1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/psimd-splat.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/gemm.h>
15
16
xnn_f32_gemm_ukernel_4x8__psimd_splat(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f32_output_params params[restrict static1])17 void xnn_f32_gemm_ukernel_4x8__psimd_splat(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const float*restrict a,
22 size_t a_stride,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_output_params params[restrict static 1])
28 {
29 assert(mr != 0);
30 assert(mr <= 4);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 const float* a0 = a;
39 float* c0 = c;
40 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
41 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr < 2) {
43 a1 = a0;
44 c1 = c0;
45 }
46 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
47 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
48 if XNN_UNPREDICTABLE(mr <= 2) {
49 a2 = a1;
50 c2 = c1;
51 }
52 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 a3 = a2;
56 c3 = c2;
57 }
58
59 do {
60 psimd_f32 vacc0x0123 = psimd_load_f32(w + 0);
61 psimd_f32 vacc0x4567 = psimd_load_f32(w + 4);
62 psimd_f32 vacc1x0123 = vacc0x0123;
63 psimd_f32 vacc1x4567 = vacc0x4567;
64 psimd_f32 vacc2x0123 = vacc0x0123;
65 psimd_f32 vacc2x4567 = vacc0x4567;
66 psimd_f32 vacc3x0123 = vacc0x0123;
67 psimd_f32 vacc3x4567 = vacc0x4567;
68 w += 8;
69
70 size_t k = kc;
71 while (k >= 4 * sizeof(float)) {
72 const psimd_f32 va0 = psimd_load_f32(a0);
73 a0 += 4;
74 const psimd_f32 va1 = psimd_load_f32(a1);
75 a1 += 4;
76 const psimd_f32 va2 = psimd_load_f32(a2);
77 a2 += 4;
78 const psimd_f32 va3 = psimd_load_f32(a3);
79 a3 += 4;
80
81 const psimd_f32 va0c0 = psimd_splat0_f32(va0);
82 const psimd_f32 va1c0 = psimd_splat0_f32(va1);
83 const psimd_f32 va2c0 = psimd_splat0_f32(va2);
84 const psimd_f32 va3c0 = psimd_splat0_f32(va3);
85
86 const psimd_f32 vb0123c0 = psimd_load_f32(w + 0);
87 const psimd_f32 vb4567c0 = psimd_load_f32(w + 4);
88
89 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c0, vb0123c0);
90 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c0, vb0123c0);
91 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c0, vb0123c0);
92 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c0, vb0123c0);
93 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c0, vb4567c0);
94 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c0, vb4567c0);
95 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c0, vb4567c0);
96 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c0, vb4567c0);
97 const psimd_f32 va0c1 = psimd_splat1_f32(va0);
98 const psimd_f32 va1c1 = psimd_splat1_f32(va1);
99 const psimd_f32 va2c1 = psimd_splat1_f32(va2);
100 const psimd_f32 va3c1 = psimd_splat1_f32(va3);
101
102 const psimd_f32 vb0123c1 = psimd_load_f32(w + 8);
103 const psimd_f32 vb4567c1 = psimd_load_f32(w + 12);
104
105 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c1, vb0123c1);
106 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c1, vb0123c1);
107 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c1, vb0123c1);
108 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c1, vb0123c1);
109 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c1, vb4567c1);
110 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c1, vb4567c1);
111 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c1, vb4567c1);
112 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c1, vb4567c1);
113 const psimd_f32 va0c2 = psimd_splat2_f32(va0);
114 const psimd_f32 va1c2 = psimd_splat2_f32(va1);
115 const psimd_f32 va2c2 = psimd_splat2_f32(va2);
116 const psimd_f32 va3c2 = psimd_splat2_f32(va3);
117
118 const psimd_f32 vb0123c2 = psimd_load_f32(w + 16);
119 const psimd_f32 vb4567c2 = psimd_load_f32(w + 20);
120
121 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c2, vb0123c2);
122 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c2, vb0123c2);
123 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c2, vb0123c2);
124 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c2, vb0123c2);
125 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c2, vb4567c2);
126 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c2, vb4567c2);
127 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c2, vb4567c2);
128 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c2, vb4567c2);
129 const psimd_f32 va0c3 = psimd_splat3_f32(va0);
130 const psimd_f32 va1c3 = psimd_splat3_f32(va1);
131 const psimd_f32 va2c3 = psimd_splat3_f32(va2);
132 const psimd_f32 va3c3 = psimd_splat3_f32(va3);
133
134 const psimd_f32 vb0123c3 = psimd_load_f32(w + 24);
135 const psimd_f32 vb4567c3 = psimd_load_f32(w + 28);
136
137 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c3, vb0123c3);
138 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1c3, vb0123c3);
139 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2c3, vb0123c3);
140 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3c3, vb0123c3);
141 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c3, vb4567c3);
142 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1c3, vb4567c3);
143 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2c3, vb4567c3);
144 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3c3, vb4567c3);
145
146 w += 32;
147 k -= 4 * sizeof(float);
148 }
149 if XNN_UNLIKELY(k != 0) {
150 do {
151 const psimd_f32 va0 = psimd_load_splat_f32(a0);
152 a0 += 1;
153 const psimd_f32 va1 = psimd_load_splat_f32(a1);
154 a1 += 1;
155 const psimd_f32 va2 = psimd_load_splat_f32(a2);
156 a2 += 1;
157 const psimd_f32 va3 = psimd_load_splat_f32(a3);
158 a3 += 1;
159
160 const psimd_f32 vb0123 = psimd_load_f32(w);
161 const psimd_f32 vb4567 = psimd_load_f32(w + 4);
162 w += 8;
163
164 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123);
165 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1, vb0123);
166 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123);
167 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3, vb0123);
168 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567);
169 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1, vb4567);
170 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567);
171 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3, vb4567);
172
173 k -= sizeof(float);
174 } while (k != 0);
175 }
176
177 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
178 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
179 vacc1x0123 = psimd_min_f32(vacc1x0123, vmax);
180 vacc2x0123 = psimd_min_f32(vacc2x0123, vmax);
181 vacc3x0123 = psimd_min_f32(vacc3x0123, vmax);
182 vacc0x4567 = psimd_min_f32(vacc0x4567, vmax);
183 vacc1x4567 = psimd_min_f32(vacc1x4567, vmax);
184 vacc2x4567 = psimd_min_f32(vacc2x4567, vmax);
185 vacc3x4567 = psimd_min_f32(vacc3x4567, vmax);
186
187 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
188 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
189 vacc1x0123 = psimd_max_f32(vacc1x0123, vmin);
190 vacc2x0123 = psimd_max_f32(vacc2x0123, vmin);
191 vacc3x0123 = psimd_max_f32(vacc3x0123, vmin);
192 vacc0x4567 = psimd_max_f32(vacc0x4567, vmin);
193 vacc1x4567 = psimd_max_f32(vacc1x4567, vmin);
194 vacc2x4567 = psimd_max_f32(vacc2x4567, vmin);
195 vacc3x4567 = psimd_max_f32(vacc3x4567, vmin);
196
197 if XNN_LIKELY(nc >= 8) {
198 psimd_store_f32(c3, vacc3x0123);
199 psimd_store_f32(c3 + 4, vacc3x4567);
200 c3 = (float*) ((uintptr_t) c3 + cn_stride);
201 psimd_store_f32(c2, vacc2x0123);
202 psimd_store_f32(c2 + 4, vacc2x4567);
203 c2 = (float*) ((uintptr_t) c2 + cn_stride);
204 psimd_store_f32(c1, vacc1x0123);
205 psimd_store_f32(c1 + 4, vacc1x4567);
206 c1 = (float*) ((uintptr_t) c1 + cn_stride);
207 psimd_store_f32(c0, vacc0x0123);
208 psimd_store_f32(c0 + 4, vacc0x4567);
209 c0 = (float*) ((uintptr_t) c0 + cn_stride);
210
211 a3 = (const float*) ((uintptr_t) a3 - kc);
212 a2 = (const float*) ((uintptr_t) a2 - kc);
213 a1 = (const float*) ((uintptr_t) a1 - kc);
214 a0 = (const float*) ((uintptr_t) a0 - kc);
215
216 nc -= 8;
217 } else {
218 if (nc & 4) {
219 psimd_store_f32(c3, vacc3x0123);
220 psimd_store_f32(c2, vacc2x0123);
221 psimd_store_f32(c1, vacc1x0123);
222 psimd_store_f32(c0, vacc0x0123);
223
224 vacc3x0123 = vacc3x4567;
225 vacc2x0123 = vacc2x4567;
226 vacc1x0123 = vacc1x4567;
227 vacc0x0123 = vacc0x4567;
228
229 c3 += 4;
230 c2 += 4;
231 c1 += 4;
232 c0 += 4;
233 }
234 if (nc & 2) {
235 psimd_store2_f32(c3, vacc3x0123);
236 psimd_store2_f32(c2, vacc2x0123);
237 psimd_store2_f32(c1, vacc1x0123);
238 psimd_store2_f32(c0, vacc0x0123);
239
240 vacc3x0123 = psimd_concat_hi_f32(vacc3x0123, vacc3x0123);
241 vacc2x0123 = psimd_concat_hi_f32(vacc2x0123, vacc2x0123);
242 vacc1x0123 = psimd_concat_hi_f32(vacc1x0123, vacc1x0123);
243 vacc0x0123 = psimd_concat_hi_f32(vacc0x0123, vacc0x0123);
244
245 c3 += 2;
246 c2 += 2;
247 c1 += 2;
248 c0 += 2;
249 }
250 if (nc & 1) {
251 psimd_store1_f32(c3, vacc3x0123);
252 psimd_store1_f32(c2, vacc2x0123);
253 psimd_store1_f32(c1, vacc1x0123);
254 psimd_store1_f32(c0, vacc0x0123);
255 }
256
257 nc = 0;
258 }
259 } while (nc != 0);
260 }
261