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_5x8__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_5x8__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 <= 5);
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
68 do {
69 float32x4_t vacc0x0123 = vld1q_f32(acc); acc += 4;
70 float32x4_t vacc0x4567 = vld1q_f32(acc); acc += 4;
71 float32x4_t vacc1x0123 = vld1q_f32(acc); acc += 4;
72 float32x4_t vacc1x4567 = vld1q_f32(acc); acc += 4;
73 float32x4_t vacc2x0123 = vld1q_f32(acc); acc += 4;
74 float32x4_t vacc2x4567 = vld1q_f32(acc); acc += 4;
75 float32x4_t vacc3x0123 = vld1q_f32(acc); acc += 4;
76 float32x4_t vacc3x4567 = vld1q_f32(acc); acc += 4;
77 float32x4_t vacc4x0123 = vld1q_f32(acc); acc += 4;
78 float32x4_t vacc4x4567 = vld1q_f32(acc); acc += 4;
79
80 size_t k = kc;
81 for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
82 const float32x2_t va0 = vld1_f32(a0); a0 += 2;
83 const float32x2_t va1 = vld1_f32(a1); a1 += 2;
84 const float32x2_t va2 = vld1_f32(a2); a2 += 2;
85 const float32x2_t va3 = vld1_f32(a3); a3 += 2;
86 const float32x2_t va4 = vld1_f32(a4); a4 += 2;
87
88 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
89 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
90
91 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, va0, 0);
92 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, va1, 0);
93 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, va2, 0);
94 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, va3, 0);
95 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c0, va4, 0);
96 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, va0, 0);
97 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, va1, 0);
98 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, va2, 0);
99 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, va3, 0);
100 vacc4x4567 = vmlaq_lane_f32(vacc4x4567, vb4567c0, va4, 0);
101 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
102 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
103
104 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, va0, 1);
105 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, va1, 1);
106 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, va2, 1);
107 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, va3, 1);
108 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c1, va4, 1);
109 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c1, va0, 1);
110 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c1, va1, 1);
111 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c1, va2, 1);
112 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c1, va3, 1);
113 vacc4x4567 = vmlaq_lane_f32(vacc4x4567, vb4567c1, va4, 1);
114 }
115 if XNN_UNLIKELY(k != 0) {
116 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
117 const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
118 const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
119 const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
120 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1;
121
122 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
123 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
124
125 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);
126 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123);
127 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123);
128 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123);
129 vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123);
130 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);
131 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567);
132 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567);
133 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567);
134 vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567);
135 }
136 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
137 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
138 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
139 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
140 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
141 vacc4x0123 = vminq_f32(vacc4x0123, vmax);
142 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
143 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
144 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
145 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
146 vacc4x4567 = vminq_f32(vacc4x4567, vmax);
147
148 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
149 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
150 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
151 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
152 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
153 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
154 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
155 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
156 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
157 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
158 vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
159
160 if XNN_LIKELY(nc >= 8) {
161 vst1q_f32(c4, vacc4x0123);
162 vst1q_f32(c4 + 4, vacc4x4567);
163 c4 = (float*) ((uintptr_t) c4 + cn_stride);
164 vst1q_f32(c3, vacc3x0123);
165 vst1q_f32(c3 + 4, vacc3x4567);
166 c3 = (float*) ((uintptr_t) c3 + cn_stride);
167 vst1q_f32(c2, vacc2x0123);
168 vst1q_f32(c2 + 4, vacc2x4567);
169 c2 = (float*) ((uintptr_t) c2 + cn_stride);
170 vst1q_f32(c1, vacc1x0123);
171 vst1q_f32(c1 + 4, vacc1x4567);
172 c1 = (float*) ((uintptr_t) c1 + cn_stride);
173 vst1q_f32(c0, vacc0x0123);
174 vst1q_f32(c0 + 4, vacc0x4567);
175 c0 = (float*) ((uintptr_t) c0 + cn_stride);
176
177 a4 = (const float*) ((uintptr_t) a4 - kc);
178 a3 = (const float*) ((uintptr_t) a3 - kc);
179 a2 = (const float*) ((uintptr_t) a2 - kc);
180 a1 = (const float*) ((uintptr_t) a1 - kc);
181 a0 = (const float*) ((uintptr_t) a0 - kc);
182
183 nc -= 8;
184
185 } else {
186 if (nc & 4) {
187 vst1q_f32(c4, vacc4x0123); c4 += 4;
188 vst1q_f32(c3, vacc3x0123); c3 += 4;
189 vst1q_f32(c2, vacc2x0123); c2 += 4;
190 vst1q_f32(c1, vacc1x0123); c1 += 4;
191 vst1q_f32(c0, vacc0x0123); c0 += 4;
192
193 vacc4x0123 = vacc4x4567;
194 vacc3x0123 = vacc3x4567;
195 vacc2x0123 = vacc2x4567;
196 vacc1x0123 = vacc1x4567;
197 vacc0x0123 = vacc0x4567;
198 }
199 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
200 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
201 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
202 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
203 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
204 if (nc & 2) {
205 vst1_f32(c4, vacc4x01); c4 += 2;
206 vst1_f32(c3, vacc3x01); c3 += 2;
207 vst1_f32(c2, vacc2x01); c2 += 2;
208 vst1_f32(c1, vacc1x01); c1 += 2;
209 vst1_f32(c0, vacc0x01); c0 += 2;
210
211 vacc4x01 = vget_high_f32(vacc4x0123);
212 vacc3x01 = vget_high_f32(vacc3x0123);
213 vacc2x01 = vget_high_f32(vacc2x0123);
214 vacc1x01 = vget_high_f32(vacc1x0123);
215 vacc0x01 = vget_high_f32(vacc0x0123);
216 }
217 if (nc & 1) {
218 vst1_lane_f32(c4, vacc4x01, 0);
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