Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x16c4-minmax-fp32-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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
18 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_4x16c4__wasmusdot(
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_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 4 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
59 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
60 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
61 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
62 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
63 //XNN_FORCE_REALIZATION(vmagic_bias);
64 //XNN_FORCE_REALIZATION(vmagic_min);
65 //XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
66 //XNN_FORCE_REALIZATION(voutput_max);
67
68 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
69 XNN_FORCE_REALIZATION(vsign_mask);
70 do {
71 v128_t vacc0x0123 = wasm_v128_load(w);
72 v128_t vacc0x4567 = wasm_v128_load((const int32_t*) w + 4);
73 v128_t vacc0x89AB = wasm_v128_load((const int32_t*) w + 8);
74 v128_t vacc0xCDEF = wasm_v128_load((const int32_t*) w + 12);
75 v128_t vacc1x0123= vacc0x0123;
76 v128_t vacc1x4567= vacc0x4567;
77 v128_t vacc1x89AB= vacc0x89AB;
78 v128_t vacc1xCDEF= vacc0xCDEF;
79 v128_t vacc2x0123= vacc0x0123;
80 v128_t vacc2x4567= vacc0x4567;
81 v128_t vacc2x89AB= vacc0x89AB;
82 v128_t vacc2xCDEF= vacc0xCDEF;
83 v128_t vacc3x0123= vacc0x0123;
84 v128_t vacc3x4567= vacc0x4567;
85 v128_t vacc3x89AB= vacc0x89AB;
86 v128_t vacc3xCDEF= vacc0xCDEF;
87 w = (const int32_t*) w + 16;
88
89 size_t p = ks;
90 do {
91 const int8_t* restrict a0 = a[0];
92 if XNN_UNPREDICTABLE(a0 != zero) {
93 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
94 }
95 const int8_t* restrict a1 = a[1];
96 if XNN_UNPREDICTABLE(a1 != zero) {
97 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
98 }
99 const int8_t* restrict a2 = a[2];
100 if XNN_UNPREDICTABLE(a2 != zero) {
101 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
102 }
103 const int8_t* restrict a3 = a[3];
104 if XNN_UNPREDICTABLE(a3 != zero) {
105 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
106 }
107 a += 4;
108
109 size_t k = kc;
110
111 while (k != 0) {
112 v128_t va0x0123 = wasm_v128_load32_splat(a0);
113 a0 += 4;
114 v128_t va1x0123 = wasm_v128_load32_splat(a1);
115 a1 += 4;
116 v128_t va2x0123 = wasm_v128_load32_splat(a2);
117 a2 += 4;
118 v128_t va3x0123 = wasm_v128_load32_splat(a3);
119 a3 += 4;
120
121 va0x0123 = wasm_v128_xor(va0x0123, vsign_mask);
122 va1x0123 = wasm_v128_xor(va1x0123, vsign_mask);
123 va2x0123 = wasm_v128_xor(va2x0123, vsign_mask);
124 va3x0123 = wasm_v128_xor(va3x0123, vsign_mask);
125
126 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
127 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
128 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
129 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
130
131 vacc0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0123);
132 vacc0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x4567);
133 vacc0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x89AB);
134 vacc0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0xCDEF);
135 vacc1x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va1x0123, vacc1x0123);
136 vacc1x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va1x0123, vacc1x4567);
137 vacc1x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va1x0123, vacc1x89AB);
138 vacc1xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va1x0123, vacc1xCDEF);
139 vacc2x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va2x0123, vacc2x0123);
140 vacc2x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va2x0123, vacc2x4567);
141 vacc2x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va2x0123, vacc2x89AB);
142 vacc2xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va2x0123, vacc2xCDEF);
143 vacc3x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va3x0123, vacc3x0123);
144 vacc3x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va3x0123, vacc3x4567);
145 vacc3x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va3x0123, vacc3x89AB);
146 vacc3xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va3x0123, vacc3xCDEF);
147
148 w = (const int8_t*) w + 64;
149 k -= 4 * sizeof(int8_t);
150 }
151
152 p -= 4 * sizeof(void*);
153 } while (p != 0);
154
155 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
156 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
157 vacc0x89AB = wasm_f32x4_convert_i32x4(vacc0x89AB);
158 vacc0xCDEF = wasm_f32x4_convert_i32x4(vacc0xCDEF);
159 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
160 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
161 vacc1x89AB = wasm_f32x4_convert_i32x4(vacc1x89AB);
162 vacc1xCDEF = wasm_f32x4_convert_i32x4(vacc1xCDEF);
163 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
164 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
165 vacc2x89AB = wasm_f32x4_convert_i32x4(vacc2x89AB);
166 vacc2xCDEF = wasm_f32x4_convert_i32x4(vacc2xCDEF);
167 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
168 vacc3x4567 = wasm_f32x4_convert_i32x4(vacc3x4567);
169 vacc3x89AB = wasm_f32x4_convert_i32x4(vacc3x89AB);
170 vacc3xCDEF = wasm_f32x4_convert_i32x4(vacc3xCDEF);
171
172 const v128_t vscale0123 = wasm_v128_load(w);
173 w = (const float*) w + 4;
174 const v128_t vscale4567 = wasm_v128_load(w);
175 w = (const float*) w + 4;
176 const v128_t vscale89AB = wasm_v128_load(w);
177 w = (const float*) w + 4;
178 const v128_t vscaleCDEF = wasm_v128_load(w);
179 w = (const float*) w + 4;
180 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
181 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
182 vacc0x89AB = wasm_f32x4_mul(vacc0x89AB, vscale89AB);
183 vacc0xCDEF = wasm_f32x4_mul(vacc0xCDEF, vscaleCDEF);
184 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
185 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
186 vacc1x89AB = wasm_f32x4_mul(vacc1x89AB, vscale89AB);
187 vacc1xCDEF = wasm_f32x4_mul(vacc1xCDEF, vscaleCDEF);
188 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
189 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vscale4567);
190 vacc2x89AB = wasm_f32x4_mul(vacc2x89AB, vscale89AB);
191 vacc2xCDEF = wasm_f32x4_mul(vacc2xCDEF, vscaleCDEF);
192 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale0123);
193 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vscale4567);
194 vacc3x89AB = wasm_f32x4_mul(vacc3x89AB, vscale89AB);
195 vacc3xCDEF = wasm_f32x4_mul(vacc3xCDEF, vscaleCDEF);
196
197 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
198 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
199 vacc0x89AB = wasm_f32x4_add(vacc0x89AB, vmagic_bias);
200 vacc0xCDEF = wasm_f32x4_add(vacc0xCDEF, vmagic_bias);
201 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
202 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
203 vacc1x89AB = wasm_f32x4_add(vacc1x89AB, vmagic_bias);
204 vacc1xCDEF = wasm_f32x4_add(vacc1xCDEF, vmagic_bias);
205 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
206 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vmagic_bias);
207 vacc2x89AB = wasm_f32x4_add(vacc2x89AB, vmagic_bias);
208 vacc2xCDEF = wasm_f32x4_add(vacc2xCDEF, vmagic_bias);
209 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
210 vacc3x4567 = wasm_f32x4_add(vacc3x4567, vmagic_bias);
211 vacc3x89AB = wasm_f32x4_add(vacc3x89AB, vmagic_bias);
212 vacc3xCDEF = wasm_f32x4_add(vacc3xCDEF, vmagic_bias);
213
214 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
215 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
216 vacc0x89AB = wasm_i32x4_max(vacc0x89AB, vmagic_min);
217 vacc0xCDEF = wasm_i32x4_max(vacc0xCDEF, vmagic_min);
218 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
219 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
220 vacc1x89AB = wasm_i32x4_max(vacc1x89AB, vmagic_min);
221 vacc1xCDEF = wasm_i32x4_max(vacc1xCDEF, vmagic_min);
222 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
223 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vmagic_min);
224 vacc2x89AB = wasm_i32x4_max(vacc2x89AB, vmagic_min);
225 vacc2xCDEF = wasm_i32x4_max(vacc2xCDEF, vmagic_min);
226 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
227 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vmagic_min);
228 vacc3x89AB = wasm_i32x4_max(vacc3x89AB, vmagic_min);
229 vacc3xCDEF = wasm_i32x4_max(vacc3xCDEF, vmagic_min);
230
231 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
232 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
233 vacc0x89AB = wasm_i32x4_sub(vacc0x89AB, vmagic_bias_less_output_zero_point);
234 vacc0xCDEF = wasm_i32x4_sub(vacc0xCDEF, vmagic_bias_less_output_zero_point);
235 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
236 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
237 vacc1x89AB = wasm_i32x4_sub(vacc1x89AB, vmagic_bias_less_output_zero_point);
238 vacc1xCDEF = wasm_i32x4_sub(vacc1xCDEF, vmagic_bias_less_output_zero_point);
239 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
240 vacc2x4567 = wasm_i32x4_sub(vacc2x4567, vmagic_bias_less_output_zero_point);
241 vacc2x89AB = wasm_i32x4_sub(vacc2x89AB, vmagic_bias_less_output_zero_point);
242 vacc2xCDEF = wasm_i32x4_sub(vacc2xCDEF, vmagic_bias_less_output_zero_point);
243 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
244 vacc3x4567 = wasm_i32x4_sub(vacc3x4567, vmagic_bias_less_output_zero_point);
245 vacc3x89AB = wasm_i32x4_sub(vacc3x89AB, vmagic_bias_less_output_zero_point);
246 vacc3xCDEF = wasm_i32x4_sub(vacc3xCDEF, vmagic_bias_less_output_zero_point);
247
248 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
249 v128_t vacc0x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc0x89AB, vacc0xCDEF);
250 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
251 v128_t vacc1x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc1x89AB, vacc1xCDEF);
252 v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x4567);
253 v128_t vacc2x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc2x89AB, vacc2xCDEF);
254 v128_t vacc3x01234567 = wasm_i16x8_narrow_i32x4(vacc3x0123, vacc3x4567);
255 v128_t vacc3x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc3x89AB, vacc3xCDEF);
256
257 vacc0x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc0x01234567);
258 vacc0x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc0x89ABCDEF, vacc0x89ABCDEF);
259 vacc1x01234567 = wasm_i8x16_narrow_i16x8(vacc1x01234567, vacc1x01234567);
260 vacc1x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc1x89ABCDEF, vacc1x89ABCDEF);
261 vacc2x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc2x01234567);
262 vacc2x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc2x89ABCDEF, vacc2x89ABCDEF);
263 vacc3x01234567 = wasm_i8x16_narrow_i16x8(vacc3x01234567, vacc3x01234567);
264 vacc3x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc3x89ABCDEF, vacc3x89ABCDEF);
265
266 vacc0x01234567 = wasm_i8x16_min(vacc0x01234567, voutput_max);
267 vacc0x89ABCDEF = wasm_i8x16_min(vacc0x89ABCDEF, voutput_max);
268 vacc1x01234567 = wasm_i8x16_min(vacc1x01234567, voutput_max);
269 vacc1x89ABCDEF = wasm_i8x16_min(vacc1x89ABCDEF, voutput_max);
270 vacc2x01234567 = wasm_i8x16_min(vacc2x01234567, voutput_max);
271 vacc2x89ABCDEF = wasm_i8x16_min(vacc2x89ABCDEF, voutput_max);
272 vacc3x01234567 = wasm_i8x16_min(vacc3x01234567, voutput_max);
273 vacc3x89ABCDEF = wasm_i8x16_min(vacc3x89ABCDEF, voutput_max);
274
275 if (nc >= 16) {
276 wasm_v128_store64_lane(c3, vacc3x01234567, 0);
277 wasm_v128_store64_lane(c3 + 8, vacc3x89ABCDEF, 0);
278 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
279 wasm_v128_store64_lane(c2 + 8, vacc2x89ABCDEF, 0);
280 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
281 wasm_v128_store64_lane(c1 + 8, vacc1x89ABCDEF, 0);
282 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
283 wasm_v128_store64_lane(c0 + 8, vacc0x89ABCDEF, 0);
284
285 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
286 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
287 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
288 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
289
290 a = (const int8_t**restrict) ((uintptr_t) a - ks);
291
292 nc -= 16;
293 } else {
294 if (nc & 8) {
295 wasm_v128_store64_lane(c3, vacc3x01234567, 0);
296 c3 += 8;
297 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
298 c2 += 8;
299 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
300 c1 += 8;
301 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
302 c0 += 8;
303
304 vacc0x01234567 = vacc0x89ABCDEF;
305 vacc1x01234567 = vacc1x89ABCDEF;
306 vacc2x01234567 = vacc2x89ABCDEF;
307 vacc3x01234567 = vacc3x89ABCDEF;
308 }
309 if (nc & 4) {
310 wasm_v128_store32_lane(c3, vacc3x01234567, 0);
311 c3 += 4;
312 wasm_v128_store32_lane(c2, vacc2x01234567, 0);
313 c2 += 4;
314 wasm_v128_store32_lane(c1, vacc1x01234567, 0);
315 c1 += 4;
316 wasm_v128_store32_lane(c0, vacc0x01234567, 0);
317 c0 += 4;
318
319 vacc0x01234567 = wasm_u64x2_shr(vacc0x01234567, 32);
320 vacc1x01234567 = wasm_u64x2_shr(vacc1x01234567, 32);
321 vacc2x01234567 = wasm_u64x2_shr(vacc2x01234567, 32);
322 vacc3x01234567 = wasm_u64x2_shr(vacc3x01234567, 32);
323 }
324 if (nc & 2) {
325 wasm_v128_store16_lane(c3, vacc3x01234567, 0);
326 c3 += 2;
327 wasm_v128_store16_lane(c2, vacc2x01234567, 0);
328 c2 += 2;
329 wasm_v128_store16_lane(c1, vacc1x01234567, 0);
330 c1 += 2;
331 wasm_v128_store16_lane(c0, vacc0x01234567, 0);
332 c0 += 2;
333
334 vacc0x01234567 = wasm_u32x4_shr(vacc0x01234567, 16);
335 vacc1x01234567 = wasm_u32x4_shr(vacc1x01234567, 16);
336 vacc2x01234567 = wasm_u32x4_shr(vacc2x01234567, 16);
337 vacc3x01234567 = wasm_u32x4_shr(vacc3x01234567, 16);
338 }
339 if (nc & 1) {
340 wasm_v128_store8_lane(c3, vacc3x01234567, 0);
341 wasm_v128_store8_lane(c2, vacc2x01234567, 0);
342 wasm_v128_store8_lane(c1, vacc1x01234567, 0);
343 wasm_v128_store8_lane(c0, vacc0x01234567, 0);
344 }
345
346 nc = 0;
347 }
348 } while (nc != 0);
349 }