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_qc8_gemm_minmax_fp32_ukernel_4x4__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)])17 void xnn_qc8_gemm_minmax_fp32_ukernel_4x4__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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 4);
31 assert(nc != 0);
32 assert(kc != 0);
33
34 const int8_t* a0 = a;
35 int8_t* c0 = c;
36 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
37 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
38 if XNN_UNPREDICTABLE(mr < 2) {
39 a1 = a0;
40 c1 = c0;
41 }
42 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
43 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
44 if XNN_UNPREDICTABLE(mr <= 2) {
45 a2 = a1;
46 c2 = c1;
47 }
48 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
49 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
50 if XNN_UNPREDICTABLE(mr != 4) {
51 a3 = a2;
52 c3 = c2;
53 }
54
55 do {
56 int32_t vacc0x0 = ((const int32_t*) w)[0];
57 int32_t vacc0x1 = ((const int32_t*) w)[1];
58 int32_t vacc0x2 = ((const int32_t*) w)[2];
59 int32_t vacc0x3 = ((const int32_t*) w)[3];
60 int32_t vacc1x0 = vacc0x0;
61 int32_t vacc1x1 = vacc0x1;
62 int32_t vacc1x2 = vacc0x2;
63 int32_t vacc1x3 = vacc0x3;
64 int32_t vacc2x0 = vacc0x0;
65 int32_t vacc2x1 = vacc0x1;
66 int32_t vacc2x2 = vacc0x2;
67 int32_t vacc2x3 = vacc0x3;
68 int32_t vacc3x0 = vacc0x0;
69 int32_t vacc3x1 = vacc0x1;
70 int32_t vacc3x2 = vacc0x2;
71 int32_t vacc3x3 = vacc0x3;
72 w = (const void*) ((const int32_t*) w + 4);
73
74 size_t k = kc;
75 do {
76 const int32_t va0 = (int32_t) *a0++;
77 const int32_t va1 = (int32_t) *a1++;
78 const int32_t va2 = (int32_t) *a2++;
79 const int32_t va3 = (int32_t) *a3++;
80
81 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
82 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
83 const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
84 const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
85 w = (const void*) ((const int8_t*) w + 4);
86
87 vacc0x0 += va0 * vb0;
88 vacc0x1 += va0 * vb1;
89 vacc0x2 += va0 * vb2;
90 vacc0x3 += va0 * vb3;
91 vacc1x0 += va1 * vb0;
92 vacc1x1 += va1 * vb1;
93 vacc1x2 += va1 * vb2;
94 vacc1x3 += va1 * vb3;
95 vacc2x0 += va2 * vb0;
96 vacc2x1 += va2 * vb1;
97 vacc2x2 += va2 * vb2;
98 vacc2x3 += va2 * vb3;
99 vacc3x0 += va3 * vb0;
100 vacc3x1 += va3 * vb1;
101 vacc3x2 += va3 * vb2;
102 vacc3x3 += va3 * vb3;
103
104 k -= sizeof(int8_t);
105 } while (k != 0);
106
107 float vfpacc0x0 = (float) vacc0x0;
108 float vfpacc0x1 = (float) vacc0x1;
109 float vfpacc0x2 = (float) vacc0x2;
110 float vfpacc0x3 = (float) vacc0x3;
111 float vfpacc1x0 = (float) vacc1x0;
112 float vfpacc1x1 = (float) vacc1x1;
113 float vfpacc1x2 = (float) vacc1x2;
114 float vfpacc1x3 = (float) vacc1x3;
115 float vfpacc2x0 = (float) vacc2x0;
116 float vfpacc2x1 = (float) vacc2x1;
117 float vfpacc2x2 = (float) vacc2x2;
118 float vfpacc2x3 = (float) vacc2x3;
119 float vfpacc3x0 = (float) vacc3x0;
120 float vfpacc3x1 = (float) vacc3x1;
121 float vfpacc3x2 = (float) vacc3x2;
122 float vfpacc3x3 = (float) vacc3x3;
123
124 const float vscale0 = ((const float*) w)[0];
125 vfpacc0x0 *= vscale0;
126 vfpacc1x0 *= vscale0;
127 vfpacc2x0 *= vscale0;
128 vfpacc3x0 *= vscale0;
129 const float vscale1 = ((const float*) w)[1];
130 vfpacc0x1 *= vscale1;
131 vfpacc1x1 *= vscale1;
132 vfpacc2x1 *= vscale1;
133 vfpacc3x1 *= vscale1;
134 const float vscale2 = ((const float*) w)[2];
135 vfpacc0x2 *= vscale2;
136 vfpacc1x2 *= vscale2;
137 vfpacc2x2 *= vscale2;
138 vfpacc3x2 *= vscale2;
139 const float vscale3 = ((const float*) w)[3];
140 vfpacc0x3 *= vscale3;
141 vfpacc1x3 *= vscale3;
142 vfpacc2x3 *= vscale3;
143 vfpacc3x3 *= vscale3;
144 w = (const void*) ((const float*) w + 4);
145
146 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
147 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
148 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
149 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
150 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
151 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
152 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
153 vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
154 vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
155 vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
156 vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
157 vfpacc2x2 = math_max_f32(vfpacc2x2, voutput_min_less_zero_point);
158 vfpacc2x3 = math_max_f32(vfpacc2x3, voutput_min_less_zero_point);
159 vfpacc3x0 = math_max_f32(vfpacc3x0, voutput_min_less_zero_point);
160 vfpacc3x1 = math_max_f32(vfpacc3x1, voutput_min_less_zero_point);
161 vfpacc3x2 = math_max_f32(vfpacc3x2, voutput_min_less_zero_point);
162 vfpacc3x3 = math_max_f32(vfpacc3x3, voutput_min_less_zero_point);
163
164 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
165 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
166 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
167 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
168 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
169 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
170 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
171 vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
172 vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
173 vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
174 vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
175 vfpacc2x2 = math_min_f32(vfpacc2x2, voutput_max_less_zero_point);
176 vfpacc2x3 = math_min_f32(vfpacc2x3, voutput_max_less_zero_point);
177 vfpacc3x0 = math_min_f32(vfpacc3x0, voutput_max_less_zero_point);
178 vfpacc3x1 = math_min_f32(vfpacc3x1, voutput_max_less_zero_point);
179 vfpacc3x2 = math_min_f32(vfpacc3x2, voutput_max_less_zero_point);
180 vfpacc3x3 = math_min_f32(vfpacc3x3, voutput_max_less_zero_point);
181
182 const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
183 const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
184 const int32_t vrndacc0x2 = (int32_t) lrintf(vfpacc0x2);
185 const int32_t vrndacc0x3 = (int32_t) lrintf(vfpacc0x3);
186 const int32_t vrndacc1x0 = (int32_t) lrintf(vfpacc1x0);
187 const int32_t vrndacc1x1 = (int32_t) lrintf(vfpacc1x1);
188 const int32_t vrndacc1x2 = (int32_t) lrintf(vfpacc1x2);
189 const int32_t vrndacc1x3 = (int32_t) lrintf(vfpacc1x3);
190 const int32_t vrndacc2x0 = (int32_t) lrintf(vfpacc2x0);
191 const int32_t vrndacc2x1 = (int32_t) lrintf(vfpacc2x1);
192 const int32_t vrndacc2x2 = (int32_t) lrintf(vfpacc2x2);
193 const int32_t vrndacc2x3 = (int32_t) lrintf(vfpacc2x3);
194 const int32_t vrndacc3x0 = (int32_t) lrintf(vfpacc3x0);
195 const int32_t vrndacc3x1 = (int32_t) lrintf(vfpacc3x1);
196 const int32_t vrndacc3x2 = (int32_t) lrintf(vfpacc3x2);
197 const int32_t vrndacc3x3 = (int32_t) lrintf(vfpacc3x3);
198
199 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
200 int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
201 int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
202 int32_t vout0x2 = vrndacc0x2 + voutput_zero_point;
203 int32_t vout0x3 = vrndacc0x3 + voutput_zero_point;
204 int32_t vout1x0 = vrndacc1x0 + voutput_zero_point;
205 int32_t vout1x1 = vrndacc1x1 + voutput_zero_point;
206 int32_t vout1x2 = vrndacc1x2 + voutput_zero_point;
207 int32_t vout1x3 = vrndacc1x3 + voutput_zero_point;
208 int32_t vout2x0 = vrndacc2x0 + voutput_zero_point;
209 int32_t vout2x1 = vrndacc2x1 + voutput_zero_point;
210 int32_t vout2x2 = vrndacc2x2 + voutput_zero_point;
211 int32_t vout2x3 = vrndacc2x3 + voutput_zero_point;
212 int32_t vout3x0 = vrndacc3x0 + voutput_zero_point;
213 int32_t vout3x1 = vrndacc3x1 + voutput_zero_point;
214 int32_t vout3x2 = vrndacc3x2 + voutput_zero_point;
215 int32_t vout3x3 = vrndacc3x3 + voutput_zero_point;
216
217 if XNN_LIKELY(nc >= 4) {
218 c0[0] = (int8_t) vout0x0;
219 c0[1] = (int8_t) vout0x1;
220 c0[2] = (int8_t) vout0x2;
221 c0[3] = (int8_t) vout0x3;
222 c1[0] = (int8_t) vout1x0;
223 c1[1] = (int8_t) vout1x1;
224 c1[2] = (int8_t) vout1x2;
225 c1[3] = (int8_t) vout1x3;
226 c2[0] = (int8_t) vout2x0;
227 c2[1] = (int8_t) vout2x1;
228 c2[2] = (int8_t) vout2x2;
229 c2[3] = (int8_t) vout2x3;
230 c3[0] = (int8_t) vout3x0;
231 c3[1] = (int8_t) vout3x1;
232 c3[2] = (int8_t) vout3x2;
233 c3[3] = (int8_t) vout3x3;
234
235 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
236 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
237 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
238 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
239
240 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
241 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
242 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
243 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
244
245 nc -= 4;
246 } else {
247 if (nc & 2) {
248 c0[0] = (int8_t) vout0x0;
249 c0[1] = (int8_t) vout0x1;
250 vout0x0 = vout0x2;
251 c0 += 2;
252 c1[0] = (int8_t) vout1x0;
253 c1[1] = (int8_t) vout1x1;
254 vout1x0 = vout1x2;
255 c1 += 2;
256 c2[0] = (int8_t) vout2x0;
257 c2[1] = (int8_t) vout2x1;
258 vout2x0 = vout2x2;
259 c2 += 2;
260 c3[0] = (int8_t) vout3x0;
261 c3[1] = (int8_t) vout3x1;
262 vout3x0 = vout3x2;
263 c3 += 2;
264 }
265 if (nc & 1) {
266 c0[0] = (int8_t) vout0x0;
267 c1[0] = (int8_t) vout1x0;
268 c2[0] = (int8_t) vout2x0;
269 c3[0] = (int8_t) vout3x0;
270 }
271
272 nc = 0;
273 }
274 } while (nc != 0);
275 }
276