1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/neon-ld64.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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include <xnnpack/igemm.h>
16
17
xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
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 size_t a_offset,
29 const float* zero,
30 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(float) == 0);
37 assert(ks != 0);
38 assert(ks % (1 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(float) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 float* c0 = c;
45
46 do {
47 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
48 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
49
50 size_t p = ks;
51 do {
52 const float* restrict a0 = a[0];
53 assert(a0 != NULL);
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const float*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
61 const float32x2_t va0 = vld1_f32(a0); a0 += 2;
62
63 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
64 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
65
66 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, va0, 0);
67 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, va0, 0);
68 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
69 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
70
71 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, va0, 1);
72 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c1, va0, 1);
73
74 }
75 if XNN_UNLIKELY(k != 0) {
76 const float32x4_t va0 = vld1q_dup_f32(a0);
77
78 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
79 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
80
81 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);
82 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);
83 }
84 p -= 1 * sizeof(void*);
85 } while (p != 0);
86
87 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
88 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
89 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
90
91 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
92 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
93 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
94
95 if XNN_LIKELY(nc >= 8) {
96 vst1q_f32(c0, vacc0x0123);
97 vst1q_f32(c0 + 4, vacc0x4567);
98 c0 = (float*) ((uintptr_t) c0 + cn_stride);
99
100 a = (const float**restrict) ((uintptr_t) a - ks);
101 nc -= 8;
102 } else {
103 if (nc & 4) {
104 vst1q_f32(c0, vacc0x0123); c0 += 4;
105
106 vacc0x0123 = vacc0x4567;
107 }
108 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
109 if (nc & 2) {
110 vst1_f32(c0, vacc0x01); c0 += 2;
111
112 vacc0x01 = vget_high_f32(vacc0x0123);
113 }
114 if (nc & 1) {
115 vst1_lane_f32(c0, vacc0x01, 0);
116 }
117
118 nc = 0;
119 }
120 } while (nc != 0);
121 }
122