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
16
xnn_qu8_igemm_minmax_fp32_ukernel_1x4__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)])17 void xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const uint8_t**restrict a,
23 const void*restrict w,
24 uint8_t*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const uint8_t* zero,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (1 * sizeof(void*)) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 uint8_t* c0 = c;
42
43 const int32_t vb_zero_point = params->fp32_scalar_lrintf.kernel_zero_point;
44 do {
45 int32_t vacc0x0 = ((const int32_t*) w)[0];
46 int32_t vacc0x1 = ((const int32_t*) w)[1];
47 int32_t vacc0x2 = ((const int32_t*) w)[2];
48 int32_t vacc0x3 = ((const int32_t*) w)[3];
49 w = (const void*) ((const int32_t*) w + 4);
50
51 size_t p = ks;
52 do {
53 const uint8_t* restrict a0 = a[0];
54 assert(a0 != NULL);
55 if XNN_UNPREDICTABLE(a0 != zero) {
56 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
57 }
58 a += 1;
59
60 size_t k = kc;
61 do {
62 const int32_t va0 = (int32_t) (uint32_t) *a0++;
63
64 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
65 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
66 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
67 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
68 w = (const void*) ((const uint8_t*) w + 4);
69
70 vacc0x0 += va0 * vb0;
71 vacc0x1 += va0 * vb1;
72 vacc0x2 += va0 * vb2;
73 vacc0x3 += va0 * vb3;
74
75 k -= sizeof(uint8_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 float vfpacc0x2 = (float) vacc0x2;
83 float vfpacc0x3 = (float) vacc0x3;
84
85 const float vscale = params->fp32_scalar_lrintf.scale;
86 vfpacc0x0 *= vscale;
87 vfpacc0x1 *= vscale;
88 vfpacc0x2 *= vscale;
89 vfpacc0x3 *= vscale;
90
91 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
92 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
93 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
94 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
95 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
96
97 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
98 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
99 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
100 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
101 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
102
103 const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
104 const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
105 const int32_t vrndacc0x2 = (int32_t) lrintf(vfpacc0x2);
106 const int32_t vrndacc0x3 = (int32_t) lrintf(vfpacc0x3);
107
108 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
109 int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
110 int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
111 int32_t vout0x2 = vrndacc0x2 + voutput_zero_point;
112 int32_t vout0x3 = vrndacc0x3 + voutput_zero_point;
113
114 if XNN_LIKELY(nc >= 4) {
115 c0[0] = (uint8_t) vout0x0;
116 c0[1] = (uint8_t) vout0x1;
117 c0[2] = (uint8_t) vout0x2;
118 c0[3] = (uint8_t) vout0x3;
119
120 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
121
122 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
123 nc -= 4;
124 } else {
125 if (nc & 2) {
126 c0[0] = (uint8_t) vout0x0;
127 c0[1] = (uint8_t) vout0x1;
128 vout0x0 = vout0x2;
129 c0 += 2;
130 }
131 if (nc & 1) {
132 c0[0] = (uint8_t) vout0x0;
133 }
134
135 nc = 0;
136 }
137 } while (nc != 0);
138 }
139