Debian Codegraph @

src/f32-ppmm/gen/f32-ppmm-4x8-minmax-neon.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__neon(
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 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123, vget_low_f32(va0123), 0);
78 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123, vget_low_f32(va0123), 1);
79 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123, vget_high_f32(va0123), 0);
80 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123, vget_high_f32(va0123), 1);
81 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567, vget_low_f32(va0123), 0);
82 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567, vget_low_f32(va0123), 1);
83 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567, vget_high_f32(va0123), 0);
84 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567, vget_high_f32(va0123), 1);
85
86 k -= sizeof(float);
87 } while (k != 0);
88
89 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
90 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
91 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
92 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
93 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
94 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
95 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
96 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
97
98 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
99 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
100 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
101 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
102 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
103 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
104 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
105 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
106
107 if XNN_LIKELY(nc >= 8) {
108 vst1q_f32(c3, vacc3x0123);
109 vst1q_f32(c3 + 4, vacc3x4567);
110 c3 = (float*) ((uintptr_t) c3 + cn_stride);
111 vst1q_f32(c2, vacc2x0123);
112 vst1q_f32(c2 + 4, vacc2x4567);
113 c2 = (float*) ((uintptr_t) c2 + cn_stride);
114 vst1q_f32(c1, vacc1x0123);
115 vst1q_f32(c1 + 4, vacc1x4567);
116 c1 = (float*) ((uintptr_t) c1 + cn_stride);
117 vst1q_f32(c0, vacc0x0123);
118 vst1q_f32(c0 + 4, vacc0x4567);
119 c0 = (float*) ((uintptr_t) c0 + cn_stride);
120
121 a = (const float*) ((uintptr_t) a - kc * 4);
122
123 nc -= 8;
124 } else {
125 if (nc & 4) {
126 vst1q_f32(c3, vacc3x0123); c3 += 4;
127 vst1q_f32(c2, vacc2x0123); c2 += 4;
128 vst1q_f32(c1, vacc1x0123); c1 += 4;
129 vst1q_f32(c0, vacc0x0123); c0 += 4;
130
131 vacc3x0123 = vacc3x4567;
132 vacc2x0123 = vacc2x4567;
133 vacc1x0123 = vacc1x4567;
134 vacc0x0123 = vacc0x4567;
135 }
136 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
137 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
138 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
139 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
140 if (nc & 2) {
141 vst1_f32(c3, vacc3x01); c3 += 2;
142 vst1_f32(c2, vacc2x01); c2 += 2;
143 vst1_f32(c1, vacc1x01); c1 += 2;
144 vst1_f32(c0, vacc0x01); c0 += 2;
145
146 vacc3x01 = vget_high_f32(vacc3x0123);
147 vacc2x01 = vget_high_f32(vacc2x0123);
148 vacc1x01 = vget_high_f32(vacc1x0123);
149 vacc0x01 = vget_high_f32(vacc0x0123);
150 }
151 if (nc & 1) {
152 vst1_lane_f32(c3, vacc3x01, 0);
153 vst1_lane_f32(c2, vacc2x01, 0);
154 vst1_lane_f32(c1, vacc1x01, 0);
155 vst1_lane_f32(c0, vacc0x01, 0);
156 }
157
158 nc = 0;
159 }
160 } while (nc != 0);
161 }