xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/1x2-minmax-fp32-scalar-imagic.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 #include <xnnpack/unaligned.h>
15 
16 
xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const int8_t**restrict a,
23     const void*restrict w,
24     int8_t*restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const int8_t* zero,
29     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31   assert(mr != 0);
32   assert(mr <= 1);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (1 * sizeof(void*)) == 0);
37   assert(a != NULL);
38   assert(w != NULL);
39   assert(c != NULL);
40 
41   int8_t* c0 = c;
42 
43   do {
44     int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
45     int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
46     w = (const void*) ((const int32_t*) w + 2);
47 
48     size_t p = ks;
49     do {
50       const int8_t* restrict a0 = a[0];
51       assert(a0 != NULL);
52       if XNN_UNPREDICTABLE(a0 != zero) {
53         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
54       }
55       a += 1;
56 
57       size_t k = kc;
58       do {
59         const int32_t va0 = (int32_t) *a0++;
60 
61         const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
62         const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
63         w = (const void*) ((const int8_t*) w + 2);
64 
65         vacc0x0 += va0 * vb0;
66         vacc0x1 += va0 * vb1;
67 
68         k -= sizeof(int8_t);
69       } while (k != 0);
70       p -= 1 * sizeof(void*);
71     } while (p != 0);
72 
73     float vfpacc0x0 = (float) vacc0x0;
74     float vfpacc0x1 = (float) vacc0x1;
75 
76     const float vscale0 = unaligned_indexed_load_f32(w, 0);
77     vfpacc0x0 *= vscale0;
78     const float vscale1 = unaligned_indexed_load_f32(w, 1);
79     vfpacc0x1 *= vscale1;
80     w = (const void*) ((const float*) w + 2);
81 
82     const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
83     vfpacc0x0 += vmagic_bias;
84     vfpacc0x1 += vmagic_bias;
85 
86     int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
87     int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1);
88 
89     const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
90     vout0x0 = math_max_s32(vout0x0, vmagic_min);
91     vout0x1 = math_max_s32(vout0x1, vmagic_min);
92 
93     const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
94     vout0x0 = math_min_s32(vout0x0, vmagic_max);
95     vout0x1 = math_min_s32(vout0x1, vmagic_max);
96 
97     const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
98     vout0x0 -= vmagic_bias_less_zero_point;
99     vout0x1 -= vmagic_bias_less_zero_point;
100 
101     if XNN_LIKELY(nc >= 2) {
102       c0[0] = (int8_t) vout0x0;
103       c0[1] = (int8_t) vout0x1;
104 
105       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
106 
107       a = (const int8_t**restrict) ((uintptr_t) a - ks);
108       nc -= 2;
109     } else {
110       if (nc & 1) {
111         c0[0] = (int8_t) vout0x0;
112       }
113 
114       nc = 0;
115     }
116   } while (nc != 0);
117 }
118