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_gemm_minmax_ukernel_6x8__neon_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_gemm_minmax_ukernel_6x8__neon_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 6);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const float* a0 = a;
40 float* c0 = c;
41 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
42 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
54 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 a3 = a2;
57 c3 = c2;
58 }
59 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
60 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 a4 = a3;
63 c4 = c3;
64 }
65 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
66 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
67 if XNN_UNPREDICTABLE(mr != 6) {
68 a5 = a4;
69 c5 = c4;
70 }
71
72 do {
73 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
74 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
75 float32x4_t vacc1x0123 = vacc0x0123;
76 float32x4_t vacc1x4567 = vacc0x4567;
77 float32x4_t vacc2x0123 = vacc0x0123;
78 float32x4_t vacc2x4567 = vacc0x4567;
79 float32x4_t vacc3x0123 = vacc0x0123;
80 float32x4_t vacc3x4567 = vacc0x4567;
81 float32x4_t vacc4x0123 = vacc0x0123;
82 float32x4_t vacc4x4567 = vacc0x4567;
83 float32x4_t vacc5x0123 = vacc0x0123;
84 float32x4_t vacc5x4567 = vacc0x4567;
85
86 size_t k = kc;
87 for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
88 const float32x2_t va0 = vld1_f32(a0); a0 += 2;
89 const float32x2_t va1 = vld1_f32(a1); a1 += 2;
90 const float32x2_t va2 = vld1_f32(a2); a2 += 2;
91 const float32x2_t va3 = vld1_f32(a3); a3 += 2;
92 const float32x2_t va4 = vld1_f32(a4); a4 += 2;
93 const float32x2_t va5 = vld1_f32(a5); a5 += 2;
94
95 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
96 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
97
98 const float32x4_t va0c0 = vdupq_lane_f32(va0, 0);
99 const float32x4_t va1c0 = vdupq_lane_f32(va1, 0);
100 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0);
101 const float32x4_t va3c0 = vdupq_lane_f32(va3, 0);
102 const float32x4_t va4c0 = vdupq_lane_f32(va4, 0);
103 const float32x4_t va5c0 = vdupq_lane_f32(va5, 0);
104 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c0, vb0123c0);
105 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c0, vb0123c0);
106 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0);
107 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c0, vb0123c0);
108 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c0, vb0123c0);
109 vacc5x0123 = vmlaq_f32(vacc5x0123, va5c0, vb0123c0);
110 vacc0x4567 = vmlaq_f32(vacc0x4567, va0c0, vb4567c0);
111 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c0, vb4567c0);
112 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0);
113 vacc3x4567 = vmlaq_f32(vacc3x4567, va3c0, vb4567c0);
114 vacc4x4567 = vmlaq_f32(vacc4x4567, va4c0, vb4567c0);
115 vacc5x4567 = vmlaq_f32(vacc5x4567, va5c0, vb4567c0);
116 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
117 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
118
119 const float32x4_t va0c1 = vdupq_lane_f32(va0, 1);
120 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1);
121 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1);
122 const float32x4_t va3c1 = vdupq_lane_f32(va3, 1);
123 const float32x4_t va4c1 = vdupq_lane_f32(va4, 1);
124 const float32x4_t va5c1 = vdupq_lane_f32(va5, 1);
125 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c1, vb0123c1);
126 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1);
127 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1);
128 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c1, vb0123c1);
129 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c1, vb0123c1);
130 vacc5x0123 = vmlaq_f32(vacc5x0123, va5c1, vb0123c1);
131 vacc0x4567 = vmlaq_f32(vacc0x4567, va0c1, vb4567c1);
132 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c1, vb4567c1);
133 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c1, vb4567c1);
134 vacc3x4567 = vmlaq_f32(vacc3x4567, va3c1, vb4567c1);
135 vacc4x4567 = vmlaq_f32(vacc4x4567, va4c1, vb4567c1);
136 vacc5x4567 = vmlaq_f32(vacc5x4567, va5c1, vb4567c1);
137 }
138 if XNN_UNLIKELY(k != 0) {
139 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
140 const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
141 const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
142 const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
143 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1;
144 const float32x4_t va5 = vld1q_dup_f32(a5); a5 += 1;
145
146 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
147 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
148
149 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);
150 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123);
151 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123);
152 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123);
153 vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123);
154 vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123);
155 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);
156 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567);
157 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567);
158 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567);
159 vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567);
160 vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567);
161 }
162 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
163 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
164 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
165 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
166 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
167 vacc4x0123 = vminq_f32(vacc4x0123, vmax);
168 vacc5x0123 = vminq_f32(vacc5x0123, vmax);
169 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
170 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
171 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
172 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
173 vacc4x4567 = vminq_f32(vacc4x4567, vmax);
174 vacc5x4567 = vminq_f32(vacc5x4567, vmax);
175
176 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
177 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
178 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
179 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
180 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
181 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
182 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
183 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
184 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
185 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
186 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
187 vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
188 vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
189
190 if XNN_LIKELY(nc >= 8) {
191 vst1q_f32(c5, vacc5x0123);
192 vst1q_f32(c5 + 4, vacc5x4567);
193 c5 = (float*) ((uintptr_t) c5 + cn_stride);
194 vst1q_f32(c4, vacc4x0123);
195 vst1q_f32(c4 + 4, vacc4x4567);
196 c4 = (float*) ((uintptr_t) c4 + cn_stride);
197 vst1q_f32(c3, vacc3x0123);
198 vst1q_f32(c3 + 4, vacc3x4567);
199 c3 = (float*) ((uintptr_t) c3 + cn_stride);
200 vst1q_f32(c2, vacc2x0123);
201 vst1q_f32(c2 + 4, vacc2x4567);
202 c2 = (float*) ((uintptr_t) c2 + cn_stride);
203 vst1q_f32(c1, vacc1x0123);
204 vst1q_f32(c1 + 4, vacc1x4567);
205 c1 = (float*) ((uintptr_t) c1 + cn_stride);
206 vst1q_f32(c0, vacc0x0123);
207 vst1q_f32(c0 + 4, vacc0x4567);
208 c0 = (float*) ((uintptr_t) c0 + cn_stride);
209
210 a5 = (const float*) ((uintptr_t) a5 - kc);
211 a4 = (const float*) ((uintptr_t) a4 - kc);
212 a3 = (const float*) ((uintptr_t) a3 - kc);
213 a2 = (const float*) ((uintptr_t) a2 - kc);
214 a1 = (const float*) ((uintptr_t) a1 - kc);
215 a0 = (const float*) ((uintptr_t) a0 - kc);
216
217 nc -= 8;
218
219 } else {
220 if (nc & 4) {
221 vst1q_f32(c5, vacc5x0123); c5 += 4;
222 vst1q_f32(c4, vacc4x0123); c4 += 4;
223 vst1q_f32(c3, vacc3x0123); c3 += 4;
224 vst1q_f32(c2, vacc2x0123); c2 += 4;
225 vst1q_f32(c1, vacc1x0123); c1 += 4;
226 vst1q_f32(c0, vacc0x0123); c0 += 4;
227
228 vacc5x0123 = vacc5x4567;
229 vacc4x0123 = vacc4x4567;
230 vacc3x0123 = vacc3x4567;
231 vacc2x0123 = vacc2x4567;
232 vacc1x0123 = vacc1x4567;
233 vacc0x0123 = vacc0x4567;
234 }
235 float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
236 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
237 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
238 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
239 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
240 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
241 if (nc & 2) {
242 vst1_f32(c5, vacc5x01); c5 += 2;
243 vst1_f32(c4, vacc4x01); c4 += 2;
244 vst1_f32(c3, vacc3x01); c3 += 2;
245 vst1_f32(c2, vacc2x01); c2 += 2;
246 vst1_f32(c1, vacc1x01); c1 += 2;
247 vst1_f32(c0, vacc0x01); c0 += 2;
248
249 vacc5x01 = vget_high_f32(vacc5x0123);
250 vacc4x01 = vget_high_f32(vacc4x0123);
251 vacc3x01 = vget_high_f32(vacc3x0123);
252 vacc2x01 = vget_high_f32(vacc2x0123);
253 vacc1x01 = vget_high_f32(vacc1x0123);
254 vacc0x01 = vget_high_f32(vacc0x0123);
255 }
256 if (nc & 1) {
257 vst1_lane_f32(c5, vacc5x01, 0);
258 vst1_lane_f32(c4, vacc4x01, 0);
259 vst1_lane_f32(c3, vacc3x01, 0);
260 vst1_lane_f32(c2, vacc2x01, 0);
261 vst1_lane_f32(c1, vacc1x01, 0);
262 vst1_lane_f32(c0, vacc0x01, 0);
263 }
264
265 nc = 0;
266 }
267 } while (nc != 0);
268 }
269