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_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_fmagic(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_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 int8_t**restrict a,
24 const void*restrict w,
25 int8_t*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_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 int8_t* c0 = c;
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 c1 = c0;
46 }
47 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
48 if XNN_UNPREDICTABLE(mr <= 2) {
49 c2 = c1;
50 }
51
52 do {
53 int32_t vacc0x0 = ((const int32_t*) w)[0];
54 int32_t vacc0x1 = ((const int32_t*) w)[1];
55 int32_t vacc0x2 = ((const int32_t*) w)[2];
56 int32_t vacc0x3 = ((const int32_t*) w)[3];
57 int32_t vacc1x0 = vacc0x0;
58 int32_t vacc1x1 = vacc0x1;
59 int32_t vacc1x2 = vacc0x2;
60 int32_t vacc1x3 = vacc0x3;
61 int32_t vacc2x0 = vacc0x0;
62 int32_t vacc2x1 = vacc0x1;
63 int32_t vacc2x2 = vacc0x2;
64 int32_t vacc2x3 = vacc0x3;
65 w = (const void*) ((const int32_t*) w + 4);
66
67 size_t p = ks;
68 do {
69 const int8_t* restrict a0 = a[0];
70 assert(a0 != NULL);
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73 }
74 const int8_t* restrict a1 = a[1];
75 assert(a1 != NULL);
76 if XNN_UNPREDICTABLE(a1 != zero) {
77 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
78 }
79 const int8_t* restrict a2 = a[2];
80 assert(a2 != NULL);
81 if XNN_UNPREDICTABLE(a2 != zero) {
82 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
83 }
84 a += 3;
85
86 size_t k = kc;
87 do {
88 const int32_t va0 = (int32_t) *a0++;
89 const int32_t va1 = (int32_t) *a1++;
90 const int32_t va2 = (int32_t) *a2++;
91
92 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
93 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
94 const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
95 const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
96 w = (const void*) ((const int8_t*) w + 4);
97
98 vacc0x0 += va0 * vb0;
99 vacc0x1 += va0 * vb1;
100 vacc0x2 += va0 * vb2;
101 vacc0x3 += va0 * vb3;
102 vacc1x0 += va1 * vb0;
103 vacc1x1 += va1 * vb1;
104 vacc1x2 += va1 * vb2;
105 vacc1x3 += va1 * vb3;
106 vacc2x0 += va2 * vb0;
107 vacc2x1 += va2 * vb1;
108 vacc2x2 += va2 * vb2;
109 vacc2x3 += va2 * vb3;
110
111 k -= sizeof(int8_t);
112 } while (k != 0);
113 p -= 3 * sizeof(void*);
114 } while (p != 0);
115
116 float vfpacc0x0 = (float) vacc0x0;
117 float vfpacc0x1 = (float) vacc0x1;
118 float vfpacc0x2 = (float) vacc0x2;
119 float vfpacc0x3 = (float) vacc0x3;
120 float vfpacc1x0 = (float) vacc1x0;
121 float vfpacc1x1 = (float) vacc1x1;
122 float vfpacc1x2 = (float) vacc1x2;
123 float vfpacc1x3 = (float) vacc1x3;
124 float vfpacc2x0 = (float) vacc2x0;
125 float vfpacc2x1 = (float) vacc2x1;
126 float vfpacc2x2 = (float) vacc2x2;
127 float vfpacc2x3 = (float) vacc2x3;
128
129 const float vscale0 = ((const float*) w)[0];
130 vfpacc0x0 *= vscale0;
131 vfpacc1x0 *= vscale0;
132 vfpacc2x0 *= vscale0;
133 const float vscale1 = ((const float*) w)[1];
134 vfpacc0x1 *= vscale1;
135 vfpacc1x1 *= vscale1;
136 vfpacc2x1 *= vscale1;
137 const float vscale2 = ((const float*) w)[2];
138 vfpacc0x2 *= vscale2;
139 vfpacc1x2 *= vscale2;
140 vfpacc2x2 *= vscale2;
141 const float vscale3 = ((const float*) w)[3];
142 vfpacc0x3 *= vscale3;
143 vfpacc1x3 *= vscale3;
144 vfpacc2x3 *= vscale3;
145 w = (const void*) ((const float*) w + 4);
146
147 const float voutput_min_less_zero_point = params->scalar_fmagic.output_min_less_zero_point;
148 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
149 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
150 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
151 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
152 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
153 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
154 vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
155 vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
156 vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
157 vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
158 vfpacc2x2 = math_max_f32(vfpacc2x2, voutput_min_less_zero_point);
159 vfpacc2x3 = math_max_f32(vfpacc2x3, voutput_min_less_zero_point);
160
161 const float voutput_max_less_zero_point = params->scalar_fmagic.output_max_less_zero_point;
162 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
163 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
164 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
165 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
166 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
167 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
168 vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
169 vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
170 vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
171 vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
172 vfpacc2x2 = math_min_f32(vfpacc2x2, voutput_max_less_zero_point);
173 vfpacc2x3 = math_min_f32(vfpacc2x3, voutput_max_less_zero_point);
174
175 const float vmagic_bias = params->scalar_fmagic.magic_bias;
176 vfpacc0x0 += vmagic_bias;
177 vfpacc0x1 += vmagic_bias;
178 vfpacc0x2 += vmagic_bias;
179 vfpacc0x3 += vmagic_bias;
180 vfpacc1x0 += vmagic_bias;
181 vfpacc1x1 += vmagic_bias;
182 vfpacc1x2 += vmagic_bias;
183 vfpacc1x3 += vmagic_bias;
184 vfpacc2x0 += vmagic_bias;
185 vfpacc2x1 += vmagic_bias;
186 vfpacc2x2 += vmagic_bias;
187 vfpacc2x3 += vmagic_bias;
188
189 const int32_t vmagic_bias_less_output_zero_point = params->scalar_fmagic.magic_bias_less_output_zero_point;
190 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0) - vmagic_bias_less_output_zero_point;
191 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1) - vmagic_bias_less_output_zero_point;
192 int32_t vout0x2 = (int32_t) fp32_to_bits(vfpacc0x2) - vmagic_bias_less_output_zero_point;
193 int32_t vout0x3 = (int32_t) fp32_to_bits(vfpacc0x3) - vmagic_bias_less_output_zero_point;
194 int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0) - vmagic_bias_less_output_zero_point;
195 int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1) - vmagic_bias_less_output_zero_point;
196 int32_t vout1x2 = (int32_t) fp32_to_bits(vfpacc1x2) - vmagic_bias_less_output_zero_point;
197 int32_t vout1x3 = (int32_t) fp32_to_bits(vfpacc1x3) - vmagic_bias_less_output_zero_point;
198 int32_t vout2x0 = (int32_t) fp32_to_bits(vfpacc2x0) - vmagic_bias_less_output_zero_point;
199 int32_t vout2x1 = (int32_t) fp32_to_bits(vfpacc2x1) - vmagic_bias_less_output_zero_point;
200 int32_t vout2x2 = (int32_t) fp32_to_bits(vfpacc2x2) - vmagic_bias_less_output_zero_point;
201 int32_t vout2x3 = (int32_t) fp32_to_bits(vfpacc2x3) - vmagic_bias_less_output_zero_point;
202
203 if XNN_LIKELY(nc >= 4) {
204 c2[0] = (int8_t) vout2x0;
205 c2[1] = (int8_t) vout2x1;
206 c2[2] = (int8_t) vout2x2;
207 c2[3] = (int8_t) vout2x3;
208 c1[0] = (int8_t) vout1x0;
209 c1[1] = (int8_t) vout1x1;
210 c1[2] = (int8_t) vout1x2;
211 c1[3] = (int8_t) vout1x3;
212 c0[0] = (int8_t) vout0x0;
213 c0[1] = (int8_t) vout0x1;
214 c0[2] = (int8_t) vout0x2;
215 c0[3] = (int8_t) vout0x3;
216
217 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
218 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
219 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
220
221 a = (const int8_t**restrict) ((uintptr_t) a - ks);
222 nc -= 4;
223 } else {
224 if (nc & 2) {
225 c2[0] = (int8_t) vout2x0;
226 c2[1] = (int8_t) vout2x1;
227 vout2x0 = vout2x2;
228 c2 += 2;
229 c1[0] = (int8_t) vout1x0;
230 c1[1] = (int8_t) vout1x1;
231 vout1x0 = vout1x2;
232 c1 += 2;
233 c0[0] = (int8_t) vout0x0;
234 c0[1] = (int8_t) vout0x1;
235 vout0x0 = vout0x2;
236 c0 += 2;
237 }
238 if (nc & 1) {
239 c2[0] = (int8_t) vout2x0;
240 c1[0] = (int8_t) vout1x0;
241 c0[0] = (int8_t) vout0x0;
242 }
243
244 nc = 0;
245 }
246 } while (nc != 0);
247 }
248