1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/psimd-loadsplat.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_gemminc_ukernel_6x8__psimd_loadsplat(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 float * restrict acc,const union xnn_f32_output_params params[restrict static1])17 void xnn_f32_gemminc_ukernel_6x8__psimd_loadsplat(
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 float*restrict acc,
28 const union xnn_f32_output_params params[restrict static 1])
29 {
30 assert(mr != 0);
31 assert(mr <= 6);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38 assert(acc != NULL);
39
40 const float* a0 = a;
41 float* c0 = c;
42 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
61 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 a4 = a3;
64 c4 = c3;
65 }
66 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
67 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
68 if XNN_UNPREDICTABLE(mr != 6) {
69 a5 = a4;
70 c5 = c4;
71 }
72
73 do {
74 psimd_f32 vacc0x0123 = psimd_load_f32(acc + 0);
75 psimd_f32 vacc0x4567 = psimd_load_f32(acc + 4);
76 psimd_f32 vacc1x0123 = psimd_load_f32(acc + 8);
77 psimd_f32 vacc1x4567 = psimd_load_f32(acc + 12);
78 psimd_f32 vacc2x0123 = psimd_load_f32(acc + 16);
79 psimd_f32 vacc2x4567 = psimd_load_f32(acc + 20);
80 psimd_f32 vacc3x0123 = psimd_load_f32(acc + 24);
81 psimd_f32 vacc3x4567 = psimd_load_f32(acc + 28);
82 psimd_f32 vacc4x0123 = psimd_load_f32(acc + 32);
83 psimd_f32 vacc4x4567 = psimd_load_f32(acc + 36);
84 psimd_f32 vacc5x0123 = psimd_load_f32(acc + 40);
85 psimd_f32 vacc5x4567 = psimd_load_f32(acc + 44);
86 acc += 48;
87
88 size_t k = kc;
89 do {
90 const psimd_f32 va0 = psimd_load_splat_f32(a0);
91 a0 += 1;
92 const psimd_f32 va1 = psimd_load_splat_f32(a1);
93 a1 += 1;
94 const psimd_f32 va2 = psimd_load_splat_f32(a2);
95 a2 += 1;
96 const psimd_f32 va3 = psimd_load_splat_f32(a3);
97 a3 += 1;
98 const psimd_f32 va4 = psimd_load_splat_f32(a4);
99 a4 += 1;
100 const psimd_f32 va5 = psimd_load_splat_f32(a5);
101 a5 += 1;
102
103 const psimd_f32 vb0123 = psimd_load_f32(w);
104 const psimd_f32 vb4567 = psimd_load_f32(w + 4);
105 w += 8;
106
107 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123);
108 vacc1x0123 = psimd_qfma_f32(vacc1x0123, va1, vb0123);
109 vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123);
110 vacc3x0123 = psimd_qfma_f32(vacc3x0123, va3, vb0123);
111 vacc4x0123 = psimd_qfma_f32(vacc4x0123, va4, vb0123);
112 vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5, vb0123);
113 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567);
114 vacc1x4567 = psimd_qfma_f32(vacc1x4567, va1, vb4567);
115 vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567);
116 vacc3x4567 = psimd_qfma_f32(vacc3x4567, va3, vb4567);
117 vacc4x4567 = psimd_qfma_f32(vacc4x4567, va4, vb4567);
118 vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5, vb4567);
119
120 k -= sizeof(float);
121 } while (k != 0);
122
123 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
124 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
125 vacc1x0123 = psimd_min_f32(vacc1x0123, vmax);
126 vacc2x0123 = psimd_min_f32(vacc2x0123, vmax);
127 vacc3x0123 = psimd_min_f32(vacc3x0123, vmax);
128 vacc4x0123 = psimd_min_f32(vacc4x0123, vmax);
129 vacc5x0123 = psimd_min_f32(vacc5x0123, vmax);
130 vacc0x4567 = psimd_min_f32(vacc0x4567, vmax);
131 vacc1x4567 = psimd_min_f32(vacc1x4567, vmax);
132 vacc2x4567 = psimd_min_f32(vacc2x4567, vmax);
133 vacc3x4567 = psimd_min_f32(vacc3x4567, vmax);
134 vacc4x4567 = psimd_min_f32(vacc4x4567, vmax);
135 vacc5x4567 = psimd_min_f32(vacc5x4567, vmax);
136
137 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
138 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
139 vacc1x0123 = psimd_max_f32(vacc1x0123, vmin);
140 vacc2x0123 = psimd_max_f32(vacc2x0123, vmin);
141 vacc3x0123 = psimd_max_f32(vacc3x0123, vmin);
142 vacc4x0123 = psimd_max_f32(vacc4x0123, vmin);
143 vacc5x0123 = psimd_max_f32(vacc5x0123, vmin);
144 vacc0x4567 = psimd_max_f32(vacc0x4567, vmin);
145 vacc1x4567 = psimd_max_f32(vacc1x4567, vmin);
146 vacc2x4567 = psimd_max_f32(vacc2x4567, vmin);
147 vacc3x4567 = psimd_max_f32(vacc3x4567, vmin);
148 vacc4x4567 = psimd_max_f32(vacc4x4567, vmin);
149 vacc5x4567 = psimd_max_f32(vacc5x4567, vmin);
150
151 if XNN_LIKELY(nc >= 8) {
152 psimd_store_f32(c5, vacc5x0123);
153 psimd_store_f32(c5 + 4, vacc5x4567);
154 c5 = (float*) ((uintptr_t) c5 + cn_stride);
155 psimd_store_f32(c4, vacc4x0123);
156 psimd_store_f32(c4 + 4, vacc4x4567);
157 c4 = (float*) ((uintptr_t) c4 + cn_stride);
158 psimd_store_f32(c3, vacc3x0123);
159 psimd_store_f32(c3 + 4, vacc3x4567);
160 c3 = (float*) ((uintptr_t) c3 + cn_stride);
161 psimd_store_f32(c2, vacc2x0123);
162 psimd_store_f32(c2 + 4, vacc2x4567);
163 c2 = (float*) ((uintptr_t) c2 + cn_stride);
164 psimd_store_f32(c1, vacc1x0123);
165 psimd_store_f32(c1 + 4, vacc1x4567);
166 c1 = (float*) ((uintptr_t) c1 + cn_stride);
167 psimd_store_f32(c0, vacc0x0123);
168 psimd_store_f32(c0 + 4, vacc0x4567);
169 c0 = (float*) ((uintptr_t) c0 + cn_stride);
170
171 a5 = (const float*) ((uintptr_t) a5 - kc);
172 a4 = (const float*) ((uintptr_t) a4 - kc);
173 a3 = (const float*) ((uintptr_t) a3 - kc);
174 a2 = (const float*) ((uintptr_t) a2 - kc);
175 a1 = (const float*) ((uintptr_t) a1 - kc);
176 a0 = (const float*) ((uintptr_t) a0 - kc);
177
178 nc -= 8;
179 } else {
180 if (nc & 4) {
181 psimd_store_f32(c5, vacc5x0123);
182 psimd_store_f32(c4, vacc4x0123);
183 psimd_store_f32(c3, vacc3x0123);
184 psimd_store_f32(c2, vacc2x0123);
185 psimd_store_f32(c1, vacc1x0123);
186 psimd_store_f32(c0, vacc0x0123);
187
188 vacc5x0123 = vacc5x4567;
189 vacc4x0123 = vacc4x4567;
190 vacc3x0123 = vacc3x4567;
191 vacc2x0123 = vacc2x4567;
192 vacc1x0123 = vacc1x4567;
193 vacc0x0123 = vacc0x4567;
194
195 c5 += 4;
196 c4 += 4;
197 c3 += 4;
198 c2 += 4;
199 c1 += 4;
200 c0 += 4;
201 }
202 if (nc & 2) {
203 psimd_store2_f32(c5, vacc5x0123);
204 psimd_store2_f32(c4, vacc4x0123);
205 psimd_store2_f32(c3, vacc3x0123);
206 psimd_store2_f32(c2, vacc2x0123);
207 psimd_store2_f32(c1, vacc1x0123);
208 psimd_store2_f32(c0, vacc0x0123);
209
210 vacc5x0123 = psimd_concat_hi_f32(vacc5x0123, vacc5x0123);
211 vacc4x0123 = psimd_concat_hi_f32(vacc4x0123, vacc4x0123);
212 vacc3x0123 = psimd_concat_hi_f32(vacc3x0123, vacc3x0123);
213 vacc2x0123 = psimd_concat_hi_f32(vacc2x0123, vacc2x0123);
214 vacc1x0123 = psimd_concat_hi_f32(vacc1x0123, vacc1x0123);
215 vacc0x0123 = psimd_concat_hi_f32(vacc0x0123, vacc0x0123);
216
217 c5 += 2;
218 c4 += 2;
219 c3 += 2;
220 c2 += 2;
221 c1 += 2;
222 c0 += 2;
223 }
224 if (nc & 1) {
225 psimd_store1_f32(c5, vacc5x0123);
226 psimd_store1_f32(c4, vacc4x0123);
227 psimd_store1_f32(c3, vacc3x0123);
228 psimd_store1_f32(c2, vacc2x0123);
229 psimd_store1_f32(c1, vacc1x0123);
230 psimd_store1_f32(c0, vacc0x0123);
231 }
232
233 nc = 0;
234 }
235 } while (nc != 0);
236 }
237