Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x8c8-minmax-fp32-wasmsdot-u2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-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_3x8c8__wasmsdot_u2(
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, 8 * 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 vacc0x01 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
66 v128_t vacc0x23 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
67 v128_t vacc0x45 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
68 v128_t vacc0x67 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
69 v128_t vacc1x01 = vacc0x01;
70 v128_t vacc2x01 = vacc0x01;
71 v128_t vacc1x23 = vacc0x23;
72 v128_t vacc2x23 = vacc0x23;
73 v128_t vacc1x45 = vacc0x45;
74 v128_t vacc2x45 = vacc0x45;
75 v128_t vacc1x67 = vacc0x67;
76 v128_t vacc2x67 = vacc0x67;
77
78 size_t p = ks;
79 do {
80 const int8_t* restrict a0 = a[0];
81 if XNN_UNPREDICTABLE(a0 != zero) {
82 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
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 }
88 const int8_t* restrict a2 = a[2];
89 if XNN_UNPREDICTABLE(a2 != zero) {
90 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
91 }
92 a += 3;
93
94 size_t k = kc;
95 while (k >= 16 * sizeof(int8_t)) {
96 const v128_t va0x01 = wasm_v128_load64_splat(a0);
97 const v128_t va0x23 = wasm_v128_load64_splat((const int8_t*) a0 + 8);
98 a0 += 16;
99 const v128_t va1x01 = wasm_v128_load64_splat(a1);
100 const v128_t va1x23 = wasm_v128_load64_splat((const int8_t*) a1 + 8);
101 a1 += 16;
102 const v128_t va2x01 = wasm_v128_load64_splat(a2);
103 const v128_t va2x23 = wasm_v128_load64_splat((const int8_t*) a2 + 8);
104 a2 += 16;
105
106 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
107 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
108 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
109 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
110 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
111 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
112 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
113 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
114
115 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
116 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
117 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
118 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
119 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
120 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
121 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
122 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
123 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01);
124 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23);
125 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45);
126 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67);
127 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
128 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
129 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
130 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
131 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va1x23, vacc1x01);
132 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va1x23, vacc1x23);
133 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va1x23, vacc1x45);
134 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va1x23, vacc1x67);
135 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va2x23, vacc2x01);
136 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va2x23, vacc2x23);
137 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va2x23, vacc2x45);
138 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va2x23, vacc2x67);
139
140 k -= 16 * sizeof(int8_t);
141 }
142
143 if (k != 0) {
144 const v128_t va0x01 = wasm_v128_load64_splat(a0);
145 a0 += 8;
146 const v128_t va1x01 = wasm_v128_load64_splat(a1);
147 a1 += 8;
148 const v128_t va2x01 = wasm_v128_load64_splat(a2);
149 a2 += 8;
150
151 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
152 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
153 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
154 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
155
156 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
157 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
158 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
159 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
160 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va1x01, vacc1x01);
161 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va1x01, vacc1x23);
162 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va1x01, vacc1x45);
163 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va1x01, vacc1x67);
164 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va2x01, vacc2x01);
165 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va2x01, vacc2x23);
166 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va2x01, vacc2x45);
167 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va2x01, vacc2x67);
168
169 k -= 8 * sizeof(int8_t);
170 }
171 p -= 3 * sizeof(void*);
172 } while (p != 0);
173
174 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
175 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
176 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
177 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
178 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7));
179 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7));
180
181 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
182 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
183 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
184 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
185 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
186 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
187
188 const v128_t vscale0123 = wasm_v128_load(w);
189 w = (const float*) w + 4;
190 const v128_t vscale4567 = wasm_v128_load(w);
191 w = (const float*) w + 4;
192 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
193 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
194 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
195 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
196 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
197 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vscale4567);
198
199 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
200 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
201 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
202 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
203 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
204 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vmagic_bias);
205
206 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
207 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
208 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
209 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
210 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
211 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vmagic_min);
212
213 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
214 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
215 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
216 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
217 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
218 vacc2x4567 = wasm_i32x4_sub(vacc2x4567, vmagic_bias_less_output_zero_point);
219
220 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
221 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
222 v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x4567);
223
224 v128_t vout0x01234567_1x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc1x01234567);
225 v128_t vout2x01234567_2x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc2x01234567);
226
227 vout0x01234567_1x01234567 = wasm_i8x16_min(vout0x01234567_1x01234567, voutput_max);
228 vout2x01234567_2x01234567 = wasm_i8x16_min(vout2x01234567_2x01234567, voutput_max);
229
230 if (nc >= 8) {
231 wasm_v128_store64_lane(c2, vout2x01234567_2x01234567, 0);
232 wasm_v128_store64_lane(c1, vout0x01234567_1x01234567, 1);
233 wasm_v128_store64_lane(c0, vout0x01234567_1x01234567, 0);
234
235 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
236 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
237 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
238
239 a = (const int8_t**restrict) ((uintptr_t) a - ks);
240
241 nc -= 8;
242 } else {
243 if (nc & 4) {
244 wasm_v128_store32_lane(c2, vout2x01234567_2x01234567, 0);
245 c2 += 4;
246 wasm_v128_store32_lane(c1, vout0x01234567_1x01234567, 2);
247 c1 += 4;
248 wasm_v128_store32_lane(c0, vout0x01234567_1x01234567, 0);
249 c0 += 4;
250
251 vout0x01234567_1x01234567 = wasm_u64x2_shr(vout0x01234567_1x01234567, 32);
252 vout2x01234567_2x01234567 = wasm_u64x2_shr(vout2x01234567_2x01234567, 32);
253 }
254 if (nc & 2) {
255 wasm_v128_store16_lane(c2, vout2x01234567_2x01234567, 0);
256 c2 += 2;
257 wasm_v128_store16_lane(c1, vout0x01234567_1x01234567, 4);
258 c1 += 2;
259 wasm_v128_store16_lane(c0, vout0x01234567_1x01234567, 0);
260 c0 += 2;
261
262 vout0x01234567_1x01234567 = wasm_u32x4_shr(vout0x01234567_1x01234567, 16);
263 vout2x01234567_2x01234567 = wasm_u32x4_shr(vout2x01234567_2x01234567, 16);
264 }
265 if (nc & 1) {
266 wasm_v128_store8_lane(c2, vout2x01234567_2x01234567, 0);
267 wasm_v128_store8_lane(c1, vout0x01234567_1x01234567, 8);
268 wasm_v128_store8_lane(c0, vout0x01234567_1x01234567, 0);
269 }
270
271 nc = 0;
272 }
273 } while (nc != 0);
274 }