1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/psimd-s4.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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 <psimd.h>
13
14 #include <xnnpack/igemm.h>
15
16
xnn_f32_igemm_ukernel_4x8s4__psimd(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_output_params params[restrict static1])17 void xnn_f32_igemm_ukernel_4x8s4__psimd(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_output_params params[restrict static 1])
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 c3 = c2;
55 }
56
57 do {
58 psimd_f32 vacc0x0123 = psimd_load_f32(w);
59 psimd_f32 vacc0x4567 = psimd_load_f32(w + 4);
60 psimd_f32 vacc1x0123 = vacc0x0123;
61 psimd_f32 vacc1x4567 = vacc0x4567;
62 psimd_f32 vacc2x0123 = vacc0x0123;
63 psimd_f32 vacc2x4567 = vacc0x4567;
64 psimd_f32 vacc3x0123 = vacc0x0123;
65 psimd_f32 vacc3x4567 = vacc0x4567;
66 w += 8;
67
68 size_t p = ks;
69 do {
70 const float* restrict a0 = a[0];
71 assert(a0 != NULL);
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const float*) ((uintptr_t) a0 + a_offset);
74 }
75 const float* restrict a1 = a[1];
76 assert(a1 != NULL);
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const float*) ((uintptr_t) a1 + a_offset);
79 }
80 const float* restrict a2 = a[2];
81 assert(a2 != NULL);
82 if XNN_UNPREDICTABLE(a2 != zero) {
83 a2 = (const float*) ((uintptr_t) a2 + a_offset);
84 }
85 const float* restrict a3 = a[3];
86 assert(a3 != NULL);
87 if XNN_UNPREDICTABLE(a3 != zero) {
88 a3 = (const float*) ((uintptr_t) a3 + a_offset);
89 }
90 a += 4;
91
92 size_t k = kc;
93 while (k >= 4 * sizeof(float)) {
94 psimd_f32 va0 = psimd_load_f32(a0);
95 a0 += 4;
96 psimd_f32 va1 = psimd_load_f32(a1);
97 a1 += 4;
98 psimd_f32 va2 = psimd_load_f32(a2);
99 a2 += 4;
100 psimd_f32 va3 = psimd_load_f32(a3);
101 a3 += 4;
102
103
104 const psimd_f32 vb0123c0 = psimd_load_f32(w + 0);
105 const psimd_f32 vb4567c0 = psimd_load_f32(w + 4);
106
107 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c0);
108 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1, vb0123c0);
109 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c0);
110 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3, vb0123c0);
111 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c0);
112 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1, vb4567c0);
113 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c0);
114 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3, vb4567c0);
115
116 #ifdef __clang__
117 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
118 va1 = __builtin_shufflevector(va1, va1, 1, 2, 3, 0);
119 va2 = __builtin_shufflevector(va2, va2, 1, 2, 3, 0);
120 va3 = __builtin_shufflevector(va3, va3, 1, 2, 3, 0);
121 #else
122 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
123 va1 = __builtin_shuffle(va1, va1, (psimd_s32) { 1, 2, 3, 0 });
124 va2 = __builtin_shuffle(va2, va2, (psimd_s32) { 1, 2, 3, 0 });
125 va3 = __builtin_shuffle(va3, va3, (psimd_s32) { 1, 2, 3, 0 });
126 #endif
127
128 const psimd_f32 vb0123c1 = psimd_load_f32(w + 8);
129 const psimd_f32 vb4567c1 = psimd_load_f32(w + 12);
130
131 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c1);
132 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1, vb0123c1);
133 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c1);
134 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3, vb0123c1);
135 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c1);
136 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1, vb4567c1);
137 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c1);
138 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3, vb4567c1);
139
140 #ifdef __clang__
141 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
142 va1 = __builtin_shufflevector(va1, va1, 1, 2, 3, 0);
143 va2 = __builtin_shufflevector(va2, va2, 1, 2, 3, 0);
144 va3 = __builtin_shufflevector(va3, va3, 1, 2, 3, 0);
145 #else
146 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
147 va1 = __builtin_shuffle(va1, va1, (psimd_s32) { 1, 2, 3, 0 });
148 va2 = __builtin_shuffle(va2, va2, (psimd_s32) { 1, 2, 3, 0 });
149 va3 = __builtin_shuffle(va3, va3, (psimd_s32) { 1, 2, 3, 0 });
150 #endif
151
152 const psimd_f32 vb0123c2 = psimd_load_f32(w + 16);
153 const psimd_f32 vb4567c2 = psimd_load_f32(w + 20);
154
155 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c2);
156 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1, vb0123c2);
157 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c2);
158 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3, vb0123c2);
159 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c2);
160 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1, vb4567c2);
161 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c2);
162 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3, vb4567c2);
163
164 #ifdef __clang__
165 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
166 va1 = __builtin_shufflevector(va1, va1, 1, 2, 3, 0);
167 va2 = __builtin_shufflevector(va2, va2, 1, 2, 3, 0);
168 va3 = __builtin_shufflevector(va3, va3, 1, 2, 3, 0);
169 #else
170 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
171 va1 = __builtin_shuffle(va1, va1, (psimd_s32) { 1, 2, 3, 0 });
172 va2 = __builtin_shuffle(va2, va2, (psimd_s32) { 1, 2, 3, 0 });
173 va3 = __builtin_shuffle(va3, va3, (psimd_s32) { 1, 2, 3, 0 });
174 #endif
175
176 const psimd_f32 vb0123c3 = psimd_load_f32(w + 24);
177 const psimd_f32 vb4567c3 = psimd_load_f32(w + 28);
178
179 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c3);
180 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1, vb0123c3);
181 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c3);
182 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3, vb0123c3);
183 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c3);
184 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1, vb4567c3);
185 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c3);
186 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3, vb4567c3);
187
188
189 w += 32;
190 k -= 4 * sizeof(float);
191 }
192 if XNN_UNLIKELY(k != 0) {
193 do {
194 const psimd_f32 vb0123 = psimd_load_f32(w);
195 const psimd_f32 vb4567 = psimd_load_f32(w + 4);
196 w += 8;
197
198 const psimd_f32 va0 = psimd_load_splat_f32(a0);
199 a0 += 1;
200 const psimd_f32 va1 = psimd_load_splat_f32(a1);
201 a1 += 1;
202 const psimd_f32 va2 = psimd_load_splat_f32(a2);
203 a2 += 1;
204 const psimd_f32 va3 = psimd_load_splat_f32(a3);
205 a3 += 1;
206
207 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123);
208 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567);
209 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1, vb0123);
210 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1, vb4567);
211 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123);
212 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567);
213 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3, vb0123);
214 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3, vb4567);
215 k -= sizeof(float);
216 } while (k != 0);
217 }
218 p -= 4 * sizeof(void*);
219 } while (p != 0);
220
221 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
222 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
223 vacc1x0123 = psimd_min_f32(vacc1x0123, vmax);
224 vacc2x0123 = psimd_min_f32(vacc2x0123, vmax);
225 vacc3x0123 = psimd_min_f32(vacc3x0123, vmax);
226 vacc0x4567 = psimd_min_f32(vacc0x4567, vmax);
227 vacc1x4567 = psimd_min_f32(vacc1x4567, vmax);
228 vacc2x4567 = psimd_min_f32(vacc2x4567, vmax);
229 vacc3x4567 = psimd_min_f32(vacc3x4567, vmax);
230
231 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
232 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
233 vacc1x0123 = psimd_max_f32(vacc1x0123, vmin);
234 vacc2x0123 = psimd_max_f32(vacc2x0123, vmin);
235 vacc3x0123 = psimd_max_f32(vacc3x0123, vmin);
236 vacc0x4567 = psimd_max_f32(vacc0x4567, vmin);
237 vacc1x4567 = psimd_max_f32(vacc1x4567, vmin);
238 vacc2x4567 = psimd_max_f32(vacc2x4567, vmin);
239 vacc3x4567 = psimd_max_f32(vacc3x4567, vmin);
240
241 if XNN_LIKELY(nc >= 8) {
242 psimd_store_f32(c3, vacc3x0123);
243 psimd_store_f32(c3 + 4, vacc3x4567);
244 c3 = (float*) ((uintptr_t) c3 + cn_stride);
245 psimd_store_f32(c2, vacc2x0123);
246 psimd_store_f32(c2 + 4, vacc2x4567);
247 c2 = (float*) ((uintptr_t) c2 + cn_stride);
248 psimd_store_f32(c1, vacc1x0123);
249 psimd_store_f32(c1 + 4, vacc1x4567);
250 c1 = (float*) ((uintptr_t) c1 + cn_stride);
251 psimd_store_f32(c0, vacc0x0123);
252 psimd_store_f32(c0 + 4, vacc0x4567);
253 c0 = (float*) ((uintptr_t) c0 + cn_stride);
254
255 a = (const float**restrict) ((uintptr_t) a - ks);
256 nc -= 8;
257 } else {
258 if (nc & 4) {
259 psimd_store_f32(c3, vacc3x0123);
260 psimd_store_f32(c2, vacc2x0123);
261 psimd_store_f32(c1, vacc1x0123);
262 psimd_store_f32(c0, vacc0x0123);
263
264 vacc3x0123 = vacc3x4567;
265 vacc2x0123 = vacc2x4567;
266 vacc1x0123 = vacc1x4567;
267 vacc0x0123 = vacc0x4567;
268
269 c3 += 4;
270 c2 += 4;
271 c1 += 4;
272 c0 += 4;
273 }
274 if (nc & 2) {
275 psimd_store2_f32(c3, vacc3x0123);
276 psimd_store2_f32(c2, vacc2x0123);
277 psimd_store2_f32(c1, vacc1x0123);
278 psimd_store2_f32(c0, vacc0x0123);
279
280 vacc3x0123 = psimd_concat_hi_f32(vacc3x0123, vacc3x0123);
281 vacc2x0123 = psimd_concat_hi_f32(vacc2x0123, vacc2x0123);
282 vacc1x0123 = psimd_concat_hi_f32(vacc1x0123, vacc1x0123);
283 vacc0x0123 = psimd_concat_hi_f32(vacc0x0123, vacc0x0123);
284
285 c3 += 2;
286 c2 += 2;
287 c1 += 2;
288 c0 += 2;
289 }
290 if (nc & 1) {
291 psimd_store1_f32(c3, vacc3x0123);
292 psimd_store1_f32(c2, vacc2x0123);
293 psimd_store1_f32(c1, vacc1x0123);
294 psimd_store1_f32(c0, vacc0x0123);
295 }
296
297 nc = 0;
298 }
299 } while (nc != 0);
300 }
301