Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x16c4-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_3x16c4__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 <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * 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
54 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
55 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
56 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
57 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
58 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
59 //XNN_FORCE_REALIZATION(vmagic_bias);
60 //XNN_FORCE_REALIZATION(vmagic_min);
61 //XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
62 //XNN_FORCE_REALIZATION(voutput_max);
63
64 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
65 XNN_FORCE_REALIZATION(vsign_mask);
66 do {
67 v128_t vacc0x0123 = wasm_v128_load(w);
68 v128_t vacc0x4567 = wasm_v128_load((const int32_t*) w + 4);
69 v128_t vacc0x89AB = wasm_v128_load((const int32_t*) w + 8);
70 v128_t vacc0xCDEF = wasm_v128_load((const int32_t*) w + 12);
71 v128_t vacc1x0123= vacc0x0123;
72 v128_t vacc1x4567= vacc0x4567;
73 v128_t vacc1x89AB= vacc0x89AB;
74 v128_t vacc1xCDEF= vacc0xCDEF;
75 v128_t vacc2x0123= vacc0x0123;
76 v128_t vacc2x4567= vacc0x4567;
77 v128_t vacc2x89AB= vacc0x89AB;
78 v128_t vacc2xCDEF= vacc0xCDEF;
79 w = (const int32_t*) w + 16;
80
81 size_t p = ks;
82 do {
83 const int8_t* restrict a0 = a[0];
84 if XNN_UNPREDICTABLE(a0 != zero) {
85 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
86 }
87 const int8_t* restrict a1 = a[1];
88 if XNN_UNPREDICTABLE(a1 != zero) {
89 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
90 }
91 const int8_t* restrict a2 = a[2];
92 if XNN_UNPREDICTABLE(a2 != zero) {
93 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
94 }
95 a += 3;
96
97 size_t k = kc;
98
99 while (k != 0) {
100 v128_t va0x0123 = wasm_v128_load32_splat(a0);
101 a0 += 4;
102 v128_t va1x0123 = wasm_v128_load32_splat(a1);
103 a1 += 4;
104 v128_t va2x0123 = wasm_v128_load32_splat(a2);
105 a2 += 4;
106
107 va0x0123 = wasm_v128_xor(va0x0123, vsign_mask);
108 va1x0123 = wasm_v128_xor(va1x0123, vsign_mask);
109 va2x0123 = wasm_v128_xor(va2x0123, vsign_mask);
110
111 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
112 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
113 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
114 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
115
116 vacc0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0123);
117 vacc0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x4567);
118 vacc0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x89AB);
119 vacc0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0xCDEF);
120 vacc1x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va1x0123, vacc1x0123);
121 vacc1x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va1x0123, vacc1x4567);
122 vacc1x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va1x0123, vacc1x89AB);
123 vacc1xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va1x0123, vacc1xCDEF);
124 vacc2x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va2x0123, vacc2x0123);
125 vacc2x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va2x0123, vacc2x4567);
126 vacc2x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va2x0123, vacc2x89AB);
127 vacc2xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va2x0123, vacc2xCDEF);
128
129 w = (const int8_t*) w + 64;
130 k -= 4 * sizeof(int8_t);
131 }
132
133 p -= 3 * sizeof(void*);
134 } while (p != 0);
135
136 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
137 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
138 vacc0x89AB = wasm_f32x4_convert_i32x4(vacc0x89AB);
139 vacc0xCDEF = wasm_f32x4_convert_i32x4(vacc0xCDEF);
140 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
141 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
142 vacc1x89AB = wasm_f32x4_convert_i32x4(vacc1x89AB);
143 vacc1xCDEF = wasm_f32x4_convert_i32x4(vacc1xCDEF);
144 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
145 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
146 vacc2x89AB = wasm_f32x4_convert_i32x4(vacc2x89AB);
147 vacc2xCDEF = wasm_f32x4_convert_i32x4(vacc2xCDEF);
148
149 const v128_t vscale0123 = wasm_v128_load(w);
150 w = (const float*) w + 4;
151 const v128_t vscale4567 = wasm_v128_load(w);
152 w = (const float*) w + 4;
153 const v128_t vscale89AB = wasm_v128_load(w);
154 w = (const float*) w + 4;
155 const v128_t vscaleCDEF = wasm_v128_load(w);
156 w = (const float*) w + 4;
157 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
158 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
159 vacc0x89AB = wasm_f32x4_mul(vacc0x89AB, vscale89AB);
160 vacc0xCDEF = wasm_f32x4_mul(vacc0xCDEF, vscaleCDEF);
161 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
162 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
163 vacc1x89AB = wasm_f32x4_mul(vacc1x89AB, vscale89AB);
164 vacc1xCDEF = wasm_f32x4_mul(vacc1xCDEF, vscaleCDEF);
165 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
166 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vscale4567);
167 vacc2x89AB = wasm_f32x4_mul(vacc2x89AB, vscale89AB);
168 vacc2xCDEF = wasm_f32x4_mul(vacc2xCDEF, vscaleCDEF);
169
170 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
171 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
172 vacc0x89AB = wasm_f32x4_add(vacc0x89AB, vmagic_bias);
173 vacc0xCDEF = wasm_f32x4_add(vacc0xCDEF, vmagic_bias);
174 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
175 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
176 vacc1x89AB = wasm_f32x4_add(vacc1x89AB, vmagic_bias);
177 vacc1xCDEF = wasm_f32x4_add(vacc1xCDEF, vmagic_bias);
178 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
179 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vmagic_bias);
180 vacc2x89AB = wasm_f32x4_add(vacc2x89AB, vmagic_bias);
181 vacc2xCDEF = wasm_f32x4_add(vacc2xCDEF, vmagic_bias);
182
183 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
184 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
185 vacc0x89AB = wasm_i32x4_max(vacc0x89AB, vmagic_min);
186 vacc0xCDEF = wasm_i32x4_max(vacc0xCDEF, vmagic_min);
187 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
188 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
189 vacc1x89AB = wasm_i32x4_max(vacc1x89AB, vmagic_min);
190 vacc1xCDEF = wasm_i32x4_max(vacc1xCDEF, vmagic_min);
191 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
192 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vmagic_min);
193 vacc2x89AB = wasm_i32x4_max(vacc2x89AB, vmagic_min);
194 vacc2xCDEF = wasm_i32x4_max(vacc2xCDEF, vmagic_min);
195
196 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
197 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
198 vacc0x89AB = wasm_i32x4_sub(vacc0x89AB, vmagic_bias_less_output_zero_point);
199 vacc0xCDEF = wasm_i32x4_sub(vacc0xCDEF, vmagic_bias_less_output_zero_point);
200 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
201 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
202 vacc1x89AB = wasm_i32x4_sub(vacc1x89AB, vmagic_bias_less_output_zero_point);
203 vacc1xCDEF = wasm_i32x4_sub(vacc1xCDEF, vmagic_bias_less_output_zero_point);
204 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
205 vacc2x4567 = wasm_i32x4_sub(vacc2x4567, vmagic_bias_less_output_zero_point);
206 vacc2x89AB = wasm_i32x4_sub(vacc2x89AB, vmagic_bias_less_output_zero_point);
207 vacc2xCDEF = wasm_i32x4_sub(vacc2xCDEF, vmagic_bias_less_output_zero_point);
208
209 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
210 v128_t vacc0x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc0x89AB, vacc0xCDEF);
211 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
212 v128_t vacc1x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc1x89AB, vacc1xCDEF);
213 v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x4567);
214 v128_t vacc2x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc2x89AB, vacc2xCDEF);
215
216 vacc0x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc0x01234567);
217 vacc0x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc0x89ABCDEF, vacc0x89ABCDEF);
218 vacc1x01234567 = wasm_i8x16_narrow_i16x8(vacc1x01234567, vacc1x01234567);
219 vacc1x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc1x89ABCDEF, vacc1x89ABCDEF);
220 vacc2x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc2x01234567);
221 vacc2x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc2x89ABCDEF, vacc2x89ABCDEF);
222
223 vacc0x01234567 = wasm_i8x16_min(vacc0x01234567, voutput_max);
224 vacc0x89ABCDEF = wasm_i8x16_min(vacc0x89ABCDEF, voutput_max);
225 vacc1x01234567 = wasm_i8x16_min(vacc1x01234567, voutput_max);
226 vacc1x89ABCDEF = wasm_i8x16_min(vacc1x89ABCDEF, voutput_max);
227 vacc2x01234567 = wasm_i8x16_min(vacc2x01234567, voutput_max);
228 vacc2x89ABCDEF = wasm_i8x16_min(vacc2x89ABCDEF, voutput_max);
229
230 if (nc >= 16) {
231 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
232 wasm_v128_store64_lane(c2 + 8, vacc2x89ABCDEF, 0);
233 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
234 wasm_v128_store64_lane(c1 + 8, vacc1x89ABCDEF, 0);
235 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
236 wasm_v128_store64_lane(c0 + 8, vacc0x89ABCDEF, 0);
237
238 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
239 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
240 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
241
242 a = (const int8_t**restrict) ((uintptr_t) a - ks);
243
244 nc -= 16;
245 } else {
246 if (nc & 8) {
247 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
248 c2 += 8;
249 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
250 c1 += 8;
251 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
252 c0 += 8;
253
254 vacc0x01234567 = vacc0x89ABCDEF;
255 vacc1x01234567 = vacc1x89ABCDEF;
256 vacc2x01234567 = vacc2x89ABCDEF;
257 }
258 if (nc & 4) {
259 wasm_v128_store32_lane(c2, vacc2x01234567, 0);
260 c2 += 4;
261 wasm_v128_store32_lane(c1, vacc1x01234567, 0);
262 c1 += 4;
263 wasm_v128_store32_lane(c0, vacc0x01234567, 0);
264 c0 += 4;
265
266 vacc0x01234567 = wasm_u64x2_shr(vacc0x01234567, 32);
267 vacc1x01234567 = wasm_u64x2_shr(vacc1x01234567, 32);
268 vacc2x01234567 = wasm_u64x2_shr(vacc2x01234567, 32);
269 }
270 if (nc & 2) {
271 wasm_v128_store16_lane(c2, vacc2x01234567, 0);
272 c2 += 2;
273 wasm_v128_store16_lane(c1, vacc1x01234567, 0);
274 c1 += 2;
275 wasm_v128_store16_lane(c0, vacc0x01234567, 0);
276 c0 += 2;
277
278 vacc0x01234567 = wasm_u32x4_shr(vacc0x01234567, 16);
279 vacc1x01234567 = wasm_u32x4_shr(vacc1x01234567, 16);
280 vacc2x01234567 = wasm_u32x4_shr(vacc2x01234567, 16);
281 }
282 if (nc & 1) {
283 wasm_v128_store8_lane(c2, vacc2x01234567, 0);
284 wasm_v128_store8_lane(c1, vacc1x01234567, 0);
285 wasm_v128_store8_lane(c0, vacc0x01234567, 0);
286 }
287
288 nc = 0;
289 }
290 } while (nc != 0);
291 }