Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-3x8c8-minmax-wasmusdot-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_qd8_f32_qc8w_igemm_minmax_ukernel_3x8c8__wasmusdot_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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 3);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (3 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55
56 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
57 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
58 XNN_FORCE_REALIZATION(vmin);
59 XNN_FORCE_REALIZATION(vmax);
60
61 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
62 XNN_FORCE_REALIZATION(vsign_mask);
63 do {
64 const v128_t vinput_zero_point = wasm_i32x4_splat((int32_t) quantization_params->zero_point + 128);
65 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
66 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
67 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
68 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
69 v128_t vsum1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
70 v128_t vacc1x01 = wasm_u64x2_extend_low_u32x4(vsum1x0123);
71 v128_t vacc1x23 = wasm_u64x2_extend_high_u32x4(vsum1x0123);
72 v128_t vsum2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
73 v128_t vacc2x01 = wasm_u64x2_extend_low_u32x4(vsum2x0123);
74 v128_t vacc2x23 = wasm_u64x2_extend_high_u32x4(vsum2x0123);
75 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
76 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
77 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
78 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
79 v128_t vsum1x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
80 v128_t vacc1x45 = wasm_u64x2_extend_low_u32x4(vsum1x4567);
81 v128_t vacc1x67 = wasm_u64x2_extend_high_u32x4(vsum1x4567);
82 v128_t vsum2x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
83 v128_t vacc2x45 = wasm_u64x2_extend_low_u32x4(vsum2x4567);
84 v128_t vacc2x67 = wasm_u64x2_extend_high_u32x4(vsum2x4567);
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 } else {
92 a0 = zero_data;
93 }
94 const int8_t* restrict a1 = a[1];
95 if XNN_UNPREDICTABLE(a1 != zero) {
96 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
97 } else {
98 a1 = zero_data;
99 }
100 const int8_t* restrict a2 = a[2];
101 if XNN_UNPREDICTABLE(a2 != zero) {
102 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
103 } else {
104 a2 = zero_data;
105 }
106 a += 3;
107
108 size_t k = kc;
109 while (k >= 16 * sizeof(int8_t)) {
110 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
111 const v128_t va0x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a0 + 8), vsign_mask);
112 a0 += 16;
113 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
114 const v128_t va1x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a1 + 8), vsign_mask);
115 a1 += 16;
116 const v128_t va2x01 = wasm_v128_xor(wasm_v128_load64_splat(a2), vsign_mask);
117 const v128_t va2x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a2 + 8), vsign_mask);
118 a2 += 16;
119
120 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
121 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
122 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
123 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
124 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
125 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
126 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
127 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
128
129 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
130 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
131 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
132 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
133 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
134 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
135 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
136 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
137 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01);
138 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23);
139 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45);
140 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67);
141 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
142 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
143 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
144 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
145 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va1x23, vacc1x01);
146 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va1x23, vacc1x23);
147 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va1x23, vacc1x45);
148 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va1x23, vacc1x67);
149 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va2x23, vacc2x01);
150 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va2x23, vacc2x23);
151 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va2x23, vacc2x45);
152 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va2x23, vacc2x67);
153
154 k -= 16 * sizeof(int8_t);
155 }
156
157 if (k != 0) {
158 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
159 a0 += 8;
160 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
161 a1 += 8;
162 const v128_t va2x01 = wasm_v128_xor(wasm_v128_load64_splat(a2), vsign_mask);
163 a2 += 8;
164
165 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
166 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
167 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
168 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
169
170 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
171 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
172 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
173 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
174 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va1x01, vacc1x01);
175 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va1x01, vacc1x23);
176 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va1x01, vacc1x45);
177 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va1x01, vacc1x67);
178 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va2x01, vacc2x01);
179 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va2x01, vacc2x23);
180 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va2x01, vacc2x45);
181 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va2x01, vacc2x67);
182
183 k -= 8 * sizeof(int8_t);
184 }
185 p -= 3 * sizeof(void*);
186 } while (p != 0);
187
188 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
189 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
190 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
191 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
192 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7));
193 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7));
194
195 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
196 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
197 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
198 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
199 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
200 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
201
202 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
203
204 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
205 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale);
206 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
207 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale);
208 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale);
209 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale);
210
211 const v128_t vfilter_output_scale0123 = wasm_v128_load(w); w = (const float*) w + 4;
212 const v128_t vfilter_output_scale4567 = wasm_v128_load(w); w = (const float*) w + 4;
213
214 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
215 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
216 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
217 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
218 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
219 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567);
220
221 const v128_t vbias0123 = wasm_v128_load(w); w = (const float*) w + 4;
222 const v128_t vbias4567 = wasm_v128_load(w); w = (const float*) w + 4;
223 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
224 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
225 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
226 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
227 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
228 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567);
229
230 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
231 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
232 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
233 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
234 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
235 vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin);
236
237 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
238 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
239 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
240 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
241 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
242 vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax);
243
244 if XNN_LIKELY(nc >= 8) {
245 wasm_v128_store(c2, vacc2x0123);
246 wasm_v128_store(c2 + 4, vacc2x4567);
247 wasm_v128_store(c1, vacc1x0123);
248 wasm_v128_store(c1 + 4, vacc1x4567);
249 wasm_v128_store(c0, vacc0x0123);
250 wasm_v128_store(c0 + 4, vacc0x4567);
251
252 c0 = (float*) ((uintptr_t) c0 + cn_stride);
253 c1 = (float*) ((uintptr_t) c1 + cn_stride);
254 c2 = (float*) ((uintptr_t) c2 + cn_stride);
255
256 a = (const int8_t**restrict) ((uintptr_t) a - ks);
257
258 nc -= 8;
259 } else {
260 if (nc & 4) {
261 wasm_v128_store(c2, vacc2x0123);
262 vacc2x0123 = vacc2x4567;
263 c2 += 4;
264 wasm_v128_store(c1, vacc1x0123);
265 vacc1x0123 = vacc1x4567;
266 c1 += 4;
267 wasm_v128_store(c0, vacc0x0123);
268 vacc0x0123 = vacc0x4567;
269 c0 += 4;
270 }
271 if (nc & 2) {
272 wasm_v128_store64_lane(c2, vacc2x0123, 0);
273 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
274 c2 += 2;
275 wasm_v128_store64_lane(c1, vacc1x0123, 0);
276 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
277 c1 += 2;
278 wasm_v128_store64_lane(c0, vacc0x0123, 0);
279 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
280 c0 += 2;
281 }
282 if (nc & 1) {
283 wasm_v128_store32_lane(c2, vacc2x0123, 0);
284 wasm_v128_store32_lane(c1, vacc1x0123, 0);
285 wasm_v128_store32_lane(c0, vacc0x0123, 0);
286 }
287 nc = 0;
288 }
289 } while (nc != 0);
290 }