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_gemm_ukernel_4x8__neon_dup_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 union xnn_f32_output_params params[restrict static1])18 void xnn_f32_gemm_ukernel_4x8__neon_dup_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 union xnn_f32_output_params params[restrict static 1])
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
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
60 do {
61 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
62 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
63 float32x4_t vacc1x0123 = vacc0x0123;
64 float32x4_t vacc1x4567 = vacc0x4567;
65 float32x4_t vacc2x0123 = vacc0x0123;
66 float32x4_t vacc2x4567 = vacc0x4567;
67 float32x4_t vacc3x0123 = vacc0x0123;
68 float32x4_t vacc3x4567 = vacc0x4567;
69
70 size_t k = kc;
71 for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
72 const float32x4_t va0 = vld1q_f32(a0); a0 += 4;
73 const float32x4_t va1 = vld1q_f32(a1); a1 += 4;
74 const float32x4_t va2 = vld1q_f32(a2); a2 += 4;
75 const float32x4_t va3 = vld1q_f32(a3); a3 += 4;
76
77
78 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
79 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
80
81 const float32x4_t va0c0 = vdupq_lane_f32(vget_low_f32(va0), 0);
82 const float32x4_t va1c0 = vdupq_lane_f32(vget_low_f32(va1), 0);
83 const float32x4_t va2c0 = vdupq_lane_f32(vget_low_f32(va2), 0);
84 const float32x4_t va3c0 = vdupq_lane_f32(vget_low_f32(va3), 0);
85 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c0, vb0123c0);
86 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c0, vb0123c0);
87 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0);
88 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c0, vb0123c0);
89 vacc0x4567 = vmlaq_f32(vacc0x4567, va0c0, vb4567c0);
90 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c0, vb4567c0);
91 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0);
92 vacc3x4567 = vmlaq_f32(vacc3x4567, va3c0, vb4567c0);
93
94 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
95 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
96
97 const float32x4_t va0c1 = vdupq_lane_f32(vget_low_f32(va0), 1);
98 const float32x4_t va1c1 = vdupq_lane_f32(vget_low_f32(va1), 1);
99 const float32x4_t va2c1 = vdupq_lane_f32(vget_low_f32(va2), 1);
100 const float32x4_t va3c1 = vdupq_lane_f32(vget_low_f32(va3), 1);
101 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c1, vb0123c1);
102 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1);
103 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1);
104 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c1, vb0123c1);
105 vacc0x4567 = vmlaq_f32(vacc0x4567, va0c1, vb4567c1);
106 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c1, vb4567c1);
107 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c1, vb4567c1);
108 vacc3x4567 = vmlaq_f32(vacc3x4567, va3c1, vb4567c1);
109
110 const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
111 const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
112
113 const float32x4_t va0c2 = vdupq_lane_f32(vget_high_f32(va0), 0);
114 const float32x4_t va1c2 = vdupq_lane_f32(vget_high_f32(va1), 0);
115 const float32x4_t va2c2 = vdupq_lane_f32(vget_high_f32(va2), 0);
116 const float32x4_t va3c2 = vdupq_lane_f32(vget_high_f32(va3), 0);
117 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c2, vb0123c2);
118 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c2, vb0123c2);
119 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c2, vb0123c2);
120 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c2, vb0123c2);
121 vacc0x4567 = vmlaq_f32(vacc0x4567, va0c2, vb4567c2);
122 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c2, vb4567c2);
123 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c2, vb4567c2);
124 vacc3x4567 = vmlaq_f32(vacc3x4567, va3c2, vb4567c2);
125
126 const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
127 const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
128
129 const float32x4_t va0c3 = vdupq_lane_f32(vget_high_f32(va0), 1);
130 const float32x4_t va1c3 = vdupq_lane_f32(vget_high_f32(va1), 1);
131 const float32x4_t va2c3 = vdupq_lane_f32(vget_high_f32(va2), 1);
132 const float32x4_t va3c3 = vdupq_lane_f32(vget_high_f32(va3), 1);
133 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c3, vb0123c3);
134 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c3, vb0123c3);
135 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c3, vb0123c3);
136 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c3, vb0123c3);
137 vacc0x4567 = vmlaq_f32(vacc0x4567, va0c3, vb4567c3);
138 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c3, vb4567c3);
139 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c3, vb4567c3);
140 vacc3x4567 = vmlaq_f32(vacc3x4567, va3c3, vb4567c3);
141 }
142 if XNN_UNLIKELY(k != 0) {
143 do {
144 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
145 const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
146 const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
147 const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
148
149 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
150 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
151
152 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);
153 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123);
154 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123);
155 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123);
156 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);
157 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567);
158 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567);
159 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567);
160
161 k -= sizeof(float);
162 } while (k != 0);
163 }
164 const float32x4_t vmax = vld1q_dup_f32(¶ms->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 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
170 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
171 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
172 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
173
174 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
175 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
176 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
177 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
178 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
179 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
180 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
181 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
182 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
183
184 if XNN_LIKELY(nc >= 8) {
185 vst1q_f32(c3, vacc3x0123);
186 vst1q_f32(c3 + 4, vacc3x4567);
187 c3 = (float*) ((uintptr_t) c3 + cn_stride);
188 vst1q_f32(c2, vacc2x0123);
189 vst1q_f32(c2 + 4, vacc2x4567);
190 c2 = (float*) ((uintptr_t) c2 + cn_stride);
191 vst1q_f32(c1, vacc1x0123);
192 vst1q_f32(c1 + 4, vacc1x4567);
193 c1 = (float*) ((uintptr_t) c1 + cn_stride);
194 vst1q_f32(c0, vacc0x0123);
195 vst1q_f32(c0 + 4, vacc0x4567);
196 c0 = (float*) ((uintptr_t) c0 + cn_stride);
197
198 a3 = (const float*) ((uintptr_t) a3 - kc);
199 a2 = (const float*) ((uintptr_t) a2 - kc);
200 a1 = (const float*) ((uintptr_t) a1 - kc);
201 a0 = (const float*) ((uintptr_t) a0 - kc);
202
203 nc -= 8;
204
205 } else {
206 if (nc & 4) {
207 vst1q_f32(c3, vacc3x0123); c3 += 4;
208 vst1q_f32(c2, vacc2x0123); c2 += 4;
209 vst1q_f32(c1, vacc1x0123); c1 += 4;
210 vst1q_f32(c0, vacc0x0123); c0 += 4;
211
212 vacc3x0123 = vacc3x4567;
213 vacc2x0123 = vacc2x4567;
214 vacc1x0123 = vacc1x4567;
215 vacc0x0123 = vacc0x4567;
216 }
217 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
218 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
219 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
220 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
221 if (nc & 2) {
222 vst1_f32(c3, vacc3x01); c3 += 2;
223 vst1_f32(c2, vacc2x01); c2 += 2;
224 vst1_f32(c1, vacc1x01); c1 += 2;
225 vst1_f32(c0, vacc0x01); c0 += 2;
226
227 vacc3x01 = vget_high_f32(vacc3x0123);
228 vacc2x01 = vget_high_f32(vacc2x0123);
229 vacc1x01 = vget_high_f32(vacc1x0123);
230 vacc0x01 = vget_high_f32(vacc0x0123);
231 }
232 if (nc & 1) {
233 vst1_lane_f32(c3, vacc3x01, 0);
234 vst1_lane_f32(c2, vacc2x01, 0);
235 vst1_lane_f32(c1, vacc1x01, 0);
236 vst1_lane_f32(c0, vacc0x01, 0);
237 }
238
239 nc = 0;
240 }
241 } while (nc != 0);
242 }
243