Debian Codegraph @

src/qu8-f32-vcvt/gen/qu8-f32-vcvt-scalar-u4.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-f32-vcvt/scalar.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 "xnnpack/common.h"
13 #include "xnnpack/math.h"
14 #include "xnnpack/vcvt.h"
15
16
17 void xnn_qu8_f32_vcvt_ukernel__scalar_u4(
18 size_t batch,
19 const uint8_t* input,
20 float* output,
21 const struct xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
22 {
23 assert(batch != 0);
24 assert(batch % sizeof(uint8_t) == 0);
25 assert(input != NULL);
26 assert(output != NULL);
27
28 const int32_t vzero_point = params->scalar.zero_point;
29 const float vscale = params->scalar.scale;
30
31 for (; batch >= 4 * sizeof(uint8_t); batch -= 4 * sizeof(uint8_t)) {
32 int32_t vx0 = (int32_t) input[0];
33 int32_t vx1 = (int32_t) input[1];
34 int32_t vx2 = (int32_t) input[2];
35 int32_t vx3 = (int32_t) input[3];
36 input += 4;
37
38 vx0 -= vzero_point;
39 vx1 -= vzero_point;
40 vx2 -= vzero_point;
41 vx3 -= vzero_point;
42
43 float vy0 = (float) vx0;
44 float vy1 = (float) vx1;
45 float vy2 = (float) vx2;
46 float vy3 = (float) vx3;
47
48 vy0 *= vscale;
49 vy1 *= vscale;
50 vy2 *= vscale;
51 vy3 *= vscale;
52
53 output[0] = vy0;
54 output[1] = vy1;
55 output[2] = vy2;
56 output[3] = vy3;
57 output += 4;
58 }
59 if XNN_UNLIKELY(batch != 0) {
60 do {
61 int32_t vx = *input++;
62 vx -= vzero_point;
63
64 float vy = (float) vx;
65 vy *= vscale;
66 *output++ = vy;
67
68 batch -= sizeof(uint8_t);
69 } while (batch != 0);
70 }
71 }