xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/1x4-minmax-fp32-scalar-fmagic.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_fmagic(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_fmagic(
17     size_t mr,
18     size_t nc,
19     size_t kc,
20     const int8_t* restrict a,
21     size_t a_stride,
22     const void* restrict w,
23     int8_t* restrict c,
24     size_t cm_stride,
25     size_t cn_stride,
26     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28   assert(mr != 0);
29   assert(mr <= 1);
30   assert(nc != 0);
31   assert(kc != 0);
32 
33   const int8_t* a0 = a;
34   int8_t* c0 = c;
35 
36   do {
37     int32_t vacc0x0 = ((const int32_t*) w)[0];
38     int32_t vacc0x1 = ((const int32_t*) w)[1];
39     int32_t vacc0x2 = ((const int32_t*) w)[2];
40     int32_t vacc0x3 = ((const int32_t*) w)[3];
41     w = (const void*) ((const int32_t*) w + 4);
42 
43     size_t k = kc;
44     do {
45       const int32_t va0 = (int32_t) *a0++;
46 
47       const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
48       const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
49       const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
50       const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
51       w = (const void*) ((const int8_t*) w + 4);
52 
53       vacc0x0 += va0 * vb0;
54       vacc0x1 += va0 * vb1;
55       vacc0x2 += va0 * vb2;
56       vacc0x3 += va0 * vb3;
57 
58       k -= sizeof(int8_t);
59     } while (k != 0);
60 
61     float vfpacc0x0 = (float) vacc0x0;
62     float vfpacc0x1 = (float) vacc0x1;
63     float vfpacc0x2 = (float) vacc0x2;
64     float vfpacc0x3 = (float) vacc0x3;
65 
66     const float vscale0 = ((const float*) w)[0];
67     vfpacc0x0 *= vscale0;
68     const float vscale1 = ((const float*) w)[1];
69     vfpacc0x1 *= vscale1;
70     const float vscale2 = ((const float*) w)[2];
71     vfpacc0x2 *= vscale2;
72     const float vscale3 = ((const float*) w)[3];
73     vfpacc0x3 *= vscale3;
74     w = (const void*) ((const float*) w + 4);
75 
76     const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
77     vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
78     vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
79     vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
80     vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
81 
82     const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
83     vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
84     vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
85     vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
86     vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
87 
88     const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
89     vfpacc0x0 += vmagic_bias;
90     vfpacc0x1 += vmagic_bias;
91     vfpacc0x2 += vmagic_bias;
92     vfpacc0x3 += vmagic_bias;
93 
94     const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
95     int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
96     int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
97     int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
98     int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
99 
100     if XNN_LIKELY(nc >= 4) {
101       c0[0] = (int8_t) vout0x0;
102       c0[1] = (int8_t) vout0x1;
103       c0[2] = (int8_t) vout0x2;
104       c0[3] = (int8_t) vout0x3;
105 
106       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
107 
108       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
109 
110       nc -= 4;
111     } else {
112       if (nc & 2) {
113         c0[0] = (int8_t) vout0x0;
114         c0[1] = (int8_t) vout0x1;
115         vout0x0 = vout0x2;
116         c0 += 2;
117       }
118       if (nc & 1) {
119         c0[0] = (int8_t) vout0x0;
120       }
121 
122       nc = 0;
123     }
124   } while (nc != 0);
125 }
126