Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x2-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_3x2__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 <= 3);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (3 * 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 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
47 if XNN_UNPREDICTABLE(mr <= 2) {
48 c2 = c1;
49 }
50
51 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
52 const int32_t output_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point;
53 const float vmagic_bias = 12582912.0f;
54 const int32_t vmagic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point);
55 const int32_t vmagic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zero_point);
56 const int32_t vmagic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point;
57
58 do {
59 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
60 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
61 int32_t vacc1x0 = vacc0x0;
62 int32_t vacc1x1 = vacc0x1;
63 int32_t vacc2x0 = vacc0x0;
64 int32_t vacc2x1 = vacc0x1;
65 w = (const void*) ((const int32_t*) w + 2);
66
67 size_t p = ks;
68 do {
69 const int8_t* restrict a0 = a[0];
70 assert(a0 != NULL);
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73 }
74 const int8_t* restrict a1 = a[1];
75 assert(a1 != NULL);
76 if XNN_UNPREDICTABLE(a1 != zero) {
77 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
78 }
79 const int8_t* restrict a2 = a[2];
80 assert(a2 != NULL);
81 if XNN_UNPREDICTABLE(a2 != zero) {
82 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
83 }
84 a += 3;
85
86 size_t k = kc;
87 do {
88 const int32_t va0 = (int32_t) *a0++;
89 const int32_t va1 = (int32_t) *a1++;
90 const int32_t va2 = (int32_t) *a2++;
91
92 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
93 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
94 w = (const void*) ((const int8_t*) w + 2);
95
96 vacc0x0 += va0 * vb0;
97 vacc0x1 += va0 * vb1;
98 vacc1x0 += va1 * vb0;
99 vacc1x1 += va1 * vb1;
100 vacc2x0 += va2 * vb0;
101 vacc2x1 += va2 * vb1;
102
103 k -= sizeof(int8_t);
104 } while (k != 0);
105 p -= 3 * sizeof(void*);
106 } while (p != 0);
107
108 float vfpacc0x0 = (float) vacc0x0;
109 float vfpacc0x1 = (float) vacc0x1;
110 float vfpacc1x0 = (float) vacc1x0;
111 float vfpacc1x1 = (float) vacc1x1;
112 float vfpacc2x0 = (float) vacc2x0;
113 float vfpacc2x1 = (float) vacc2x1;
114
115 const float vscale0 = unaligned_indexed_load_f32(w, 0);
116 vfpacc0x0 *= vscale0;
117 vfpacc1x0 *= vscale0;
118 vfpacc2x0 *= vscale0;
119 const float vscale1 = unaligned_indexed_load_f32(w, 1);
120 vfpacc0x1 *= vscale1;
121 vfpacc1x1 *= vscale1;
122 vfpacc2x1 *= vscale1;
123 w = (const void*) ((const float*) w + 2);
124
125 vfpacc0x0 += vmagic_bias;
126 vfpacc0x1 += vmagic_bias;
127 vfpacc1x0 += vmagic_bias;
128 vfpacc1x1 += vmagic_bias;
129 vfpacc2x0 += vmagic_bias;
130 vfpacc2x1 += vmagic_bias;
131
132 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
133 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1);
134 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0);
135 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1);
136 int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0);
137 int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1);
138
139 vout0x0 = math_max_s32(vout0x0, vmagic_min);
140 vout0x1 = math_max_s32(vout0x1, vmagic_min);
141 vout1x0 = math_max_s32(vout1x0, vmagic_min);
142 vout1x1 = math_max_s32(vout1x1, vmagic_min);
143 vout2x0 = math_max_s32(vout2x0, vmagic_min);
144 vout2x1 = math_max_s32(vout2x1, vmagic_min);
145
146 vout0x0 = math_min_s32(vout0x0, vmagic_max);
147 vout0x1 = math_min_s32(vout0x1, vmagic_max);
148 vout1x0 = math_min_s32(vout1x0, vmagic_max);
149 vout1x1 = math_min_s32(vout1x1, vmagic_max);
150 vout2x0 = math_min_s32(vout2x0, vmagic_max);
151 vout2x1 = math_min_s32(vout2x1, vmagic_max);
152
153 vout0x0 -= vmagic_bias_less_zero_point;
154 vout0x1 -= vmagic_bias_less_zero_point;
155 vout1x0 -= vmagic_bias_less_zero_point;
156 vout1x1 -= vmagic_bias_less_zero_point;
157 vout2x0 -= vmagic_bias_less_zero_point;
158 vout2x1 -= vmagic_bias_less_zero_point;
159
160 if XNN_LIKELY(nc >= 2) {
161 c2[0] = (int8_t) vout2x0;
162 c2[1] = (int8_t) vout2x1;
163 c1[0] = (int8_t) vout1x0;
164 c1[1] = (int8_t) vout1x1;
165 c0[0] = (int8_t) vout0x0;
166 c0[1] = (int8_t) vout0x1;
167
168 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
169 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
170 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
171
172 a = (const int8_t**restrict) ((uintptr_t) a - ks);
173 nc -= 2;
174 } else {
175 if (nc & 1) {
176 c2[0] = (int8_t) vout2x0;
177 c1[0] = (int8_t) vout1x0;
178 c0[0] = (int8_t) vout0x0;
179 }
180
181 nc = 0;
182 }
183 } while (nc != 0);
184 }