1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/neon-shuffle.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/gemm.h>
16
17
xnn_f32_gemm_minmax_ukernel_1x8s4__neonfma(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,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_gemm_minmax_ukernel_1x8s4__neonfma(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const float* restrict a,
23 size_t a_stride,
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 <= 1);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const float* a0 = a;
40 float* c0 = c;
41
42 do {
43 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
44 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
45
46 size_t k = kc;
47 while (k >= 4 * sizeof(float)) {
48 float32x4_t va0 = vld1q_f32(a0); a0 += 4;
49
50
51 const float32x4_t vb0123c0 = vld1q_f32(w + 0);
52 const float32x4_t vb4567c0 = vld1q_f32(w + 4);
53
54 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c0);
55 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c0);
56
57 va0 = vextq_f32(va0, va0, 1);
58
59 const float32x4_t vb0123c1 = vld1q_f32(w + 8);
60 const float32x4_t vb4567c1 = vld1q_f32(w + 12);
61
62 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c1);
63 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c1);
64
65 va0 = vextq_f32(va0, va0, 1);
66
67 const float32x4_t vb0123c2 = vld1q_f32(w + 16);
68 const float32x4_t vb4567c2 = vld1q_f32(w + 20);
69
70 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c2);
71 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c2);
72
73 va0 = vextq_f32(va0, va0, 1);
74
75 const float32x4_t vb0123c3 = vld1q_f32(w + 24);
76 const float32x4_t vb4567c3 = vld1q_f32(w + 28);
77
78 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c3);
79 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c3);
80
81
82 w += 32;
83 k -= 4 * sizeof(float);
84 }
85 if XNN_UNLIKELY(k != 0) {
86 do {
87 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
88
89 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
90 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
91
92 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123);
93 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567);
94
95 k -= sizeof(float);
96 } while (k != 0);
97 }
98 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
99 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
100 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
101
102 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
103 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
104 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
105
106 if XNN_LIKELY(nc >= 8) {
107 vst1q_f32(c0, vacc0x0123);
108 vst1q_f32(c0 + 4, vacc0x4567);
109 c0 = (float*) ((uintptr_t) c0 + cn_stride);
110
111 a0 = (const float*) ((uintptr_t) a0 - kc);
112
113 nc -= 8;
114
115 } else {
116 if (nc & 4) {
117 vst1q_f32(c0, vacc0x0123); c0 += 4;
118
119 vacc0x0123 = vacc0x4567;
120 }
121 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
122 if (nc & 2) {
123 vst1_f32(c0, vacc0x01); c0 += 2;
124
125 vacc0x01 = vget_high_f32(vacc0x0123);
126 }
127 if (nc & 1) {
128 vst1_lane_f32(c0, vacc0x01, 0);
129 }
130
131 nc = 0;
132 }
133 } while (nc != 0);
134 }
135