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_4x4__neonfma_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_4x4__neonfma_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 <= 4);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(kc % sizeof(float) == 0);
37   assert(ks != 0);
38   assert(ks % (4 * 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   float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr < 2) {
47     c1 = c0;
48   }
49   float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     c2 = c1;
52   }
53   float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54   if XNN_UNPREDICTABLE(mr != 4) {
55     c3 = c2;
56   }
57 
58   do {
59     float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
60     float32x4_t vacc1x0123 = vacc0x0123;
61     float32x4_t vacc2x0123 = vacc0x0123;
62     float32x4_t vacc3x0123 = vacc0x0123;
63 
64     size_t p = ks;
65     do {
66       const float* restrict a0 = a[0];
67       assert(a0 != NULL);
68       if XNN_UNPREDICTABLE(a0 != zero) {
69         a0 = (const float*) ((uintptr_t) a0 + a_offset);
70       }
71       const float* restrict a1 = a[1];
72       assert(a1 != NULL);
73       if XNN_UNPREDICTABLE(a1 != zero) {
74         a1 = (const float*) ((uintptr_t) a1 + a_offset);
75       }
76       const float* restrict a2 = a[2];
77       assert(a2 != NULL);
78       if XNN_UNPREDICTABLE(a2 != zero) {
79         a2 = (const float*) ((uintptr_t) a2 + a_offset);
80       }
81       const float* restrict a3 = a[3];
82       assert(a3 != NULL);
83       if XNN_UNPREDICTABLE(a3 != zero) {
84         a3 = (const float*) ((uintptr_t) a3 + a_offset);
85       }
86       a += 4;
87 
88       size_t k = kc;
89       for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
90         const float32x2_t va0 = vld1_f32(a0); a0 += 2;
91         const float32x2_t va1 = vld1_f32(a1); a1 += 2;
92         const float32x2_t va2 = vld1_f32(a2); a2 += 2;
93         const float32x2_t va3 = vld1_f32(a3); a3 += 2;
94 
95         const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
96 
97         vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, va0, 0);
98         vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, va1, 0);
99         vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, va2, 0);
100         vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, va3, 0);
101         const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
102 
103         vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, va0, 1);
104         vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, va1, 1);
105         vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, va2, 1);
106         vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, va3, 1);
107 
108       }
109       if XNN_UNLIKELY(k != 0) {
110         const float32x4_t va0 = vld1q_dup_f32(a0);
111         const float32x4_t va1 = vld1q_dup_f32(a1);
112         const float32x4_t va2 = vld1q_dup_f32(a2);
113         const float32x4_t va3 = vld1q_dup_f32(a3);
114 
115         const float32x4_t vb0123 = vld1q_f32(w); w += 4;
116 
117         vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123);
118         vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123);
119         vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123);
120         vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123);
121       }
122       p -= 4 * sizeof(void*);
123     } while (p != 0);
124 
125     const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
126     vacc0x0123 = vminq_f32(vacc0x0123, vmax);
127     vacc1x0123 = vminq_f32(vacc1x0123, vmax);
128     vacc2x0123 = vminq_f32(vacc2x0123, vmax);
129     vacc3x0123 = vminq_f32(vacc3x0123, vmax);
130 
131     const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
132     vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
133     vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
134     vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
135     vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
136 
137     if XNN_LIKELY(nc >= 4) {
138       vst1q_f32(c3, vacc3x0123);
139       c3 = (float*) ((uintptr_t) c3 + cn_stride);
140       vst1q_f32(c2, vacc2x0123);
141       c2 = (float*) ((uintptr_t) c2 + cn_stride);
142       vst1q_f32(c1, vacc1x0123);
143       c1 = (float*) ((uintptr_t) c1 + cn_stride);
144       vst1q_f32(c0, vacc0x0123);
145       c0 = (float*) ((uintptr_t) c0 + cn_stride);
146 
147       a = (const float**restrict) ((uintptr_t) a - ks);
148       nc -= 4;
149     } else {
150       float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
151       float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
152       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
153       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
154       if (nc & 2) {
155         vst1_f32(c3, vacc3x01); c3 += 2;
156         vst1_f32(c2, vacc2x01); c2 += 2;
157         vst1_f32(c1, vacc1x01); c1 += 2;
158         vst1_f32(c0, vacc0x01); c0 += 2;
159 
160         vacc3x01 = vget_high_f32(vacc3x0123);
161         vacc2x01 = vget_high_f32(vacc2x0123);
162         vacc1x01 = vget_high_f32(vacc1x0123);
163         vacc0x01 = vget_high_f32(vacc0x0123);
164       }
165       if (nc & 1) {
166         vst1_lane_f32(c3, vacc3x01, 0);
167         vst1_lane_f32(c2, vacc2x01, 0);
168         vst1_lane_f32(c1, vacc1x01, 0);
169         vst1_lane_f32(c0, vacc0x01, 0);
170       }
171 
172       nc = 0;
173     }
174   } while (nc != 0);
175 }
176