Debian Codegraph @

src/f32-ppmm/gen/f32-ppmm-4x8-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_4x8__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 <= 4);
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
59
60 do {
61 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
62 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
63 float32x4_t vacc1x0123 = vacc0x0123;
64 float32x4_t vacc1x4567 = vacc0x4567;
65 float32x4_t vacc2x0123 = vacc0x0123;
66 float32x4_t vacc2x4567 = vacc0x4567;
67 float32x4_t vacc3x0123 = vacc0x0123;
68 float32x4_t vacc3x4567 = vacc0x4567;
69
70 size_t k = kc;
71 do {
72 const float32x4_t va0123 = vld1q_f32(a); a += 4;
73
74 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
75 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
76
77 #if XNN_ARCH_ARM64
78 vacc0x0123 = vfmaq_laneq_f32(vacc0x0123, vb0123, va0123, 0);
79 vacc1x0123 = vfmaq_laneq_f32(vacc1x0123, vb0123, va0123, 1);
80 vacc2x0123 = vfmaq_laneq_f32(vacc2x0123, vb0123, va0123, 2);
81 vacc3x0123 = vfmaq_laneq_f32(vacc3x0123, vb0123, va0123, 3);
82 vacc0x4567 = vfmaq_laneq_f32(vacc0x4567, vb4567, va0123, 0);
83 vacc1x4567 = vfmaq_laneq_f32(vacc1x4567, vb4567, va0123, 1);
84 vacc2x4567 = vfmaq_laneq_f32(vacc2x4567, vb4567, va0123, 2);
85 vacc3x4567 = vfmaq_laneq_f32(vacc3x4567, vb4567, va0123, 3);
86 #else
87 const float32x4_t va0000 = vdupq_lane_f32(vget_low_f32(va0123), 0);
88 const float32x4_t va1111 = vdupq_lane_f32(vget_low_f32(va0123), 1);
89 const float32x4_t va2222 = vdupq_lane_f32(vget_high_f32(va0123), 0);
90 const float32x4_t va3333 = vdupq_lane_f32(vget_high_f32(va0123), 1);
91
92 vacc0x0123 = vfmaq_f32(vacc0x0123, va0000, vb0123);
93 vacc1x0123 = vfmaq_f32(vacc1x0123, va1111, vb0123);
94 vacc2x0123 = vfmaq_f32(vacc2x0123, va2222, vb0123);
95 vacc3x0123 = vfmaq_f32(vacc3x0123, va3333, vb0123);
96 vacc0x4567 = vfmaq_f32(vacc0x4567, va0000, vb4567);
97 vacc1x4567 = vfmaq_f32(vacc1x4567, va1111, vb4567);
98 vacc2x4567 = vfmaq_f32(vacc2x4567, va2222, vb4567);
99 vacc3x4567 = vfmaq_f32(vacc3x4567, va3333, vb4567);
100 #endif
101
102 k -= sizeof(float);
103 } while (k != 0);
104
105 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
106 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
107 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
108 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
109 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
110 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
111 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
112 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
113
114 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
115 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
116 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
117 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
118 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
119 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
120 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
121 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
122
123 if XNN_LIKELY(nc >= 8) {
124 vst1q_f32(c3, vacc3x0123);
125 vst1q_f32(c3 + 4, vacc3x4567);
126 c3 = (float*) ((uintptr_t) c3 + cn_stride);
127 vst1q_f32(c2, vacc2x0123);
128 vst1q_f32(c2 + 4, vacc2x4567);
129 c2 = (float*) ((uintptr_t) c2 + cn_stride);
130 vst1q_f32(c1, vacc1x0123);
131 vst1q_f32(c1 + 4, vacc1x4567);
132 c1 = (float*) ((uintptr_t) c1 + cn_stride);
133 vst1q_f32(c0, vacc0x0123);
134 vst1q_f32(c0 + 4, vacc0x4567);
135 c0 = (float*) ((uintptr_t) c0 + cn_stride);
136
137 a = (const float*) ((uintptr_t) a - kc * 4);
138
139 nc -= 8;
140 } else {
141 if (nc & 4) {
142 vst1q_f32(c3, vacc3x0123); c3 += 4;
143 vst1q_f32(c2, vacc2x0123); c2 += 4;
144 vst1q_f32(c1, vacc1x0123); c1 += 4;
145 vst1q_f32(c0, vacc0x0123); c0 += 4;
146
147 vacc3x0123 = vacc3x4567;
148 vacc2x0123 = vacc2x4567;
149 vacc1x0123 = vacc1x4567;
150 vacc0x0123 = vacc0x4567;
151 }
152 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
153 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
154 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
155 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
156 if (nc & 2) {
157 vst1_f32(c3, vacc3x01); c3 += 2;
158 vst1_f32(c2, vacc2x01); c2 += 2;
159 vst1_f32(c1, vacc1x01); c1 += 2;
160 vst1_f32(c0, vacc0x01); c0 += 2;
161
162 vacc3x01 = vget_high_f32(vacc3x0123);
163 vacc2x01 = vget_high_f32(vacc2x0123);
164 vacc1x01 = vget_high_f32(vacc1x0123);
165 vacc0x01 = vget_high_f32(vacc0x0123);
166 }
167 if (nc & 1) {
168 vst1_lane_f32(c3, vacc3x01, 0);
169 vst1_lane_f32(c2, vacc2x01, 0);
170 vst1_lane_f32(c1, vacc1x01, 0);
171 vst1_lane_f32(c0, vacc0x01, 0);
172 }
173
174 nc = 0;
175 }
176 } while (nc != 0);
177 }