Debian Codegraph @

src/f32-ppmm/gen/f32-ppmm-8x8-minmax-aarch64-neonfma.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/f32-ppmm/neon.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 <arm_neon.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/ppmm.h"
16
17
18 void xnn_f32_ppmm_minmax_ukernel_8x8__aarch64_neonfma(
19 size_t mr,
20 size_t nc,
21 size_t kc,
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 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 8);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34
35 #if XNN_ARCH_ARM64
36 const float32x4x2_t vminmax = vld2q_dup_f32(&params->scalar.min);
37 const float32x4_t vmin = vminmax.val[0];
38 const float32x4_t vmax = vminmax.val[1];
39 #else
40 const float32x2x2_t vminmax = vld2_dup_f32(&params->scalar.min);
41 const float32x4_t vmin = vcombine_f32(vminmax.val[0], vminmax.val[0]);
42 const float32x4_t vmax = vcombine_f32(vminmax.val[1], vminmax.val[1]);
43 #endif
44
45 float* c0 = c;
46 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 c3 = c2;
57 }
58 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
59 if XNN_UNPREDICTABLE(mr <= 4) {
60 c4 = c3;
61 }
62 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
63 if XNN_UNPREDICTABLE(mr < 6) {
64 c5 = c4;
65 }
66 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 6) {
68 c6 = c5;
69 }
70 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
71 if XNN_UNPREDICTABLE(mr != 8) {
72 c7 = c6;
73 }
74
75
76 do {
77 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
78 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
79 float32x4_t vacc1x0123 = vacc0x0123;
80 float32x4_t vacc1x4567 = vacc0x4567;
81 float32x4_t vacc2x0123 = vacc0x0123;
82 float32x4_t vacc2x4567 = vacc0x4567;
83 float32x4_t vacc3x0123 = vacc0x0123;
84 float32x4_t vacc3x4567 = vacc0x4567;
85 float32x4_t vacc4x0123 = vacc0x0123;
86 float32x4_t vacc4x4567 = vacc0x4567;
87 float32x4_t vacc5x0123 = vacc0x0123;
88 float32x4_t vacc5x4567 = vacc0x4567;
89 float32x4_t vacc6x0123 = vacc0x0123;
90 float32x4_t vacc6x4567 = vacc0x4567;
91 float32x4_t vacc7x0123 = vacc0x0123;
92 float32x4_t vacc7x4567 = vacc0x4567;
93
94 size_t k = kc;
95 do {
96 const float32x4_t va0123 = vld1q_f32(a); a += 4;
97 const float32x4_t va4567 = vld1q_f32(a); a += 4;
98
99 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
100 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
101
102 #if XNN_ARCH_ARM64
103 vacc0x0123 = vfmaq_laneq_f32(vacc0x0123, vb0123, va0123, 0);
104 vacc1x0123 = vfmaq_laneq_f32(vacc1x0123, vb0123, va0123, 1);
105 vacc2x0123 = vfmaq_laneq_f32(vacc2x0123, vb0123, va0123, 2);
106 vacc3x0123 = vfmaq_laneq_f32(vacc3x0123, vb0123, va0123, 3);
107 vacc4x0123 = vfmaq_laneq_f32(vacc4x0123, vb0123, va4567, 0);
108 vacc5x0123 = vfmaq_laneq_f32(vacc5x0123, vb0123, va4567, 1);
109 vacc6x0123 = vfmaq_laneq_f32(vacc6x0123, vb0123, va4567, 2);
110 vacc7x0123 = vfmaq_laneq_f32(vacc7x0123, vb0123, va4567, 3);
111 vacc0x4567 = vfmaq_laneq_f32(vacc0x4567, vb4567, va0123, 0);
112 vacc1x4567 = vfmaq_laneq_f32(vacc1x4567, vb4567, va0123, 1);
113 vacc2x4567 = vfmaq_laneq_f32(vacc2x4567, vb4567, va0123, 2);
114 vacc3x4567 = vfmaq_laneq_f32(vacc3x4567, vb4567, va0123, 3);
115 vacc4x4567 = vfmaq_laneq_f32(vacc4x4567, vb4567, va4567, 0);
116 vacc5x4567 = vfmaq_laneq_f32(vacc5x4567, vb4567, va4567, 1);
117 vacc6x4567 = vfmaq_laneq_f32(vacc6x4567, vb4567, va4567, 2);
118 vacc7x4567 = vfmaq_laneq_f32(vacc7x4567, vb4567, va4567, 3);
119 #else
120 const float32x4_t va0000 = vdupq_lane_f32(vget_low_f32(va0123), 0);
121 const float32x4_t va1111 = vdupq_lane_f32(vget_low_f32(va0123), 1);
122 const float32x4_t va2222 = vdupq_lane_f32(vget_high_f32(va0123), 0);
123 const float32x4_t va3333 = vdupq_lane_f32(vget_high_f32(va0123), 1);
124 const float32x4_t va4444 = vdupq_lane_f32(vget_low_f32(va4567), 0);
125 const float32x4_t va5555 = vdupq_lane_f32(vget_low_f32(va4567), 1);
126 const float32x4_t va6666 = vdupq_lane_f32(vget_high_f32(va4567), 0);
127 const float32x4_t va7777 = vdupq_lane_f32(vget_high_f32(va4567), 1);
128
129 vacc0x0123 = vfmaq_f32(vacc0x0123, va0000, vb0123);
130 vacc1x0123 = vfmaq_f32(vacc1x0123, va1111, vb0123);
131 vacc2x0123 = vfmaq_f32(vacc2x0123, va2222, vb0123);
132 vacc3x0123 = vfmaq_f32(vacc3x0123, va3333, vb0123);
133 vacc4x0123 = vfmaq_f32(vacc4x0123, va4444, vb0123);
134 vacc5x0123 = vfmaq_f32(vacc5x0123, va5555, vb0123);
135 vacc6x0123 = vfmaq_f32(vacc6x0123, va6666, vb0123);
136 vacc7x0123 = vfmaq_f32(vacc7x0123, va7777, vb0123);
137 vacc0x4567 = vfmaq_f32(vacc0x4567, va0000, vb4567);
138 vacc1x4567 = vfmaq_f32(vacc1x4567, va1111, vb4567);
139 vacc2x4567 = vfmaq_f32(vacc2x4567, va2222, vb4567);
140 vacc3x4567 = vfmaq_f32(vacc3x4567, va3333, vb4567);
141 vacc4x4567 = vfmaq_f32(vacc4x4567, va4444, vb4567);
142 vacc5x4567 = vfmaq_f32(vacc5x4567, va5555, vb4567);
143 vacc6x4567 = vfmaq_f32(vacc6x4567, va6666, vb4567);
144 vacc7x4567 = vfmaq_f32(vacc7x4567, va7777, vb4567);
145 #endif
146
147 k -= sizeof(float);
148 } while (k != 0);
149
150 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
151 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
152 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
153 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
154 vacc4x0123 = vminq_f32(vacc4x0123, vmax);
155 vacc5x0123 = vminq_f32(vacc5x0123, vmax);
156 vacc6x0123 = vminq_f32(vacc6x0123, vmax);
157 vacc7x0123 = vminq_f32(vacc7x0123, vmax);
158 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
159 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
160 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
161 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
162 vacc4x4567 = vminq_f32(vacc4x4567, vmax);
163 vacc5x4567 = vminq_f32(vacc5x4567, vmax);
164 vacc6x4567 = vminq_f32(vacc6x4567, vmax);
165 vacc7x4567 = vminq_f32(vacc7x4567, vmax);
166
167 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
168 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
169 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
170 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
171 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
172 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
173 vacc6x0123 = vmaxq_f32(vacc6x0123, vmin);
174 vacc7x0123 = vmaxq_f32(vacc7x0123, vmin);
175 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
176 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
177 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
178 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
179 vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
180 vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
181 vacc6x4567 = vmaxq_f32(vacc6x4567, vmin);
182 vacc7x4567 = vmaxq_f32(vacc7x4567, vmin);
183
184 if XNN_LIKELY(nc >= 8) {
185 vst1q_f32(c7, vacc7x0123);
186 vst1q_f32(c7 + 4, vacc7x4567);
187 c7 = (float*) ((uintptr_t) c7 + cn_stride);
188 vst1q_f32(c6, vacc6x0123);
189 vst1q_f32(c6 + 4, vacc6x4567);
190 c6 = (float*) ((uintptr_t) c6 + cn_stride);
191 vst1q_f32(c5, vacc5x0123);
192 vst1q_f32(c5 + 4, vacc5x4567);
193 c5 = (float*) ((uintptr_t) c5 + cn_stride);
194 vst1q_f32(c4, vacc4x0123);
195 vst1q_f32(c4 + 4, vacc4x4567);
196 c4 = (float*) ((uintptr_t) c4 + cn_stride);
197 vst1q_f32(c3, vacc3x0123);
198 vst1q_f32(c3 + 4, vacc3x4567);
199 c3 = (float*) ((uintptr_t) c3 + cn_stride);
200 vst1q_f32(c2, vacc2x0123);
201 vst1q_f32(c2 + 4, vacc2x4567);
202 c2 = (float*) ((uintptr_t) c2 + cn_stride);
203 vst1q_f32(c1, vacc1x0123);
204 vst1q_f32(c1 + 4, vacc1x4567);
205 c1 = (float*) ((uintptr_t) c1 + cn_stride);
206 vst1q_f32(c0, vacc0x0123);
207 vst1q_f32(c0 + 4, vacc0x4567);
208 c0 = (float*) ((uintptr_t) c0 + cn_stride);
209
210 a = (const float*) ((uintptr_t) a - kc * 8);
211
212 nc -= 8;
213 } else {
214 if (nc & 4) {
215 vst1q_f32(c7, vacc7x0123); c7 += 4;
216 vst1q_f32(c6, vacc6x0123); c6 += 4;
217 vst1q_f32(c5, vacc5x0123); c5 += 4;
218 vst1q_f32(c4, vacc4x0123); c4 += 4;
219 vst1q_f32(c3, vacc3x0123); c3 += 4;
220 vst1q_f32(c2, vacc2x0123); c2 += 4;
221 vst1q_f32(c1, vacc1x0123); c1 += 4;
222 vst1q_f32(c0, vacc0x0123); c0 += 4;
223
224 vacc7x0123 = vacc7x4567;
225 vacc6x0123 = vacc6x4567;
226 vacc5x0123 = vacc5x4567;
227 vacc4x0123 = vacc4x4567;
228 vacc3x0123 = vacc3x4567;
229 vacc2x0123 = vacc2x4567;
230 vacc1x0123 = vacc1x4567;
231 vacc0x0123 = vacc0x4567;
232 }
233 float32x2_t vacc7x01 = vget_low_f32(vacc7x0123);
234 float32x2_t vacc6x01 = vget_low_f32(vacc6x0123);
235 float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
236 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
237 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
238 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
239 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
240 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
241 if (nc & 2) {
242 vst1_f32(c7, vacc7x01); c7 += 2;
243 vst1_f32(c6, vacc6x01); c6 += 2;
244 vst1_f32(c5, vacc5x01); c5 += 2;
245 vst1_f32(c4, vacc4x01); c4 += 2;
246 vst1_f32(c3, vacc3x01); c3 += 2;
247 vst1_f32(c2, vacc2x01); c2 += 2;
248 vst1_f32(c1, vacc1x01); c1 += 2;
249 vst1_f32(c0, vacc0x01); c0 += 2;
250
251 vacc7x01 = vget_high_f32(vacc7x0123);
252 vacc6x01 = vget_high_f32(vacc6x0123);
253 vacc5x01 = vget_high_f32(vacc5x0123);
254 vacc4x01 = vget_high_f32(vacc4x0123);
255 vacc3x01 = vget_high_f32(vacc3x0123);
256 vacc2x01 = vget_high_f32(vacc2x0123);
257 vacc1x01 = vget_high_f32(vacc1x0123);
258 vacc0x01 = vget_high_f32(vacc0x0123);
259 }
260 if (nc & 1) {
261 vst1_lane_f32(c7, vacc7x01, 0);
262 vst1_lane_f32(c6, vacc6x01, 0);
263 vst1_lane_f32(c5, vacc5x01, 0);
264 vst1_lane_f32(c4, vacc4x01, 0);
265 vst1_lane_f32(c3, vacc3x01, 0);
266 vst1_lane_f32(c2, vacc2x01, 0);
267 vst1_lane_f32(c1, vacc1x01, 0);
268 vst1_lane_f32(c0, vacc0x01, 0);
269 }
270
271 nc = 0;
272 }
273 } while (nc != 0);
274 }