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 <xnnpack/math.h>
13 #include <xnnpack/gemm.h>
14
15
xnn_qu8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic(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)])16 void xnn_qu8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 const uint8_t* restrict a,
21 size_t a_stride,
22 const void* restrict w,
23 uint8_t* restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(mr != 0);
29 assert(mr <= 4);
30 assert(nc != 0);
31 assert(kc != 0);
32
33 const uint8_t* a0 = a;
34 uint8_t* c0 = c;
35 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
36 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
37 if XNN_UNPREDICTABLE(mr < 2) {
38 a1 = a0;
39 c1 = c0;
40 }
41 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
42 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
43 if XNN_UNPREDICTABLE(mr <= 2) {
44 a2 = a1;
45 c2 = c1;
46 }
47 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
48 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
49 if XNN_UNPREDICTABLE(mr != 4) {
50 a3 = a2;
51 c3 = c2;
52 }
53
54 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
55 do {
56 int32_t vacc0x0 = ((const int32_t*) w)[0];
57 int32_t vacc0x1 = ((const int32_t*) w)[1];
58 int32_t vacc0x2 = ((const int32_t*) w)[2];
59 int32_t vacc0x3 = ((const int32_t*) w)[3];
60 int32_t vacc1x0 = vacc0x0;
61 int32_t vacc1x1 = vacc0x1;
62 int32_t vacc1x2 = vacc0x2;
63 int32_t vacc1x3 = vacc0x3;
64 int32_t vacc2x0 = vacc0x0;
65 int32_t vacc2x1 = vacc0x1;
66 int32_t vacc2x2 = vacc0x2;
67 int32_t vacc2x3 = vacc0x3;
68 int32_t vacc3x0 = vacc0x0;
69 int32_t vacc3x1 = vacc0x1;
70 int32_t vacc3x2 = vacc0x2;
71 int32_t vacc3x3 = vacc0x3;
72 w = (const void*) ((const int32_t*) w + 4);
73
74 size_t k = kc;
75 do {
76 const int32_t va0 = (int32_t) (uint32_t) *a0++;
77 const int32_t va1 = (int32_t) (uint32_t) *a1++;
78 const int32_t va2 = (int32_t) (uint32_t) *a2++;
79 const int32_t va3 = (int32_t) (uint32_t) *a3++;
80
81 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
82 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
83 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
84 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
85 w = (const void*) ((const uint8_t*) w + 4);
86
87 vacc0x0 += va0 * vb0;
88 vacc0x1 += va0 * vb1;
89 vacc0x2 += va0 * vb2;
90 vacc0x3 += va0 * vb3;
91 vacc1x0 += va1 * vb0;
92 vacc1x1 += va1 * vb1;
93 vacc1x2 += va1 * vb2;
94 vacc1x3 += va1 * vb3;
95 vacc2x0 += va2 * vb0;
96 vacc2x1 += va2 * vb1;
97 vacc2x2 += va2 * vb2;
98 vacc2x3 += va2 * vb3;
99 vacc3x0 += va3 * vb0;
100 vacc3x1 += va3 * vb1;
101 vacc3x2 += va3 * vb2;
102 vacc3x3 += va3 * vb3;
103
104 k -= sizeof(uint8_t);
105 } while (k != 0);
106
107 float vfpacc0x0 = (float) vacc0x0;
108 float vfpacc0x1 = (float) vacc0x1;
109 float vfpacc0x2 = (float) vacc0x2;
110 float vfpacc0x3 = (float) vacc0x3;
111 float vfpacc1x0 = (float) vacc1x0;
112 float vfpacc1x1 = (float) vacc1x1;
113 float vfpacc1x2 = (float) vacc1x2;
114 float vfpacc1x3 = (float) vacc1x3;
115 float vfpacc2x0 = (float) vacc2x0;
116 float vfpacc2x1 = (float) vacc2x1;
117 float vfpacc2x2 = (float) vacc2x2;
118 float vfpacc2x3 = (float) vacc2x3;
119 float vfpacc3x0 = (float) vacc3x0;
120 float vfpacc3x1 = (float) vacc3x1;
121 float vfpacc3x2 = (float) vacc3x2;
122 float vfpacc3x3 = (float) vacc3x3;
123
124 const float vscale = params->fp32_scalar_fmagic.scale;
125 vfpacc0x0 *= vscale;
126 vfpacc0x1 *= vscale;
127 vfpacc0x2 *= vscale;
128 vfpacc0x3 *= vscale;
129 vfpacc1x0 *= vscale;
130 vfpacc1x1 *= vscale;
131 vfpacc1x2 *= vscale;
132 vfpacc1x3 *= vscale;
133 vfpacc2x0 *= vscale;
134 vfpacc2x1 *= vscale;
135 vfpacc2x2 *= vscale;
136 vfpacc2x3 *= vscale;
137 vfpacc3x0 *= vscale;
138 vfpacc3x1 *= vscale;
139 vfpacc3x2 *= vscale;
140 vfpacc3x3 *= vscale;
141
142 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
143 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
144 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
145 vfpacc0x2 = __builtin_wasm_max_f32(vfpacc0x2, voutput_min_less_zero_point);
146 vfpacc0x3 = __builtin_wasm_max_f32(vfpacc0x3, voutput_min_less_zero_point);
147 vfpacc1x0 = __builtin_wasm_max_f32(vfpacc1x0, voutput_min_less_zero_point);
148 vfpacc1x1 = __builtin_wasm_max_f32(vfpacc1x1, voutput_min_less_zero_point);
149 vfpacc1x2 = __builtin_wasm_max_f32(vfpacc1x2, voutput_min_less_zero_point);
150 vfpacc1x3 = __builtin_wasm_max_f32(vfpacc1x3, voutput_min_less_zero_point);
151 vfpacc2x0 = __builtin_wasm_max_f32(vfpacc2x0, voutput_min_less_zero_point);
152 vfpacc2x1 = __builtin_wasm_max_f32(vfpacc2x1, voutput_min_less_zero_point);
153 vfpacc2x2 = __builtin_wasm_max_f32(vfpacc2x2, voutput_min_less_zero_point);
154 vfpacc2x3 = __builtin_wasm_max_f32(vfpacc2x3, voutput_min_less_zero_point);
155 vfpacc3x0 = __builtin_wasm_max_f32(vfpacc3x0, voutput_min_less_zero_point);
156 vfpacc3x1 = __builtin_wasm_max_f32(vfpacc3x1, voutput_min_less_zero_point);
157 vfpacc3x2 = __builtin_wasm_max_f32(vfpacc3x2, voutput_min_less_zero_point);
158 vfpacc3x3 = __builtin_wasm_max_f32(vfpacc3x3, voutput_min_less_zero_point);
159
160 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
161 vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
162 vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
163 vfpacc0x2 = __builtin_wasm_min_f32(vfpacc0x2, voutput_max_less_zero_point);
164 vfpacc0x3 = __builtin_wasm_min_f32(vfpacc0x3, voutput_max_less_zero_point);
165 vfpacc1x0 = __builtin_wasm_min_f32(vfpacc1x0, voutput_max_less_zero_point);
166 vfpacc1x1 = __builtin_wasm_min_f32(vfpacc1x1, voutput_max_less_zero_point);
167 vfpacc1x2 = __builtin_wasm_min_f32(vfpacc1x2, voutput_max_less_zero_point);
168 vfpacc1x3 = __builtin_wasm_min_f32(vfpacc1x3, voutput_max_less_zero_point);
169 vfpacc2x0 = __builtin_wasm_min_f32(vfpacc2x0, voutput_max_less_zero_point);
170 vfpacc2x1 = __builtin_wasm_min_f32(vfpacc2x1, voutput_max_less_zero_point);
171 vfpacc2x2 = __builtin_wasm_min_f32(vfpacc2x2, voutput_max_less_zero_point);
172 vfpacc2x3 = __builtin_wasm_min_f32(vfpacc2x3, voutput_max_less_zero_point);
173 vfpacc3x0 = __builtin_wasm_min_f32(vfpacc3x0, voutput_max_less_zero_point);
174 vfpacc3x1 = __builtin_wasm_min_f32(vfpacc3x1, voutput_max_less_zero_point);
175 vfpacc3x2 = __builtin_wasm_min_f32(vfpacc3x2, voutput_max_less_zero_point);
176 vfpacc3x3 = __builtin_wasm_min_f32(vfpacc3x3, voutput_max_less_zero_point);
177
178 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
179 vfpacc0x0 += vmagic_bias;
180 vfpacc0x1 += vmagic_bias;
181 vfpacc0x2 += vmagic_bias;
182 vfpacc0x3 += vmagic_bias;
183 vfpacc1x0 += vmagic_bias;
184 vfpacc1x1 += vmagic_bias;
185 vfpacc1x2 += vmagic_bias;
186 vfpacc1x3 += vmagic_bias;
187 vfpacc2x0 += vmagic_bias;
188 vfpacc2x1 += vmagic_bias;
189 vfpacc2x2 += vmagic_bias;
190 vfpacc2x3 += vmagic_bias;
191 vfpacc3x0 += vmagic_bias;
192 vfpacc3x1 += vmagic_bias;
193 vfpacc3x2 += vmagic_bias;
194 vfpacc3x3 += vmagic_bias;
195
196 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
197 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
198 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
199 int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
200 int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
201 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
202 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
203 int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2) - vmagic_bias_less_output_zero_point;
204 int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3) - vmagic_bias_less_output_zero_point;
205 int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0) - vmagic_bias_less_output_zero_point;
206 int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1) - vmagic_bias_less_output_zero_point;
207 int32_t vout2x2 = (int32_t) float_as_uint32(vfpacc2x2) - vmagic_bias_less_output_zero_point;
208 int32_t vout2x3 = (int32_t) float_as_uint32(vfpacc2x3) - vmagic_bias_less_output_zero_point;
209 int32_t vout3x0 = (int32_t) float_as_uint32(vfpacc3x0) - vmagic_bias_less_output_zero_point;
210 int32_t vout3x1 = (int32_t) float_as_uint32(vfpacc3x1) - vmagic_bias_less_output_zero_point;
211 int32_t vout3x2 = (int32_t) float_as_uint32(vfpacc3x2) - vmagic_bias_less_output_zero_point;
212 int32_t vout3x3 = (int32_t) float_as_uint32(vfpacc3x3) - vmagic_bias_less_output_zero_point;
213
214 if XNN_LIKELY(nc >= 4) {
215 c0[0] = (uint8_t) vout0x0;
216 c0[1] = (uint8_t) vout0x1;
217 c0[2] = (uint8_t) vout0x2;
218 c0[3] = (uint8_t) vout0x3;
219 c1[0] = (uint8_t) vout1x0;
220 c1[1] = (uint8_t) vout1x1;
221 c1[2] = (uint8_t) vout1x2;
222 c1[3] = (uint8_t) vout1x3;
223 c2[0] = (uint8_t) vout2x0;
224 c2[1] = (uint8_t) vout2x1;
225 c2[2] = (uint8_t) vout2x2;
226 c2[3] = (uint8_t) vout2x3;
227 c3[0] = (uint8_t) vout3x0;
228 c3[1] = (uint8_t) vout3x1;
229 c3[2] = (uint8_t) vout3x2;
230 c3[3] = (uint8_t) vout3x3;
231
232 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
233 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
234 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
235 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
236
237 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
238 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
239 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
240 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
241
242 nc -= 4;
243 } else {
244 if (nc & 2) {
245 c0[0] = (uint8_t) vout0x0;
246 c0[1] = (uint8_t) vout0x1;
247 vout0x0 = vout0x2;
248 c0 += 2;
249 c1[0] = (uint8_t) vout1x0;
250 c1[1] = (uint8_t) vout1x1;
251 vout1x0 = vout1x2;
252 c1 += 2;
253 c2[0] = (uint8_t) vout2x0;
254 c2[1] = (uint8_t) vout2x1;
255 vout2x0 = vout2x2;
256 c2 += 2;
257 c3[0] = (uint8_t) vout3x0;
258 c3[1] = (uint8_t) vout3x1;
259 vout3x0 = vout3x2;
260 c3 += 2;
261 }
262 if (nc & 1) {
263 c0[0] = (uint8_t) vout0x0;
264 c1[0] = (uint8_t) vout1x0;
265 c2[0] = (uint8_t) vout2x0;
266 c3[0] = (uint8_t) vout3x0;
267 }
268
269 nc = 0;
270 }
271 } while (nc != 0);
272 }
273