• 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_ukernel_6x8__neonfma_dup_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_output_params params[restrict static1])18 void xnn_f32_gemminc_ukernel_6x8__neonfma_dup_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_output_params params[restrict static 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       const float32x4_t va0c0 = vdupq_lane_f32(va0, 0);
101       const float32x4_t va1c0 = vdupq_lane_f32(va1, 0);
102       const float32x4_t va2c0 = vdupq_lane_f32(va2, 0);
103       const float32x4_t va3c0 = vdupq_lane_f32(va3, 0);
104       const float32x4_t va4c0 = vdupq_lane_f32(va4, 0);
105       const float32x4_t va5c0 = vdupq_lane_f32(va5, 0);
106       vacc0x0123 = vfmaq_f32(vacc0x0123, va0c0, vb0123c0);
107       vacc1x0123 = vfmaq_f32(vacc1x0123, va1c0, vb0123c0);
108       vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0);
109       vacc3x0123 = vfmaq_f32(vacc3x0123, va3c0, vb0123c0);
110       vacc4x0123 = vfmaq_f32(vacc4x0123, va4c0, vb0123c0);
111       vacc5x0123 = vfmaq_f32(vacc5x0123, va5c0, vb0123c0);
112       vacc0x4567 = vfmaq_f32(vacc0x4567, va0c0, vb4567c0);
113       vacc1x4567 = vfmaq_f32(vacc1x4567, va1c0, vb4567c0);
114       vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0);
115       vacc3x4567 = vfmaq_f32(vacc3x4567, va3c0, vb4567c0);
116       vacc4x4567 = vfmaq_f32(vacc4x4567, va4c0, vb4567c0);
117       vacc5x4567 = vfmaq_f32(vacc5x4567, va5c0, vb4567c0);
118       const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
119       const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
120 
121       const float32x4_t va0c1 = vdupq_lane_f32(va0, 1);
122       const float32x4_t va1c1 = vdupq_lane_f32(va1, 1);
123       const float32x4_t va2c1 = vdupq_lane_f32(va2, 1);
124       const float32x4_t va3c1 = vdupq_lane_f32(va3, 1);
125       const float32x4_t va4c1 = vdupq_lane_f32(va4, 1);
126       const float32x4_t va5c1 = vdupq_lane_f32(va5, 1);
127       vacc0x0123 = vfmaq_f32(vacc0x0123, va0c1, vb0123c1);
128       vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1);
129       vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1);
130       vacc3x0123 = vfmaq_f32(vacc3x0123, va3c1, vb0123c1);
131       vacc4x0123 = vfmaq_f32(vacc4x0123, va4c1, vb0123c1);
132       vacc5x0123 = vfmaq_f32(vacc5x0123, va5c1, vb0123c1);
133       vacc0x4567 = vfmaq_f32(vacc0x4567, va0c1, vb4567c1);
134       vacc1x4567 = vfmaq_f32(vacc1x4567, va1c1, vb4567c1);
135       vacc2x4567 = vfmaq_f32(vacc2x4567, va2c1, vb4567c1);
136       vacc3x4567 = vfmaq_f32(vacc3x4567, va3c1, vb4567c1);
137       vacc4x4567 = vfmaq_f32(vacc4x4567, va4c1, vb4567c1);
138       vacc5x4567 = vfmaq_f32(vacc5x4567, va5c1, vb4567c1);
139     }
140     if XNN_UNLIKELY(k != 0) {
141       const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
142       const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
143       const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
144       const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
145       const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1;
146       const float32x4_t va5 = vld1q_dup_f32(a5); a5 += 1;
147 
148       const float32x4_t vb0123 = vld1q_f32(w); w += 4;
149       const float32x4_t vb4567 = vld1q_f32(w); w += 4;
150 
151       vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123);
152       vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123);
153       vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123);
154       vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123);
155       vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123);
156       vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123);
157       vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567);
158       vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567);
159       vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567);
160       vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567);
161       vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567);
162       vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567);
163     }
164     const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
165     vacc0x0123 = vminq_f32(vacc0x0123, vmax);
166     vacc1x0123 = vminq_f32(vacc1x0123, vmax);
167     vacc2x0123 = vminq_f32(vacc2x0123, vmax);
168     vacc3x0123 = vminq_f32(vacc3x0123, vmax);
169     vacc4x0123 = vminq_f32(vacc4x0123, vmax);
170     vacc5x0123 = vminq_f32(vacc5x0123, vmax);
171     vacc0x4567 = vminq_f32(vacc0x4567, vmax);
172     vacc1x4567 = vminq_f32(vacc1x4567, vmax);
173     vacc2x4567 = vminq_f32(vacc2x4567, vmax);
174     vacc3x4567 = vminq_f32(vacc3x4567, vmax);
175     vacc4x4567 = vminq_f32(vacc4x4567, vmax);
176     vacc5x4567 = vminq_f32(vacc5x4567, vmax);
177 
178     const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
179     vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
180     vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
181     vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
182     vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
183     vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
184     vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
185     vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
186     vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
187     vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
188     vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
189     vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
190     vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
191 
192     if XNN_LIKELY(nc >= 8) {
193       vst1q_f32(c5, vacc5x0123);
194       vst1q_f32(c5 + 4, vacc5x4567);
195       c5 = (float*) ((uintptr_t) c5 + cn_stride);
196       vst1q_f32(c4, vacc4x0123);
197       vst1q_f32(c4 + 4, vacc4x4567);
198       c4 = (float*) ((uintptr_t) c4 + cn_stride);
199       vst1q_f32(c3, vacc3x0123);
200       vst1q_f32(c3 + 4, vacc3x4567);
201       c3 = (float*) ((uintptr_t) c3 + cn_stride);
202       vst1q_f32(c2, vacc2x0123);
203       vst1q_f32(c2 + 4, vacc2x4567);
204       c2 = (float*) ((uintptr_t) c2 + cn_stride);
205       vst1q_f32(c1, vacc1x0123);
206       vst1q_f32(c1 + 4, vacc1x4567);
207       c1 = (float*) ((uintptr_t) c1 + cn_stride);
208       vst1q_f32(c0, vacc0x0123);
209       vst1q_f32(c0 + 4, vacc0x4567);
210       c0 = (float*) ((uintptr_t) c0 + cn_stride);
211 
212       a5 = (const float*) ((uintptr_t) a5 - kc);
213       a4 = (const float*) ((uintptr_t) a4 - kc);
214       a3 = (const float*) ((uintptr_t) a3 - kc);
215       a2 = (const float*) ((uintptr_t) a2 - kc);
216       a1 = (const float*) ((uintptr_t) a1 - kc);
217       a0 = (const float*) ((uintptr_t) a0 - kc);
218 
219       nc -= 8;
220 
221     } else {
222       if (nc & 4) {
223         vst1q_f32(c5, vacc5x0123); c5 += 4;
224         vst1q_f32(c4, vacc4x0123); c4 += 4;
225         vst1q_f32(c3, vacc3x0123); c3 += 4;
226         vst1q_f32(c2, vacc2x0123); c2 += 4;
227         vst1q_f32(c1, vacc1x0123); c1 += 4;
228         vst1q_f32(c0, vacc0x0123); c0 += 4;
229 
230         vacc5x0123 = vacc5x4567;
231         vacc4x0123 = vacc4x4567;
232         vacc3x0123 = vacc3x4567;
233         vacc2x0123 = vacc2x4567;
234         vacc1x0123 = vacc1x4567;
235         vacc0x0123 = vacc0x4567;
236       }
237       float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
238       float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
239       float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
240       float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
241       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
242       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
243       if (nc & 2) {
244         vst1_f32(c5, vacc5x01); c5 += 2;
245         vst1_f32(c4, vacc4x01); c4 += 2;
246         vst1_f32(c3, vacc3x01); c3 += 2;
247         vst1_f32(c2, vacc2x01); c2 += 2;
248         vst1_f32(c1, vacc1x01); c1 += 2;
249         vst1_f32(c0, vacc0x01); c0 += 2;
250 
251         vacc5x01 = vget_high_f32(vacc5x0123);
252         vacc4x01 = vget_high_f32(vacc4x0123);
253         vacc3x01 = vget_high_f32(vacc3x0123);
254         vacc2x01 = vget_high_f32(vacc2x0123);
255         vacc1x01 = vget_high_f32(vacc1x0123);
256         vacc0x01 = vget_high_f32(vacc0x0123);
257       }
258       if (nc & 1) {
259         vst1_lane_f32(c5, vacc5x01, 0);
260         vst1_lane_f32(c4, vacc4x01, 0);
261         vst1_lane_f32(c3, vacc3x01, 0);
262         vst1_lane_f32(c2, vacc2x01, 0);
263         vst1_lane_f32(c1, vacc1x01, 0);
264         vst1_lane_f32(c0, vacc0x01, 0);
265       }
266 
267       nc = 0;
268     }
269   } while (nc != 0);
270 }
271