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 #include <xnnpack/unaligned.h>
16
17
xnn_qc8_gemm_minmax_fp32_ukernel_4x2__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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
32 assert(nc != 0);
33 assert(kc != 0);
34
35 const int8_t* a0 = a;
36 int8_t* c0 = c;
37 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
38 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
39 if XNN_UNPREDICTABLE(mr < 2) {
40 a1 = a0;
41 c1 = c0;
42 }
43 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
44 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
45 if XNN_UNPREDICTABLE(mr <= 2) {
46 a2 = a1;
47 c2 = c1;
48 }
49 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
50 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
51 if XNN_UNPREDICTABLE(mr != 4) {
52 a3 = a2;
53 c3 = c2;
54 }
55
56 do {
57 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
58 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
59 int32_t vacc1x0 = vacc0x0;
60 int32_t vacc1x1 = vacc0x1;
61 int32_t vacc2x0 = vacc0x0;
62 int32_t vacc2x1 = vacc0x1;
63 int32_t vacc3x0 = vacc0x0;
64 int32_t vacc3x1 = vacc0x1;
65 w = (const void*) ((const int32_t*) w + 2);
66
67 size_t k = kc;
68 do {
69 const int32_t va0 = (int32_t) *a0++;
70 const int32_t va1 = (int32_t) *a1++;
71 const int32_t va2 = (int32_t) *a2++;
72 const int32_t va3 = (int32_t) *a3++;
73
74 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
75 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
76 w = (const void*) ((const int8_t*) w + 2);
77
78 vacc0x0 += va0 * vb0;
79 vacc0x1 += va0 * vb1;
80 vacc1x0 += va1 * vb0;
81 vacc1x1 += va1 * vb1;
82 vacc2x0 += va2 * vb0;
83 vacc2x1 += va2 * vb1;
84 vacc3x0 += va3 * vb0;
85 vacc3x1 += va3 * vb1;
86
87 k -= sizeof(int8_t);
88 } while (k != 0);
89
90 float vfpacc0x0 = (float) vacc0x0;
91 float vfpacc0x1 = (float) vacc0x1;
92 float vfpacc1x0 = (float) vacc1x0;
93 float vfpacc1x1 = (float) vacc1x1;
94 float vfpacc2x0 = (float) vacc2x0;
95 float vfpacc2x1 = (float) vacc2x1;
96 float vfpacc3x0 = (float) vacc3x0;
97 float vfpacc3x1 = (float) vacc3x1;
98
99 const float vscale0 = unaligned_indexed_load_f32(w, 0);
100 vfpacc0x0 *= vscale0;
101 vfpacc1x0 *= vscale0;
102 vfpacc2x0 *= vscale0;
103 vfpacc3x0 *= vscale0;
104 const float vscale1 = unaligned_indexed_load_f32(w, 1);
105 vfpacc0x1 *= vscale1;
106 vfpacc1x1 *= vscale1;
107 vfpacc2x1 *= vscale1;
108 vfpacc3x1 *= vscale1;
109 w = (const void*) ((const float*) w + 2);
110
111 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
112 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
113 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
114 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
115 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
116 vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
117 vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
118 vfpacc3x0 = math_max_f32(vfpacc3x0, voutput_min_less_zero_point);
119 vfpacc3x1 = math_max_f32(vfpacc3x1, voutput_min_less_zero_point);
120
121 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
122 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
123 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
124 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
125 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
126 vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
127 vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
128 vfpacc3x0 = math_min_f32(vfpacc3x0, voutput_max_less_zero_point);
129 vfpacc3x1 = math_min_f32(vfpacc3x1, voutput_max_less_zero_point);
130
131 const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
132 const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
133 const int32_t vrndacc1x0 = (int32_t) lrintf(vfpacc1x0);
134 const int32_t vrndacc1x1 = (int32_t) lrintf(vfpacc1x1);
135 const int32_t vrndacc2x0 = (int32_t) lrintf(vfpacc2x0);
136 const int32_t vrndacc2x1 = (int32_t) lrintf(vfpacc2x1);
137 const int32_t vrndacc3x0 = (int32_t) lrintf(vfpacc3x0);
138 const int32_t vrndacc3x1 = (int32_t) lrintf(vfpacc3x1);
139
140 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
141 int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
142 int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
143 int32_t vout1x0 = vrndacc1x0 + voutput_zero_point;
144 int32_t vout1x1 = vrndacc1x1 + voutput_zero_point;
145 int32_t vout2x0 = vrndacc2x0 + voutput_zero_point;
146 int32_t vout2x1 = vrndacc2x1 + voutput_zero_point;
147 int32_t vout3x0 = vrndacc3x0 + voutput_zero_point;
148 int32_t vout3x1 = vrndacc3x1 + voutput_zero_point;
149
150 if XNN_LIKELY(nc >= 2) {
151 c0[0] = (int8_t) vout0x0;
152 c0[1] = (int8_t) vout0x1;
153 c1[0] = (int8_t) vout1x0;
154 c1[1] = (int8_t) vout1x1;
155 c2[0] = (int8_t) vout2x0;
156 c2[1] = (int8_t) vout2x1;
157 c3[0] = (int8_t) vout3x0;
158 c3[1] = (int8_t) vout3x1;
159
160 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
161 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
162 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
163 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
164
165 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
166 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
167 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
168 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
169
170 nc -= 2;
171 } else {
172 if (nc & 1) {
173 c0[0] = (int8_t) vout0x0;
174 c1[0] = (int8_t) vout1x0;
175 c2[0] = (int8_t) vout2x0;
176 c3[0] = (int8_t) vout3x0;
177 }
178
179 nc = 0;
180 }
181 } while (nc != 0);
182 }
183