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 <fp16.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_qu8_igemm_minmax_fp32_ukernel_3x4__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)])18 void xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_fmagic(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t**restrict a,
24 const void*restrict w,
25 uint8_t*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * sizeof(void*)) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 uint8_t* c0 = c;
43 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 c1 = c0;
46 }
47 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
48 if XNN_UNPREDICTABLE(mr <= 2) {
49 c2 = c1;
50 }
51
52 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
53 do {
54 int32_t vacc0x0 = ((const int32_t*) w)[0];
55 int32_t vacc0x1 = ((const int32_t*) w)[1];
56 int32_t vacc0x2 = ((const int32_t*) w)[2];
57 int32_t vacc0x3 = ((const int32_t*) w)[3];
58 int32_t vacc1x0 = vacc0x0;
59 int32_t vacc1x1 = vacc0x1;
60 int32_t vacc1x2 = vacc0x2;
61 int32_t vacc1x3 = vacc0x3;
62 int32_t vacc2x0 = vacc0x0;
63 int32_t vacc2x1 = vacc0x1;
64 int32_t vacc2x2 = vacc0x2;
65 int32_t vacc2x3 = vacc0x3;
66 w = (const void*) ((const int32_t*) w + 4);
67
68 size_t p = ks;
69 do {
70 const uint8_t* restrict a0 = a[0];
71 assert(a0 != NULL);
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
74 }
75 const uint8_t* restrict a1 = a[1];
76 assert(a1 != NULL);
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
79 }
80 const uint8_t* restrict a2 = a[2];
81 assert(a2 != NULL);
82 if XNN_UNPREDICTABLE(a2 != zero) {
83 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
84 }
85 a += 3;
86
87 size_t k = kc;
88 do {
89 const int32_t va0 = (int32_t) (uint32_t) *a0++;
90 const int32_t va1 = (int32_t) (uint32_t) *a1++;
91 const int32_t va2 = (int32_t) (uint32_t) *a2++;
92
93 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
94 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
95 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
96 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
97 w = (const void*) ((const uint8_t*) w + 4);
98
99 vacc0x0 += va0 * vb0;
100 vacc0x1 += va0 * vb1;
101 vacc0x2 += va0 * vb2;
102 vacc0x3 += va0 * vb3;
103 vacc1x0 += va1 * vb0;
104 vacc1x1 += va1 * vb1;
105 vacc1x2 += va1 * vb2;
106 vacc1x3 += va1 * vb3;
107 vacc2x0 += va2 * vb0;
108 vacc2x1 += va2 * vb1;
109 vacc2x2 += va2 * vb2;
110 vacc2x3 += va2 * vb3;
111
112 k -= sizeof(uint8_t);
113 } while (k != 0);
114 p -= 3 * sizeof(void*);
115 } while (p != 0);
116
117 float vfpacc0x0 = (float) vacc0x0;
118 float vfpacc0x1 = (float) vacc0x1;
119 float vfpacc0x2 = (float) vacc0x2;
120 float vfpacc0x3 = (float) vacc0x3;
121 float vfpacc1x0 = (float) vacc1x0;
122 float vfpacc1x1 = (float) vacc1x1;
123 float vfpacc1x2 = (float) vacc1x2;
124 float vfpacc1x3 = (float) vacc1x3;
125 float vfpacc2x0 = (float) vacc2x0;
126 float vfpacc2x1 = (float) vacc2x1;
127 float vfpacc2x2 = (float) vacc2x2;
128 float vfpacc2x3 = (float) vacc2x3;
129
130 const float vscale = params->fp32_scalar_fmagic.scale;
131 vfpacc0x0 *= vscale;
132 vfpacc0x1 *= vscale;
133 vfpacc0x2 *= vscale;
134 vfpacc0x3 *= vscale;
135 vfpacc1x0 *= vscale;
136 vfpacc1x1 *= vscale;
137 vfpacc1x2 *= vscale;
138 vfpacc1x3 *= vscale;
139 vfpacc2x0 *= vscale;
140 vfpacc2x1 *= vscale;
141 vfpacc2x2 *= vscale;
142 vfpacc2x3 *= vscale;
143
144 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
145 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
146 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
147 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
148 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
149 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
150 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
151 vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
152 vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
153 vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
154 vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
155 vfpacc2x2 = math_max_f32(vfpacc2x2, voutput_min_less_zero_point);
156 vfpacc2x3 = math_max_f32(vfpacc2x3, voutput_min_less_zero_point);
157
158 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
159 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
160 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
161 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
162 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
163 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
164 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
165 vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
166 vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
167 vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
168 vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
169 vfpacc2x2 = math_min_f32(vfpacc2x2, voutput_max_less_zero_point);
170 vfpacc2x3 = math_min_f32(vfpacc2x3, voutput_max_less_zero_point);
171
172 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
173 vfpacc0x0 += vmagic_bias;
174 vfpacc0x1 += vmagic_bias;
175 vfpacc0x2 += vmagic_bias;
176 vfpacc0x3 += vmagic_bias;
177 vfpacc1x0 += vmagic_bias;
178 vfpacc1x1 += vmagic_bias;
179 vfpacc1x2 += vmagic_bias;
180 vfpacc1x3 += vmagic_bias;
181 vfpacc2x0 += vmagic_bias;
182 vfpacc2x1 += vmagic_bias;
183 vfpacc2x2 += vmagic_bias;
184 vfpacc2x3 += vmagic_bias;
185
186 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
187 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0) - vmagic_bias_less_output_zero_point;
188 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1) - vmagic_bias_less_output_zero_point;
189 int32_t vout0x2 = (int32_t) fp32_to_bits(vfpacc0x2) - vmagic_bias_less_output_zero_point;
190 int32_t vout0x3 = (int32_t) fp32_to_bits(vfpacc0x3) - vmagic_bias_less_output_zero_point;
191 int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0) - vmagic_bias_less_output_zero_point;
192 int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1) - vmagic_bias_less_output_zero_point;
193 int32_t vout1x2 = (int32_t) fp32_to_bits(vfpacc1x2) - vmagic_bias_less_output_zero_point;
194 int32_t vout1x3 = (int32_t) fp32_to_bits(vfpacc1x3) - vmagic_bias_less_output_zero_point;
195 int32_t vout2x0 = (int32_t) fp32_to_bits(vfpacc2x0) - vmagic_bias_less_output_zero_point;
196 int32_t vout2x1 = (int32_t) fp32_to_bits(vfpacc2x1) - vmagic_bias_less_output_zero_point;
197 int32_t vout2x2 = (int32_t) fp32_to_bits(vfpacc2x2) - vmagic_bias_less_output_zero_point;
198 int32_t vout2x3 = (int32_t) fp32_to_bits(vfpacc2x3) - vmagic_bias_less_output_zero_point;
199
200 if XNN_LIKELY(nc >= 4) {
201 c2[0] = (uint8_t) vout2x0;
202 c2[1] = (uint8_t) vout2x1;
203 c2[2] = (uint8_t) vout2x2;
204 c2[3] = (uint8_t) vout2x3;
205 c1[0] = (uint8_t) vout1x0;
206 c1[1] = (uint8_t) vout1x1;
207 c1[2] = (uint8_t) vout1x2;
208 c1[3] = (uint8_t) vout1x3;
209 c0[0] = (uint8_t) vout0x0;
210 c0[1] = (uint8_t) vout0x1;
211 c0[2] = (uint8_t) vout0x2;
212 c0[3] = (uint8_t) vout0x3;
213
214 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
215 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
216 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
217
218 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
219 nc -= 4;
220 } else {
221 if (nc & 2) {
222 c2[0] = (uint8_t) vout2x0;
223 c2[1] = (uint8_t) vout2x1;
224 vout2x0 = vout2x2;
225 c2 += 2;
226 c1[0] = (uint8_t) vout1x0;
227 c1[1] = (uint8_t) vout1x1;
228 vout1x0 = vout1x2;
229 c1 += 2;
230 c0[0] = (uint8_t) vout0x0;
231 c0[1] = (uint8_t) vout0x1;
232 vout0x0 = vout0x2;
233 c0 += 2;
234 }
235 if (nc & 1) {
236 c2[0] = (uint8_t) vout2x0;
237 c1[0] = (uint8_t) vout1x0;
238 c0[0] = (uint8_t) vout0x0;
239 }
240
241 nc = 0;
242 }
243 } while (nc != 0);
244 }
245