xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/1x2-minmax-fp32-scalar-lrintf.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/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 #include <math.h>
12 
13 #include <xnnpack/math.h>
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/unaligned.h>
16 
17 
xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const int8_t**restrict a,
24     const void*restrict w,
25     int8_t*restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const int8_t* zero,
30     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32   assert(mr != 0);
33   assert(mr <= 1);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (1 * sizeof(void*)) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   int8_t* c0 = c;
43 
44   do {
45     int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
46     int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
47     w = (const void*) ((const int32_t*) w + 2);
48 
49     size_t p = ks;
50     do {
51       const int8_t* restrict a0 = a[0];
52       assert(a0 != NULL);
53       if XNN_UNPREDICTABLE(a0 != zero) {
54         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
55       }
56       a += 1;
57 
58       size_t k = kc;
59       do {
60         const int32_t va0 = (int32_t) *a0++;
61 
62         const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
63         const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
64         w = (const void*) ((const int8_t*) w + 2);
65 
66         vacc0x0 += va0 * vb0;
67         vacc0x1 += va0 * vb1;
68 
69         k -= sizeof(int8_t);
70       } while (k != 0);
71       p -= 1 * sizeof(void*);
72     } while (p != 0);
73 
74     float vfpacc0x0 = (float) vacc0x0;
75     float vfpacc0x1 = (float) vacc0x1;
76 
77     const float vscale0 = unaligned_indexed_load_f32(w, 0);
78     vfpacc0x0 *= vscale0;
79     const float vscale1 = unaligned_indexed_load_f32(w, 1);
80     vfpacc0x1 *= vscale1;
81     w = (const void*) ((const float*) w + 2);
82 
83     const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
84     vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
85     vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
86 
87     const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
88     vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
89     vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
90 
91     const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
92     const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
93 
94     const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
95     int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
96     int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
97 
98     if XNN_LIKELY(nc >= 2) {
99       c0[0] = (int8_t) vout0x0;
100       c0[1] = (int8_t) vout0x1;
101 
102       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
103 
104       a = (const int8_t**restrict) ((uintptr_t) a - ks);
105       nc -= 2;
106     } else {
107       if (nc & 1) {
108         c0[0] = (int8_t) vout0x0;
109       }
110 
111       nc = 0;
112     }
113   } while (nc != 0);
114 }
115