0123 = wasm_f32x4_mul(vy0123, vscale);
51 vy4567 = wasm_f32x4_mul(vy4567, vscale);
52 vy89AB = wasm_f32x4_mul(vy89AB, vscale);
53 vyCDEF = wasm_f32x4_mul(vyCDEF, vscale);
54
55 wasm_v128_store(y, vy0123);
56 wasm_v128_store(y + 4, vy4567);
57 wasm_v128_store(y + 8, vy89AB);
58 wasm_v128_store(y + 12, vyCDEF);
59 y += 16;
60 }
61 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
62 v128_t vx = wasm_u16x8_load8x8(x);
63 vx = wasm_i16x8_add(vx, vminus_zero_point);
64 x += 8;
65
66 v128_t vy_lo = wasm_i32x4_extend_low_i16x8(vx);
67 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx);
68
69 vy_lo = wasm_f32x4_convert_i32x4(vy_lo);
70 vy_hi = wasm_f32x4_convert_i32x4(vy_hi);
71
72 vy_lo = wasm_f32x4_mul(vy_lo, vscale);
73 vy_hi = wasm_f32x4_mul(vy_hi, vscale);
74
75 wasm_v128_store(y, vy_lo);
76 wasm_v128_store(y + 4, vy_hi);
77 y += 8;
78 }
79 if XNN_UNLIKELY(n != 0) {
80 assert(n >= 1 * sizeof(uint8_t));
81 assert(n <= 7 * sizeof(uint8_t));
82
83 v128_t vx = wasm_u16x8_load8x8(x);
84 vx = wasm_i16x8_add(vx, vminus_zero_point);
85 x += 8;
86
87 v128_t vy = wasm_i32x4_extend_low_i16x8(vx);
88 vy = wasm_f32x4_convert_i32x4(vy);
89 vy = wasm_f32x4_mul(vy, vscale);
90
91 if (n & (4 * sizeof(uint8_t))) {
92 wasm_v128_store(y, vy); y += 4;
93 vy = wasm_i32x4_extend_high_i16x8(vx);
94 vy = wasm_f32x4_convert_i32x4(vy);
95 vy = wasm_f32x4_mul(vy, vscale);
96 }
97 if (n & (2 * sizeof(uint8_t))) {
98 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
99 vy = wasm_v64x2_shuffle(vy, vy, 1, 1);
100 y += 2;
101 }
102 if (n & (1 * sizeof(uint8_t))) {
103 *y = wasm_f32x4_extract_lane(vy, 0);
104 }
105 }
106 }
107