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_x24(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_x24(
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 >= 24 * sizeof(uint8_t); n -= 24 * sizeof(uint8_t)) {
33 v128_t vx01234567 = wasm_u16x8_load8x8(x);
34 v128_t vx89ABCDEF = wasm_u16x8_load8x8(x + 8);
35 v128_t vxGHIJKLMN = wasm_u16x8_load8x8(x + 16);
36 x += 24;
37
38 vx01234567 = wasm_i16x8_add(vx01234567, vminus_zero_point);
39 vx89ABCDEF = wasm_i16x8_add(vx89ABCDEF, vminus_zero_point);
40 vxGHIJKLMN = wasm_i16x8_add(vxGHIJKLMN, vminus_zero_point);
41
42 v128_t vy0123 = wasm_i32x4_extend_low_i16x8(vx01234567);
43 v128_t vy4567 = wasm_i32x4_extend_high_i16x8(vx01234567);
44 v128_t vy89AB = wasm_i32x4_extend_low_i16x8(vx89ABCDEF);
45 v128_t vyCDEF = wasm_i32x4_extend_high_i16x8(vx89ABCDEF);
46 v128_t vyGHIJ = wasm_i32x4_extend_low_i16x8(vxGHIJKLMN);
47 v128_t vyKLMN = wasm_i32x4_extend_high_i16x8(vxGHIJKLMN);
48
49 vy0123 = wasm_f32x4_convert_i32x4(vy0123);
50 vy4567 = wasm_f32x4_convert_i32x4(vy4567);
51 vy89AB = wasm_f32x4_convert_i32x4(vy89AB);
52 vyCDEF = wasm_f32x4_convert_i32x4(vyCDEF);
53 vyGHIJ = wasm_f32x4_convert_i32x4(vyGHIJ);
54 vyKLMN = wasm_f32x4_convert_i32x4(vyKLMN);
55
56 vy0123 = wasm_f32x4_mul(vy0123, vscale);
57 vy4567 = wasm_f32x4_mul(vy4567, vscale);
58 vy89AB = wasm_f32x4_mul(vy89AB, vscale);
59 vyCDEF = wasm_f32x4_mul(vyCDEF, vscale);
60 vyGHIJ = wasm_f32x4_mul(vyGHIJ, vscale);
61 vyKLMN = wasm_f32x4_mul(vyKLMN, vscale);
62
63 wasm_v128_store(y, vy0123);
64 wasm_v128_store(y + 4, vy4567);
65 wasm_v128_store(y + 8, vy89AB);
66 wasm_v128_store(y + 12, vyCDEF);
67 wasm_v128_store(y + 16, vyGHIJ);
68 wasm_v128_store(y + 20, vyKLMN);
69 y += 24;
70 }
71 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
72 v128_t vx = wasm_u16x8_load8x8(x);
73 vx = wasm_i16x8_add(vx, vminus_zero_point);
74 x += 8;
75
76 v128_t vy_lo = wasm_i32x4_extend_low_i16x8(vx);
77 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx);
78
79 vy_lo = wasm_f32x4_convert_i32x4(vy_lo);
80 vy_hi = wasm_f32x4_convert_i32x4(vy_hi);
81
82 vy_lo = wasm_f32x4_mul(vy_lo, vscale);
83 vy_hi = wasm_f32x4_mul(vy_hi, vscale);
84
85 wasm_v128_store(y, vy_lo);
86 wasm_v128_store(y + 4, vy_hi);
87 y += 8;
88 }
89 if XNN_UNLIKELY(n != 0) {
90 assert(n >= 1 * sizeof(uint8_t));
91 assert(n <= 7 * sizeof(uint8_t));
92
93 v128_t vx = wasm_u16x8_load8x8(x);
94 vx = wasm_i16x8_add(vx, vminus_zero_point);
95 x += 8;
96
97 v128_t vy = wasm_i32x4_extend_low_i16x8(vx);
98 vy = wasm_f32x4_convert_i32x4(vy);
99 vy = wasm_f32x4_mul(vy, vscale);
100
101 if (n & (4 * sizeof(uint8_t))) {
102 wasm_v128_store(y, vy); y += 4;
103 vy = wasm_i32x4_extend_high_i16x8(vx);
104 vy = wasm_f32x4_convert_i32x4(vy);
105 vy = wasm_f32x4_mul(vy, vscale);
106 }
107 if (n & (2 * sizeof(uint8_t))) {
108 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
109 vy = wasm_v64x2_shuffle(vy, vy, 1, 1);
110 y += 2;
111 }
112 if (n & (1 * sizeof(uint8_t))) {
113 *y = wasm_f32x4_extract_lane(vy, 0);
114 }
115 }
116 }
117