1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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
16
xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
31 assert(nc != 0);
32 assert(kc != 0);
33
34 const int8_t* a0 = a;
35 int8_t* c0 = c;
36
37 do {
38 int32_t vacc0x0 = ((const int32_t*) w)[0];
39 int32_t vacc0x1 = ((const int32_t*) w)[1];
40 int32_t vacc0x2 = ((const int32_t*) w)[2];
41 int32_t vacc0x3 = ((const int32_t*) w)[3];
42 w = (const void*) ((const int32_t*) w + 4);
43
44 size_t k = kc;
45 do {
46 const int32_t va0 = (int32_t) *a0++;
47
48 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
49 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
50 const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
51 const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
52 w = (const void*) ((const int8_t*) w + 4);
53
54 vacc0x0 += va0 * vb0;
55 vacc0x1 += va0 * vb1;
56 vacc0x2 += va0 * vb2;
57 vacc0x3 += va0 * vb3;
58
59 k -= sizeof(int8_t);
60 } while (k != 0);
61
62 float vfpacc0x0 = (float) vacc0x0;
63 float vfpacc0x1 = (float) vacc0x1;
64 float vfpacc0x2 = (float) vacc0x2;
65 float vfpacc0x3 = (float) vacc0x3;
66
67 const float vscale = params->fp32_scalar_lrintf.scale;
68 vfpacc0x0 *= vscale;
69 vfpacc0x1 *= vscale;
70 vfpacc0x2 *= vscale;
71 vfpacc0x3 *= vscale;
72
73 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
74 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
75 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
76 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
77 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
78
79 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
80 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
81 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
82 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
83 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
84
85 const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
86 const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
87 const int32_t vrndacc0x2 = (int32_t) lrintf(vfpacc0x2);
88 const int32_t vrndacc0x3 = (int32_t) lrintf(vfpacc0x3);
89
90 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
91 int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
92 int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
93 int32_t vout0x2 = vrndacc0x2 + voutput_zero_point;
94 int32_t vout0x3 = vrndacc0x3 + voutput_zero_point;
95
96 if XNN_LIKELY(nc >= 4) {
97 c0[0] = (int8_t) vout0x0;
98 c0[1] = (int8_t) vout0x1;
99 c0[2] = (int8_t) vout0x2;
100 c0[3] = (int8_t) vout0x3;
101
102 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
103
104 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
105
106 nc -= 4;
107 } else {
108 if (nc & 2) {
109 c0[0] = (int8_t) vout0x0;
110 c0[1] = (int8_t) vout0x1;
111 vout0x0 = vout0x2;
112 c0 += 2;
113 }
114 if (nc & 1) {
115 c0[0] = (int8_t) vout0x0;
116 }
117
118 nc = 0;
119 }
120 } while (nc != 0);
121 }
122