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
xnn_f32_ppmm_minmax_ukernel_4x8__neonfma(size_t mr,size_t nc,size_t kc,const float * restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_ppmm_minmax_ukernel_4x8__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 float* c0 = c;
36 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
37 if XNN_UNPREDICTABLE(mr < 2) {
38 c1 = c0;
39 }
40 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
41 if XNN_UNPREDICTABLE(mr <= 2) {
42 c2 = c1;
43 }
44 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 4) {
46 c3 = c2;
47 }
48
49 do {
50 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
51 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
52 float32x4_t vacc1x0123 = vacc0x0123;
53 float32x4_t vacc1x4567 = vacc0x4567;
54 float32x4_t vacc2x0123 = vacc0x0123;
55 float32x4_t vacc2x4567 = vacc0x4567;
56 float32x4_t vacc3x0123 = vacc0x0123;
57 float32x4_t vacc3x4567 = vacc0x4567;
58
59 size_t k = kc;
60 do {
61 const float32x4_t va0123 = vld1q_f32(a); a += 4;
62
63 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
64 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
65
66 #if XNN_ARCH_ARM64
67 vacc0x0123 = vfmaq_laneq_f32(vacc0x0123, vb0123, va0123, 0);
68 vacc1x0123 = vfmaq_laneq_f32(vacc1x0123, vb0123, va0123, 1);
69 vacc2x0123 = vfmaq_laneq_f32(vacc2x0123, vb0123, va0123, 2);
70 vacc3x0123 = vfmaq_laneq_f32(vacc3x0123, vb0123, va0123, 3);
71 vacc0x4567 = vfmaq_laneq_f32(vacc0x4567, vb4567, va0123, 0);
72 vacc1x4567 = vfmaq_laneq_f32(vacc1x4567, vb4567, va0123, 1);
73 vacc2x4567 = vfmaq_laneq_f32(vacc2x4567, vb4567, va0123, 2);
74 vacc3x4567 = vfmaq_laneq_f32(vacc3x4567, vb4567, va0123, 3);
75 #else
76 const float32x4_t va0000 = vdupq_lane_f32(vget_low_f32(va0123), 0);
77 const float32x4_t va1111 = vdupq_lane_f32(vget_low_f32(va0123), 1);
78 const float32x4_t va2222 = vdupq_lane_f32(vget_high_f32(va0123), 0);
79 const float32x4_t va3333 = vdupq_lane_f32(vget_high_f32(va0123), 1);
80
81 vacc0x0123 = vfmaq_f32(vacc0x0123, va0000, vb0123);
82 vacc1x0123 = vfmaq_f32(vacc1x0123, va1111, vb0123);
83 vacc2x0123 = vfmaq_f32(vacc2x0123, va2222, vb0123);
84 vacc3x0123 = vfmaq_f32(vacc3x0123, va3333, vb0123);
85 vacc0x4567 = vfmaq_f32(vacc0x4567, va0000, vb4567);
86 vacc1x4567 = vfmaq_f32(vacc1x4567, va1111, vb4567);
87 vacc2x4567 = vfmaq_f32(vacc2x4567, va2222, vb4567);
88 vacc3x4567 = vfmaq_f32(vacc3x4567, va3333, vb4567);
89 #endif
90
91 k -= sizeof(float);
92 } while (k != 0);
93
94 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
95 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
96 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
97 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
98 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
99 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
100 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
101 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
102 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
103
104 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
105 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
106 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
107 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
108 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
109 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
110 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
111 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
112 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
113
114 if XNN_LIKELY(nc >= 8) {
115 vst1q_f32(c3, vacc3x0123);
116 vst1q_f32(c3 + 4, vacc3x4567);
117 c3 = (float*) ((uintptr_t) c3 + cn_stride);
118 vst1q_f32(c2, vacc2x0123);
119 vst1q_f32(c2 + 4, vacc2x4567);
120 c2 = (float*) ((uintptr_t) c2 + cn_stride);
121 vst1q_f32(c1, vacc1x0123);
122 vst1q_f32(c1 + 4, vacc1x4567);
123 c1 = (float*) ((uintptr_t) c1 + cn_stride);
124 vst1q_f32(c0, vacc0x0123);
125 vst1q_f32(c0 + 4, vacc0x4567);
126 c0 = (float*) ((uintptr_t) c0 + cn_stride);
127
128 a = (const float*) ((uintptr_t) a - kc * 4);
129
130 nc -= 8;
131 } else {
132 if (nc & 4) {
133 vst1q_f32(c3, vacc3x0123); c3 += 4;
134 vst1q_f32(c2, vacc2x0123); c2 += 4;
135 vst1q_f32(c1, vacc1x0123); c1 += 4;
136 vst1q_f32(c0, vacc0x0123); c0 += 4;
137
138 vacc3x0123 = vacc3x4567;
139 vacc2x0123 = vacc2x4567;
140 vacc1x0123 = vacc1x4567;
141 vacc0x0123 = vacc0x4567;
142 }
143 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
144 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
145 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
146 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
147 if (nc & 2) {
148 vst1_f32(c3, vacc3x01); c3 += 2;
149 vst1_f32(c2, vacc2x01); c2 += 2;
150 vst1_f32(c1, vacc1x01); c1 += 2;
151 vst1_f32(c0, vacc0x01); c0 += 2;
152
153 vacc3x01 = vget_high_f32(vacc3x0123);
154 vacc2x01 = vget_high_f32(vacc2x0123);
155 vacc1x01 = vget_high_f32(vacc1x0123);
156 vacc0x01 = vget_high_f32(vacc0x0123);
157 }
158 if (nc & 1) {
159 vst1_lane_f32(c3, vacc3x01, 0);
160 vst1_lane_f32(c2, vacc2x01, 0);
161 vst1_lane_f32(c1, vacc1x01, 0);
162 vst1_lane_f32(c0, vacc0x01, 0);
163 }
164
165 nc = 0;
166 }
167 } while (nc != 0);
168 }
169