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
12 #include <xnnpack/math.h>
13 #include <xnnpack/gemm.h>
14
15
xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic(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)])16 void xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 const int8_t* restrict a,
21 size_t a_stride,
22 const void* restrict w,
23 int8_t* restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(mr != 0);
29 assert(mr <= 4);
30 assert(nc != 0);
31 assert(kc != 0);
32
33 const int8_t* a0 = a;
34 int8_t* c0 = c;
35 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
36 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
37 if XNN_UNPREDICTABLE(mr < 2) {
38 a1 = a0;
39 c1 = c0;
40 }
41 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
42 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
43 if XNN_UNPREDICTABLE(mr <= 2) {
44 a2 = a1;
45 c2 = c1;
46 }
47 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
48 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
49 if XNN_UNPREDICTABLE(mr != 4) {
50 a3 = a2;
51 c3 = c2;
52 }
53
54 do {
55 int32_t vacc0x0 = ((const int32_t*) w)[0];
56 int32_t vacc0x1 = ((const int32_t*) w)[1];
57 int32_t vacc0x2 = ((const int32_t*) w)[2];
58 int32_t vacc0x3 = ((const int32_t*) w)[3];
59 int32_t vacc1x0 = vacc0x0;
60 int32_t vacc1x1 = vacc0x1;
61 int32_t vacc1x2 = vacc0x2;
62 int32_t vacc1x3 = vacc0x3;
63 int32_t vacc2x0 = vacc0x0;
64 int32_t vacc2x1 = vacc0x1;
65 int32_t vacc2x2 = vacc0x2;
66 int32_t vacc2x3 = vacc0x3;
67 int32_t vacc3x0 = vacc0x0;
68 int32_t vacc3x1 = vacc0x1;
69 int32_t vacc3x2 = vacc0x2;
70 int32_t vacc3x3 = vacc0x3;
71 w = (const void*) ((const int32_t*) w + 4);
72
73 size_t k = kc;
74 do {
75 const int32_t va0 = (int32_t) *a0++;
76 const int32_t va1 = (int32_t) *a1++;
77 const int32_t va2 = (int32_t) *a2++;
78 const int32_t va3 = (int32_t) *a3++;
79
80 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
81 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
82 const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
83 const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
84 w = (const void*) ((const int8_t*) w + 4);
85
86 vacc0x0 += va0 * vb0;
87 vacc0x1 += va0 * vb1;
88 vacc0x2 += va0 * vb2;
89 vacc0x3 += va0 * vb3;
90 vacc1x0 += va1 * vb0;
91 vacc1x1 += va1 * vb1;
92 vacc1x2 += va1 * vb2;
93 vacc1x3 += va1 * vb3;
94 vacc2x0 += va2 * vb0;
95 vacc2x1 += va2 * vb1;
96 vacc2x2 += va2 * vb2;
97 vacc2x3 += va2 * vb3;
98 vacc3x0 += va3 * vb0;
99 vacc3x1 += va3 * vb1;
100 vacc3x2 += va3 * vb2;
101 vacc3x3 += va3 * vb3;
102
103 k -= sizeof(int8_t);
104 } while (k != 0);
105
106 float vfpacc0x0 = (float) vacc0x0;
107 float vfpacc0x1 = (float) vacc0x1;
108 float vfpacc0x2 = (float) vacc0x2;
109 float vfpacc0x3 = (float) vacc0x3;
110 float vfpacc1x0 = (float) vacc1x0;
111 float vfpacc1x1 = (float) vacc1x1;
112 float vfpacc1x2 = (float) vacc1x2;
113 float vfpacc1x3 = (float) vacc1x3;
114 float vfpacc2x0 = (float) vacc2x0;
115 float vfpacc2x1 = (float) vacc2x1;
116 float vfpacc2x2 = (float) vacc2x2;
117 float vfpacc2x3 = (float) vacc2x3;
118 float vfpacc3x0 = (float) vacc3x0;
119 float vfpacc3x1 = (float) vacc3x1;
120 float vfpacc3x2 = (float) vacc3x2;
121 float vfpacc3x3 = (float) vacc3x3;
122
123 const float vscale = params->fp32_scalar_fmagic.scale;
124 vfpacc0x0 *= vscale;
125 vfpacc0x1 *= vscale;
126 vfpacc0x2 *= vscale;
127 vfpacc0x3 *= vscale;
128 vfpacc1x0 *= vscale;
129 vfpacc1x1 *= vscale;
130 vfpacc1x2 *= vscale;
131 vfpacc1x3 *= vscale;
132 vfpacc2x0 *= vscale;
133 vfpacc2x1 *= vscale;
134 vfpacc2x2 *= vscale;
135 vfpacc2x3 *= vscale;
136 vfpacc3x0 *= vscale;
137 vfpacc3x1 *= vscale;
138 vfpacc3x2 *= vscale;
139 vfpacc3x3 *= vscale;
140
141 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
142 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
143 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
144 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
145 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
146 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
147 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
148 vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
149 vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
150 vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
151 vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
152 vfpacc2x2 = math_max_f32(vfpacc2x2, voutput_min_less_zero_point);
153 vfpacc2x3 = math_max_f32(vfpacc2x3, voutput_min_less_zero_point);
154 vfpacc3x0 = math_max_f32(vfpacc3x0, voutput_min_less_zero_point);
155 vfpacc3x1 = math_max_f32(vfpacc3x1, voutput_min_less_zero_point);
156 vfpacc3x2 = math_max_f32(vfpacc3x2, voutput_min_less_zero_point);
157 vfpacc3x3 = math_max_f32(vfpacc3x3, voutput_min_less_zero_point);
158
159 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
160 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
161 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
162 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
163 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
164 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
165 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
166 vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
167 vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
168 vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
169 vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
170 vfpacc2x2 = math_min_f32(vfpacc2x2, voutput_max_less_zero_point);
171 vfpacc2x3 = math_min_f32(vfpacc2x3, voutput_max_less_zero_point);
172 vfpacc3x0 = math_min_f32(vfpacc3x0, voutput_max_less_zero_point);
173 vfpacc3x1 = math_min_f32(vfpacc3x1, voutput_max_less_zero_point);
174 vfpacc3x2 = math_min_f32(vfpacc3x2, voutput_max_less_zero_point);
175 vfpacc3x3 = math_min_f32(vfpacc3x3, voutput_max_less_zero_point);
176
177 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
178 vfpacc0x0 += vmagic_bias;
179 vfpacc0x1 += vmagic_bias;
180 vfpacc0x2 += vmagic_bias;
181 vfpacc0x3 += vmagic_bias;
182 vfpacc1x0 += vmagic_bias;
183 vfpacc1x1 += vmagic_bias;
184 vfpacc1x2 += vmagic_bias;
185 vfpacc1x3 += vmagic_bias;
186 vfpacc2x0 += vmagic_bias;
187 vfpacc2x1 += vmagic_bias;
188 vfpacc2x2 += vmagic_bias;
189 vfpacc2x3 += vmagic_bias;
190 vfpacc3x0 += vmagic_bias;
191 vfpacc3x1 += vmagic_bias;
192 vfpacc3x2 += vmagic_bias;
193 vfpacc3x3 += vmagic_bias;
194
195 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
196 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
197 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
198 int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
199 int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
200 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
201 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
202 int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2) - vmagic_bias_less_output_zero_point;
203 int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3) - vmagic_bias_less_output_zero_point;
204 int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0) - vmagic_bias_less_output_zero_point;
205 int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1) - vmagic_bias_less_output_zero_point;
206 int32_t vout2x2 = (int32_t) float_as_uint32(vfpacc2x2) - vmagic_bias_less_output_zero_point;
207 int32_t vout2x3 = (int32_t) float_as_uint32(vfpacc2x3) - vmagic_bias_less_output_zero_point;
208 int32_t vout3x0 = (int32_t) float_as_uint32(vfpacc3x0) - vmagic_bias_less_output_zero_point;
209 int32_t vout3x1 = (int32_t) float_as_uint32(vfpacc3x1) - vmagic_bias_less_output_zero_point;
210 int32_t vout3x2 = (int32_t) float_as_uint32(vfpacc3x2) - vmagic_bias_less_output_zero_point;
211 int32_t vout3x3 = (int32_t) float_as_uint32(vfpacc3x3) - vmagic_bias_less_output_zero_point;
212
213 if XNN_LIKELY(nc >= 4) {
214 c0[0] = (int8_t) vout0x0;
215 c0[1] = (int8_t) vout0x1;
216 c0[2] = (int8_t) vout0x2;
217 c0[3] = (int8_t) vout0x3;
218 c1[0] = (int8_t) vout1x0;
219 c1[1] = (int8_t) vout1x1;
220 c1[2] = (int8_t) vout1x2;
221 c1[3] = (int8_t) vout1x3;
222 c2[0] = (int8_t) vout2x0;
223 c2[1] = (int8_t) vout2x1;
224 c2[2] = (int8_t) vout2x2;
225 c2[3] = (int8_t) vout2x3;
226 c3[0] = (int8_t) vout3x0;
227 c3[1] = (int8_t) vout3x1;
228 c3[2] = (int8_t) vout3x2;
229 c3[3] = (int8_t) vout3x3;
230
231 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
232 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
233 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
234 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
235
236 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
237 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
238 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
239 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
240
241 nc -= 4;
242 } else {
243 if (nc & 2) {
244 c0[0] = (int8_t) vout0x0;
245 c0[1] = (int8_t) vout0x1;
246 vout0x0 = vout0x2;
247 c0 += 2;
248 c1[0] = (int8_t) vout1x0;
249 c1[1] = (int8_t) vout1x1;
250 vout1x0 = vout1x2;
251 c1 += 2;
252 c2[0] = (int8_t) vout2x0;
253 c2[1] = (int8_t) vout2x1;
254 vout2x0 = vout2x2;
255 c2 += 2;
256 c3[0] = (int8_t) vout3x0;
257 c3[1] = (int8_t) vout3x1;
258 vout3x0 = vout3x2;
259 c3 += 2;
260 }
261 if (nc & 1) {
262 c0[0] = (int8_t) vout0x0;
263 c1[0] = (int8_t) vout1x0;
264 c2[0] = (int8_t) vout2x0;
265 c3[0] = (int8_t) vout3x0;
266 }
267
268 nc = 0;
269 }
270 } while (nc != 0);
271 }
272