Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x16c4-minmax-fp32-wasmsdot.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__wasmsdot(
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 do {
69 v128_t vacc0x0123 = wasm_v128_load(w);
70 v128_t vacc0x4567 = wasm_v128_load((const int32_t*) w + 4);
71 v128_t vacc0x89AB = wasm_v128_load((const int32_t*) w + 8);
72 v128_t vacc0xCDEF = wasm_v128_load((const int32_t*) w + 12);
73 v128_t vacc1x0123= vacc0x0123;
74 v128_t vacc1x4567= vacc0x4567;
75 v128_t vacc1x89AB= vacc0x89AB;
76 v128_t vacc1xCDEF= vacc0xCDEF;
77 v128_t vacc2x0123= vacc0x0123;
78 v128_t vacc2x4567= vacc0x4567;
79 v128_t vacc2x89AB= vacc0x89AB;
80 v128_t vacc2xCDEF= vacc0xCDEF;
81 v128_t vacc3x0123= vacc0x0123;
82 v128_t vacc3x4567= vacc0x4567;
83 v128_t vacc3x89AB= vacc0x89AB;
84 v128_t vacc3xCDEF= vacc0xCDEF;
85 w = (const int32_t*) w + 16;
86
87 size_t p = ks;
88 do {
89 const int8_t* restrict a0 = a[0];
90 if XNN_UNPREDICTABLE(a0 != zero) {
91 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
92 }
93 const int8_t* restrict a1 = a[1];
94 if XNN_UNPREDICTABLE(a1 != zero) {
95 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
96 }
97 const int8_t* restrict a2 = a[2];
98 if XNN_UNPREDICTABLE(a2 != zero) {
99 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
100 }
101 const int8_t* restrict a3 = a[3];
102 if XNN_UNPREDICTABLE(a3 != zero) {
103 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
104 }
105 a += 4;
106
107 size_t k = kc;
108
109 while (k != 0) {
110 v128_t va0x0123 = wasm_v128_load32_splat(a0);
111 a0 += 4;
112 v128_t va1x0123 = wasm_v128_load32_splat(a1);
113 a1 += 4;
114 v128_t va2x0123 = wasm_v128_load32_splat(a2);
115 a2 += 4;
116 v128_t va3x0123 = wasm_v128_load32_splat(a3);
117 a3 += 4;
118
119
120 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
121 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
122 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
123 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
124
125 vacc0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0123);
126 vacc0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x4567);
127 vacc0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x89AB);
128 vacc0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0xCDEF);
129 vacc1x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va1x0123, vacc1x0123);
130 vacc1x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va1x0123, vacc1x4567);
131 vacc1x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va1x0123, vacc1x89AB);
132 vacc1xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va1x0123, vacc1xCDEF);
133 vacc2x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va2x0123, vacc2x0123);
134 vacc2x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va2x0123, vacc2x4567);
135 vacc2x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va2x0123, vacc2x89AB);
136 vacc2xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va2x0123, vacc2xCDEF);
137 vacc3x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va3x0123, vacc3x0123);
138 vacc3x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va3x0123, vacc3x4567);
139 vacc3x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va3x0123, vacc3x89AB);
140 vacc3xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va3x0123, vacc3xCDEF);
141
142 w = (const int8_t*) w + 64;
143 k -= 4 * sizeof(int8_t);
144 }
145
146 p -= 4 * sizeof(void*);
147 } while (p != 0);
148
149 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
150 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
151 vacc0x89AB = wasm_f32x4_convert_i32x4(vacc0x89AB);
152 vacc0xCDEF = wasm_f32x4_convert_i32x4(vacc0xCDEF);
153 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
154 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
155 vacc1x89AB = wasm_f32x4_convert_i32x4(vacc1x89AB);
156 vacc1xCDEF = wasm_f32x4_convert_i32x4(vacc1xCDEF);
157 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
158 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
159 vacc2x89AB = wasm_f32x4_convert_i32x4(vacc2x89AB);
160 vacc2xCDEF = wasm_f32x4_convert_i32x4(vacc2xCDEF);
161 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
162 vacc3x4567 = wasm_f32x4_convert_i32x4(vacc3x4567);
163 vacc3x89AB = wasm_f32x4_convert_i32x4(vacc3x89AB);
164 vacc3xCDEF = wasm_f32x4_convert_i32x4(vacc3xCDEF);
165
166 const v128_t vscale0123 = wasm_v128_load(w);
167 w = (const float*) w + 4;
168 const v128_t vscale4567 = wasm_v128_load(w);
169 w = (const float*) w + 4;
170 const v128_t vscale89AB = wasm_v128_load(w);
171 w = (const float*) w + 4;
172 const v128_t vscaleCDEF = wasm_v128_load(w);
173 w = (const float*) w + 4;
174 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
175 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
176 vacc0x89AB = wasm_f32x4_mul(vacc0x89AB, vscale89AB);
177 vacc0xCDEF = wasm_f32x4_mul(vacc0xCDEF, vscaleCDEF);
178 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
179 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
180 vacc1x89AB = wasm_f32x4_mul(vacc1x89AB, vscale89AB);
181 vacc1xCDEF = wasm_f32x4_mul(vacc1xCDEF, vscaleCDEF);
182 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
183 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vscale4567);
184 vacc2x89AB = wasm_f32x4_mul(vacc2x89AB, vscale89AB);
185 vacc2xCDEF = wasm_f32x4_mul(vacc2xCDEF, vscaleCDEF);
186 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale0123);
187 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vscale4567);
188 vacc3x89AB = wasm_f32x4_mul(vacc3x89AB, vscale89AB);
189 vacc3xCDEF = wasm_f32x4_mul(vacc3xCDEF, vscaleCDEF);
190
191 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
192 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
193 vacc0x89AB = wasm_f32x4_add(vacc0x89AB, vmagic_bias);
194 vacc0xCDEF = wasm_f32x4_add(vacc0xCDEF, vmagic_bias);
195 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
196 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
197 vacc1x89AB = wasm_f32x4_add(vacc1x89AB, vmagic_bias);
198 vacc1xCDEF = wasm_f32x4_add(vacc1xCDEF, vmagic_bias);
199 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
200 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vmagic_bias);
201 vacc2x89AB = wasm_f32x4_add(vacc2x89AB, vmagic_bias);
202 vacc2xCDEF = wasm_f32x4_add(vacc2xCDEF, vmagic_bias);
203 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
204 vacc3x4567 = wasm_f32x4_add(vacc3x4567, vmagic_bias);
205 vacc3x89AB = wasm_f32x4_add(vacc3x89AB, vmagic_bias);
206 vacc3xCDEF = wasm_f32x4_add(vacc3xCDEF, vmagic_bias);
207
208 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
209 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
210 vacc0x89AB = wasm_i32x4_max(vacc0x89AB, vmagic_min);
211 vacc0xCDEF = wasm_i32x4_max(vacc0xCDEF, vmagic_min);
212 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
213 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
214 vacc1x89AB = wasm_i32x4_max(vacc1x89AB, vmagic_min);
215 vacc1xCDEF = wasm_i32x4_max(vacc1xCDEF, vmagic_min);
216 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
217 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vmagic_min);
218 vacc2x89AB = wasm_i32x4_max(vacc2x89AB, vmagic_min);
219 vacc2xCDEF = wasm_i32x4_max(vacc2xCDEF, vmagic_min);
220 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
221 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vmagic_min);
222 vacc3x89AB = wasm_i32x4_max(vacc3x89AB, vmagic_min);
223 vacc3xCDEF = wasm_i32x4_max(vacc3xCDEF, vmagic_min);
224
225 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
226 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
227 vacc0x89AB = wasm_i32x4_sub(vacc0x89AB, vmagic_bias_less_output_zero_point);
228 vacc0xCDEF = wasm_i32x4_sub(vacc0xCDEF, vmagic_bias_less_output_zero_point);
229 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
230 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
231 vacc1x89AB = wasm_i32x4_sub(vacc1x89AB, vmagic_bias_less_output_zero_point);
232 vacc1xCDEF = wasm_i32x4_sub(vacc1xCDEF, vmagic_bias_less_output_zero_point);
233 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
234 vacc2x4567 = wasm_i32x4_sub(vacc2x4567, vmagic_bias_less_output_zero_point);
235 vacc2x89AB = wasm_i32x4_sub(vacc2x89AB, vmagic_bias_less_output_zero_point);
236 vacc2xCDEF = wasm_i32x4_sub(vacc2xCDEF, vmagic_bias_less_output_zero_point);
237 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
238 vacc3x4567 = wasm_i32x4_sub(vacc3x4567, vmagic_bias_less_output_zero_point);
239 vacc3x89AB = wasm_i32x4_sub(vacc3x89AB, vmagic_bias_less_output_zero_point);
240 vacc3xCDEF = wasm_i32x4_sub(vacc3xCDEF, vmagic_bias_less_output_zero_point);
241
242 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
243 v128_t vacc0x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc0x89AB, vacc0xCDEF);
244 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
245 v128_t vacc1x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc1x89AB, vacc1xCDEF);
246 v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x4567);
247 v128_t vacc2x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc2x89AB, vacc2xCDEF);
248 v128_t vacc3x01234567 = wasm_i16x8_narrow_i32x4(vacc3x0123, vacc3x4567);
249 v128_t vacc3x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc3x89AB, vacc3xCDEF);
250
251 vacc0x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc0x01234567);
252 vacc0x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc0x89ABCDEF, vacc0x89ABCDEF);
253 vacc1x01234567 = wasm_i8x16_narrow_i16x8(vacc1x01234567, vacc1x01234567);
254 vacc1x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc1x89ABCDEF, vacc1x89ABCDEF);
255 vacc2x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc2x01234567);
256 vacc2x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc2x89ABCDEF, vacc2x89ABCDEF);
257 vacc3x01234567 = wasm_i8x16_narrow_i16x8(vacc3x01234567, vacc3x01234567);
258 vacc3x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc3x89ABCDEF, vacc3x89ABCDEF);
259
260 vacc0x01234567 = wasm_i8x16_min(vacc0x01234567, voutput_max);
261 vacc0x89ABCDEF = wasm_i8x16_min(vacc0x89ABCDEF, voutput_max);
262 vacc1x01234567 = wasm_i8x16_min(vacc1x01234567, voutput_max);
263 vacc1x89ABCDEF = wasm_i8x16_min(vacc1x89ABCDEF, voutput_max);
264 vacc2x01234567 = wasm_i8x16_min(vacc2x01234567, voutput_max);
265 vacc2x89ABCDEF = wasm_i8x16_min(vacc2x89ABCDEF, voutput_max);
266 vacc3x01234567 = wasm_i8x16_min(vacc3x01234567, voutput_max);
267 vacc3x89ABCDEF = wasm_i8x16_min(vacc3x89ABCDEF, voutput_max);
268
269 if (nc >= 16) {
270 wasm_v128_store64_lane(c3, vacc3x01234567, 0);
271 wasm_v128_store64_lane(c3 + 8, vacc3x89ABCDEF, 0);
272 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
273 wasm_v128_store64_lane(c2 + 8, vacc2x89ABCDEF, 0);
274 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
275 wasm_v128_store64_lane(c1 + 8, vacc1x89ABCDEF, 0);
276 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
277 wasm_v128_store64_lane(c0 + 8, vacc0x89ABCDEF, 0);
278
279 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
280 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
281 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
282 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
283
284 a = (const int8_t**restrict) ((uintptr_t) a - ks);
285
286 nc -= 16;
287 } else {
288 if (nc & 8) {
289 wasm_v128_store64_lane(c3, vacc3x01234567, 0);
290 c3 += 8;
291 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
292 c2 += 8;
293 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
294 c1 += 8;
295 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
296 c0 += 8;
297
298 vacc0x01234567 = vacc0x89ABCDEF;
299 vacc1x01234567 = vacc1x89ABCDEF;
300 vacc2x01234567 = vacc2x89ABCDEF;
301 vacc3x01234567 = vacc3x89ABCDEF;
302 }
303 if (nc & 4) {
304 wasm_v128_store32_lane(c3, vacc3x01234567, 0);
305 c3 += 4;
306 wasm_v128_store32_lane(c2, vacc2x01234567, 0);
307 c2 += 4;
308 wasm_v128_store32_lane(c1, vacc1x01234567, 0);
309 c1 += 4;
310 wasm_v128_store32_lane(c0, vacc0x01234567, 0);
311 c0 += 4;
312
313 vacc0x01234567 = wasm_u64x2_shr(vacc0x01234567, 32);
314 vacc1x01234567 = wasm_u64x2_shr(vacc1x01234567, 32);
315 vacc2x01234567 = wasm_u64x2_shr(vacc2x01234567, 32);
316 vacc3x01234567 = wasm_u64x2_shr(vacc3x01234567, 32);
317 }
318 if (nc & 2) {
319 wasm_v128_store16_lane(c3, vacc3x01234567, 0);
320 c3 += 2;
321 wasm_v128_store16_lane(c2, vacc2x01234567, 0);
322 c2 += 2;
323 wasm_v128_store16_lane(c1, vacc1x01234567, 0);
324 c1 += 2;
325 wasm_v128_store16_lane(c0, vacc0x01234567, 0);
326 c0 += 2;
327
328 vacc0x01234567 = wasm_u32x4_shr(vacc0x01234567, 16);
329 vacc1x01234567 = wasm_u32x4_shr(vacc1x01234567, 16);
330 vacc2x01234567 = wasm_u32x4_shr(vacc2x01234567, 16);
331 vacc3x01234567 = wasm_u32x4_shr(vacc3x01234567, 16);
332 }
333 if (nc & 1) {
334 wasm_v128_store8_lane(c3, vacc3x01234567, 0);
335 wasm_v128_store8_lane(c2, vacc2x01234567, 0);
336 wasm_v128_store8_lane(c1, vacc1x01234567, 0);
337 wasm_v128_store8_lane(c0, vacc0x01234567, 0);
338 }
339
340 nc = 0;
341 }
342 } while (nc != 0);
343 }