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
xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf(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_1x2__scalar_lrintf(
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 <= 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 uint8_t* c0 = c;
43
44 const int32_t vb_zero_point = params->fp32_scalar_lrintf.kernel_zero_point;
45 do {
46 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
47 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
48 w = (const void*) ((const int32_t*) w + 2);
49
50 size_t p = ks;
51 do {
52 const uint8_t* restrict a0 = a[0];
53 assert(a0 != NULL);
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 do {
61 const int32_t va0 = (int32_t) (uint32_t) *a0++;
62
63 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
64 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
65 w = (const void*) ((const uint8_t*) w + 2);
66
67 vacc0x0 += va0 * vb0;
68 vacc0x1 += va0 * vb1;
69
70 k -= sizeof(uint8_t);
71 } while (k != 0);
72 p -= 1 * sizeof(void*);
73 } while (p != 0);
74
75 float vfpacc0x0 = (float) vacc0x0;
76 float vfpacc0x1 = (float) vacc0x1;
77
78 const float vscale = params->fp32_scalar_lrintf.scale;
79 vfpacc0x0 *= vscale;
80 vfpacc0x1 *= vscale;
81
82 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
83 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
84 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
85
86 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
87 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
88 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
89
90 const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
91 const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
92
93 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
94 int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
95 int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
96
97 if XNN_LIKELY(nc >= 2) {
98 c0[0] = (uint8_t) vout0x0;
99 c0[1] = (uint8_t) vout0x1;
100
101 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
102
103 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
104 nc -= 2;
105 } else {
106 if (nc & 1) {
107 c0[0] = (uint8_t) vout0x0;
108 }
109
110 nc = 0;
111 }
112 } while (nc != 0);
113 }
114