1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
16
17
xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_output_params params[restrict static1])18 void xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const float**restrict a,
24 const float*restrict w,
25 float*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const float* zero,
30 const union xnn_f32_output_params params[restrict static 1])
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(float) == 0);
37 assert(ks != 0);
38 assert(ks % (4 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(float) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 float* c0 = c;
45 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 do {
59 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
60 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
61 float32x4_t vacc1x0123 = vacc0x0123;
62 float32x4_t vacc1x4567 = vacc0x4567;
63 float32x4_t vacc2x0123 = vacc0x0123;
64 float32x4_t vacc2x4567 = vacc0x4567;
65 float32x4_t vacc3x0123 = vacc0x0123;
66 float32x4_t vacc3x4567 = vacc0x4567;
67
68 size_t p = ks;
69 do {
70 const float* restrict a0 = a[0];
71 assert(a0 != NULL);
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const float*) ((uintptr_t) a0 + a_offset);
74 }
75 const float* restrict a1 = a[1];
76 assert(a1 != NULL);
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const float*) ((uintptr_t) a1 + a_offset);
79 }
80 const float* restrict a2 = a[2];
81 assert(a2 != NULL);
82 if XNN_UNPREDICTABLE(a2 != zero) {
83 a2 = (const float*) ((uintptr_t) a2 + a_offset);
84 }
85 const float* restrict a3 = a[3];
86 assert(a3 != NULL);
87 if XNN_UNPREDICTABLE(a3 != zero) {
88 a3 = (const float*) ((uintptr_t) a3 + a_offset);
89 }
90 a += 4;
91
92 size_t k = kc;
93 for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
94 const float32x4_t va0 = vld1q_f32(a0); a0 += 4;
95 const float32x4_t va1 = vld1q_f32(a1); a1 += 4;
96 const float32x4_t va2 = vld1q_f32(a2); a2 += 4;
97 const float32x4_t va3 = vld1q_f32(a3); a3 += 4;
98
99
100 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
101 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
102
103 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, vget_low_f32(va0), 0);
104 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, vget_low_f32(va1), 0);
105 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, vget_low_f32(va2), 0);
106 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, vget_low_f32(va3), 0);
107 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, vget_low_f32(va0), 0);
108 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, vget_low_f32(va1), 0);
109 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, vget_low_f32(va2), 0);
110 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, vget_low_f32(va3), 0);
111
112 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
113 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
114
115 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, vget_low_f32(va0), 1);
116 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, vget_low_f32(va1), 1);
117 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, vget_low_f32(va2), 1);
118 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, vget_low_f32(va3), 1);
119 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c1, vget_low_f32(va0), 1);
120 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c1, vget_low_f32(va1), 1);
121 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c1, vget_low_f32(va2), 1);
122 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c1, vget_low_f32(va3), 1);
123
124 const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
125 const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
126
127 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c2, vget_high_f32(va0), 0);
128 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c2, vget_high_f32(va1), 0);
129 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c2, vget_high_f32(va2), 0);
130 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c2, vget_high_f32(va3), 0);
131 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c2, vget_high_f32(va0), 0);
132 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c2, vget_high_f32(va1), 0);
133 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c2, vget_high_f32(va2), 0);
134 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c2, vget_high_f32(va3), 0);
135
136 const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
137 const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
138
139 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c3, vget_high_f32(va0), 1);
140 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c3, vget_high_f32(va1), 1);
141 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c3, vget_high_f32(va2), 1);
142 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c3, vget_high_f32(va3), 1);
143 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c3, vget_high_f32(va0), 1);
144 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c3, vget_high_f32(va1), 1);
145 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c3, vget_high_f32(va2), 1);
146 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c3, vget_high_f32(va3), 1);
147 }
148 if XNN_UNLIKELY(k != 0) {
149 do {
150 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
151 const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
152 const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
153 const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
154
155 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
156 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
157
158 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123);
159 vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123);
160 vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123);
161 vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123);
162 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567);
163 vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567);
164 vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567);
165 vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567);
166
167 k -= sizeof(float);
168 } while (k != 0);
169 }
170
171 p -= 4 * sizeof(void*);
172 } while (p != 0);
173
174 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
175 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
176 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
177 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
178 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
179 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
180 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
181 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
182 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
183
184 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
185 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
186 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
187 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
188 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
189 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
190 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
191 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
192 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
193
194 if XNN_LIKELY(nc >= 8) {
195 vst1q_f32(c3, vacc3x0123);
196 vst1q_f32(c3 + 4, vacc3x4567);
197 c3 = (float*) ((uintptr_t) c3 + cn_stride);
198 vst1q_f32(c2, vacc2x0123);
199 vst1q_f32(c2 + 4, vacc2x4567);
200 c2 = (float*) ((uintptr_t) c2 + cn_stride);
201 vst1q_f32(c1, vacc1x0123);
202 vst1q_f32(c1 + 4, vacc1x4567);
203 c1 = (float*) ((uintptr_t) c1 + cn_stride);
204 vst1q_f32(c0, vacc0x0123);
205 vst1q_f32(c0 + 4, vacc0x4567);
206 c0 = (float*) ((uintptr_t) c0 + cn_stride);
207
208 a = (const float**restrict) ((uintptr_t) a - ks);
209 nc -= 8;
210 } else {
211 if (nc & 4) {
212 vst1q_f32(c3, vacc3x0123); c3 += 4;
213 vst1q_f32(c2, vacc2x0123); c2 += 4;
214 vst1q_f32(c1, vacc1x0123); c1 += 4;
215 vst1q_f32(c0, vacc0x0123); c0 += 4;
216
217 vacc3x0123 = vacc3x4567;
218 vacc2x0123 = vacc2x4567;
219 vacc1x0123 = vacc1x4567;
220 vacc0x0123 = vacc0x4567;
221 }
222 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
223 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
224 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
225 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
226 if (nc & 2) {
227 vst1_f32(c3, vacc3x01); c3 += 2;
228 vst1_f32(c2, vacc2x01); c2 += 2;
229 vst1_f32(c1, vacc1x01); c1 += 2;
230 vst1_f32(c0, vacc0x01); c0 += 2;
231
232 vacc3x01 = vget_high_f32(vacc3x0123);
233 vacc2x01 = vget_high_f32(vacc2x0123);
234 vacc1x01 = vget_high_f32(vacc1x0123);
235 vacc0x01 = vget_high_f32(vacc0x0123);
236 }
237 if (nc & 1) {
238 vst1_lane_f32(c3, vacc3x01, 0);
239 vst1_lane_f32(c2, vacc2x01, 0);
240 vst1_lane_f32(c1, vacc1x01, 0);
241 vst1_lane_f32(c0, vacc0x01, 0);
242 }
243
244 nc = 0;
245 }
246 } while (nc != 0);
247 }
248