xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/1x4-minmax-fp32-scalar-fmagic.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 
12 #include <xnnpack/math.h>
13 #include <xnnpack/gemm.h>
14 
15 
xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_fmagic(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)])16 void xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_fmagic(
17     size_t mr,
18     size_t nc,
19     size_t kc,
20     size_t ks,
21     const int8_t**restrict a,
22     const void*restrict w,
23     int8_t*restrict c,
24     size_t cm_stride,
25     size_t cn_stride,
26     size_t a_offset,
27     const int8_t* zero,
28     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30   assert(mr != 0);
31   assert(mr <= 1);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(ks != 0);
35   assert(ks % (1 * sizeof(void*)) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39 
40   int8_t* c0 = c;
41 
42   do {
43     int32_t vacc0x0 = ((const int32_t*) w)[0];
44     int32_t vacc0x1 = ((const int32_t*) w)[1];
45     int32_t vacc0x2 = ((const int32_t*) w)[2];
46     int32_t vacc0x3 = ((const int32_t*) w)[3];
47     w = (const void*) ((const int32_t*) w + 4);
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         const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
65         const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
66         w = (const void*) ((const int8_t*) w + 4);
67 
68         vacc0x0 += va0 * vb0;
69         vacc0x1 += va0 * vb1;
70         vacc0x2 += va0 * vb2;
71         vacc0x3 += va0 * vb3;
72 
73         k -= sizeof(int8_t);
74       } while (k != 0);
75       p -= 1 * sizeof(void*);
76     } while (p != 0);
77 
78     float vfpacc0x0 = (float) vacc0x0;
79     float vfpacc0x1 = (float) vacc0x1;
80     float vfpacc0x2 = (float) vacc0x2;
81     float vfpacc0x3 = (float) vacc0x3;
82 
83     const float vscale = params->fp32_scalar_fmagic.scale;
84     vfpacc0x0 *= vscale;
85     vfpacc0x1 *= vscale;
86     vfpacc0x2 *= vscale;
87     vfpacc0x3 *= vscale;
88 
89     const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
90     vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
91     vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
92     vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
93     vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
94 
95     const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
96     vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
97     vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
98     vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
99     vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
100 
101     const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
102     vfpacc0x0 += vmagic_bias;
103     vfpacc0x1 += vmagic_bias;
104     vfpacc0x2 += vmagic_bias;
105     vfpacc0x3 += vmagic_bias;
106 
107     const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
108     int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
109     int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
110     int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
111     int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
112 
113     if XNN_LIKELY(nc >= 4) {
114       c0[0] = (int8_t) vout0x0;
115       c0[1] = (int8_t) vout0x1;
116       c0[2] = (int8_t) vout0x2;
117       c0[3] = (int8_t) vout0x3;
118 
119       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
120 
121       a = (const int8_t**restrict) ((uintptr_t) a - ks);
122       nc -= 4;
123     } else {
124       if (nc & 2) {
125         c0[0] = (int8_t) vout0x0;
126         c0[1] = (int8_t) vout0x1;
127         vout0x0 = vout0x2;
128         c0 += 2;
129       }
130       if (nc & 1) {
131         c0[0] = (int8_t) vout0x0;
132       }
133 
134       nc = 0;
135     }
136   } while (nc != 0);
137 }
138