Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c8-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_4x8c8__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 <= 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, 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 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
59 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
60 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
61 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
62 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
63 XNN_FORCE_REALIZATION(vmagic_bias);
64 XNN_FORCE_REALIZATION(vmagic_min);
65 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
66 XNN_FORCE_REALIZATION(voutput_max);
67
68 do {
69 v128_t vacc0x01 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
70 v128_t vacc0x23 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
71 v128_t vacc0x45 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
72 v128_t vacc0x67 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
73 v128_t vacc1x01 = vacc0x01;
74 v128_t vacc2x01 = vacc0x01;
75 v128_t vacc3x01 = vacc0x01;
76 v128_t vacc1x23 = vacc0x23;
77 v128_t vacc2x23 = vacc0x23;
78 v128_t vacc3x23 = vacc0x23;
79 v128_t vacc1x45 = vacc0x45;
80 v128_t vacc2x45 = vacc0x45;
81 v128_t vacc3x45 = vacc0x45;
82 v128_t vacc1x67 = vacc0x67;
83 v128_t vacc2x67 = vacc0x67;
84 v128_t vacc3x67 = vacc0x67;
85
86 size_t p = ks;
87 do {
88 const int8_t* restrict a0 = a[0];
89 if XNN_UNPREDICTABLE(a0 != zero) {
90 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
91 }
92 const int8_t* restrict a1 = a[1];
93 if XNN_UNPREDICTABLE(a1 != zero) {
94 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
95 }
96 const int8_t* restrict a2 = a[2];
97 if XNN_UNPREDICTABLE(a2 != zero) {
98 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
99 }
100 const int8_t* restrict a3 = a[3];
101 if XNN_UNPREDICTABLE(a3 != zero) {
102 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
103 }
104 a += 4;
105
106 size_t k = kc;
107 while (k >= 16 * sizeof(int8_t)) {
108 const v128_t va0x01 = wasm_v128_load64_splat(a0);
109 const v128_t va0x23 = wasm_v128_load64_splat((const int8_t*) a0 + 8);
110 a0 += 16;
111 const v128_t va1x01 = wasm_v128_load64_splat(a1);
112 const v128_t va1x23 = wasm_v128_load64_splat((const int8_t*) a1 + 8);
113 a1 += 16;
114 const v128_t va2x01 = wasm_v128_load64_splat(a2);
115 const v128_t va2x23 = wasm_v128_load64_splat((const int8_t*) a2 + 8);
116 a2 += 16;
117 const v128_t va3x01 = wasm_v128_load64_splat(a3);
118 const v128_t va3x23 = wasm_v128_load64_splat((const int8_t*) a3 + 8);
119 a3 += 16;
120
121 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
122 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
123 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
124 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
125 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
126 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
127 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
128 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
129
130 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
131 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
132 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
133 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
134 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
135 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
136 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
137 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
138 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01);
139 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23);
140 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45);
141 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67);
142 vacc3x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va3x01, vacc3x01);
143 vacc3x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va3x01, vacc3x23);
144 vacc3x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va3x01, vacc3x45);
145 vacc3x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va3x01, vacc3x67);
146 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
147 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
148 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
149 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
150 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va1x23, vacc1x01);
151 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va1x23, vacc1x23);
152 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va1x23, vacc1x45);
153 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va1x23, vacc1x67);
154 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va2x23, vacc2x01);
155 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va2x23, vacc2x23);
156 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va2x23, vacc2x45);
157 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va2x23, vacc2x67);
158 vacc3x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va3x23, vacc3x01);
159 vacc3x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va3x23, vacc3x23);
160 vacc3x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va3x23, vacc3x45);
161 vacc3x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va3x23, vacc3x67);
162
163 k -= 16 * sizeof(int8_t);
164 }
165
166 if (k != 0) {
167 const v128_t va0x01 = wasm_v128_load64_splat(a0);
168 a0 += 8;
169 const v128_t va1x01 = wasm_v128_load64_splat(a1);
170 a1 += 8;
171 const v128_t va2x01 = wasm_v128_load64_splat(a2);
172 a2 += 8;
173 const v128_t va3x01 = wasm_v128_load64_splat(a3);
174 a3 += 8;
175
176 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
177 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
178 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
179 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
180
181 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
182 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
183 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
184 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
185 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va1x01, vacc1x01);
186 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va1x01, vacc1x23);
187 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va1x01, vacc1x45);
188 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va1x01, vacc1x67);
189 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va2x01, vacc2x01);
190 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va2x01, vacc2x23);
191 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va2x01, vacc2x45);
192 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va2x01, vacc2x67);
193 vacc3x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va3x01, vacc3x01);
194 vacc3x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va3x01, vacc3x23);
195 vacc3x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va3x01, vacc3x45);
196 vacc3x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va3x01, vacc3x67);
197
198 k -= 8 * sizeof(int8_t);
199 }
200 p -= 4 * sizeof(void*);
201 } while (p != 0);
202
203 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
204 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
205 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
206 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
207 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7));
208 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7));
209 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x01, vacc3x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc3x01, vacc3x23, 1, 3, 5, 7));
210 v128_t vacc3x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x45, vacc3x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc3x45, vacc3x67, 1, 3, 5, 7));
211
212 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
213 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
214 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
215 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
216 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
217 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
218 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
219 vacc3x4567 = wasm_f32x4_convert_i32x4(vacc3x4567);
220
221 const v128_t vscale0123 = wasm_v128_load(w);
222 w = (const float*) w + 4;
223 const v128_t vscale4567 = wasm_v128_load(w);
224 w = (const float*) w + 4;
225 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
226 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
227 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
228 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
229 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
230 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vscale4567);
231 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale0123);
232 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vscale4567);
233
234 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
235 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
236 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
237 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
238 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
239 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vmagic_bias);
240 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
241 vacc3x4567 = wasm_f32x4_add(vacc3x4567, vmagic_bias);
242
243 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
244 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
245 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
246 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
247 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
248 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vmagic_min);
249 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
250 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vmagic_min);
251
252 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
253 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
254 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
255 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
256 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
257 vacc2x4567 = wasm_i32x4_sub(vacc2x4567, vmagic_bias_less_output_zero_point);
258 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
259 vacc3x4567 = wasm_i32x4_sub(vacc3x4567, vmagic_bias_less_output_zero_point);
260
261 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
262 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
263 v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x4567);
264 v128_t vacc3x01234567 = wasm_i16x8_narrow_i32x4(vacc3x0123, vacc3x4567);
265
266 v128_t vout0x01234567_1x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc1x01234567);
267 v128_t vout2x01234567_3x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc3x01234567);
268
269 vout0x01234567_1x01234567 = wasm_i8x16_min(vout0x01234567_1x01234567, voutput_max);
270 vout2x01234567_3x01234567 = wasm_i8x16_min(vout2x01234567_3x01234567, voutput_max);
271
272 if (nc >= 8) {
273 wasm_v128_store64_lane(c3, vout2x01234567_3x01234567, 1);
274 wasm_v128_store64_lane(c2, vout2x01234567_3x01234567, 0);
275 wasm_v128_store64_lane(c1, vout0x01234567_1x01234567, 1);
276 wasm_v128_store64_lane(c0, vout0x01234567_1x01234567, 0);
277
278 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
279 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
280 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
281 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
282
283 a = (const int8_t**restrict) ((uintptr_t) a - ks);
284
285 nc -= 8;
286 } else {
287 if (nc & 4) {
288 wasm_v128_store32_lane(c3, vout2x01234567_3x01234567, 2);
289 c3 += 4;
290 wasm_v128_store32_lane(c2, vout2x01234567_3x01234567, 0);
291 c2 += 4;
292 wasm_v128_store32_lane(c1, vout0x01234567_1x01234567, 2);
293 c1 += 4;
294 wasm_v128_store32_lane(c0, vout0x01234567_1x01234567, 0);
295 c0 += 4;
296
297 vout0x01234567_1x01234567 = wasm_u64x2_shr(vout0x01234567_1x01234567, 32);
298 vout2x01234567_3x01234567 = wasm_u64x2_shr(vout2x01234567_3x01234567, 32);
299 }
300 if (nc & 2) {
301 wasm_v128_store16_lane(c3, vout2x01234567_3x01234567, 4);
302 c3 += 2;
303 wasm_v128_store16_lane(c2, vout2x01234567_3x01234567, 0);
304 c2 += 2;
305 wasm_v128_store16_lane(c1, vout0x01234567_1x01234567, 4);
306 c1 += 2;
307 wasm_v128_store16_lane(c0, vout0x01234567_1x01234567, 0);
308 c0 += 2;
309
310 vout0x01234567_1x01234567 = wasm_u32x4_shr(vout0x01234567_1x01234567, 16);
311 vout2x01234567_3x01234567 = wasm_u32x4_shr(vout2x01234567_3x01234567, 16);
312 }
313 if (nc & 1) {
314 wasm_v128_store8_lane(c3, vout2x01234567_3x01234567, 8);
315 wasm_v128_store8_lane(c2, vout2x01234567_3x01234567, 0);
316 wasm_v128_store8_lane(c1, vout0x01234567_1x01234567, 8);
317 wasm_v128_store8_lane(c0, vout0x01234567_1x01234567, 0);
318 }
319
320 nc = 0;
321 }
322 } while (nc != 0);
323 }