xref: /aosp_15_r20/external/XNNPACK/src/f32-qs8-vcvt/gen/vcvt-scalar-lrintf-x3.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_qs8_vcvt_ukernel__scalar_lrintf_x3(size_t n,const float * x,int8_t * y,const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_qs8_vcvt_ukernel__scalar_lrintf_x3(
19     size_t n,
20     const float* x,
21     int8_t* y,
22     const union xnn_f32_qs8_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 >= 3 * sizeof(float); n -= 3 * sizeof(float)) {
35     float vx0 = x[0];
36     float vx1 = x[1];
37     float vx2 = x[2];
38     x += 3;
39 
40     vx0 *= vscale;
41     vx1 *= vscale;
42     vx2 *= vscale;
43 
44     vx0 = math_max_f32(vx0, voutput_min_less_zero_point);
45     vx1 = math_max_f32(vx1, voutput_min_less_zero_point);
46     vx2 = math_max_f32(vx2, voutput_min_less_zero_point);
47 
48     vx0 = math_min_f32(vx0, voutput_max_less_zero_point);
49     vx1 = math_min_f32(vx1, voutput_max_less_zero_point);
50     vx2 = math_min_f32(vx2, voutput_max_less_zero_point);
51 
52     int32_t vy0 = (int32_t) lrintf(vx0);
53     int32_t vy1 = (int32_t) lrintf(vx1);
54     int32_t vy2 = (int32_t) lrintf(vx2);
55 
56     vy0 += voutput_zero_point;
57     vy1 += voutput_zero_point;
58     vy2 += voutput_zero_point;
59 
60     y[0] = (int8_t) vy0;
61     y[1] = (int8_t) vy1;
62     y[2] = (int8_t) vy2;
63     y += 3;
64   }
65   if XNN_UNLIKELY(n != 0) {
66     do {
67       float vx = *x++;
68       vx *= vscale;
69       vx = math_max_f32(vx, voutput_min_less_zero_point);
70       vx = math_min_f32(vx, voutput_max_less_zero_point);
71 
72       int32_t vy = (int32_t) lrintf(vx);
73       vy += voutput_zero_point;
74 
75       *y++ = (int8_t) vy;
76 
77       n -= sizeof(float);
78     } while (n != 0);
79   }
80 }
81