Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-4x4c2-minmax-wasmsimd-dot16x2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-wasmsimd-dot16x2.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 <wasm_simd128.h>
13
14 #include "xnnpack/igemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_4x4c2__wasmsimd_dot16x2_ld64(
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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (4 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 2 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
61 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
62 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
63 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
64 XNN_FORCE_REALIZATION(vmin);
65 XNN_FORCE_REALIZATION(vmax);
66
67
68 do {
69 const v128_t vksum0123 = wasm_v128_load(w);
70 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
71 v128_t vacc1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
72 v128_t vacc2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
73 v128_t vacc3x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
74 w = (const void*) ((const int32_t*) w + 4);
75
76 size_t p = ks;
77 do {
78 const int8_t* restrict a0 = a[0];
79 if XNN_UNPREDICTABLE(a0 != zero) {
80 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
81 } else {
82 a0 = zero_data;
83 }
84 const int8_t* restrict a1 = a[1];
85 if XNN_UNPREDICTABLE(a1 != zero) {
86 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
87 } else {
88 a1 = zero_data;
89 }
90 const int8_t* restrict a2 = a[2];
91 if XNN_UNPREDICTABLE(a2 != zero) {
92 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
93 } else {
94 a2 = zero_data;
95 }
96 const int8_t* restrict a3 = a[3];
97 if XNN_UNPREDICTABLE(a3 != zero) {
98 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
99 } else {
100 a3 = zero_data;
101 }
102 a += 4;
103
104 size_t k = kc;
105 while (k >= 8 * sizeof(int8_t)) {
106 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
107 a0 += 8;
108 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
109 a1 += 8;
110 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
111 a2 += 8;
112 const v128_t vxa3 = wasm_i16x8_load8x8(a3);
113 a3 += 8;
114
115 const v128_t vxb0 = wasm_i16x8_load8x8(w);
116
117 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
118 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
119 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
120 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
121 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
122 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
123 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
124 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
125 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
126
127 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
128 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
129 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
130 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
131 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
132 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
133 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
134 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
135 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
136
137 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
138 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
139 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
140 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
141 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
142 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
143 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
144 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
145 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
146
147 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
148 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
149 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
150 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
151 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
152 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 3, 3, 3, 3), vxb3));
153 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
154 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 3, 3, 3, 3), vxb3));
155
156 w = (const void*) ((const int8_t*) w + 32);
157 k -= 8 * sizeof(int8_t);
158 }
159 if (k != 0) {
160 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
161 a0 = (const int8_t*) ((uintptr_t) a0 + k);
162 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
163 a1 = (const int8_t*) ((uintptr_t) a1 + k);
164 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
165 a2 = (const int8_t*) ((uintptr_t) a2 + k);
166 const v128_t vxa3 = wasm_i16x8_load8x8(a3);
167 a3 = (const int8_t*) ((uintptr_t) a3 + k);
168
169 const v128_t vxb0 = wasm_i16x8_load8x8(w);
170 w = (const int8_t*) w + 8;
171
172 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
173 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
174 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
175 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
176 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
177 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
178 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
179 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
180
181 if (k > 2 * sizeof(int8_t)) {
182 const v128_t vxb1 = wasm_i16x8_load8x8(w);
183 w = (const int8_t*) w + 8;
184
185 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
186 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
187 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
188 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
189 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
190 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
191 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
192 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
193
194 if (k > 4 * sizeof(int8_t)) {
195 const v128_t vxb2 = wasm_i16x8_load8x8(w);
196 w = (const int8_t*) w + 8;
197
198 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
199 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
200 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
201 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
202 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
203 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
204 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
205 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
206 }
207 }
208 }
209 p -= 4 * sizeof(void*);
210 } while (p != 0);
211
212 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
213 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
214 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
215 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
216
217 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
218 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
219 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale);
220 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale);
221
222 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
223 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
224 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
225 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
226 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
227 w = (const float*) w + 4;
228
229 const v128_t vbias0123 = wasm_v128_load(w);
230 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
231 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
232 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
233 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
234 w = (const float*) w + 4;
235
236 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
237 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
238 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
239 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
240
241 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
242 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
243 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
244 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
245
246 if XNN_LIKELY(nc >= 4) {
247 wasm_v128_store(c3, vacc3x0123);
248 wasm_v128_store(c2, vacc2x0123);
249 wasm_v128_store(c1, vacc1x0123);
250 wasm_v128_store(c0, vacc0x0123);
251
252 a = (const int8_t**restrict) ((uintptr_t) a - ks);
253
254 c0 = (float*) ((uintptr_t) c0 + cn_stride);
255 c1 = (float*) ((uintptr_t) c1 + cn_stride);
256 c2 = (float*) ((uintptr_t) c2 + cn_stride);
257 c3 = (float*) ((uintptr_t) c3 + cn_stride);
258
259 nc -= 4;
260 } else {
261 if (nc & 2) {
262 wasm_v128_store64_lane(c3, vacc3x0123, 0);
263 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
264 c3 += 2;
265 wasm_v128_store64_lane(c2, vacc2x0123, 0);
266 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
267 c2 += 2;
268 wasm_v128_store64_lane(c1, vacc1x0123, 0);
269 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
270 c1 += 2;
271 wasm_v128_store64_lane(c0, vacc0x0123, 0);
272 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
273 c0 += 2;
274 }
275 if (nc & 1) {
276 wasm_v128_store32_lane(c3, vacc3x0123, 0);
277 wasm_v128_store32_lane(c2, vacc2x0123, 0);
278 wasm_v128_store32_lane(c1, vacc1x0123, 0);
279 wasm_v128_store32_lane(c0, vacc0x0123, 0);
280 }
281 nc = 0;
282 }
283 } while (nc != 0);
284 }