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_3x4__scalar_imagic(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_3x4__scalar_imagic(
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 <= 3);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(ks != 0);
35 assert(ks % (3 * 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
50 const int32_t vb_zero_point = params->fp32_scalar_imagic.kernel_zero_point;
51 do {
52 int32_t vacc0x0 = ((const int32_t*) w)[0];
53 int32_t vacc0x1 = ((const int32_t*) w)[1];
54 int32_t vacc0x2 = ((const int32_t*) w)[2];
55 int32_t vacc0x3 = ((const int32_t*) w)[3];
56 int32_t vacc1x0 = vacc0x0;
57 int32_t vacc1x1 = vacc0x1;
58 int32_t vacc1x2 = vacc0x2;
59 int32_t vacc1x3 = vacc0x3;
60 int32_t vacc2x0 = vacc0x0;
61 int32_t vacc2x1 = vacc0x1;
62 int32_t vacc2x2 = vacc0x2;
63 int32_t vacc2x3 = vacc0x3;
64 w = (const void*) ((const int32_t*) w + 4);
65
66 size_t p = ks;
67 do {
68 const uint8_t* restrict a0 = a[0];
69 assert(a0 != NULL);
70 if XNN_UNPREDICTABLE(a0 != zero) {
71 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
72 }
73 const uint8_t* restrict a1 = a[1];
74 assert(a1 != NULL);
75 if XNN_UNPREDICTABLE(a1 != zero) {
76 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
77 }
78 const uint8_t* restrict a2 = a[2];
79 assert(a2 != NULL);
80 if XNN_UNPREDICTABLE(a2 != zero) {
81 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
82 }
83 a += 3;
84
85 size_t k = kc;
86 do {
87 const int32_t va0 = (int32_t) (uint32_t) *a0++;
88 const int32_t va1 = (int32_t) (uint32_t) *a1++;
89 const int32_t va2 = (int32_t) (uint32_t) *a2++;
90
91 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
92 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
93 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
94 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
95 w = (const void*) ((const uint8_t*) w + 4);
96
97 vacc0x0 += va0 * vb0;
98 vacc0x1 += va0 * vb1;
99 vacc0x2 += va0 * vb2;
100 vacc0x3 += va0 * vb3;
101 vacc1x0 += va1 * vb0;
102 vacc1x1 += va1 * vb1;
103 vacc1x2 += va1 * vb2;
104 vacc1x3 += va1 * vb3;
105 vacc2x0 += va2 * vb0;
106 vacc2x1 += va2 * vb1;
107 vacc2x2 += va2 * vb2;
108 vacc2x3 += va2 * vb3;
109
110 k -= sizeof(uint8_t);
111 } while (k != 0);
112 p -= 3 * sizeof(void*);
113 } while (p != 0);
114
115 float vfpacc0x0 = (float) vacc0x0;
116 float vfpacc0x1 = (float) vacc0x1;
117 float vfpacc0x2 = (float) vacc0x2;
118 float vfpacc0x3 = (float) vacc0x3;
119 float vfpacc1x0 = (float) vacc1x0;
120 float vfpacc1x1 = (float) vacc1x1;
121 float vfpacc1x2 = (float) vacc1x2;
122 float vfpacc1x3 = (float) vacc1x3;
123 float vfpacc2x0 = (float) vacc2x0;
124 float vfpacc2x1 = (float) vacc2x1;
125 float vfpacc2x2 = (float) vacc2x2;
126 float vfpacc2x3 = (float) vacc2x3;
127
128 const float vscale = params->fp32_scalar_imagic.scale;
129 vfpacc0x0 *= vscale;
130 vfpacc0x1 *= vscale;
131 vfpacc0x2 *= vscale;
132 vfpacc0x3 *= vscale;
133 vfpacc1x0 *= vscale;
134 vfpacc1x1 *= vscale;
135 vfpacc1x2 *= vscale;
136 vfpacc1x3 *= vscale;
137 vfpacc2x0 *= vscale;
138 vfpacc2x1 *= vscale;
139 vfpacc2x2 *= vscale;
140 vfpacc2x3 *= vscale;
141
142 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
143 vfpacc0x0 += vmagic_bias;
144 vfpacc0x1 += vmagic_bias;
145 vfpacc0x2 += vmagic_bias;
146 vfpacc0x3 += vmagic_bias;
147 vfpacc1x0 += vmagic_bias;
148 vfpacc1x1 += vmagic_bias;
149 vfpacc1x2 += vmagic_bias;
150 vfpacc1x3 += vmagic_bias;
151 vfpacc2x0 += vmagic_bias;
152 vfpacc2x1 += vmagic_bias;
153 vfpacc2x2 += vmagic_bias;
154 vfpacc2x3 += vmagic_bias;
155
156 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
157 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1);
158 int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2);
159 int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3);
160 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0);
161 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1);
162 int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2);
163 int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3);
164 int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0);
165 int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1);
166 int32_t vout2x2 = (int32_t) float_as_uint32(vfpacc2x2);
167 int32_t vout2x3 = (int32_t) float_as_uint32(vfpacc2x3);
168
169 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
170 vout0x0 = math_max_s32(vout0x0, vmagic_min);
171 vout0x1 = math_max_s32(vout0x1, vmagic_min);
172 vout0x2 = math_max_s32(vout0x2, vmagic_min);
173 vout0x3 = math_max_s32(vout0x3, vmagic_min);
174 vout1x0 = math_max_s32(vout1x0, vmagic_min);
175 vout1x1 = math_max_s32(vout1x1, vmagic_min);
176 vout1x2 = math_max_s32(vout1x2, vmagic_min);
177 vout1x3 = math_max_s32(vout1x3, vmagic_min);
178 vout2x0 = math_max_s32(vout2x0, vmagic_min);
179 vout2x1 = math_max_s32(vout2x1, vmagic_min);
180 vout2x2 = math_max_s32(vout2x2, vmagic_min);
181 vout2x3 = math_max_s32(vout2x3, vmagic_min);
182
183 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
184 vout0x0 = math_min_s32(vout0x0, vmagic_max);
185 vout0x1 = math_min_s32(vout0x1, vmagic_max);
186 vout0x2 = math_min_s32(vout0x2, vmagic_max);
187 vout0x3 = math_min_s32(vout0x3, vmagic_max);
188 vout1x0 = math_min_s32(vout1x0, vmagic_max);
189 vout1x1 = math_min_s32(vout1x1, vmagic_max);
190 vout1x2 = math_min_s32(vout1x2, vmagic_max);
191 vout1x3 = math_min_s32(vout1x3, vmagic_max);
192 vout2x0 = math_min_s32(vout2x0, vmagic_max);
193 vout2x1 = math_min_s32(vout2x1, vmagic_max);
194 vout2x2 = math_min_s32(vout2x2, vmagic_max);
195 vout2x3 = math_min_s32(vout2x3, vmagic_max);
196
197 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
198 vout0x0 -= vmagic_bias_less_zero_point;
199 vout0x1 -= vmagic_bias_less_zero_point;
200 vout0x2 -= vmagic_bias_less_zero_point;
201 vout0x3 -= vmagic_bias_less_zero_point;
202 vout1x0 -= vmagic_bias_less_zero_point;
203 vout1x1 -= vmagic_bias_less_zero_point;
204 vout1x2 -= vmagic_bias_less_zero_point;
205 vout1x3 -= vmagic_bias_less_zero_point;
206 vout2x0 -= vmagic_bias_less_zero_point;
207 vout2x1 -= vmagic_bias_less_zero_point;
208 vout2x2 -= vmagic_bias_less_zero_point;
209 vout2x3 -= vmagic_bias_less_zero_point;
210
211 if XNN_LIKELY(nc >= 4) {
212 c2[0] = (uint8_t) vout2x0;
213 c2[1] = (uint8_t) vout2x1;
214 c2[2] = (uint8_t) vout2x2;
215 c2[3] = (uint8_t) vout2x3;
216 c1[0] = (uint8_t) vout1x0;
217 c1[1] = (uint8_t) vout1x1;
218 c1[2] = (uint8_t) vout1x2;
219 c1[3] = (uint8_t) vout1x3;
220 c0[0] = (uint8_t) vout0x0;
221 c0[1] = (uint8_t) vout0x1;
222 c0[2] = (uint8_t) vout0x2;
223 c0[3] = (uint8_t) vout0x3;
224
225 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
226 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
227 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
228
229 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
230 nc -= 4;
231 } else {
232 if (nc & 2) {
233 c2[0] = (uint8_t) vout2x0;
234 c2[1] = (uint8_t) vout2x1;
235 vout2x0 = vout2x2;
236 c2 += 2;
237 c1[0] = (uint8_t) vout1x0;
238 c1[1] = (uint8_t) vout1x1;
239 vout1x0 = vout1x2;
240 c1 += 2;
241 c0[0] = (uint8_t) vout0x0;
242 c0[1] = (uint8_t) vout0x1;
243 vout0x0 = vout0x2;
244 c0 += 2;
245 }
246 if (nc & 1) {
247 c2[0] = (uint8_t) vout2x0;
248 c1[0] = (uint8_t) vout1x0;
249 c0[0] = (uint8_t) vout0x0;
250 }
251
252 nc = 0;
253 }
254 } while (nc != 0);
255 }
256