xref: /aosp_15_r20/external/XNNPACK/src/qu8-igemm/gen/2x4-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 
16 
xnn_qu8_igemm_minmax_fp32_ukernel_2x4__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)])17 void xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const uint8_t**restrict a,
23     const void*restrict w,
24     uint8_t*restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const uint8_t* zero,
29     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31   assert(mr != 0);
32   assert(mr <= 2);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (2 * sizeof(void*)) == 0);
37   assert(a != NULL);
38   assert(w != NULL);
39   assert(c != NULL);
40 
41   uint8_t* c0 = c;
42   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr != 2) {
44     c1 = c0;
45   }
46 
47   const int32_t vb_zero_point = params->fp32_scalar_lrintf.kernel_zero_point;
48   do {
49     int32_t vacc0x0 = ((const int32_t*) w)[0];
50     int32_t vacc0x1 = ((const int32_t*) w)[1];
51     int32_t vacc0x2 = ((const int32_t*) w)[2];
52     int32_t vacc0x3 = ((const int32_t*) w)[3];
53     int32_t vacc1x0 = vacc0x0;
54     int32_t vacc1x1 = vacc0x1;
55     int32_t vacc1x2 = vacc0x2;
56     int32_t vacc1x3 = vacc0x3;
57     w = (const void*) ((const int32_t*) w + 4);
58 
59     size_t p = ks;
60     do {
61       const uint8_t* restrict a0 = a[0];
62       assert(a0 != NULL);
63       if XNN_UNPREDICTABLE(a0 != zero) {
64         a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
65       }
66       const uint8_t* restrict a1 = a[1];
67       assert(a1 != NULL);
68       if XNN_UNPREDICTABLE(a1 != zero) {
69         a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
70       }
71       a += 2;
72 
73       size_t k = kc;
74       do {
75         const int32_t va0 = (int32_t) (uint32_t) *a0++;
76         const int32_t va1 = (int32_t) (uint32_t) *a1++;
77 
78         const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
79         const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
80         const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
81         const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
82         w = (const void*) ((const uint8_t*) w + 4);
83 
84         vacc0x0 += va0 * vb0;
85         vacc0x1 += va0 * vb1;
86         vacc0x2 += va0 * vb2;
87         vacc0x3 += va0 * vb3;
88         vacc1x0 += va1 * vb0;
89         vacc1x1 += va1 * vb1;
90         vacc1x2 += va1 * vb2;
91         vacc1x3 += va1 * vb3;
92 
93         k -= sizeof(uint8_t);
94       } while (k != 0);
95       p -= 2 * sizeof(void*);
96     } while (p != 0);
97 
98     float vfpacc0x0 = (float) vacc0x0;
99     float vfpacc0x1 = (float) vacc0x1;
100     float vfpacc0x2 = (float) vacc0x2;
101     float vfpacc0x3 = (float) vacc0x3;
102     float vfpacc1x0 = (float) vacc1x0;
103     float vfpacc1x1 = (float) vacc1x1;
104     float vfpacc1x2 = (float) vacc1x2;
105     float vfpacc1x3 = (float) vacc1x3;
106 
107     const float vscale = params->fp32_scalar_lrintf.scale;
108     vfpacc0x0 *= vscale;
109     vfpacc0x1 *= vscale;
110     vfpacc0x2 *= vscale;
111     vfpacc0x3 *= vscale;
112     vfpacc1x0 *= vscale;
113     vfpacc1x1 *= vscale;
114     vfpacc1x2 *= vscale;
115     vfpacc1x3 *= vscale;
116 
117     const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
118     vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
119     vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
120     vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
121     vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
122     vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
123     vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
124     vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
125     vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
126 
127     const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
128     vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
129     vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
130     vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
131     vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
132     vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
133     vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
134     vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
135     vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
136 
137     const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
138     const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
139     const int32_t vrndacc0x2 = (int32_t) lrintf(vfpacc0x2);
140     const int32_t vrndacc0x3 = (int32_t) lrintf(vfpacc0x3);
141     const int32_t vrndacc1x0 = (int32_t) lrintf(vfpacc1x0);
142     const int32_t vrndacc1x1 = (int32_t) lrintf(vfpacc1x1);
143     const int32_t vrndacc1x2 = (int32_t) lrintf(vfpacc1x2);
144     const int32_t vrndacc1x3 = (int32_t) lrintf(vfpacc1x3);
145 
146     const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
147     int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
148     int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
149     int32_t vout0x2 = vrndacc0x2 + voutput_zero_point;
150     int32_t vout0x3 = vrndacc0x3 + voutput_zero_point;
151     int32_t vout1x0 = vrndacc1x0 + voutput_zero_point;
152     int32_t vout1x1 = vrndacc1x1 + voutput_zero_point;
153     int32_t vout1x2 = vrndacc1x2 + voutput_zero_point;
154     int32_t vout1x3 = vrndacc1x3 + voutput_zero_point;
155 
156     if XNN_LIKELY(nc >= 4) {
157       c1[0] = (uint8_t) vout1x0;
158       c1[1] = (uint8_t) vout1x1;
159       c1[2] = (uint8_t) vout1x2;
160       c1[3] = (uint8_t) vout1x3;
161       c0[0] = (uint8_t) vout0x0;
162       c0[1] = (uint8_t) vout0x1;
163       c0[2] = (uint8_t) vout0x2;
164       c0[3] = (uint8_t) vout0x3;
165 
166       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
167       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
168 
169       a = (const uint8_t**restrict) ((uintptr_t) a - ks);
170       nc -= 4;
171     } else {
172       if (nc & 2) {
173         c1[0] = (uint8_t) vout1x0;
174         c1[1] = (uint8_t) vout1x1;
175         vout1x0 = vout1x2;
176         c1 += 2;
177         c0[0] = (uint8_t) vout0x0;
178         c0[1] = (uint8_t) vout0x1;
179         vout0x0 = vout0x2;
180         c0 += 2;
181       }
182       if (nc & 1) {
183         c1[0] = (uint8_t) vout1x0;
184         c0[0] = (uint8_t) vout0x0;
185       }
186 
187       nc = 0;
188     }
189   } while (nc != 0);
190 }
191