Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x2-minmax-fp32-scalar-imagic.c

raw
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
17 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_2x2__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_qs8_qc8w_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 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
48 const int32_t output_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point;
49 const float vmagic_bias = 12582912.0f;
50 const int32_t vmagic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point);
51 const int32_t vmagic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zero_point);
52 const int32_t vmagic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point;
53
54 do {
55 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
56 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
57 int32_t vacc1x0 = vacc0x0;
58 int32_t vacc1x1 = vacc0x1;
59 w = (const void*) ((const int32_t*) w + 2);
60
61 size_t p = ks;
62 do {
63 const int8_t* restrict a0 = a[0];
64 assert(a0 != NULL);
65 if XNN_UNPREDICTABLE(a0 != zero) {
66 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
67 }
68 const int8_t* restrict a1 = a[1];
69 assert(a1 != NULL);
70 if XNN_UNPREDICTABLE(a1 != zero) {
71 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
72 }
73 a += 2;
74
75 size_t k = kc;
76 do {
77 const int32_t va0 = (int32_t) *a0++;
78 const int32_t va1 = (int32_t) *a1++;
79
80 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
81 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
82 w = (const void*) ((const int8_t*) w + 2);
83
84 vacc0x0 += va0 * vb0;
85 vacc0x1 += va0 * vb1;
86 vacc1x0 += va1 * vb0;
87 vacc1x1 += va1 * vb1;
88
89 k -= sizeof(int8_t);
90 } while (k != 0);
91 p -= 2 * sizeof(void*);
92 } while (p != 0);
93
94 float vfpacc0x0 = (float) vacc0x0;
95 float vfpacc0x1 = (float) vacc0x1;
96 float vfpacc1x0 = (float) vacc1x0;
97 float vfpacc1x1 = (float) vacc1x1;
98
99 const float vscale0 = unaligned_indexed_load_f32(w, 0);
100 vfpacc0x0 *= vscale0;
101 vfpacc1x0 *= vscale0;
102 const float vscale1 = unaligned_indexed_load_f32(w, 1);
103 vfpacc0x1 *= vscale1;
104 vfpacc1x1 *= vscale1;
105 w = (const void*) ((const float*) w + 2);
106
107 vfpacc0x0 += vmagic_bias;
108 vfpacc0x1 += vmagic_bias;
109 vfpacc1x0 += vmagic_bias;
110 vfpacc1x1 += vmagic_bias;
111
112 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
113 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1);
114 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0);
115 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1);
116
117 vout0x0 = math_max_s32(vout0x0, vmagic_min);
118 vout0x1 = math_max_s32(vout0x1, vmagic_min);
119 vout1x0 = math_max_s32(vout1x0, vmagic_min);
120 vout1x1 = math_max_s32(vout1x1, vmagic_min);
121
122 vout0x0 = math_min_s32(vout0x0, vmagic_max);
123 vout0x1 = math_min_s32(vout0x1, vmagic_max);
124 vout1x0 = math_min_s32(vout1x0, vmagic_max);
125 vout1x1 = math_min_s32(vout1x1, vmagic_max);
126
127 vout0x0 -= vmagic_bias_less_zero_point;
128 vout0x1 -= vmagic_bias_less_zero_point;
129 vout1x0 -= vmagic_bias_less_zero_point;
130 vout1x1 -= vmagic_bias_less_zero_point;
131
132 if XNN_LIKELY(nc >= 2) {
133 c1[0] = (int8_t) vout1x0;
134 c1[1] = (int8_t) vout1x1;
135 c0[0] = (int8_t) vout0x0;
136 c0[1] = (int8_t) vout0x1;
137
138 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
139 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
140
141 a = (const int8_t**restrict) ((uintptr_t) a - ks);
142 nc -= 2;
143 } else {
144 if (nc & 1) {
145 c1[0] = (int8_t) vout1x0;
146 c0[0] = (int8_t) vout0x0;
147 }
148
149 nc = 0;
150 }
151 } while (nc != 0);
152 }