Debian Codegraph @

src/f32-ppmm/gen/f32-ppmm-4x8-minmax-aarch64-neonfma-prfm.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 #include "xnnpack/prefetch.h"
17
18
19 void xnn_f32_ppmm_minmax_ukernel_4x8__aarch64_neonfma_prfm(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const float* restrict a,
24 const float* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35
36 #if XNN_ARCH_ARM64
37 const float32x4x2_t vminmax = vld2q_dup_f32(&params->scalar.min);
38 const float32x4_t vmin = vminmax.val[0];
39 const float32x4_t vmax = vminmax.val[1];
40 #else
41 const float32x2x2_t vminmax = vld2_dup_f32(&params->scalar.min);
42 const float32x4_t vmin = vcombine_f32(vminmax.val[0], vminmax.val[0]);
43 const float32x4_t vmax = vcombine_f32(vminmax.val[1], vminmax.val[1]);
44 #endif
45
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 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 xnn_prefetch_to_l1((const int8_t*) w + 0);
61 xnn_prefetch_to_l1((const int8_t*) w + 64);
62 xnn_prefetch_to_l1((const int8_t*) w + 128);
63 xnn_prefetch_to_l1((const int8_t*) w + 192);
64
65 do {
66 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
67 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
68 float32x4_t vacc1x0123 = vacc0x0123;
69 float32x4_t vacc1x4567 = vacc0x4567;
70 float32x4_t vacc2x0123 = vacc0x0123;
71 float32x4_t vacc2x4567 = vacc0x4567;
72 float32x4_t vacc3x0123 = vacc0x0123;
73 float32x4_t vacc3x4567 = vacc0x4567;
74
75 size_t k = kc;
76 do {
77 const float32x4_t va0123 = vld1q_f32(a); a += 4;
78
79 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
80 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
81
82 #if XNN_ARCH_ARM64
83 vacc0x0123 = vfmaq_laneq_f32(vacc0x0123, vb0123, va0123, 0);
84 vacc1x0123 = vfmaq_laneq_f32(vacc1x0123, vb0123, va0123, 1);
85 vacc2x0123 = vfmaq_laneq_f32(vacc2x0123, vb0123, va0123, 2);
86 vacc3x0123 = vfmaq_laneq_f32(vacc3x0123, vb0123, va0123, 3);
87 xnn_prefetch_to_l1((const int8_t*) w + 192);
88 vacc0x4567 = vfmaq_laneq_f32(vacc0x4567, vb4567, va0123, 0);
89 vacc1x4567 = vfmaq_laneq_f32(vacc1x4567, vb4567, va0123, 1);
90 vacc2x4567 = vfmaq_laneq_f32(vacc2x4567, vb4567, va0123, 2);
91 vacc3x4567 = vfmaq_laneq_f32(vacc3x4567, vb4567, va0123, 3);
92 #else
93 const float32x4_t va0000 = vdupq_lane_f32(vget_low_f32(va0123), 0);
94 const float32x4_t va1111 = vdupq_lane_f32(vget_low_f32(va0123), 1);
95 const float32x4_t va2222 = vdupq_lane_f32(vget_high_f32(va0123), 0);
96 const float32x4_t va3333 = vdupq_lane_f32(vget_high_f32(va0123), 1);
97
98 vacc0x0123 = vfmaq_f32(vacc0x0123, va0000, vb0123);
99 vacc1x0123 = vfmaq_f32(vacc1x0123, va1111, vb0123);
100 vacc2x0123 = vfmaq_f32(vacc2x0123, va2222, vb0123);
101 vacc3x0123 = vfmaq_f32(vacc3x0123, va3333, vb0123);
102 xnn_prefetch_to_l1((const int8_t*) w + 192);
103 vacc0x4567 = vfmaq_f32(vacc0x4567, va0000, vb4567);
104 vacc1x4567 = vfmaq_f32(vacc1x4567, va1111, vb4567);
105 vacc2x4567 = vfmaq_f32(vacc2x4567, va2222, vb4567);
106 vacc3x4567 = vfmaq_f32(vacc3x4567, va3333, vb4567);
107 #endif
108
109 k -= sizeof(float);
110 } while (k != 0);
111
112 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
113 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
114 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
115 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
116 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
117 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
118 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
119 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
120
121 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
122 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
123 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
124 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
125 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
126 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
127 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
128 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
129
130 if XNN_LIKELY(nc >= 8) {
131 vst1q_f32(c3, vacc3x0123);
132 vst1q_f32(c3 + 4, vacc3x4567);
133 c3 = (float*) ((uintptr_t) c3 + cn_stride);
134 vst1q_f32(c2, vacc2x0123);
135 vst1q_f32(c2 + 4, vacc2x4567);
136 c2 = (float*) ((uintptr_t) c2 + cn_stride);
137 vst1q_f32(c1, vacc1x0123);
138 vst1q_f32(c1 + 4, vacc1x4567);
139 c1 = (float*) ((uintptr_t) c1 + cn_stride);
140 vst1q_f32(c0, vacc0x0123);
141 vst1q_f32(c0 + 4, vacc0x4567);
142 c0 = (float*) ((uintptr_t) c0 + cn_stride);
143
144 a = (const float*) ((uintptr_t) a - kc * 4);
145
146 nc -= 8;
147 } else {
148 if (nc & 4) {
149 vst1q_f32(c3, vacc3x0123); c3 += 4;
150 vst1q_f32(c2, vacc2x0123); c2 += 4;
151 vst1q_f32(c1, vacc1x0123); c1 += 4;
152 vst1q_f32(c0, vacc0x0123); c0 += 4;
153
154 vacc3x0123 = vacc3x4567;
155 vacc2x0123 = vacc2x4567;
156 vacc1x0123 = vacc1x4567;
157 vacc0x0123 = vacc0x4567;
158 }
159 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
160 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
161 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
162 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
163 if (nc & 2) {
164 vst1_f32(c3, vacc3x01); c3 += 2;
165 vst1_f32(c2, vacc2x01); c2 += 2;
166 vst1_f32(c1, vacc1x01); c1 += 2;
167 vst1_f32(c0, vacc0x01); c0 += 2;
168
169 vacc3x01 = vget_high_f32(vacc3x0123);
170 vacc2x01 = vget_high_f32(vacc2x0123);
171 vacc1x01 = vget_high_f32(vacc1x0123);
172 vacc0x01 = vget_high_f32(vacc0x0123);
173 }
174 if (nc & 1) {
175 vst1_lane_f32(c3, vacc3x01, 0);
176 vst1_lane_f32(c2, vacc2x01, 0);
177 vst1_lane_f32(c1, vacc1x01, 0);
178 vst1_lane_f32(c0, vacc0x01, 0);
179 }
180
181 nc = 0;
182 }
183 } while (nc != 0);
184 }