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 <fp16.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_qc8_igemm_minmax_fp32_ukernel_2x2__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_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic(
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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * 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
48 do {
49 int32_t vacc0x0 = ((const int32_t*) w)[0];
50 int32_t vacc0x1 = ((const int32_t*) w)[1];
51 int32_t vacc1x0 = vacc0x0;
52 int32_t vacc1x1 = vacc0x1;
53 w = (const void*) ((const int32_t*) w + 2);
54
55 size_t p = ks;
56 do {
57 const int8_t* restrict a0 = a[0];
58 assert(a0 != NULL);
59 if XNN_UNPREDICTABLE(a0 != zero) {
60 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
61 }
62 const int8_t* restrict a1 = a[1];
63 assert(a1 != NULL);
64 if XNN_UNPREDICTABLE(a1 != zero) {
65 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
66 }
67 a += 2;
68
69 size_t k = kc;
70 do {
71 const int32_t va0 = (int32_t) *a0++;
72 const int32_t va1 = (int32_t) *a1++;
73
74 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
75 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
76 w = (const void*) ((const int8_t*) w + 2);
77
78 vacc0x0 += va0 * vb0;
79 vacc0x1 += va0 * vb1;
80 vacc1x0 += va1 * vb0;
81 vacc1x1 += va1 * vb1;
82
83 k -= sizeof(int8_t);
84 } while (k != 0);
85 p -= 2 * sizeof(void*);
86 } while (p != 0);
87
88 float vfpacc0x0 = (float) vacc0x0;
89 float vfpacc0x1 = (float) vacc0x1;
90 float vfpacc1x0 = (float) vacc1x0;
91 float vfpacc1x1 = (float) vacc1x1;
92
93 typedef XNN_UNALIGNED float unaligned_float;
94 const float vscale0 = ((const unaligned_float*) w)[0];
95 vfpacc0x0 *= vscale0;
96 vfpacc1x0 *= vscale0;
97 const float vscale1 = ((const unaligned_float*) w)[1];
98 vfpacc0x1 *= vscale1;
99 vfpacc1x1 *= vscale1;
100 w = (const void*) ((const float*) w + 2);
101
102 const float vmagic_bias = params->scalar_imagic.magic_bias;
103 vfpacc0x0 += vmagic_bias;
104 vfpacc0x1 += vmagic_bias;
105 vfpacc1x0 += vmagic_bias;
106 vfpacc1x1 += vmagic_bias;
107
108 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0);
109 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1);
110 int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0);
111 int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1);
112
113 const int32_t vmagic_min = params->scalar_imagic.magic_min;
114 vout0x0 = math_max_s32(vout0x0, vmagic_min);
115 vout0x1 = math_max_s32(vout0x1, vmagic_min);
116 vout1x0 = math_max_s32(vout1x0, vmagic_min);
117 vout1x1 = math_max_s32(vout1x1, vmagic_min);
118
119 const int32_t vmagic_max = params->scalar_imagic.magic_max;
120 vout0x0 = math_min_s32(vout0x0, vmagic_max);
121 vout0x1 = math_min_s32(vout0x1, vmagic_max);
122 vout1x0 = math_min_s32(vout1x0, vmagic_max);
123 vout1x1 = math_min_s32(vout1x1, vmagic_max);
124
125 const int32_t vmagic_bias_less_zero_point = params->scalar_imagic.magic_bias_less_zero_point;
126 vout0x0 -= vmagic_bias_less_zero_point;
127 vout0x1 -= vmagic_bias_less_zero_point;
128 vout1x0 -= vmagic_bias_less_zero_point;
129 vout1x1 -= vmagic_bias_less_zero_point;
130
131 if XNN_LIKELY(nc >= 2) {
132 c1[0] = (int8_t) vout1x0;
133 c1[1] = (int8_t) vout1x1;
134 c0[0] = (int8_t) vout0x0;
135 c0[1] = (int8_t) vout0x1;
136
137 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
138 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
139
140 a = (const int8_t**restrict) ((uintptr_t) a - ks);
141 nc -= 2;
142 } else {
143 if (nc & 1) {
144 c1[0] = (int8_t) vout1x0;
145 c0[0] = (int8_t) vout0x0;
146 }
147
148 nc = 0;
149 }
150 } while (nc != 0);
151 }
152