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 <fp16.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const uint8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
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
35 const uint8_t* a0 = a;
36 uint8_t* c0 = c;
37 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
38 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
39 if XNN_UNPREDICTABLE(mr < 2) {
40 a1 = a0;
41 c1 = c0;
42 }
43 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
44 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
45 if XNN_UNPREDICTABLE(mr <= 2) {
46 a2 = a1;
47 c2 = c1;
48 }
49 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
50 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
51 if XNN_UNPREDICTABLE(mr != 4) {
52 a3 = a2;
53 c3 = c2;
54 }
55
56 const int32_t vb_zero_point = params->fp32_scalar_imagic.kernel_zero_point;
57 do {
58 int32_t vacc0x0 = ((const int32_t*) w)[0];
59 int32_t vacc0x1 = ((const int32_t*) w)[1];
60 int32_t vacc0x2 = ((const int32_t*) w)[2];
61 int32_t vacc0x3 = ((const int32_t*) w)[3];
62 int32_t vacc1x0 = vacc0x0;
63 int32_t vacc1x1 = vacc0x1;
64 int32_t vacc1x2 = vacc0x2;
65 int32_t vacc1x3 = vacc0x3;
66 int32_t vacc2x0 = vacc0x0;
67 int32_t vacc2x1 = vacc0x1;
68 int32_t vacc2x2 = vacc0x2;
69 int32_t vacc2x3 = vacc0x3;
70 int32_t vacc3x0 = vacc0x0;
71 int32_t vacc3x1 = vacc0x1;
72 int32_t vacc3x2 = vacc0x2;
73 int32_t vacc3x3 = vacc0x3;
74 w = (const void*) ((const int32_t*) w + 4);
75
76 size_t k = kc;
77 do {
78 const int32_t va0 = (int32_t) (uint32_t) *a0++;
79 const int32_t va1 = (int32_t) (uint32_t) *a1++;
80 const int32_t va2 = (int32_t) (uint32_t) *a2++;
81 const int32_t va3 = (int32_t) (uint32_t) *a3++;
82
83 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
84 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
85 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
86 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
87 w = (const void*) ((const uint8_t*) w + 4);
88
89 vacc0x0 += va0 * vb0;
90 vacc0x1 += va0 * vb1;
91 vacc0x2 += va0 * vb2;
92 vacc0x3 += va0 * vb3;
93 vacc1x0 += va1 * vb0;
94 vacc1x1 += va1 * vb1;
95 vacc1x2 += va1 * vb2;
96 vacc1x3 += va1 * vb3;
97 vacc2x0 += va2 * vb0;
98 vacc2x1 += va2 * vb1;
99 vacc2x2 += va2 * vb2;
100 vacc2x3 += va2 * vb3;
101 vacc3x0 += va3 * vb0;
102 vacc3x1 += va3 * vb1;
103 vacc3x2 += va3 * vb2;
104 vacc3x3 += va3 * vb3;
105
106 k -= sizeof(uint8_t);
107 } while (k != 0);
108
109 float vfpacc0x0 = (float) vacc0x0;
110 float vfpacc0x1 = (float) vacc0x1;
111 float vfpacc0x2 = (float) vacc0x2;
112 float vfpacc0x3 = (float) vacc0x3;
113 float vfpacc1x0 = (float) vacc1x0;
114 float vfpacc1x1 = (float) vacc1x1;
115 float vfpacc1x2 = (float) vacc1x2;
116 float vfpacc1x3 = (float) vacc1x3;
117 float vfpacc2x0 = (float) vacc2x0;
118 float vfpacc2x1 = (float) vacc2x1;
119 float vfpacc2x2 = (float) vacc2x2;
120 float vfpacc2x3 = (float) vacc2x3;
121 float vfpacc3x0 = (float) vacc3x0;
122 float vfpacc3x1 = (float) vacc3x1;
123 float vfpacc3x2 = (float) vacc3x2;
124 float vfpacc3x3 = (float) vacc3x3;
125
126 const float vscale = params->fp32_scalar_imagic.scale;
127 vfpacc0x0 *= vscale;
128 vfpacc0x1 *= vscale;
129 vfpacc0x2 *= vscale;
130 vfpacc0x3 *= vscale;
131 vfpacc1x0 *= vscale;
132 vfpacc1x1 *= vscale;
133 vfpacc1x2 *= vscale;
134 vfpacc1x3 *= vscale;
135 vfpacc2x0 *= vscale;
136 vfpacc2x1 *= vscale;
137 vfpacc2x2 *= vscale;
138 vfpacc2x3 *= vscale;
139 vfpacc3x0 *= vscale;
140 vfpacc3x1 *= vscale;
141 vfpacc3x2 *= vscale;
142 vfpacc3x3 *= vscale;
143
144 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
145 vfpacc0x0 += vmagic_bias;
146 vfpacc0x1 += vmagic_bias;
147 vfpacc0x2 += vmagic_bias;
148 vfpacc0x3 += vmagic_bias;
149 vfpacc1x0 += vmagic_bias;
150 vfpacc1x1 += vmagic_bias;
151 vfpacc1x2 += vmagic_bias;
152 vfpacc1x3 += vmagic_bias;
153 vfpacc2x0 += vmagic_bias;
154 vfpacc2x1 += vmagic_bias;
155 vfpacc2x2 += vmagic_bias;
156 vfpacc2x3 += vmagic_bias;
157 vfpacc3x0 += vmagic_bias;
158 vfpacc3x1 += vmagic_bias;
159 vfpacc3x2 += vmagic_bias;
160 vfpacc3x3 += vmagic_bias;
161
162 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0);
163 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1);
164 int32_t vout0x2 = (int32_t) fp32_to_bits(vfpacc0x2);
165 int32_t vout0x3 = (int32_t) fp32_to_bits(vfpacc0x3);
166 int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0);
167 int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1);
168 int32_t vout1x2 = (int32_t) fp32_to_bits(vfpacc1x2);
169 int32_t vout1x3 = (int32_t) fp32_to_bits(vfpacc1x3);
170 int32_t vout2x0 = (int32_t) fp32_to_bits(vfpacc2x0);
171 int32_t vout2x1 = (int32_t) fp32_to_bits(vfpacc2x1);
172 int32_t vout2x2 = (int32_t) fp32_to_bits(vfpacc2x2);
173 int32_t vout2x3 = (int32_t) fp32_to_bits(vfpacc2x3);
174 int32_t vout3x0 = (int32_t) fp32_to_bits(vfpacc3x0);
175 int32_t vout3x1 = (int32_t) fp32_to_bits(vfpacc3x1);
176 int32_t vout3x2 = (int32_t) fp32_to_bits(vfpacc3x2);
177 int32_t vout3x3 = (int32_t) fp32_to_bits(vfpacc3x3);
178
179 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
180 vout0x0 = math_max_s32(vout0x0, vmagic_min);
181 vout0x1 = math_max_s32(vout0x1, vmagic_min);
182 vout0x2 = math_max_s32(vout0x2, vmagic_min);
183 vout0x3 = math_max_s32(vout0x3, vmagic_min);
184 vout1x0 = math_max_s32(vout1x0, vmagic_min);
185 vout1x1 = math_max_s32(vout1x1, vmagic_min);
186 vout1x2 = math_max_s32(vout1x2, vmagic_min);
187 vout1x3 = math_max_s32(vout1x3, vmagic_min);
188 vout2x0 = math_max_s32(vout2x0, vmagic_min);
189 vout2x1 = math_max_s32(vout2x1, vmagic_min);
190 vout2x2 = math_max_s32(vout2x2, vmagic_min);
191 vout2x3 = math_max_s32(vout2x3, vmagic_min);
192 vout3x0 = math_max_s32(vout3x0, vmagic_min);
193 vout3x1 = math_max_s32(vout3x1, vmagic_min);
194 vout3x2 = math_max_s32(vout3x2, vmagic_min);
195 vout3x3 = math_max_s32(vout3x3, vmagic_min);
196
197 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
198 vout0x0 = math_min_s32(vout0x0, vmagic_max);
199 vout0x1 = math_min_s32(vout0x1, vmagic_max);
200 vout0x2 = math_min_s32(vout0x2, vmagic_max);
201 vout0x3 = math_min_s32(vout0x3, vmagic_max);
202 vout1x0 = math_min_s32(vout1x0, vmagic_max);
203 vout1x1 = math_min_s32(vout1x1, vmagic_max);
204 vout1x2 = math_min_s32(vout1x2, vmagic_max);
205 vout1x3 = math_min_s32(vout1x3, vmagic_max);
206 vout2x0 = math_min_s32(vout2x0, vmagic_max);
207 vout2x1 = math_min_s32(vout2x1, vmagic_max);
208 vout2x2 = math_min_s32(vout2x2, vmagic_max);
209 vout2x3 = math_min_s32(vout2x3, vmagic_max);
210 vout3x0 = math_min_s32(vout3x0, vmagic_max);
211 vout3x1 = math_min_s32(vout3x1, vmagic_max);
212 vout3x2 = math_min_s32(vout3x2, vmagic_max);
213 vout3x3 = math_min_s32(vout3x3, vmagic_max);
214
215 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
216 vout0x0 -= vmagic_bias_less_zero_point;
217 vout0x1 -= vmagic_bias_less_zero_point;
218 vout0x2 -= vmagic_bias_less_zero_point;
219 vout0x3 -= vmagic_bias_less_zero_point;
220 vout1x0 -= vmagic_bias_less_zero_point;
221 vout1x1 -= vmagic_bias_less_zero_point;
222 vout1x2 -= vmagic_bias_less_zero_point;
223 vout1x3 -= vmagic_bias_less_zero_point;
224 vout2x0 -= vmagic_bias_less_zero_point;
225 vout2x1 -= vmagic_bias_less_zero_point;
226 vout2x2 -= vmagic_bias_less_zero_point;
227 vout2x3 -= vmagic_bias_less_zero_point;
228 vout3x0 -= vmagic_bias_less_zero_point;
229 vout3x1 -= vmagic_bias_less_zero_point;
230 vout3x2 -= vmagic_bias_less_zero_point;
231 vout3x3 -= vmagic_bias_less_zero_point;
232
233 if XNN_LIKELY(nc >= 4) {
234 c0[0] = (uint8_t) vout0x0;
235 c0[1] = (uint8_t) vout0x1;
236 c0[2] = (uint8_t) vout0x2;
237 c0[3] = (uint8_t) vout0x3;
238 c1[0] = (uint8_t) vout1x0;
239 c1[1] = (uint8_t) vout1x1;
240 c1[2] = (uint8_t) vout1x2;
241 c1[3] = (uint8_t) vout1x3;
242 c2[0] = (uint8_t) vout2x0;
243 c2[1] = (uint8_t) vout2x1;
244 c2[2] = (uint8_t) vout2x2;
245 c2[3] = (uint8_t) vout2x3;
246 c3[0] = (uint8_t) vout3x0;
247 c3[1] = (uint8_t) vout3x1;
248 c3[2] = (uint8_t) vout3x2;
249 c3[3] = (uint8_t) vout3x3;
250
251 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
252 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
253 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
254 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
255
256 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
257 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
258 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
259 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
260
261 nc -= 4;
262 } else {
263 if (nc & 2) {
264 c0[0] = (uint8_t) vout0x0;
265 c0[1] = (uint8_t) vout0x1;
266 vout0x0 = vout0x2;
267 c0 += 2;
268 c1[0] = (uint8_t) vout1x0;
269 c1[1] = (uint8_t) vout1x1;
270 vout1x0 = vout1x2;
271 c1 += 2;
272 c2[0] = (uint8_t) vout2x0;
273 c2[1] = (uint8_t) vout2x1;
274 vout2x0 = vout2x2;
275 c2 += 2;
276 c3[0] = (uint8_t) vout3x0;
277 c3[1] = (uint8_t) vout3x1;
278 vout3x0 = vout3x2;
279 c3 += 2;
280 }
281 if (nc & 1) {
282 c0[0] = (uint8_t) vout0x0;
283 c1[0] = (uint8_t) vout1x0;
284 c2[0] = (uint8_t) vout2x0;
285 c3[0] = (uint8_t) vout3x0;
286 }
287
288 nc = 0;
289 }
290 } while (nc != 0);
291 }
292