Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x2-minmax-fp32-scalar-lrintf.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 #include <math.h>
12
13 #include "xnnpack/math.h"
14 #include "xnnpack/gemm.h"
15 #include "xnnpack/unaligned.h"
16
17
18 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf(
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_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * sizeof(void*)) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 int8_t* c0 = c;
43
44 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
45 const int32_t output_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point;
46 const float voutput_min_less_zero_point = output_min_less_zero_point;
47 const float voutput_max_less_zero_point = output_max_less_zero_point;
48 const int32_t voutput_zero_point = params->fp32_scalar.output_zero_point;
49
50 do {
51 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
52 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
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 a += 1;
63
64 size_t k = kc;
65 do {
66 const int32_t va0 = (int32_t) *a0++;
67
68 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
69 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
70 w = (const void*) ((const int8_t*) w + 2);
71
72 vacc0x0 += va0 * vb0;
73 vacc0x1 += va0 * vb1;
74
75 k -= sizeof(int8_t);
76 } while (k != 0);
77 p -= 1 * sizeof(void*);
78 } while (p != 0);
79
80 float vfpacc0x0 = (float) vacc0x0;
81 float vfpacc0x1 = (float) vacc0x1;
82
83 const float vscale0 = unaligned_indexed_load_f32(w, 0);
84 vfpacc0x0 *= vscale0;
85 const float vscale1 = unaligned_indexed_load_f32(w, 1);
86 vfpacc0x1 *= vscale1;
87 w = (const void*) ((const float*) w + 2);
88
89 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
90 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
91
92 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
93 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
94
95 const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
96 const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
97
98 int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
99 int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
100
101 if XNN_LIKELY(nc >= 2) {
102 c0[0] = (int8_t) vout0x0;
103 c0[1] = (int8_t) vout0x1;
104
105 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
106
107 a = (const int8_t**restrict) ((uintptr_t) a - ks);
108 nc -= 2;
109 } else {
110 if (nc & 1) {
111 c0[0] = (int8_t) vout0x0;
112 }
113
114 nc = 0;
115 }
116 } while (nc != 0);
117 }