Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x8c16-minmax-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c16-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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_3x8c16__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 <= 3);
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, 16 * 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
55 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
56 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
57 XNN_FORCE_REALIZATION(vmin);
58 XNN_FORCE_REALIZATION(vmax);
59
60 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
61 do {
62 v128_t vksum0 = wasm_v128_load32_zero(w);
63 v128_t vksum1 = wasm_v128_load32_zero((const int32_t*) w + 1);
64 v128_t vksum2 = wasm_v128_load32_zero((const int32_t*) w + 2);
65 v128_t vksum3 = wasm_v128_load32_zero((const int32_t*) w + 3);
66 v128_t vksum4 = wasm_v128_load32_zero((const int32_t*) w + 4);
67 v128_t vksum5 = wasm_v128_load32_zero((const int32_t*) w + 5);
68 v128_t vksum6 = wasm_v128_load32_zero((const int32_t*) w + 6);
69 v128_t vksum7 = wasm_v128_load32_zero((const int32_t*) w + 7);
70 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
71 const v128_t vinput_zero_point1 = wasm_i32x4_splat((int32_t) quantization_params[1].zero_point + 128);
72 const v128_t vinput_zero_point2 = wasm_i32x4_splat((int32_t) quantization_params[2].zero_point + 128);
73 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point0);
74 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point0);
75 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point0);
76 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point0);
77 v128_t vacc0x4 = wasm_i32x4_mul(vksum4, vinput_zero_point0);
78 v128_t vacc0x5 = wasm_i32x4_mul(vksum5, vinput_zero_point0);
79 v128_t vacc0x6 = wasm_i32x4_mul(vksum6, vinput_zero_point0);
80 v128_t vacc0x7 = wasm_i32x4_mul(vksum7, vinput_zero_point0);
81 v128_t vacc1x0 = wasm_i32x4_mul(vksum0, vinput_zero_point1);
82 v128_t vacc1x1 = wasm_i32x4_mul(vksum1, vinput_zero_point1);
83 v128_t vacc1x2 = wasm_i32x4_mul(vksum2, vinput_zero_point1);
84 v128_t vacc1x3 = wasm_i32x4_mul(vksum3, vinput_zero_point1);
85 v128_t vacc1x4 = wasm_i32x4_mul(vksum4, vinput_zero_point1);
86 v128_t vacc1x5 = wasm_i32x4_mul(vksum5, vinput_zero_point1);
87 v128_t vacc1x6 = wasm_i32x4_mul(vksum6, vinput_zero_point1);
88 v128_t vacc1x7 = wasm_i32x4_mul(vksum7, vinput_zero_point1);
89 v128_t vacc2x0 = wasm_i32x4_mul(vksum0, vinput_zero_point2);
90 v128_t vacc2x1 = wasm_i32x4_mul(vksum1, vinput_zero_point2);
91 v128_t vacc2x2 = wasm_i32x4_mul(vksum2, vinput_zero_point2);
92 v128_t vacc2x3 = wasm_i32x4_mul(vksum3, vinput_zero_point2);
93 v128_t vacc2x4 = wasm_i32x4_mul(vksum4, vinput_zero_point2);
94 v128_t vacc2x5 = wasm_i32x4_mul(vksum5, vinput_zero_point2);
95 v128_t vacc2x6 = wasm_i32x4_mul(vksum6, vinput_zero_point2);
96 v128_t vacc2x7 = wasm_i32x4_mul(vksum7, vinput_zero_point2);
97 w = (const int32_t*) w + 8;
98
99 size_t k = kc;
100 do {
101 const v128_t va0 = wasm_v128_xor(wasm_v128_load(a0), vsign_mask);
102 a0 += 16;
103 const v128_t va1 = wasm_v128_xor(wasm_v128_load(a1), vsign_mask);
104 a1 += 16;
105 const v128_t va2 = wasm_v128_xor(wasm_v128_load(a2), vsign_mask);
106 a2 += 16;
107
108 const v128_t vb0 = wasm_v128_load(w);
109
110 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
111 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
112 vacc2x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va2, vacc2x0);
113 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
114
115 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
116 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
117 vacc2x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va2, vacc2x1);
118 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
119
120 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
121 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
122 vacc2x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va2, vacc2x2);
123 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
124
125 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
126 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
127 vacc2x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va2, vacc2x3);
128 const v128_t vb4 = wasm_v128_load((const int8_t*) w + 64);
129
130 vacc0x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va0, vacc0x4);
131 vacc1x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va1, vacc1x4);
132 vacc2x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va2, vacc2x4);
133 const v128_t vb5 = wasm_v128_load((const int8_t*) w + 80);
134
135 vacc0x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va0, vacc0x5);
136 vacc1x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va1, vacc1x5);
137 vacc2x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va2, vacc2x5);
138 const v128_t vb6 = wasm_v128_load((const int8_t*) w + 96);
139
140 vacc0x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va0, vacc0x6);
141 vacc1x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va1, vacc1x6);
142 vacc2x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va2, vacc2x6);
143 const v128_t vb7 = wasm_v128_load((const int8_t*) w + 112);
144
145 vacc0x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va0, vacc0x7);
146 vacc1x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va1, vacc1x7);
147 vacc2x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va2, vacc2x7);
148
149 w = (const int8_t*) w + 128;
150 k -= 16 * sizeof(int8_t);
151 } while (k != 0);
152
153 const v128_t vacc0x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x0, vacc0x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x0, vacc0x2, 2, 6, 3, 7));
154 const v128_t vacc0x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x1, vacc0x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x1, vacc0x3, 2, 6, 3, 7));
155 const v128_t vacc0x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x4, vacc0x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x4, vacc0x6, 2, 6, 3, 7));
156 const v128_t vacc0x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x5, vacc0x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x5, vacc0x7, 2, 6, 3, 7));
157 const v128_t vacc1x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x0, vacc1x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x0, vacc1x2, 2, 6, 3, 7));
158 const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x1, vacc1x3, 2, 6, 3, 7));
159 const v128_t vacc1x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x4, vacc1x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x4, vacc1x6, 2, 6, 3, 7));
160 const v128_t vacc1x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x5, vacc1x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x5, vacc1x7, 2, 6, 3, 7));
161 const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7));
162 const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7));
163 const v128_t vacc2x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x4, vacc2x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x4, vacc2x6, 2, 6, 3, 7));
164 const v128_t vacc2x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x5, vacc2x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x5, vacc2x7, 2, 6, 3, 7));
165
166 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
167 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x46, vacc0x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x46, vacc0x57, 2, 6, 3, 7));
168 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
169 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x46, vacc1x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x46, vacc1x57, 2, 6, 3, 7));
170 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
171 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x46, vacc2x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x46, vacc2x57, 2, 6, 3, 7));
172
173 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
174 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
175 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
176 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
177 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
178 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
179
180 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
181 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
182 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
183
184 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
185 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
186 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
187 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1);
188 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
189 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale2);
190
191 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
192 w = (const float*) w + 4;
193 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
194 w = (const float*) w + 4;
195 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
196 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
197 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
198 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
199 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
200 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567);
201
202 const v128_t vbias0123 = wasm_v128_load(w);
203 w = (const float*) w + 4;
204 const v128_t vbias4567 = wasm_v128_load(w);
205 w = (const float*) w + 4;
206 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
207 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
208 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
209 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
210 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
211 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567);
212
213 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
214 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
215 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
216 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
217 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
218 vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin);
219
220 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
221 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
222 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
223 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
224 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
225 vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax);
226
227 if XNN_LIKELY(nc >= 8) {
228 wasm_v128_store(c0, vacc0x0123);
229 wasm_v128_store(c0 + 4, vacc0x4567);
230 wasm_v128_store(c1, vacc1x0123);
231 wasm_v128_store(c1 + 4, vacc1x4567);
232 wasm_v128_store(c2, vacc2x0123);
233 wasm_v128_store(c2 + 4, vacc2x4567);
234
235 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
236 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
237 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
238
239 c0 = (float*) ((uintptr_t) c0 + cn_stride);
240 c1 = (float*) ((uintptr_t) c1 + cn_stride);
241 c2 = (float*) ((uintptr_t) c2 + cn_stride);
242
243 nc -= 8;
244 } else {
245 if (nc & 4) {
246 wasm_v128_store(c0, vacc0x0123);
247 vacc0x0123 = vacc0x4567;
248 c0 += 4;
249 wasm_v128_store(c1, vacc1x0123);
250 vacc1x0123 = vacc1x4567;
251 c1 += 4;
252 wasm_v128_store(c2, vacc2x0123);
253 vacc2x0123 = vacc2x4567;
254 c2 += 4;
255 }
256 if (nc & 2) {
257 wasm_v128_store64_lane(c0, vacc0x0123, 0);
258 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
259 c0 += 2;
260 wasm_v128_store64_lane(c1, vacc1x0123, 0);
261 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
262 c1 += 2;
263 wasm_v128_store64_lane(c2, vacc2x0123, 0);
264 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
265 c2 += 2;
266 }
267 if (nc & 1) {
268 wasm_v128_store32_lane(c0, vacc0x0123, 0);
269 wasm_v128_store32_lane(c1, vacc1x0123, 0);
270 wasm_v128_store32_lane(c2, vacc2x0123, 0);
271 }
272 nc = 0;
273 }
274 } while (nc != 0);
275 }