1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-gemm/neon-ld128.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_4x8__neonfma_lane_ld128(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_4x8__neonfma_lane_ld128(
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 <= 4);
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   const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
44   float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     a1 = a0;
47     c1 = c0;
48   }
49   const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
50   float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51   if XNN_UNPREDICTABLE(mr <= 2) {
52     a2 = a1;
53     c2 = c1;
54   }
55   const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
56   float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57   if XNN_UNPREDICTABLE(mr != 4) {
58     a3 = a2;
59     c3 = c2;
60   }
61 
62   do {
63     float32x4_t vacc0x0123 = vld1q_f32(acc); acc += 4;
64     float32x4_t vacc0x4567 = vld1q_f32(acc); acc += 4;
65     float32x4_t vacc1x0123 = vld1q_f32(acc); acc += 4;
66     float32x4_t vacc1x4567 = vld1q_f32(acc); acc += 4;
67     float32x4_t vacc2x0123 = vld1q_f32(acc); acc += 4;
68     float32x4_t vacc2x4567 = vld1q_f32(acc); acc += 4;
69     float32x4_t vacc3x0123 = vld1q_f32(acc); acc += 4;
70     float32x4_t vacc3x4567 = vld1q_f32(acc); acc += 4;
71 
72     size_t k = kc;
73     for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
74       const float32x4_t va0 = vld1q_f32(a0); a0 += 4;
75       const float32x4_t va1 = vld1q_f32(a1); a1 += 4;
76       const float32x4_t va2 = vld1q_f32(a2); a2 += 4;
77       const float32x4_t va3 = vld1q_f32(a3); a3 += 4;
78 
79 
80       const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
81       const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
82 
83       vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, vget_low_f32(va0), 0);
84       vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, vget_low_f32(va1), 0);
85       vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, vget_low_f32(va2), 0);
86       vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, vget_low_f32(va3), 0);
87       vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, vget_low_f32(va0), 0);
88       vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, vget_low_f32(va1), 0);
89       vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, vget_low_f32(va2), 0);
90       vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, vget_low_f32(va3), 0);
91 
92       const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
93       const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
94 
95       vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, vget_low_f32(va0), 1);
96       vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, vget_low_f32(va1), 1);
97       vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, vget_low_f32(va2), 1);
98       vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, vget_low_f32(va3), 1);
99       vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c1, vget_low_f32(va0), 1);
100       vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c1, vget_low_f32(va1), 1);
101       vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c1, vget_low_f32(va2), 1);
102       vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c1, vget_low_f32(va3), 1);
103 
104       const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
105       const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
106 
107       vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c2, vget_high_f32(va0), 0);
108       vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c2, vget_high_f32(va1), 0);
109       vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c2, vget_high_f32(va2), 0);
110       vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c2, vget_high_f32(va3), 0);
111       vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c2, vget_high_f32(va0), 0);
112       vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c2, vget_high_f32(va1), 0);
113       vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c2, vget_high_f32(va2), 0);
114       vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c2, vget_high_f32(va3), 0);
115 
116       const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
117       const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
118 
119       vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c3, vget_high_f32(va0), 1);
120       vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c3, vget_high_f32(va1), 1);
121       vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c3, vget_high_f32(va2), 1);
122       vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c3, vget_high_f32(va3), 1);
123       vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c3, vget_high_f32(va0), 1);
124       vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c3, vget_high_f32(va1), 1);
125       vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c3, vget_high_f32(va2), 1);
126       vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c3, vget_high_f32(va3), 1);
127     }
128     if XNN_UNLIKELY(k != 0) {
129       do {
130         const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
131         const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
132         const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
133         const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
134 
135         const float32x4_t vb0123 = vld1q_f32(w); w += 4;
136         const float32x4_t vb4567 = vld1q_f32(w); w += 4;
137 
138         vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123);
139         vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123);
140         vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123);
141         vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123);
142         vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567);
143         vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567);
144         vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567);
145         vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567);
146 
147         k -= sizeof(float);
148       } while (k != 0);
149     }
150     const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
151     vacc0x0123 = vminq_f32(vacc0x0123, vmax);
152     vacc1x0123 = vminq_f32(vacc1x0123, vmax);
153     vacc2x0123 = vminq_f32(vacc2x0123, vmax);
154     vacc3x0123 = vminq_f32(vacc3x0123, vmax);
155     vacc0x4567 = vminq_f32(vacc0x4567, vmax);
156     vacc1x4567 = vminq_f32(vacc1x4567, vmax);
157     vacc2x4567 = vminq_f32(vacc2x4567, vmax);
158     vacc3x4567 = vminq_f32(vacc3x4567, vmax);
159 
160     const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
161     vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
162     vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
163     vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
164     vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
165     vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
166     vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
167     vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
168     vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
169 
170     if XNN_LIKELY(nc >= 8) {
171       vst1q_f32(c3, vacc3x0123);
172       vst1q_f32(c3 + 4, vacc3x4567);
173       c3 = (float*) ((uintptr_t) c3 + cn_stride);
174       vst1q_f32(c2, vacc2x0123);
175       vst1q_f32(c2 + 4, vacc2x4567);
176       c2 = (float*) ((uintptr_t) c2 + cn_stride);
177       vst1q_f32(c1, vacc1x0123);
178       vst1q_f32(c1 + 4, vacc1x4567);
179       c1 = (float*) ((uintptr_t) c1 + cn_stride);
180       vst1q_f32(c0, vacc0x0123);
181       vst1q_f32(c0 + 4, vacc0x4567);
182       c0 = (float*) ((uintptr_t) c0 + cn_stride);
183 
184       a3 = (const float*) ((uintptr_t) a3 - kc);
185       a2 = (const float*) ((uintptr_t) a2 - kc);
186       a1 = (const float*) ((uintptr_t) a1 - kc);
187       a0 = (const float*) ((uintptr_t) a0 - kc);
188 
189       nc -= 8;
190 
191     } else {
192       if (nc & 4) {
193         vst1q_f32(c3, vacc3x0123); c3 += 4;
194         vst1q_f32(c2, vacc2x0123); c2 += 4;
195         vst1q_f32(c1, vacc1x0123); c1 += 4;
196         vst1q_f32(c0, vacc0x0123); c0 += 4;
197 
198         vacc3x0123 = vacc3x4567;
199         vacc2x0123 = vacc2x4567;
200         vacc1x0123 = vacc1x4567;
201         vacc0x0123 = vacc0x4567;
202       }
203       float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
204       float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
205       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
206       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
207       if (nc & 2) {
208         vst1_f32(c3, vacc3x01); c3 += 2;
209         vst1_f32(c2, vacc2x01); c2 += 2;
210         vst1_f32(c1, vacc1x01); c1 += 2;
211         vst1_f32(c0, vacc0x01); c0 += 2;
212 
213         vacc3x01 = vget_high_f32(vacc3x0123);
214         vacc2x01 = vget_high_f32(vacc2x0123);
215         vacc1x01 = vget_high_f32(vacc1x0123);
216         vacc0x01 = vget_high_f32(vacc0x0123);
217       }
218       if (nc & 1) {
219         vst1_lane_f32(c3, vacc3x01, 0);
220         vst1_lane_f32(c2, vacc2x01, 0);
221         vst1_lane_f32(c1, vacc1x01, 0);
222         vst1_lane_f32(c0, vacc0x01, 0);
223       }
224 
225       nc = 0;
226     }
227   } while (nc != 0);
228 }
229