xref: /aosp_15_r20/external/XNNPACK/src/qu8-igemm/gen/2x2-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_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const uint8_t**restrict a,
24     const void*restrict w,
25     uint8_t*restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const uint8_t* zero,
30     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32   assert(mr != 0);
33   assert(mr <= 2);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (2 * sizeof(void*)) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   uint8_t* c0 = c;
43   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr != 2) {
45     c1 = c0;
46   }
47 
48   const int32_t vb_zero_point = params->fp32_scalar_lrintf.kernel_zero_point;
49   do {
50     int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
51     int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
52     int32_t vacc1x0 = vacc0x0;
53     int32_t vacc1x1 = vacc0x1;
54     w = (const void*) ((const int32_t*) w + 2);
55 
56     size_t p = ks;
57     do {
58       const uint8_t* restrict a0 = a[0];
59       assert(a0 != NULL);
60       if XNN_UNPREDICTABLE(a0 != zero) {
61         a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
62       }
63       const uint8_t* restrict a1 = a[1];
64       assert(a1 != NULL);
65       if XNN_UNPREDICTABLE(a1 != zero) {
66         a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
67       }
68       a += 2;
69 
70       size_t k = kc;
71       do {
72         const int32_t va0 = (int32_t) (uint32_t) *a0++;
73         const int32_t va1 = (int32_t) (uint32_t) *a1++;
74 
75         const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
76         const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
77         w = (const void*) ((const uint8_t*) w + 2);
78 
79         vacc0x0 += va0 * vb0;
80         vacc0x1 += va0 * vb1;
81         vacc1x0 += va1 * vb0;
82         vacc1x1 += va1 * vb1;
83 
84         k -= sizeof(uint8_t);
85       } while (k != 0);
86       p -= 2 * sizeof(void*);
87     } while (p != 0);
88 
89     float vfpacc0x0 = (float) vacc0x0;
90     float vfpacc0x1 = (float) vacc0x1;
91     float vfpacc1x0 = (float) vacc1x0;
92     float vfpacc1x1 = (float) vacc1x1;
93 
94     const float vscale = params->fp32_scalar_lrintf.scale;
95     vfpacc0x0 *= vscale;
96     vfpacc0x1 *= vscale;
97     vfpacc1x0 *= vscale;
98     vfpacc1x1 *= vscale;
99 
100     const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
101     vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
102     vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
103     vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
104     vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
105 
106     const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
107     vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
108     vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
109     vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
110     vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
111 
112     const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
113     const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
114     const int32_t vrndacc1x0 = (int32_t) lrintf(vfpacc1x0);
115     const int32_t vrndacc1x1 = (int32_t) lrintf(vfpacc1x1);
116 
117     const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
118     int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
119     int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
120     int32_t vout1x0 = vrndacc1x0 + voutput_zero_point;
121     int32_t vout1x1 = vrndacc1x1 + voutput_zero_point;
122 
123     if XNN_LIKELY(nc >= 2) {
124       c1[0] = (uint8_t) vout1x0;
125       c1[1] = (uint8_t) vout1x1;
126       c0[0] = (uint8_t) vout0x0;
127       c0[1] = (uint8_t) vout0x1;
128 
129       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
130       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
131 
132       a = (const uint8_t**restrict) ((uintptr_t) a - ks);
133       nc -= 2;
134     } else {
135       if (nc & 1) {
136         c1[0] = (uint8_t) vout1x0;
137         c0[0] = (uint8_t) vout0x0;
138       }
139 
140       nc = 0;
141     }
142   } while (nc != 0);
143 }
144