Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x4c2-minmax-fp32-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_qs8_qc8w_igemm_minmax_fp32_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 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, 2 * 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
59 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
60 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
61 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
62 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
63 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
64 XNN_FORCE_REALIZATION(vmagic_bias);
65 XNN_FORCE_REALIZATION(vmagic_min);
66 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
67 XNN_FORCE_REALIZATION(voutput_max);
68
69
70 do {
71 v128_t vacc0x0123 = wasm_v128_load(w);
72 v128_t vacc1x0123 = vacc0x0123;
73 v128_t vacc2x0123 = vacc0x0123;
74 v128_t vacc3x0123 = vacc0x0123;
75 w = (const void*) ((const int32_t*) w + 4);
76
77 size_t p = ks;
78 do {
79 const int8_t* restrict a0 = a[0];
80 if XNN_UNPREDICTABLE(a0 != zero) {
81 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
82 }
83 const int8_t* restrict a1 = a[1];
84 if XNN_UNPREDICTABLE(a1 != zero) {
85 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
86 }
87 const int8_t* restrict a2 = a[2];
88 if XNN_UNPREDICTABLE(a2 != zero) {
89 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
90 }
91 const int8_t* restrict a3 = a[3];
92 if XNN_UNPREDICTABLE(a3 != zero) {
93 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
94 }
95 a += 4;
96
97 size_t k = kc;
98 while (k >= 8 * sizeof(int8_t)) {
99 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
100 a0 += 8;
101 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
102 a1 += 8;
103 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
104 a2 += 8;
105 const v128_t vxa3 = wasm_i16x8_load8x8(a3);
106 a3 += 8;
107
108 const v128_t vxb0 = wasm_i16x8_load8x8(w);
109
110 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
111 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
112 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
113 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
114 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
115 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
116 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
117 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
118 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
119
120 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
121 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
122 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
123 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
124 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
125 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
126 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
127 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
128 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
129
130 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
131 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
132 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
133 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
134 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
135 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
136 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
137 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
138 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
139
140 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
141 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
142 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
143 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
144 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
145 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 3, 3, 3, 3), vxb3));
146 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
147 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 3, 3, 3, 3), vxb3));
148
149 w = (const void*) ((const int8_t*) w + 32);
150 k -= 8 * sizeof(int8_t);
151 }
152 if (k != 0) {
153 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
154 a0 = (const int8_t*) ((uintptr_t) a0 + k);
155 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
156 a1 = (const int8_t*) ((uintptr_t) a1 + k);
157 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
158 a2 = (const int8_t*) ((uintptr_t) a2 + k);
159 const v128_t vxa3 = wasm_i16x8_load8x8(a3);
160 a3 = (const int8_t*) ((uintptr_t) a3 + k);
161
162 const v128_t vxb0 = wasm_i16x8_load8x8(w);
163 w = (const int8_t*) w + 8;
164
165 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
166 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
167 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
168 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
169 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
170 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
171 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
172 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
173
174 if (k > 2 * sizeof(int8_t)) {
175 const v128_t vxb1 = wasm_i16x8_load8x8(w);
176 w = (const int8_t*) w + 8;
177
178 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
179 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
180 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
181 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
182 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
183 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
184 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
185 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
186
187 if (k > 4 * sizeof(int8_t)) {
188 const v128_t vxb2 = wasm_i16x8_load8x8(w);
189 w = (const int8_t*) w + 8;
190
191 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
192 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
193 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
194 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
195 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
196 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
197 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
198 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
199 }
200 }
201 }
202 p -= 4 * sizeof(void*);
203 } while (p != 0);
204
205 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
206 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
207 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
208 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
209
210 const v128_t vscale0123 = wasm_v128_load(w);
211 w = (const float*) w + 4;
212 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
213 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
214 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
215 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale0123);
216
217 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
218 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
219 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
220 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
221
222 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
223 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
224 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
225 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
226
227 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
228 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
229 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
230 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
231
232 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
233 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
234
235 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
236
237 vout = wasm_i8x16_min(vout, voutput_max);
238
239 if (nc >= 4) {
240 wasm_v128_store32_lane(c3, vout, 3);
241 wasm_v128_store32_lane(c2, vout, 2);
242 wasm_v128_store32_lane(c1, vout, 1);
243 wasm_v128_store32_lane(c0, vout, 0);
244
245 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
246 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
247 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
248 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
249
250 a = (const int8_t**restrict) ((uintptr_t) a - ks);
251
252 nc -= 4;
253 } else {
254 if (nc & 2) {
255 wasm_v128_store16_lane(c3, vout, 6);
256 c3 += 2;
257 wasm_v128_store16_lane(c2, vout, 4);
258 c2 += 2;
259 wasm_v128_store16_lane(c1, vout, 2);
260 c1 += 2;
261 wasm_v128_store16_lane(c0, vout, 0);
262 c0 += 2;
263
264 vout = wasm_u32x4_shr(vout, 16);
265 }
266 if (nc & 1) {
267 wasm_v128_store8_lane(c3, vout, 12);
268 wasm_v128_store8_lane(c2, vout, 8);
269 wasm_v128_store8_lane(c1, vout, 4);
270 wasm_v128_store8_lane(c0, vout, 0);
271 }
272
273 nc = 0;
274 }
275 } while (nc != 0);
276 }