Debian Codegraph @

src/qu8-f32-vcvt/gen/qu8-f32-vcvt-wasmsimd-u8.c

raw
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
19 void xnn_qu8_f32_vcvt_ukernel__wasmsimd_u8(
20 size_t batch,
21 const uint8_t* input,
22 float* output,
23 const struct xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(uint8_t) == 0);
27 assert(input != NULL);
28 assert(output != NULL);
29
30 const v128_t vzero_point = wasm_v128_load16_splat(&params->scalar.zero_point);
31 const v128_t vscale = wasm_v128_load32_splat(&params->scalar.scale);
32 XNN_FORCE_REALIZATION(vzero_point);
33 XNN_FORCE_REALIZATION(vscale);
34 for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) {
35 v128_t vx = wasm_u16x8_load8x8(input);
36 vx = wasm_i16x8_sub(vx, vzero_point);
37 input += 8;
38
39 v128_t vy_lo = wasm_i32x4_extend_low_i16x8(vx);
40 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx);
41
42 vy_lo = wasm_f32x4_convert_i32x4(vy_lo);
43 vy_hi = wasm_f32x4_convert_i32x4(vy_hi);
44
45 vy_lo = wasm_f32x4_mul(vy_lo, vscale);
46 vy_hi = wasm_f32x4_mul(vy_hi, vscale);
47
48 wasm_v128_store(output, vy_lo);
49 wasm_v128_store(output + 4, vy_hi);
50 output += 8;
51 }
52 if XNN_UNLIKELY(batch != 0) {
53 assert(batch >= 1 * sizeof(uint8_t));
54 assert(batch <= 7 * sizeof(uint8_t));
55
56 v128_t vx = wasm_u16x8_load8x8(input);
57 vx = wasm_i16x8_sub(vx, vzero_point);
58 input += 8;
59
60 v128_t vy = wasm_i32x4_extend_low_i16x8(vx);
61 vy = wasm_f32x4_convert_i32x4(vy);
62 vy = wasm_f32x4_mul(vy, vscale);
63
64 if (batch & (4 * sizeof(uint8_t))) {
65 wasm_v128_store(output, vy); output += 4;
66 vy = wasm_i32x4_extend_high_i16x8(vx);
67 vy = wasm_f32x4_convert_i32x4(vy);
68 vy = wasm_f32x4_mul(vy, vscale);
69 }
70 if (batch & (2 * sizeof(uint8_t))) {
71 wasm_v128_store64_lane(output, vy, 0);
72 vy = wasm_v64x2_shuffle(vy, vy, 1, 1);
73 output += 2;
74 }
75 if (batch & (1 * sizeof(uint8_t))) {
76 wasm_v128_store32_lane(output, vy, 0);
77 }
78 }
79 }