• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-gemm/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/gemm.h>
16 
17 
xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64(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_6x8__neon_lane_ld64(
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 <= 6);
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   const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
62   float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
63   if XNN_UNPREDICTABLE(mr <= 4) {
64     a4 = a3;
65     c4 = c3;
66   }
67   const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
68   float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
69   if XNN_UNPREDICTABLE(mr != 6) {
70     a5 = a4;
71     c5 = c4;
72   }
73 
74   do {
75     float32x4_t vacc0x0123 = vld1q_f32(acc); acc += 4;
76     float32x4_t vacc0x4567 = vld1q_f32(acc); acc += 4;
77     float32x4_t vacc1x0123 = vld1q_f32(acc); acc += 4;
78     float32x4_t vacc1x4567 = vld1q_f32(acc); acc += 4;
79     float32x4_t vacc2x0123 = vld1q_f32(acc); acc += 4;
80     float32x4_t vacc2x4567 = vld1q_f32(acc); acc += 4;
81     float32x4_t vacc3x0123 = vld1q_f32(acc); acc += 4;
82     float32x4_t vacc3x4567 = vld1q_f32(acc); acc += 4;
83     float32x4_t vacc4x0123 = vld1q_f32(acc); acc += 4;
84     float32x4_t vacc4x4567 = vld1q_f32(acc); acc += 4;
85     float32x4_t vacc5x0123 = vld1q_f32(acc); acc += 4;
86     float32x4_t vacc5x4567 = vld1q_f32(acc); acc += 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       const float32x2_t va4 = vld1_f32(a4); a4 += 2;
95       const float32x2_t va5 = vld1_f32(a5); a5 += 2;
96 
97       const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
98       const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
99 
100       vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, va0, 0);
101       vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, va1, 0);
102       vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, va2, 0);
103       vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, va3, 0);
104       vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c0, va4, 0);
105       vacc5x0123 = vmlaq_lane_f32(vacc5x0123, vb0123c0, va5, 0);
106       vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, va0, 0);
107       vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, va1, 0);
108       vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, va2, 0);
109       vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, va3, 0);
110       vacc4x4567 = vmlaq_lane_f32(vacc4x4567, vb4567c0, va4, 0);
111       vacc5x4567 = vmlaq_lane_f32(vacc5x4567, vb4567c0, va5, 0);
112       const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
113       const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
114 
115       vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, va0, 1);
116       vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, va1, 1);
117       vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, va2, 1);
118       vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, va3, 1);
119       vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c1, va4, 1);
120       vacc5x0123 = vmlaq_lane_f32(vacc5x0123, vb0123c1, va5, 1);
121       vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c1, va0, 1);
122       vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c1, va1, 1);
123       vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c1, va2, 1);
124       vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c1, va3, 1);
125       vacc4x4567 = vmlaq_lane_f32(vacc4x4567, vb4567c1, va4, 1);
126       vacc5x4567 = vmlaq_lane_f32(vacc5x4567, vb4567c1, va5, 1);
127     }
128     if XNN_UNLIKELY(k != 0) {
129       const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
130       const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
131       const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
132       const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
133       const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1;
134       const float32x4_t va5 = vld1q_dup_f32(a5); a5 += 1;
135 
136       const float32x4_t vb0123 = vld1q_f32(w); w += 4;
137       const float32x4_t vb4567 = vld1q_f32(w); w += 4;
138 
139       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);
140       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123);
141       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123);
142       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123);
143       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123);
144       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123);
145       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);
146       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567);
147       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567);
148       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567);
149       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567);
150       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567);
151     }
152     const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
153     vacc0x0123 = vminq_f32(vacc0x0123, vmax);
154     vacc1x0123 = vminq_f32(vacc1x0123, vmax);
155     vacc2x0123 = vminq_f32(vacc2x0123, vmax);
156     vacc3x0123 = vminq_f32(vacc3x0123, vmax);
157     vacc4x0123 = vminq_f32(vacc4x0123, vmax);
158     vacc5x0123 = vminq_f32(vacc5x0123, vmax);
159     vacc0x4567 = vminq_f32(vacc0x4567, vmax);
160     vacc1x4567 = vminq_f32(vacc1x4567, vmax);
161     vacc2x4567 = vminq_f32(vacc2x4567, vmax);
162     vacc3x4567 = vminq_f32(vacc3x4567, vmax);
163     vacc4x4567 = vminq_f32(vacc4x4567, vmax);
164     vacc5x4567 = vminq_f32(vacc5x4567, vmax);
165 
166     const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
167     vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
168     vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
169     vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
170     vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
171     vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
172     vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
173     vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
174     vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
175     vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
176     vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
177     vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
178     vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
179 
180     if XNN_LIKELY(nc >= 8) {
181       vst1q_f32(c5, vacc5x0123);
182       vst1q_f32(c5 + 4, vacc5x4567);
183       c5 = (float*) ((uintptr_t) c5 + cn_stride);
184       vst1q_f32(c4, vacc4x0123);
185       vst1q_f32(c4 + 4, vacc4x4567);
186       c4 = (float*) ((uintptr_t) c4 + cn_stride);
187       vst1q_f32(c3, vacc3x0123);
188       vst1q_f32(c3 + 4, vacc3x4567);
189       c3 = (float*) ((uintptr_t) c3 + cn_stride);
190       vst1q_f32(c2, vacc2x0123);
191       vst1q_f32(c2 + 4, vacc2x4567);
192       c2 = (float*) ((uintptr_t) c2 + cn_stride);
193       vst1q_f32(c1, vacc1x0123);
194       vst1q_f32(c1 + 4, vacc1x4567);
195       c1 = (float*) ((uintptr_t) c1 + cn_stride);
196       vst1q_f32(c0, vacc0x0123);
197       vst1q_f32(c0 + 4, vacc0x4567);
198       c0 = (float*) ((uintptr_t) c0 + cn_stride);
199 
200       a5 = (const float*) ((uintptr_t) a5 - kc);
201       a4 = (const float*) ((uintptr_t) a4 - kc);
202       a3 = (const float*) ((uintptr_t) a3 - kc);
203       a2 = (const float*) ((uintptr_t) a2 - kc);
204       a1 = (const float*) ((uintptr_t) a1 - kc);
205       a0 = (const float*) ((uintptr_t) a0 - kc);
206 
207       nc -= 8;
208 
209     } else {
210       if (nc & 4) {
211         vst1q_f32(c5, vacc5x0123); c5 += 4;
212         vst1q_f32(c4, vacc4x0123); c4 += 4;
213         vst1q_f32(c3, vacc3x0123); c3 += 4;
214         vst1q_f32(c2, vacc2x0123); c2 += 4;
215         vst1q_f32(c1, vacc1x0123); c1 += 4;
216         vst1q_f32(c0, vacc0x0123); c0 += 4;
217 
218         vacc5x0123 = vacc5x4567;
219         vacc4x0123 = vacc4x4567;
220         vacc3x0123 = vacc3x4567;
221         vacc2x0123 = vacc2x4567;
222         vacc1x0123 = vacc1x4567;
223         vacc0x0123 = vacc0x4567;
224       }
225       float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
226       float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
227       float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
228       float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
229       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
230       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
231       if (nc & 2) {
232         vst1_f32(c5, vacc5x01); c5 += 2;
233         vst1_f32(c4, vacc4x01); c4 += 2;
234         vst1_f32(c3, vacc3x01); c3 += 2;
235         vst1_f32(c2, vacc2x01); c2 += 2;
236         vst1_f32(c1, vacc1x01); c1 += 2;
237         vst1_f32(c0, vacc0x01); c0 += 2;
238 
239         vacc5x01 = vget_high_f32(vacc5x0123);
240         vacc4x01 = vget_high_f32(vacc4x0123);
241         vacc3x01 = vget_high_f32(vacc3x0123);
242         vacc2x01 = vget_high_f32(vacc2x0123);
243         vacc1x01 = vget_high_f32(vacc1x0123);
244         vacc0x01 = vget_high_f32(vacc0x0123);
245       }
246       if (nc & 1) {
247         vst1_lane_f32(c5, vacc5x01, 0);
248         vst1_lane_f32(c4, vacc4x01, 0);
249         vst1_lane_f32(c3, vacc3x01, 0);
250         vst1_lane_f32(c2, vacc2x01, 0);
251         vst1_lane_f32(c1, vacc1x01, 0);
252         vst1_lane_f32(c0, vacc0x01, 0);
253       }
254 
255       nc = 0;
256     }
257   } while (nc != 0);
258 }
259