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_2x2__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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_fmagic(
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 <= 2);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (2 * sizeof(void*)) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 int8_t* c0 = c;
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 c1 = c0;
45 }
46
47 do {
48 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
49 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
50 int32_t vacc1x0 = vacc0x0;
51 int32_t vacc1x1 = vacc0x1;
52 w = (const void*) ((const int32_t*) w + 2);
53
54 size_t p = ks;
55 do {
56 const int8_t* restrict a0 = a[0];
57 assert(a0 != NULL);
58 if XNN_UNPREDICTABLE(a0 != zero) {
59 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60 }
61 const int8_t* restrict a1 = a[1];
62 assert(a1 != NULL);
63 if XNN_UNPREDICTABLE(a1 != zero) {
64 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
65 }
66 a += 2;
67
68 size_t k = kc;
69 do {
70 const int32_t va0 = (int32_t) *a0++;
71 const int32_t va1 = (int32_t) *a1++;
72
73 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
74 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
75 w = (const void*) ((const int8_t*) w + 2);
76
77 vacc0x0 += va0 * vb0;
78 vacc0x1 += va0 * vb1;
79 vacc1x0 += va1 * vb0;
80 vacc1x1 += va1 * vb1;
81
82 k -= sizeof(int8_t);
83 } while (k != 0);
84 p -= 2 * sizeof(void*);
85 } while (p != 0);
86
87 float vfpacc0x0 = (float) vacc0x0;
88 float vfpacc0x1 = (float) vacc0x1;
89 float vfpacc1x0 = (float) vacc1x0;
90 float vfpacc1x1 = (float) vacc1x1;
91
92 const float vscale0 = unaligned_indexed_load_f32(w, 0);
93 vfpacc0x0 *= vscale0;
94 vfpacc1x0 *= vscale0;
95 const float vscale1 = unaligned_indexed_load_f32(w, 1);
96 vfpacc0x1 *= vscale1;
97 vfpacc1x1 *= vscale1;
98 w = (const void*) ((const float*) w + 2);
99
100 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
101 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
102 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
103 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
104 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
105
106 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
107 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
108 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
109 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
110 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
111
112 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
113 vfpacc0x0 += vmagic_bias;
114 vfpacc0x1 += vmagic_bias;
115 vfpacc1x0 += vmagic_bias;
116 vfpacc1x1 += vmagic_bias;
117
118 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
119 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
120 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
121 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
122 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
123
124 if XNN_LIKELY(nc >= 2) {
125 c1[0] = (int8_t) vout1x0;
126 c1[1] = (int8_t) vout1x1;
127 c0[0] = (int8_t) vout0x0;
128 c0[1] = (int8_t) vout0x1;
129
130 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
131 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
132
133 a = (const int8_t**restrict) ((uintptr_t) a - ks);
134 nc -= 2;
135 } else {
136 if (nc & 1) {
137 c1[0] = (int8_t) vout1x0;
138 c0[0] = (int8_t) vout0x0;
139 }
140
141 nc = 0;
142 }
143 } while (nc != 0);
144 }
145