xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/3x2-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_qs8_igemm_minmax_fp32_ukernel_3x2__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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_fp32_ukernel_3x2__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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32   assert(mr != 0);
33   assert(mr <= 3);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (3 * sizeof(void*)) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   int8_t* c0 = c;
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr < 2) {
45     c1 = c0;
46   }
47   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
48   if XNN_UNPREDICTABLE(mr <= 2) {
49     c2 = c1;
50   }
51 
52   do {
53     int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
54     int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
55     int32_t vacc1x0 = vacc0x0;
56     int32_t vacc1x1 = vacc0x1;
57     int32_t vacc2x0 = vacc0x0;
58     int32_t vacc2x1 = vacc0x1;
59     w = (const void*) ((const int32_t*) w + 2);
60 
61     size_t p = ks;
62     do {
63       const int8_t* restrict a0 = a[0];
64       assert(a0 != NULL);
65       if XNN_UNPREDICTABLE(a0 != zero) {
66         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
67       }
68       const int8_t* restrict a1 = a[1];
69       assert(a1 != NULL);
70       if XNN_UNPREDICTABLE(a1 != zero) {
71         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
72       }
73       const int8_t* restrict a2 = a[2];
74       assert(a2 != NULL);
75       if XNN_UNPREDICTABLE(a2 != zero) {
76         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
77       }
78       a += 3;
79 
80       size_t k = kc;
81       do {
82         const int32_t va0 = (int32_t) *a0++;
83         const int32_t va1 = (int32_t) *a1++;
84         const int32_t va2 = (int32_t) *a2++;
85 
86         const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
87         const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
88         w = (const void*) ((const int8_t*) w + 2);
89 
90         vacc0x0 += va0 * vb0;
91         vacc0x1 += va0 * vb1;
92         vacc1x0 += va1 * vb0;
93         vacc1x1 += va1 * vb1;
94         vacc2x0 += va2 * vb0;
95         vacc2x1 += va2 * vb1;
96 
97         k -= sizeof(int8_t);
98       } while (k != 0);
99       p -= 3 * sizeof(void*);
100     } while (p != 0);
101 
102     float vfpacc0x0 = (float) vacc0x0;
103     float vfpacc0x1 = (float) vacc0x1;
104     float vfpacc1x0 = (float) vacc1x0;
105     float vfpacc1x1 = (float) vacc1x1;
106     float vfpacc2x0 = (float) vacc2x0;
107     float vfpacc2x1 = (float) vacc2x1;
108 
109     const float vscale = params->fp32_scalar_lrintf.scale;
110     vfpacc0x0 *= vscale;
111     vfpacc0x1 *= vscale;
112     vfpacc1x0 *= vscale;
113     vfpacc1x1 *= vscale;
114     vfpacc2x0 *= vscale;
115     vfpacc2x1 *= 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     vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
121     vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
122     vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
123     vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
124 
125     const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
126     vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
127     vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
128     vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
129     vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
130     vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
131     vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
132 
133     const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
134     const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
135     const int32_t vrndacc1x0 = (int32_t) lrintf(vfpacc1x0);
136     const int32_t vrndacc1x1 = (int32_t) lrintf(vfpacc1x1);
137     const int32_t vrndacc2x0 = (int32_t) lrintf(vfpacc2x0);
138     const int32_t vrndacc2x1 = (int32_t) lrintf(vfpacc2x1);
139 
140     const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
141     int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
142     int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
143     int32_t vout1x0 = vrndacc1x0 + voutput_zero_point;
144     int32_t vout1x1 = vrndacc1x1 + voutput_zero_point;
145     int32_t vout2x0 = vrndacc2x0 + voutput_zero_point;
146     int32_t vout2x1 = vrndacc2x1 + voutput_zero_point;
147 
148     if XNN_LIKELY(nc >= 2) {
149       c2[0] = (int8_t) vout2x0;
150       c2[1] = (int8_t) vout2x1;
151       c1[0] = (int8_t) vout1x0;
152       c1[1] = (int8_t) vout1x1;
153       c0[0] = (int8_t) vout0x0;
154       c0[1] = (int8_t) vout0x1;
155 
156       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
157       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
158       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
159 
160       a = (const int8_t**restrict) ((uintptr_t) a - ks);
161       nc -= 2;
162     } else {
163       if (nc & 1) {
164         c2[0] = (int8_t) vout2x0;
165         c1[0] = (int8_t) vout1x0;
166         c0[0] = (int8_t) vout0x0;
167       }
168 
169       nc = 0;
170     }
171   } while (nc != 0);
172 }
173