Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x16c4-minmax-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c4-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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 <wasm_simd128.h>
13
14 #include "xnnpack/gemm.h"
15 #include "xnnpack/math.h"
16
17 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x16c4__wasmusdot(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 float* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
28 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 4 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 float* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60
61 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
62 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
63 XNN_FORCE_REALIZATION(vmin);
64 XNN_FORCE_REALIZATION(vmax);
65
66 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
67 XNN_FORCE_REALIZATION(vsign_mask);
68
69 do {
70 v128_t vksum0123 = wasm_v128_load((const int32_t*) w);
71 v128_t vksum4567 = wasm_v128_load((const int32_t*) w + 4);
72 v128_t vksum89AB = wasm_v128_load((const int32_t*) w + 8);
73 v128_t vksumCDEF = wasm_v128_load((const int32_t*) w + 12);
74 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
75 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
76 v128_t vacc0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0);
77 v128_t vacc0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point0);
78 v128_t vacc0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point0);
79 const v128_t vinput_zero_point1 = wasm_i32x4_splat((int32_t) quantization_params[1].zero_point + 128);
80 v128_t vacc1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
81 v128_t vacc1x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point1);
82 v128_t vacc1x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point1);
83 v128_t vacc1xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point1);
84 const v128_t vinput_zero_point2 = wasm_i32x4_splat((int32_t) quantization_params[2].zero_point + 128);
85 v128_t vacc2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
86 v128_t vacc2x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point2);
87 v128_t vacc2x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point2);
88 v128_t vacc2xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point2);
89 const v128_t vinput_zero_point3 = wasm_i32x4_splat((int32_t) quantization_params[3].zero_point + 128);
90 v128_t vacc3x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point3);
91 v128_t vacc3x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point3);
92 v128_t vacc3x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point3);
93 v128_t vacc3xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point3);
94 w = (const int32_t*) w + 16;
95
96 size_t k = kc;
97
98 while (k != 0) {
99 v128_t va0x0123 = wasm_v128_load32_splat(a0);
100 a0 += 4;
101 v128_t va1x0123 = wasm_v128_load32_splat(a1);
102 a1 += 4;
103 v128_t va2x0123 = wasm_v128_load32_splat(a2);
104 a2 += 4;
105 v128_t va3x0123 = wasm_v128_load32_splat(a3);
106 a3 += 4;
107
108 va0x0123 = wasm_v128_xor(va0x0123, vsign_mask);
109 va1x0123 = wasm_v128_xor(va1x0123, vsign_mask);
110 va2x0123 = wasm_v128_xor(va2x0123, vsign_mask);
111 va3x0123 = wasm_v128_xor(va3x0123, vsign_mask);
112
113 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
114 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
115 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
116 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
117
118 vacc0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0123);
119 vacc0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x4567);
120 vacc0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x89AB);
121 vacc0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0xCDEF);
122 vacc1x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va1x0123, vacc1x0123);
123 vacc1x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va1x0123, vacc1x4567);
124 vacc1x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va1x0123, vacc1x89AB);
125 vacc1xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va1x0123, vacc1xCDEF);
126 vacc2x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va2x0123, vacc2x0123);
127 vacc2x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va2x0123, vacc2x4567);
128 vacc2x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va2x0123, vacc2x89AB);
129 vacc2xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va2x0123, vacc2xCDEF);
130 vacc3x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va3x0123, vacc3x0123);
131 vacc3x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va3x0123, vacc3x4567);
132 vacc3x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va3x0123, vacc3x89AB);
133 vacc3xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va3x0123, vacc3xCDEF);
134
135 w = (const int8_t*) w + 64;
136 k -= 4 * sizeof(int8_t);
137 }
138
139 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
140 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
141 vacc0x89AB = wasm_f32x4_convert_i32x4(vacc0x89AB);
142 vacc0xCDEF = wasm_f32x4_convert_i32x4(vacc0xCDEF);
143 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
144 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
145 vacc1x89AB = wasm_f32x4_convert_i32x4(vacc1x89AB);
146 vacc1xCDEF = wasm_f32x4_convert_i32x4(vacc1xCDEF);
147 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
148 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
149 vacc2x89AB = wasm_f32x4_convert_i32x4(vacc2x89AB);
150 vacc2xCDEF = wasm_f32x4_convert_i32x4(vacc2xCDEF);
151 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
152 vacc3x4567 = wasm_f32x4_convert_i32x4(vacc3x4567);
153 vacc3x89AB = wasm_f32x4_convert_i32x4(vacc3x89AB);
154 vacc3xCDEF = wasm_f32x4_convert_i32x4(vacc3xCDEF);
155
156 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
157 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
158 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
159 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
160
161 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
162 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
163 vacc0x89AB = wasm_f32x4_mul(vacc0x89AB, vinput_scale0);
164 vacc0xCDEF = wasm_f32x4_mul(vacc0xCDEF, vinput_scale0);
165 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
166 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1);
167 vacc1x89AB = wasm_f32x4_mul(vacc1x89AB, vinput_scale1);
168 vacc1xCDEF = wasm_f32x4_mul(vacc1xCDEF, vinput_scale1);
169 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
170 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale2);
171 vacc2x89AB = wasm_f32x4_mul(vacc2x89AB, vinput_scale2);
172 vacc2xCDEF = wasm_f32x4_mul(vacc2xCDEF, vinput_scale2);
173 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale3);
174 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vinput_scale3);
175 vacc3x89AB = wasm_f32x4_mul(vacc3x89AB, vinput_scale3);
176 vacc3xCDEF = wasm_f32x4_mul(vacc3xCDEF, vinput_scale3);
177
178 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
179 w = (const float*) w + 4;
180 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
181 w = (const float*) w + 4;
182 const v128_t vfilter_output_scale89AB = wasm_v128_load(w);
183 w = (const float*) w + 4;
184 const v128_t vfilter_output_scaleCDEF = wasm_v128_load(w);
185 w = (const float*) w + 4;
186 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
187 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
188 vacc0x89AB = wasm_f32x4_mul(vacc0x89AB, vfilter_output_scale89AB);
189 vacc0xCDEF = wasm_f32x4_mul(vacc0xCDEF, vfilter_output_scaleCDEF);
190 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
191 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
192 vacc1x89AB = wasm_f32x4_mul(vacc1x89AB, vfilter_output_scale89AB);
193 vacc1xCDEF = wasm_f32x4_mul(vacc1xCDEF, vfilter_output_scaleCDEF);
194 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
195 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567);
196 vacc2x89AB = wasm_f32x4_mul(vacc2x89AB, vfilter_output_scale89AB);
197 vacc2xCDEF = wasm_f32x4_mul(vacc2xCDEF, vfilter_output_scaleCDEF);
198 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
199 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vfilter_output_scale4567);
200 vacc3x89AB = wasm_f32x4_mul(vacc3x89AB, vfilter_output_scale89AB);
201 vacc3xCDEF = wasm_f32x4_mul(vacc3xCDEF, vfilter_output_scaleCDEF);
202
203 const v128_t vbias0123 = wasm_v128_load(w);
204 w = (const float*) w + 4;
205 const v128_t vbias4567 = wasm_v128_load(w);
206 w = (const float*) w + 4;
207 const v128_t vbias89AB = wasm_v128_load(w);
208 w = (const float*) w + 4;
209 const v128_t vbiasCDEF = wasm_v128_load(w);
210 w = (const float*) w + 4;
211 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
212 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
213 vacc0x89AB = wasm_f32x4_add(vacc0x89AB, vbias89AB);
214 vacc0xCDEF = wasm_f32x4_add(vacc0xCDEF, vbiasCDEF);
215 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
216 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
217 vacc1x89AB = wasm_f32x4_add(vacc1x89AB, vbias89AB);
218 vacc1xCDEF = wasm_f32x4_add(vacc1xCDEF, vbiasCDEF);
219 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
220 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567);
221 vacc2x89AB = wasm_f32x4_add(vacc2x89AB, vbias89AB);
222 vacc2xCDEF = wasm_f32x4_add(vacc2xCDEF, vbiasCDEF);
223 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
224 vacc3x4567 = wasm_f32x4_add(vacc3x4567, vbias4567);
225 vacc3x89AB = wasm_f32x4_add(vacc3x89AB, vbias89AB);
226 vacc3xCDEF = wasm_f32x4_add(vacc3xCDEF, vbiasCDEF);
227
228 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
229 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
230 vacc0x89AB = wasm_f32x4_pmax(vacc0x89AB, vmin);
231 vacc0xCDEF = wasm_f32x4_pmax(vacc0xCDEF, vmin);
232 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
233 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
234 vacc1x89AB = wasm_f32x4_pmax(vacc1x89AB, vmin);
235 vacc1xCDEF = wasm_f32x4_pmax(vacc1xCDEF, vmin);
236 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
237 vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin);
238 vacc2x89AB = wasm_f32x4_pmax(vacc2x89AB, vmin);
239 vacc2xCDEF = wasm_f32x4_pmax(vacc2xCDEF, vmin);
240 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
241 vacc3x4567 = wasm_f32x4_pmax(vacc3x4567, vmin);
242 vacc3x89AB = wasm_f32x4_pmax(vacc3x89AB, vmin);
243 vacc3xCDEF = wasm_f32x4_pmax(vacc3xCDEF, vmin);
244
245 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
246 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
247 vacc0x89AB = wasm_f32x4_pmin(vacc0x89AB, vmax);
248 vacc0xCDEF = wasm_f32x4_pmin(vacc0xCDEF, vmax);
249 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
250 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
251 vacc1x89AB = wasm_f32x4_pmin(vacc1x89AB, vmax);
252 vacc1xCDEF = wasm_f32x4_pmin(vacc1xCDEF, vmax);
253 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
254 vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax);
255 vacc2x89AB = wasm_f32x4_pmin(vacc2x89AB, vmax);
256 vacc2xCDEF = wasm_f32x4_pmin(vacc2xCDEF, vmax);
257 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
258 vacc3x4567 = wasm_f32x4_pmin(vacc3x4567, vmax);
259 vacc3x89AB = wasm_f32x4_pmin(vacc3x89AB, vmax);
260 vacc3xCDEF = wasm_f32x4_pmin(vacc3xCDEF, vmax);
261
262 if XNN_LIKELY(nc >= 16) {
263 wasm_v128_store(c0, vacc0x0123);
264 wasm_v128_store(c0 + 4, vacc0x4567);
265 wasm_v128_store(c0 + 8, vacc0x89AB);
266 wasm_v128_store(c0 + 12, vacc0xCDEF);
267 wasm_v128_store(c1, vacc1x0123);
268 wasm_v128_store(c1 + 4, vacc1x4567);
269 wasm_v128_store(c1 + 8, vacc1x89AB);
270 wasm_v128_store(c1 + 12, vacc1xCDEF);
271 wasm_v128_store(c2, vacc2x0123);
272 wasm_v128_store(c2 + 4, vacc2x4567);
273 wasm_v128_store(c2 + 8, vacc2x89AB);
274 wasm_v128_store(c2 + 12, vacc2xCDEF);
275 wasm_v128_store(c3, vacc3x0123);
276 wasm_v128_store(c3 + 4, vacc3x4567);
277 wasm_v128_store(c3 + 8, vacc3x89AB);
278 wasm_v128_store(c3 + 12, vacc3xCDEF);
279
280 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
281 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
282 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
283 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
284
285 c0 = (float*) ((uintptr_t) c0 + cn_stride);
286 c1 = (float*) ((uintptr_t) c1 + cn_stride);
287 c2 = (float*) ((uintptr_t) c2 + cn_stride);
288 c3 = (float*) ((uintptr_t) c3 + cn_stride);
289
290 nc -= 16;
291 } else {
292 if (nc & 8) {
293 wasm_v128_store(c0, vacc0x0123);
294 vacc0x0123 = vacc0x89AB;
295 c0 += 4;
296 wasm_v128_store(c0, vacc0x4567);
297 vacc0x4567 = vacc0xCDEF;
298 c0 += 4;
299 wasm_v128_store(c1, vacc1x0123);
300 vacc1x0123 = vacc1x89AB;
301 c1 += 4;
302 wasm_v128_store(c1, vacc1x4567);
303 vacc1x4567 = vacc1xCDEF;
304 c1 += 4;
305 wasm_v128_store(c2, vacc2x0123);
306 vacc2x0123 = vacc2x89AB;
307 c2 += 4;
308 wasm_v128_store(c2, vacc2x4567);
309 vacc2x4567 = vacc2xCDEF;
310 c2 += 4;
311 wasm_v128_store(c3, vacc3x0123);
312 vacc3x0123 = vacc3x89AB;
313 c3 += 4;
314 wasm_v128_store(c3, vacc3x4567);
315 vacc3x4567 = vacc3xCDEF;
316 c3 += 4;
317 }
318 if (nc & 4) {
319 wasm_v128_store(c0, vacc0x0123);
320 vacc0x0123 = vacc0x4567;
321 c0 += 4;
322 wasm_v128_store(c1, vacc1x0123);
323 vacc1x0123 = vacc1x4567;
324 c1 += 4;
325 wasm_v128_store(c2, vacc2x0123);
326 vacc2x0123 = vacc2x4567;
327 c2 += 4;
328 wasm_v128_store(c3, vacc3x0123);
329 vacc3x0123 = vacc3x4567;
330 c3 += 4;
331 }
332 if (nc & 2) {
333 wasm_v128_store64_lane(c0, vacc0x0123, 0);
334 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
335 c0 += 2;
336 wasm_v128_store64_lane(c1, vacc1x0123, 0);
337 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
338 c1 += 2;
339 wasm_v128_store64_lane(c2, vacc2x0123, 0);
340 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
341 c2 += 2;
342 wasm_v128_store64_lane(c3, vacc3x0123, 0);
343 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
344 c3 += 2;
345 }
346 if (nc & 1) {
347 wasm_v128_store32_lane(c0, vacc0x0123, 0);
348 wasm_v128_store32_lane(c1, vacc1x0123, 0);
349 wasm_v128_store32_lane(c2, vacc2x0123, 0);
350 wasm_v128_store32_lane(c3, vacc3x0123, 0);
351 }
352 nc = 0;
353 }
354 } while (nc != 0);
355 }