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_gemminc_minmax_ukernel_1x8s4__neon(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 float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_gemminc_minmax_ukernel_1x8s4__neon(
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 float*restrict acc,
29     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31   assert(mr != 0);
32   assert(mr <= 1);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(float) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39   assert(acc != NULL);
40 
41   const float* a0 = a;
42   float* c0 = c;
43 
44   do {
45     float32x4_t vacc0x0123 = vld1q_f32(acc); acc += 4;
46     float32x4_t vacc0x4567 = vld1q_f32(acc); acc += 4;
47 
48     size_t k = kc;
49     while (k >= 4 * sizeof(float)) {
50       float32x4_t va0 = vld1q_f32(a0); a0 += 4;
51 
52 
53       const float32x4_t vb0123c0 = vld1q_f32(w + 0);
54       const float32x4_t vb4567c0 = vld1q_f32(w + 4);
55 
56       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c0);
57       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c0);
58 
59       va0 = vextq_f32(va0, va0, 1);
60 
61       const float32x4_t vb0123c1 = vld1q_f32(w + 8);
62       const float32x4_t vb4567c1 = vld1q_f32(w + 12);
63 
64       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c1);
65       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c1);
66 
67       va0 = vextq_f32(va0, va0, 1);
68 
69       const float32x4_t vb0123c2 = vld1q_f32(w + 16);
70       const float32x4_t vb4567c2 = vld1q_f32(w + 20);
71 
72       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c2);
73       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c2);
74 
75       va0 = vextq_f32(va0, va0, 1);
76 
77       const float32x4_t vb0123c3 = vld1q_f32(w + 24);
78       const float32x4_t vb4567c3 = vld1q_f32(w + 28);
79 
80       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c3);
81       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c3);
82 
83 
84       w += 32;
85       k -= 4 * sizeof(float);
86     }
87     if XNN_UNLIKELY(k != 0) {
88       do {
89         const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
90 
91         const float32x4_t vb0123 = vld1q_f32(w); w += 4;
92         const float32x4_t vb4567 = vld1q_f32(w); w += 4;
93 
94         vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);
95         vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);
96 
97         k -= sizeof(float);
98       } while (k != 0);
99     }
100     const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
101     vacc0x0123 = vminq_f32(vacc0x0123, vmax);
102     vacc0x4567 = vminq_f32(vacc0x4567, vmax);
103 
104     const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
105     vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
106     vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
107 
108     if XNN_LIKELY(nc >= 8) {
109       vst1q_f32(c0, vacc0x0123);
110       vst1q_f32(c0 + 4, vacc0x4567);
111       c0 = (float*) ((uintptr_t) c0 + cn_stride);
112 
113       a0 = (const float*) ((uintptr_t) a0 - kc);
114 
115       nc -= 8;
116 
117     } else {
118       if (nc & 4) {
119         vst1q_f32(c0, vacc0x0123); c0 += 4;
120 
121         vacc0x0123 = vacc0x4567;
122       }
123       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
124       if (nc & 2) {
125         vst1_f32(c0, vacc0x01); c0 += 2;
126 
127         vacc0x01 = vget_high_f32(vacc0x0123);
128       }
129       if (nc & 1) {
130         vst1_lane_f32(c0, vacc0x01, 0);
131       }
132 
133       nc = 0;
134     }
135   } while (nc != 0);
136 }
137