Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x8c16-minmax-fp32-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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
18 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x8c16__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, 16 * 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 vacc0x0 = wasm_v128_load32_zero(w);
68 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
69 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
70 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
71 v128_t vacc0x4 = wasm_v128_load32_zero((const int32_t*) w + 4);
72 v128_t vacc0x5 = wasm_v128_load32_zero((const int32_t*) w + 5);
73 v128_t vacc0x6 = wasm_v128_load32_zero((const int32_t*) w + 6);
74 v128_t vacc0x7 = wasm_v128_load32_zero((const int32_t*) w + 7);
75 v128_t vacc1x0 = vacc0x0;
76 v128_t vacc1x1 = vacc0x1;
77 v128_t vacc1x2 = vacc0x2;
78 v128_t vacc1x3 = vacc0x3;
79 v128_t vacc1x4 = vacc0x4;
80 v128_t vacc1x5 = vacc0x5;
81 v128_t vacc1x6 = vacc0x6;
82 v128_t vacc1x7 = vacc0x7;
83 v128_t vacc2x0 = vacc0x0;
84 v128_t vacc2x1 = vacc0x1;
85 v128_t vacc2x2 = vacc0x2;
86 v128_t vacc2x3 = vacc0x3;
87 v128_t vacc2x4 = vacc0x4;
88 v128_t vacc2x5 = vacc0x5;
89 v128_t vacc2x6 = vacc0x6;
90 v128_t vacc2x7 = vacc0x7;
91 w = (const int32_t*) w + 8;
92
93 size_t p = ks;
94 do {
95 const int8_t* restrict a0 = a[0];
96 if XNN_UNPREDICTABLE(a0 != zero) {
97 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
98 }
99 const int8_t* restrict a1 = a[1];
100 if XNN_UNPREDICTABLE(a1 != zero) {
101 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
102 }
103 const int8_t* restrict a2 = a[2];
104 if XNN_UNPREDICTABLE(a2 != zero) {
105 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
106 }
107 a += 3;
108
109 size_t k = kc;
110 do {
111 const v128_t va0 = wasm_v128_xor(wasm_v128_load(a0), vsign_mask);
112 a0 += 16;
113 const v128_t va1 = wasm_v128_xor(wasm_v128_load(a1), vsign_mask);
114 a1 += 16;
115 const v128_t va2 = wasm_v128_xor(wasm_v128_load(a2), vsign_mask);
116 a2 += 16;
117
118 const v128_t vb0 = wasm_v128_load(w);
119 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
120 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
121 vacc2x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va2, vacc2x0);
122 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
123 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
124 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
125 vacc2x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va2, vacc2x1);
126 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
127 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
128 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
129 vacc2x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va2, vacc2x2);
130 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
131 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
132 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
133 vacc2x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va2, vacc2x3);
134 const v128_t vb4 = wasm_v128_load((const int8_t*) w + 64);
135 vacc0x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va0, vacc0x4);
136 vacc1x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va1, vacc1x4);
137 vacc2x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va2, vacc2x4);
138 const v128_t vb5 = wasm_v128_load((const int8_t*) w + 80);
139 vacc0x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va0, vacc0x5);
140 vacc1x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va1, vacc1x5);
141 vacc2x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va2, vacc2x5);
142 const v128_t vb6 = wasm_v128_load((const int8_t*) w + 96);
143 vacc0x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va0, vacc0x6);
144 vacc1x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va1, vacc1x6);
145 vacc2x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va2, vacc2x6);
146 const v128_t vb7 = wasm_v128_load((const int8_t*) w + 112);
147 vacc0x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va0, vacc0x7);
148 vacc1x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va1, vacc1x7);
149 vacc2x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va2, vacc2x7);
150
151 w = (const int8_t*) w + 128;
152 k -= 16 * sizeof(int8_t);
153 } while (k != 0);
154 p -= 3 * sizeof(void*);
155 } while (p != 0);
156
157 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));
158 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));
159 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));
160 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));
161 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));
162 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));
163 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));
164 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));
165 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));
166 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));
167 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));
168 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));
169
170 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
171 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x46, vacc0x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x46, vacc0x57, 2, 6, 3, 7));
172 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
173 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x46, vacc1x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x46, vacc1x57, 2, 6, 3, 7));
174 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
175 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x46, vacc2x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x46, vacc2x57, 2, 6, 3, 7));
176
177 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
178 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
179 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
180 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
181 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
182 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
183
184 const v128_t vscale0123 = wasm_v128_load(w);
185 w = (const float*) w + 4;
186 const v128_t vscale4567 = wasm_v128_load(w);
187 w = (const float*) w + 4;
188 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
189 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
190 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
191 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
192 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
193 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vscale4567);
194
195 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
196 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
197 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
198 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
199 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
200 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vmagic_bias);
201
202 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
203 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
204 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
205 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
206 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
207 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vmagic_min);
208
209 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
210 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
211 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
212 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
213 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
214 vacc2x4567 = wasm_i32x4_sub(vacc2x4567, vmagic_bias_less_output_zero_point);
215
216 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
217 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
218 v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x4567);
219
220 v128_t vout0x01234567_1x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc1x01234567);
221 v128_t vout2x01234567_2x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc2x01234567);
222
223 vout0x01234567_1x01234567 = wasm_i8x16_min(vout0x01234567_1x01234567, voutput_max);
224 vout2x01234567_2x01234567 = wasm_i8x16_min(vout2x01234567_2x01234567, voutput_max);
225
226 if (nc >= 8) {
227 wasm_v128_store64_lane(c2, vout2x01234567_2x01234567, 0);
228 wasm_v128_store64_lane(c1, vout0x01234567_1x01234567, 1);
229 wasm_v128_store64_lane(c0, vout0x01234567_1x01234567, 0);
230
231 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
232 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
233 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
234
235 a = (const int8_t**restrict) ((uintptr_t) a - ks);
236
237 nc -= 8;
238 } else {
239 if (nc & 4) {
240 wasm_v128_store32_lane(c2, vout2x01234567_2x01234567, 0);
241 c2 += 4;
242 wasm_v128_store32_lane(c1, vout0x01234567_1x01234567, 2);
243 c1 += 4;
244 wasm_v128_store32_lane(c0, vout0x01234567_1x01234567, 0);
245 c0 += 4;
246
247 vout0x01234567_1x01234567 = wasm_u64x2_shr(vout0x01234567_1x01234567, 32);
248 vout2x01234567_2x01234567 = wasm_u64x2_shr(vout2x01234567_2x01234567, 32);
249 }
250 if (nc & 2) {
251 wasm_v128_store16_lane(c2, vout2x01234567_2x01234567, 0);
252 c2 += 2;
253 wasm_v128_store16_lane(c1, vout0x01234567_1x01234567, 4);
254 c1 += 2;
255 wasm_v128_store16_lane(c0, vout0x01234567_1x01234567, 0);
256 c0 += 2;
257
258 vout0x01234567_1x01234567 = wasm_u32x4_shr(vout0x01234567_1x01234567, 16);
259 vout2x01234567_2x01234567 = wasm_u32x4_shr(vout2x01234567_2x01234567, 16);
260 }
261 if (nc & 1) {
262 wasm_v128_store8_lane(c2, vout2x01234567_2x01234567, 0);
263 wasm_v128_store8_lane(c1, vout0x01234567_1x01234567, 8);
264 wasm_v128_store8_lane(c0, vout0x01234567_1x01234567, 0);
265 }
266
267 nc = 0;
268 }
269 } while (nc != 0);
270 }