Debian Codegraph @

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