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
12 #include <xnnpack/math.h>
13 #include <xnnpack/gemm.h>
14
15
xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_fmagic(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)])16 void xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_fmagic(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 size_t ks,
21 const uint8_t**restrict a,
22 const void*restrict w,
23 uint8_t*restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 size_t a_offset,
27 const uint8_t* zero,
28 const union xnn_qu8_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 assert(ks != 0);
35 assert(ks % (4 * sizeof(void*)) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 uint8_t* c0 = c;
41 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr < 2) {
43 c1 = c0;
44 }
45 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
46 if XNN_UNPREDICTABLE(mr <= 2) {
47 c2 = c1;
48 }
49 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
50 if XNN_UNPREDICTABLE(mr != 4) {
51 c3 = c2;
52 }
53
54 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
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 p = ks;
75 do {
76 const uint8_t* restrict a0 = a[0];
77 assert(a0 != NULL);
78 if XNN_UNPREDICTABLE(a0 != zero) {
79 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
80 }
81 const uint8_t* restrict a1 = a[1];
82 assert(a1 != NULL);
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
85 }
86 const uint8_t* restrict a2 = a[2];
87 assert(a2 != NULL);
88 if XNN_UNPREDICTABLE(a2 != zero) {
89 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
90 }
91 const uint8_t* restrict a3 = a[3];
92 assert(a3 != NULL);
93 if XNN_UNPREDICTABLE(a3 != zero) {
94 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
95 }
96 a += 4;
97
98 size_t k = kc;
99 do {
100 const int32_t va0 = (int32_t) (uint32_t) *a0++;
101 const int32_t va1 = (int32_t) (uint32_t) *a1++;
102 const int32_t va2 = (int32_t) (uint32_t) *a2++;
103 const int32_t va3 = (int32_t) (uint32_t) *a3++;
104
105 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
106 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
107 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
108 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
109 w = (const void*) ((const uint8_t*) w + 4);
110
111 vacc0x0 += va0 * vb0;
112 vacc0x1 += va0 * vb1;
113 vacc0x2 += va0 * vb2;
114 vacc0x3 += va0 * vb3;
115 vacc1x0 += va1 * vb0;
116 vacc1x1 += va1 * vb1;
117 vacc1x2 += va1 * vb2;
118 vacc1x3 += va1 * vb3;
119 vacc2x0 += va2 * vb0;
120 vacc2x1 += va2 * vb1;
121 vacc2x2 += va2 * vb2;
122 vacc2x3 += va2 * vb3;
123 vacc3x0 += va3 * vb0;
124 vacc3x1 += va3 * vb1;
125 vacc3x2 += va3 * vb2;
126 vacc3x3 += va3 * vb3;
127
128 k -= sizeof(uint8_t);
129 } while (k != 0);
130 p -= 4 * sizeof(void*);
131 } while (p != 0);
132
133 float vfpacc0x0 = (float) vacc0x0;
134 float vfpacc0x1 = (float) vacc0x1;
135 float vfpacc0x2 = (float) vacc0x2;
136 float vfpacc0x3 = (float) vacc0x3;
137 float vfpacc1x0 = (float) vacc1x0;
138 float vfpacc1x1 = (float) vacc1x1;
139 float vfpacc1x2 = (float) vacc1x2;
140 float vfpacc1x3 = (float) vacc1x3;
141 float vfpacc2x0 = (float) vacc2x0;
142 float vfpacc2x1 = (float) vacc2x1;
143 float vfpacc2x2 = (float) vacc2x2;
144 float vfpacc2x3 = (float) vacc2x3;
145 float vfpacc3x0 = (float) vacc3x0;
146 float vfpacc3x1 = (float) vacc3x1;
147 float vfpacc3x2 = (float) vacc3x2;
148 float vfpacc3x3 = (float) vacc3x3;
149
150 const float vscale = params->fp32_scalar_fmagic.scale;
151 vfpacc0x0 *= vscale;
152 vfpacc0x1 *= vscale;
153 vfpacc0x2 *= vscale;
154 vfpacc0x3 *= vscale;
155 vfpacc1x0 *= vscale;
156 vfpacc1x1 *= vscale;
157 vfpacc1x2 *= vscale;
158 vfpacc1x3 *= vscale;
159 vfpacc2x0 *= vscale;
160 vfpacc2x1 *= vscale;
161 vfpacc2x2 *= vscale;
162 vfpacc2x3 *= vscale;
163 vfpacc3x0 *= vscale;
164 vfpacc3x1 *= vscale;
165 vfpacc3x2 *= vscale;
166 vfpacc3x3 *= vscale;
167
168 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
169 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
170 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
171 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
172 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
173 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
174 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
175 vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
176 vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
177 vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
178 vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
179 vfpacc2x2 = math_max_f32(vfpacc2x2, voutput_min_less_zero_point);
180 vfpacc2x3 = math_max_f32(vfpacc2x3, voutput_min_less_zero_point);
181 vfpacc3x0 = math_max_f32(vfpacc3x0, voutput_min_less_zero_point);
182 vfpacc3x1 = math_max_f32(vfpacc3x1, voutput_min_less_zero_point);
183 vfpacc3x2 = math_max_f32(vfpacc3x2, voutput_min_less_zero_point);
184 vfpacc3x3 = math_max_f32(vfpacc3x3, voutput_min_less_zero_point);
185
186 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
187 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
188 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
189 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
190 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
191 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
192 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
193 vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
194 vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
195 vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
196 vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
197 vfpacc2x2 = math_min_f32(vfpacc2x2, voutput_max_less_zero_point);
198 vfpacc2x3 = math_min_f32(vfpacc2x3, voutput_max_less_zero_point);
199 vfpacc3x0 = math_min_f32(vfpacc3x0, voutput_max_less_zero_point);
200 vfpacc3x1 = math_min_f32(vfpacc3x1, voutput_max_less_zero_point);
201 vfpacc3x2 = math_min_f32(vfpacc3x2, voutput_max_less_zero_point);
202 vfpacc3x3 = math_min_f32(vfpacc3x3, voutput_max_less_zero_point);
203
204 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
205 vfpacc0x0 += vmagic_bias;
206 vfpacc0x1 += vmagic_bias;
207 vfpacc0x2 += vmagic_bias;
208 vfpacc0x3 += vmagic_bias;
209 vfpacc1x0 += vmagic_bias;
210 vfpacc1x1 += vmagic_bias;
211 vfpacc1x2 += vmagic_bias;
212 vfpacc1x3 += vmagic_bias;
213 vfpacc2x0 += vmagic_bias;
214 vfpacc2x1 += vmagic_bias;
215 vfpacc2x2 += vmagic_bias;
216 vfpacc2x3 += vmagic_bias;
217 vfpacc3x0 += vmagic_bias;
218 vfpacc3x1 += vmagic_bias;
219 vfpacc3x2 += vmagic_bias;
220 vfpacc3x3 += vmagic_bias;
221
222 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
223 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
224 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
225 int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
226 int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
227 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
228 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
229 int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2) - vmagic_bias_less_output_zero_point;
230 int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3) - vmagic_bias_less_output_zero_point;
231 int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0) - vmagic_bias_less_output_zero_point;
232 int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1) - vmagic_bias_less_output_zero_point;
233 int32_t vout2x2 = (int32_t) float_as_uint32(vfpacc2x2) - vmagic_bias_less_output_zero_point;
234 int32_t vout2x3 = (int32_t) float_as_uint32(vfpacc2x3) - vmagic_bias_less_output_zero_point;
235 int32_t vout3x0 = (int32_t) float_as_uint32(vfpacc3x0) - vmagic_bias_less_output_zero_point;
236 int32_t vout3x1 = (int32_t) float_as_uint32(vfpacc3x1) - vmagic_bias_less_output_zero_point;
237 int32_t vout3x2 = (int32_t) float_as_uint32(vfpacc3x2) - vmagic_bias_less_output_zero_point;
238 int32_t vout3x3 = (int32_t) float_as_uint32(vfpacc3x3) - vmagic_bias_less_output_zero_point;
239
240 if XNN_LIKELY(nc >= 4) {
241 c3[0] = (uint8_t) vout3x0;
242 c3[1] = (uint8_t) vout3x1;
243 c3[2] = (uint8_t) vout3x2;
244 c3[3] = (uint8_t) vout3x3;
245 c2[0] = (uint8_t) vout2x0;
246 c2[1] = (uint8_t) vout2x1;
247 c2[2] = (uint8_t) vout2x2;
248 c2[3] = (uint8_t) vout2x3;
249 c1[0] = (uint8_t) vout1x0;
250 c1[1] = (uint8_t) vout1x1;
251 c1[2] = (uint8_t) vout1x2;
252 c1[3] = (uint8_t) vout1x3;
253 c0[0] = (uint8_t) vout0x0;
254 c0[1] = (uint8_t) vout0x1;
255 c0[2] = (uint8_t) vout0x2;
256 c0[3] = (uint8_t) vout0x3;
257
258 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
259 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
260 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
261 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
262
263 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
264 nc -= 4;
265 } else {
266 if (nc & 2) {
267 c3[0] = (uint8_t) vout3x0;
268 c3[1] = (uint8_t) vout3x1;
269 vout3x0 = vout3x2;
270 c3 += 2;
271 c2[0] = (uint8_t) vout2x0;
272 c2[1] = (uint8_t) vout2x1;
273 vout2x0 = vout2x2;
274 c2 += 2;
275 c1[0] = (uint8_t) vout1x0;
276 c1[1] = (uint8_t) vout1x1;
277 vout1x0 = vout1x2;
278 c1 += 2;
279 c0[0] = (uint8_t) vout0x0;
280 c0[1] = (uint8_t) vout0x1;
281 vout0x0 = vout0x2;
282 c0 += 2;
283 }
284 if (nc & 1) {
285 c3[0] = (uint8_t) vout3x0;
286 c2[0] = (uint8_t) vout2x0;
287 c1[0] = (uint8_t) vout1x0;
288 c0[0] = (uint8_t) vout0x0;
289 }
290
291 nc = 0;
292 }
293 } while (nc != 0);
294 }
295