xref: /aosp_15_r20/external/XNNPACK/src/f32-qu8-vcvt/gen/vcvt-scalar-lrintf-x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-qs8-vcvt/scalar-lrintf.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 #include <math.h>
12 
13 #include <xnnpack/common.h>
14 #include <xnnpack/math.h>
15 #include <xnnpack/vcvt.h>
16 
17 
xnn_f32_qu8_vcvt_ukernel__scalar_lrintf_x4(size_t n,const float * x,uint8_t * y,const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_qu8_vcvt_ukernel__scalar_lrintf_x4(
19     size_t n,
20     const float* x,
21     uint8_t* y,
22     const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24   assert(n != 0);
25   assert(n % sizeof(float) == 0);
26   assert(x != NULL);
27   assert(y != NULL);
28 
29   const float vscale = params->scalar_lrintf.scale;
30   const float voutput_min_less_zero_point = params->scalar_lrintf.output_min_less_zero_point;
31   const float voutput_max_less_zero_point = params->scalar_lrintf.output_max_less_zero_point;
32   const int32_t voutput_zero_point = params->scalar_lrintf.output_zero_point;
33 
34   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
35     float vx0 = x[0];
36     float vx1 = x[1];
37     float vx2 = x[2];
38     float vx3 = x[3];
39     x += 4;
40 
41     vx0 *= vscale;
42     vx1 *= vscale;
43     vx2 *= vscale;
44     vx3 *= vscale;
45 
46     vx0 = math_max_f32(vx0, voutput_min_less_zero_point);
47     vx1 = math_max_f32(vx1, voutput_min_less_zero_point);
48     vx2 = math_max_f32(vx2, voutput_min_less_zero_point);
49     vx3 = math_max_f32(vx3, voutput_min_less_zero_point);
50 
51     vx0 = math_min_f32(vx0, voutput_max_less_zero_point);
52     vx1 = math_min_f32(vx1, voutput_max_less_zero_point);
53     vx2 = math_min_f32(vx2, voutput_max_less_zero_point);
54     vx3 = math_min_f32(vx3, voutput_max_less_zero_point);
55 
56     int32_t vy0 = (int32_t) lrintf(vx0);
57     int32_t vy1 = (int32_t) lrintf(vx1);
58     int32_t vy2 = (int32_t) lrintf(vx2);
59     int32_t vy3 = (int32_t) lrintf(vx3);
60 
61     vy0 += voutput_zero_point;
62     vy1 += voutput_zero_point;
63     vy2 += voutput_zero_point;
64     vy3 += voutput_zero_point;
65 
66     y[0] = (uint8_t) vy0;
67     y[1] = (uint8_t) vy1;
68     y[2] = (uint8_t) vy2;
69     y[3] = (uint8_t) vy3;
70     y += 4;
71   }
72   if XNN_UNLIKELY(n != 0) {
73     do {
74       float vx = *x++;
75       vx *= vscale;
76       vx = math_max_f32(vx, voutput_min_less_zero_point);
77       vx = math_min_f32(vx, voutput_max_less_zero_point);
78 
79       int32_t vy = (int32_t) lrintf(vx);
80       vy += voutput_zero_point;
81 
82       *y++ = (uint8_t) vy;
83 
84       n -= sizeof(float);
85     } while (n != 0);
86   }
87 }
88