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_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf(
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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 2);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (2 * 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
47 do {
48 int32_t vacc0x0 = ((const int32_t*) w)[0];
49 int32_t vacc0x1 = ((const int32_t*) w)[1];
50 int32_t vacc0x2 = ((const int32_t*) w)[2];
51 int32_t vacc0x3 = ((const int32_t*) w)[3];
52 int32_t vacc1x0 = vacc0x0;
53 int32_t vacc1x1 = vacc0x1;
54 int32_t vacc1x2 = vacc0x2;
55 int32_t vacc1x3 = vacc0x3;
56 w = (const void*) ((const int32_t*) w + 4);
57
58 size_t p = ks;
59 do {
60 const int8_t* restrict a0 = a[0];
61 assert(a0 != NULL);
62 if XNN_UNPREDICTABLE(a0 != zero) {
63 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
64 }
65 const int8_t* restrict a1 = a[1];
66 assert(a1 != NULL);
67 if XNN_UNPREDICTABLE(a1 != zero) {
68 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
69 }
70 a += 2;
71
72 size_t k = kc;
73 do {
74 const int32_t va0 = (int32_t) *a0++;
75 const int32_t va1 = (int32_t) *a1++;
76
77 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
78 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
79 const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
80 const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
81 w = (const void*) ((const int8_t*) w + 4);
82
83 vacc0x0 += va0 * vb0;
84 vacc0x1 += va0 * vb1;
85 vacc0x2 += va0 * vb2;
86 vacc0x3 += va0 * vb3;
87 vacc1x0 += va1 * vb0;
88 vacc1x1 += va1 * vb1;
89 vacc1x2 += va1 * vb2;
90 vacc1x3 += va1 * vb3;
91
92 k -= sizeof(int8_t);
93 } while (k != 0);
94 p -= 2 * sizeof(void*);
95 } while (p != 0);
96
97 float vfpacc0x0 = (float) vacc0x0;
98 float vfpacc0x1 = (float) vacc0x1;
99 float vfpacc0x2 = (float) vacc0x2;
100 float vfpacc0x3 = (float) vacc0x3;
101 float vfpacc1x0 = (float) vacc1x0;
102 float vfpacc1x1 = (float) vacc1x1;
103 float vfpacc1x2 = (float) vacc1x2;
104 float vfpacc1x3 = (float) vacc1x3;
105
106 const float vscale0 = ((const float*) w)[0];
107 vfpacc0x0 *= vscale0;
108 vfpacc1x0 *= vscale0;
109 const float vscale1 = ((const float*) w)[1];
110 vfpacc0x1 *= vscale1;
111 vfpacc1x1 *= vscale1;
112 const float vscale2 = ((const float*) w)[2];
113 vfpacc0x2 *= vscale2;
114 vfpacc1x2 *= vscale2;
115 const float vscale3 = ((const float*) w)[3];
116 vfpacc0x3 *= vscale3;
117 vfpacc1x3 *= vscale3;
118 w = (const void*) ((const float*) w + 4);
119
120 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
121 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
122 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
123 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
124 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
125 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
126 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
127 vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
128 vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
129
130 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
131 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
132 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
133 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
134 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
135 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
136 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
137 vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
138 vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
139
140 const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
141 const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
142 const int32_t vrndacc0x2 = (int32_t) lrintf(vfpacc0x2);
143 const int32_t vrndacc0x3 = (int32_t) lrintf(vfpacc0x3);
144 const int32_t vrndacc1x0 = (int32_t) lrintf(vfpacc1x0);
145 const int32_t vrndacc1x1 = (int32_t) lrintf(vfpacc1x1);
146 const int32_t vrndacc1x2 = (int32_t) lrintf(vfpacc1x2);
147 const int32_t vrndacc1x3 = (int32_t) lrintf(vfpacc1x3);
148
149 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
150 int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
151 int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
152 int32_t vout0x2 = vrndacc0x2 + voutput_zero_point;
153 int32_t vout0x3 = vrndacc0x3 + voutput_zero_point;
154 int32_t vout1x0 = vrndacc1x0 + voutput_zero_point;
155 int32_t vout1x1 = vrndacc1x1 + voutput_zero_point;
156 int32_t vout1x2 = vrndacc1x2 + voutput_zero_point;
157 int32_t vout1x3 = vrndacc1x3 + voutput_zero_point;
158
159 if XNN_LIKELY(nc >= 4) {
160 c1[0] = (int8_t) vout1x0;
161 c1[1] = (int8_t) vout1x1;
162 c1[2] = (int8_t) vout1x2;
163 c1[3] = (int8_t) vout1x3;
164 c0[0] = (int8_t) vout0x0;
165 c0[1] = (int8_t) vout0x1;
166 c0[2] = (int8_t) vout0x2;
167 c0[3] = (int8_t) vout0x3;
168
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 -= 4;
174 } else {
175 if (nc & 2) {
176 c1[0] = (int8_t) vout1x0;
177 c1[1] = (int8_t) vout1x1;
178 vout1x0 = vout1x2;
179 c1 += 2;
180 c0[0] = (int8_t) vout0x0;
181 c0[1] = (int8_t) vout0x1;
182 vout0x0 = vout0x2;
183 c0 += 2;
184 }
185 if (nc & 1) {
186 c1[0] = (int8_t) vout1x0;
187 c0[0] = (int8_t) vout0x0;
188 }
189
190 nc = 0;
191 }
192 } while (nc != 0);
193 }
194