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_qs8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_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 int8_t* a0 = a;
36 int8_t* c0 = c;
37 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
38 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
39 if XNN_UNPREDICTABLE(mr < 2) {
40 a1 = a0;
41 c1 = c0;
42 }
43 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
44 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
45 if XNN_UNPREDICTABLE(mr <= 2) {
46 a2 = a1;
47 c2 = c1;
48 }
49 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
50 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
51 if XNN_UNPREDICTABLE(mr != 4) {
52 a3 = a2;
53 c3 = c2;
54 }
55
56 do {
57 int32_t vacc0x0 = ((const int32_t*) w)[0];
58 int32_t vacc0x1 = ((const int32_t*) w)[1];
59 int32_t vacc0x2 = ((const int32_t*) w)[2];
60 int32_t vacc0x3 = ((const int32_t*) w)[3];
61 int32_t vacc1x0 = vacc0x0;
62 int32_t vacc1x1 = vacc0x1;
63 int32_t vacc1x2 = vacc0x2;
64 int32_t vacc1x3 = vacc0x3;
65 int32_t vacc2x0 = vacc0x0;
66 int32_t vacc2x1 = vacc0x1;
67 int32_t vacc2x2 = vacc0x2;
68 int32_t vacc2x3 = vacc0x3;
69 int32_t vacc3x0 = vacc0x0;
70 int32_t vacc3x1 = vacc0x1;
71 int32_t vacc3x2 = vacc0x2;
72 int32_t vacc3x3 = vacc0x3;
73 w = (const void*) ((const int32_t*) w + 4);
74
75 size_t k = kc;
76 do {
77 const int32_t va0 = (int32_t) *a0++;
78 const int32_t va1 = (int32_t) *a1++;
79 const int32_t va2 = (int32_t) *a2++;
80 const int32_t va3 = (int32_t) *a3++;
81
82 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
83 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
84 const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
85 const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
86 w = (const void*) ((const int8_t*) w + 4);
87
88 vacc0x0 += va0 * vb0;
89 vacc0x1 += va0 * vb1;
90 vacc0x2 += va0 * vb2;
91 vacc0x3 += va0 * vb3;
92 vacc1x0 += va1 * vb0;
93 vacc1x1 += va1 * vb1;
94 vacc1x2 += va1 * vb2;
95 vacc1x3 += va1 * vb3;
96 vacc2x0 += va2 * vb0;
97 vacc2x1 += va2 * vb1;
98 vacc2x2 += va2 * vb2;
99 vacc2x3 += va2 * vb3;
100 vacc3x0 += va3 * vb0;
101 vacc3x1 += va3 * vb1;
102 vacc3x2 += va3 * vb2;
103 vacc3x3 += va3 * vb3;
104
105 k -= sizeof(int8_t);
106 } while (k != 0);
107
108 float vfpacc0x0 = (float) vacc0x0;
109 float vfpacc0x1 = (float) vacc0x1;
110 float vfpacc0x2 = (float) vacc0x2;
111 float vfpacc0x3 = (float) vacc0x3;
112 float vfpacc1x0 = (float) vacc1x0;
113 float vfpacc1x1 = (float) vacc1x1;
114 float vfpacc1x2 = (float) vacc1x2;
115 float vfpacc1x3 = (float) vacc1x3;
116 float vfpacc2x0 = (float) vacc2x0;
117 float vfpacc2x1 = (float) vacc2x1;
118 float vfpacc2x2 = (float) vacc2x2;
119 float vfpacc2x3 = (float) vacc2x3;
120 float vfpacc3x0 = (float) vacc3x0;
121 float vfpacc3x1 = (float) vacc3x1;
122 float vfpacc3x2 = (float) vacc3x2;
123 float vfpacc3x3 = (float) vacc3x3;
124
125 const float vscale = params->fp32_scalar_fmagic.scale;
126 vfpacc0x0 *= vscale;
127 vfpacc0x1 *= vscale;
128 vfpacc0x2 *= vscale;
129 vfpacc0x3 *= vscale;
130 vfpacc1x0 *= vscale;
131 vfpacc1x1 *= vscale;
132 vfpacc1x2 *= vscale;
133 vfpacc1x3 *= vscale;
134 vfpacc2x0 *= vscale;
135 vfpacc2x1 *= vscale;
136 vfpacc2x2 *= vscale;
137 vfpacc2x3 *= vscale;
138 vfpacc3x0 *= vscale;
139 vfpacc3x1 *= vscale;
140 vfpacc3x2 *= vscale;
141 vfpacc3x3 *= vscale;
142
143 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
144 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
145 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
146 vfpacc0x2 = __builtin_wasm_max_f32(vfpacc0x2, voutput_min_less_zero_point);
147 vfpacc0x3 = __builtin_wasm_max_f32(vfpacc0x3, voutput_min_less_zero_point);
148 vfpacc1x0 = __builtin_wasm_max_f32(vfpacc1x0, voutput_min_less_zero_point);
149 vfpacc1x1 = __builtin_wasm_max_f32(vfpacc1x1, voutput_min_less_zero_point);
150 vfpacc1x2 = __builtin_wasm_max_f32(vfpacc1x2, voutput_min_less_zero_point);
151 vfpacc1x3 = __builtin_wasm_max_f32(vfpacc1x3, voutput_min_less_zero_point);
152 vfpacc2x0 = __builtin_wasm_max_f32(vfpacc2x0, voutput_min_less_zero_point);
153 vfpacc2x1 = __builtin_wasm_max_f32(vfpacc2x1, voutput_min_less_zero_point);
154 vfpacc2x2 = __builtin_wasm_max_f32(vfpacc2x2, voutput_min_less_zero_point);
155 vfpacc2x3 = __builtin_wasm_max_f32(vfpacc2x3, voutput_min_less_zero_point);
156 vfpacc3x0 = __builtin_wasm_max_f32(vfpacc3x0, voutput_min_less_zero_point);
157 vfpacc3x1 = __builtin_wasm_max_f32(vfpacc3x1, voutput_min_less_zero_point);
158 vfpacc3x2 = __builtin_wasm_max_f32(vfpacc3x2, voutput_min_less_zero_point);
159 vfpacc3x3 = __builtin_wasm_max_f32(vfpacc3x3, voutput_min_less_zero_point);
160
161 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
162 vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
163 vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
164 vfpacc0x2 = __builtin_wasm_min_f32(vfpacc0x2, voutput_max_less_zero_point);
165 vfpacc0x3 = __builtin_wasm_min_f32(vfpacc0x3, voutput_max_less_zero_point);
166 vfpacc1x0 = __builtin_wasm_min_f32(vfpacc1x0, voutput_max_less_zero_point);
167 vfpacc1x1 = __builtin_wasm_min_f32(vfpacc1x1, voutput_max_less_zero_point);
168 vfpacc1x2 = __builtin_wasm_min_f32(vfpacc1x2, voutput_max_less_zero_point);
169 vfpacc1x3 = __builtin_wasm_min_f32(vfpacc1x3, voutput_max_less_zero_point);
170 vfpacc2x0 = __builtin_wasm_min_f32(vfpacc2x0, voutput_max_less_zero_point);
171 vfpacc2x1 = __builtin_wasm_min_f32(vfpacc2x1, voutput_max_less_zero_point);
172 vfpacc2x2 = __builtin_wasm_min_f32(vfpacc2x2, voutput_max_less_zero_point);
173 vfpacc2x3 = __builtin_wasm_min_f32(vfpacc2x3, voutput_max_less_zero_point);
174 vfpacc3x0 = __builtin_wasm_min_f32(vfpacc3x0, voutput_max_less_zero_point);
175 vfpacc3x1 = __builtin_wasm_min_f32(vfpacc3x1, voutput_max_less_zero_point);
176 vfpacc3x2 = __builtin_wasm_min_f32(vfpacc3x2, voutput_max_less_zero_point);
177 vfpacc3x3 = __builtin_wasm_min_f32(vfpacc3x3, voutput_max_less_zero_point);
178
179 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
180 vfpacc0x0 += vmagic_bias;
181 vfpacc0x1 += vmagic_bias;
182 vfpacc0x2 += vmagic_bias;
183 vfpacc0x3 += vmagic_bias;
184 vfpacc1x0 += vmagic_bias;
185 vfpacc1x1 += vmagic_bias;
186 vfpacc1x2 += vmagic_bias;
187 vfpacc1x3 += vmagic_bias;
188 vfpacc2x0 += vmagic_bias;
189 vfpacc2x1 += vmagic_bias;
190 vfpacc2x2 += vmagic_bias;
191 vfpacc2x3 += vmagic_bias;
192 vfpacc3x0 += vmagic_bias;
193 vfpacc3x1 += vmagic_bias;
194 vfpacc3x2 += vmagic_bias;
195 vfpacc3x3 += vmagic_bias;
196
197 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
198 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0) - vmagic_bias_less_output_zero_point;
199 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1) - vmagic_bias_less_output_zero_point;
200 int32_t vout0x2 = (int32_t) fp32_to_bits(vfpacc0x2) - vmagic_bias_less_output_zero_point;
201 int32_t vout0x3 = (int32_t) fp32_to_bits(vfpacc0x3) - vmagic_bias_less_output_zero_point;
202 int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0) - vmagic_bias_less_output_zero_point;
203 int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1) - vmagic_bias_less_output_zero_point;
204 int32_t vout1x2 = (int32_t) fp32_to_bits(vfpacc1x2) - vmagic_bias_less_output_zero_point;
205 int32_t vout1x3 = (int32_t) fp32_to_bits(vfpacc1x3) - vmagic_bias_less_output_zero_point;
206 int32_t vout2x0 = (int32_t) fp32_to_bits(vfpacc2x0) - vmagic_bias_less_output_zero_point;
207 int32_t vout2x1 = (int32_t) fp32_to_bits(vfpacc2x1) - vmagic_bias_less_output_zero_point;
208 int32_t vout2x2 = (int32_t) fp32_to_bits(vfpacc2x2) - vmagic_bias_less_output_zero_point;
209 int32_t vout2x3 = (int32_t) fp32_to_bits(vfpacc2x3) - vmagic_bias_less_output_zero_point;
210 int32_t vout3x0 = (int32_t) fp32_to_bits(vfpacc3x0) - vmagic_bias_less_output_zero_point;
211 int32_t vout3x1 = (int32_t) fp32_to_bits(vfpacc3x1) - vmagic_bias_less_output_zero_point;
212 int32_t vout3x2 = (int32_t) fp32_to_bits(vfpacc3x2) - vmagic_bias_less_output_zero_point;
213 int32_t vout3x3 = (int32_t) fp32_to_bits(vfpacc3x3) - vmagic_bias_less_output_zero_point;
214
215 if XNN_LIKELY(nc >= 4) {
216 c0[0] = (int8_t) vout0x0;
217 c0[1] = (int8_t) vout0x1;
218 c0[2] = (int8_t) vout0x2;
219 c0[3] = (int8_t) vout0x3;
220 c1[0] = (int8_t) vout1x0;
221 c1[1] = (int8_t) vout1x1;
222 c1[2] = (int8_t) vout1x2;
223 c1[3] = (int8_t) vout1x3;
224 c2[0] = (int8_t) vout2x0;
225 c2[1] = (int8_t) vout2x1;
226 c2[2] = (int8_t) vout2x2;
227 c2[3] = (int8_t) vout2x3;
228 c3[0] = (int8_t) vout3x0;
229 c3[1] = (int8_t) vout3x1;
230 c3[2] = (int8_t) vout3x2;
231 c3[3] = (int8_t) vout3x3;
232
233 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
234 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
235 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
236 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
237
238 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
239 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
240 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
241 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
242
243 nc -= 4;
244 } else {
245 if (nc & 2) {
246 c0[0] = (int8_t) vout0x0;
247 c0[1] = (int8_t) vout0x1;
248 vout0x0 = vout0x2;
249 c0 += 2;
250 c1[0] = (int8_t) vout1x0;
251 c1[1] = (int8_t) vout1x1;
252 vout1x0 = vout1x2;
253 c1 += 2;
254 c2[0] = (int8_t) vout2x0;
255 c2[1] = (int8_t) vout2x1;
256 vout2x0 = vout2x2;
257 c2 += 2;
258 c3[0] = (int8_t) vout3x0;
259 c3[1] = (int8_t) vout3x1;
260 vout3x0 = vout3x2;
261 c3 += 2;
262 }
263 if (nc & 1) {
264 c0[0] = (int8_t) vout0x0;
265 c1[0] = (int8_t) vout1x0;
266 c2[0] = (int8_t) vout2x0;
267 c3[0] = (int8_t) vout3x0;
268 }
269
270 nc = 0;
271 }
272 } while (nc != 0);
273 }
274