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_5x8__neonfma_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_gemm_minmax_ukernel_5x8__neonfma_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 5);
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
66 do {
67 float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
68 float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
69 float32x4_t vacc1x0123 = vacc0x0123;
70 float32x4_t vacc1x4567 = vacc0x4567;
71 float32x4_t vacc2x0123 = vacc0x0123;
72 float32x4_t vacc2x4567 = vacc0x4567;
73 float32x4_t vacc3x0123 = vacc0x0123;
74 float32x4_t vacc3x4567 = vacc0x4567;
75 float32x4_t vacc4x0123 = vacc0x0123;
76 float32x4_t vacc4x4567 = vacc0x4567;
77
78 size_t k = kc;
79 for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
80 const float32x2_t va0 = vld1_f32(a0); a0 += 2;
81 const float32x2_t va1 = vld1_f32(a1); a1 += 2;
82 const float32x2_t va2 = vld1_f32(a2); a2 += 2;
83 const float32x2_t va3 = vld1_f32(a3); a3 += 2;
84 const float32x2_t va4 = vld1_f32(a4); a4 += 2;
85
86 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
87 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
88
89 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, va0, 0);
90 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, va1, 0);
91 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, va2, 0);
92 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, va3, 0);
93 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c0, va4, 0);
94 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, va0, 0);
95 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, va1, 0);
96 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, va2, 0);
97 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, va3, 0);
98 vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c0, va4, 0);
99 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
100 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
101
102 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, va0, 1);
103 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, va1, 1);
104 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, va2, 1);
105 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, va3, 1);
106 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c1, va4, 1);
107 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c1, va0, 1);
108 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c1, va1, 1);
109 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c1, va2, 1);
110 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c1, va3, 1);
111 vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c1, va4, 1);
112 }
113 if XNN_UNLIKELY(k != 0) {
114 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
115 const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
116 const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
117 const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
118 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1;
119
120 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
121 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
122
123 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123);
124 vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123);
125 vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123);
126 vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123);
127 vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123);
128 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567);
129 vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567);
130 vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567);
131 vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567);
132 vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567);
133 }
134 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
135 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
136 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
137 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
138 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
139 vacc4x0123 = vminq_f32(vacc4x0123, vmax);
140 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
141 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
142 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
143 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
144 vacc4x4567 = vminq_f32(vacc4x4567, vmax);
145
146 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
147 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
148 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
149 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
150 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
151 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
152 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
153 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
154 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
155 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
156 vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
157
158 if XNN_LIKELY(nc >= 8) {
159 vst1q_f32(c4, vacc4x0123);
160 vst1q_f32(c4 + 4, vacc4x4567);
161 c4 = (float*) ((uintptr_t) c4 + cn_stride);
162 vst1q_f32(c3, vacc3x0123);
163 vst1q_f32(c3 + 4, vacc3x4567);
164 c3 = (float*) ((uintptr_t) c3 + cn_stride);
165 vst1q_f32(c2, vacc2x0123);
166 vst1q_f32(c2 + 4, vacc2x4567);
167 c2 = (float*) ((uintptr_t) c2 + cn_stride);
168 vst1q_f32(c1, vacc1x0123);
169 vst1q_f32(c1 + 4, vacc1x4567);
170 c1 = (float*) ((uintptr_t) c1 + cn_stride);
171 vst1q_f32(c0, vacc0x0123);
172 vst1q_f32(c0 + 4, vacc0x4567);
173 c0 = (float*) ((uintptr_t) c0 + cn_stride);
174
175 a4 = (const float*) ((uintptr_t) a4 - kc);
176 a3 = (const float*) ((uintptr_t) a3 - kc);
177 a2 = (const float*) ((uintptr_t) a2 - kc);
178 a1 = (const float*) ((uintptr_t) a1 - kc);
179 a0 = (const float*) ((uintptr_t) a0 - kc);
180
181 nc -= 8;
182
183 } else {
184 if (nc & 4) {
185 vst1q_f32(c4, vacc4x0123); c4 += 4;
186 vst1q_f32(c3, vacc3x0123); c3 += 4;
187 vst1q_f32(c2, vacc2x0123); c2 += 4;
188 vst1q_f32(c1, vacc1x0123); c1 += 4;
189 vst1q_f32(c0, vacc0x0123); c0 += 4;
190
191 vacc4x0123 = vacc4x4567;
192 vacc3x0123 = vacc3x4567;
193 vacc2x0123 = vacc2x4567;
194 vacc1x0123 = vacc1x4567;
195 vacc0x0123 = vacc0x4567;
196 }
197 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
198 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
199 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
200 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
201 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
202 if (nc & 2) {
203 vst1_f32(c4, vacc4x01); c4 += 2;
204 vst1_f32(c3, vacc3x01); c3 += 2;
205 vst1_f32(c2, vacc2x01); c2 += 2;
206 vst1_f32(c1, vacc1x01); c1 += 2;
207 vst1_f32(c0, vacc0x01); c0 += 2;
208
209 vacc4x01 = vget_high_f32(vacc4x0123);
210 vacc3x01 = vget_high_f32(vacc3x0123);
211 vacc2x01 = vget_high_f32(vacc2x0123);
212 vacc1x01 = vget_high_f32(vacc1x0123);
213 vacc0x01 = vget_high_f32(vacc0x0123);
214 }
215 if (nc & 1) {
216 vst1_lane_f32(c4, vacc4x01, 0);
217 vst1_lane_f32(c3, vacc3x01, 0);
218 vst1_lane_f32(c2, vacc2x01, 0);
219 vst1_lane_f32(c1, vacc1x01, 0);
220 vst1_lane_f32(c0, vacc0x01, 0);
221 }
222
223 nc = 0;
224 }
225 } while (nc != 0);
226 }
227