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_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_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 uint8_t**restrict a,
24 const void*restrict w,
25 uint8_t*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_conv_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 uint8_t* c0 = c;
43 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 2) {
45 c1 = c0;
46 }
47
48 const int32_t vb_zero_point = params->fp32_scalar_imagic.kernel_zero_point;
49 do {
50 int32_t vacc0x0 = ((const int32_t*) w)[0];
51 int32_t vacc0x1 = ((const int32_t*) w)[1];
52 int32_t vacc1x0 = vacc0x0;
53 int32_t vacc1x1 = vacc0x1;
54 w = (const void*) ((const int32_t*) w + 2);
55
56 size_t p = ks;
57 do {
58 const uint8_t* restrict a0 = a[0];
59 assert(a0 != NULL);
60 if XNN_UNPREDICTABLE(a0 != zero) {
61 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
62 }
63 const uint8_t* restrict a1 = a[1];
64 assert(a1 != NULL);
65 if XNN_UNPREDICTABLE(a1 != zero) {
66 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
67 }
68 a += 2;
69
70 size_t k = kc;
71 do {
72 const int32_t va0 = (int32_t) (uint32_t) *a0++;
73 const int32_t va1 = (int32_t) (uint32_t) *a1++;
74
75 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
76 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
77 w = (const void*) ((const uint8_t*) w + 2);
78
79 vacc0x0 += va0 * vb0;
80 vacc0x1 += va0 * vb1;
81 vacc1x0 += va1 * vb0;
82 vacc1x1 += va1 * vb1;
83
84 k -= sizeof(uint8_t);
85 } while (k != 0);
86 p -= 2 * sizeof(void*);
87 } while (p != 0);
88
89 float vfpacc0x0 = (float) vacc0x0;
90 float vfpacc0x1 = (float) vacc0x1;
91 float vfpacc1x0 = (float) vacc1x0;
92 float vfpacc1x1 = (float) vacc1x1;
93
94 const float vscale = params->fp32_scalar_imagic.scale;
95 vfpacc0x0 *= vscale;
96 vfpacc0x1 *= vscale;
97 vfpacc1x0 *= vscale;
98 vfpacc1x1 *= vscale;
99
100 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
101 vfpacc0x0 += vmagic_bias;
102 vfpacc0x1 += vmagic_bias;
103 vfpacc1x0 += vmagic_bias;
104 vfpacc1x1 += vmagic_bias;
105
106 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0);
107 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1);
108 int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0);
109 int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1);
110
111 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
112 vout0x0 = math_max_s32(vout0x0, vmagic_min);
113 vout0x1 = math_max_s32(vout0x1, vmagic_min);
114 vout1x0 = math_max_s32(vout1x0, vmagic_min);
115 vout1x1 = math_max_s32(vout1x1, vmagic_min);
116
117 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
118 vout0x0 = math_min_s32(vout0x0, vmagic_max);
119 vout0x1 = math_min_s32(vout0x1, vmagic_max);
120 vout1x0 = math_min_s32(vout1x0, vmagic_max);
121 vout1x1 = math_min_s32(vout1x1, vmagic_max);
122
123 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
124 vout0x0 -= vmagic_bias_less_zero_point;
125 vout0x1 -= vmagic_bias_less_zero_point;
126 vout1x0 -= vmagic_bias_less_zero_point;
127 vout1x1 -= vmagic_bias_less_zero_point;
128
129 if XNN_LIKELY(nc >= 2) {
130 c1[0] = (uint8_t) vout1x0;
131 c1[1] = (uint8_t) vout1x1;
132 c0[0] = (uint8_t) vout0x0;
133 c0[1] = (uint8_t) vout0x1;
134
135 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
136 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
137
138 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
139 nc -= 2;
140 } else {
141 if (nc & 1) {
142 c1[0] = (uint8_t) vout1x0;
143 c0[0] = (uint8_t) vout0x0;
144 }
145
146 nc = 0;
147 }
148 } while (nc != 0);
149 }
150