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_minmax_ukernel_4x8__neon_lane_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_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_minmax_params params[restrict XNN_MIN_ELEMENTS(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 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, vget_low_f32(va0), 0);
82 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, vget_low_f32(va1), 0);
83 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, vget_low_f32(va2), 0);
84 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, vget_low_f32(va3), 0);
85 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, vget_low_f32(va0), 0);
86 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, vget_low_f32(va1), 0);
87 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, vget_low_f32(va2), 0);
88 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, vget_low_f32(va3), 0);
89
90 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
91 const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
92
93 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, vget_low_f32(va0), 1);
94 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, vget_low_f32(va1), 1);
95 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, vget_low_f32(va2), 1);
96 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, vget_low_f32(va3), 1);
97 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c1, vget_low_f32(va0), 1);
98 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c1, vget_low_f32(va1), 1);
99 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c1, vget_low_f32(va2), 1);
100 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c1, vget_low_f32(va3), 1);
101
102 const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
103 const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
104
105 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c2, vget_high_f32(va0), 0);
106 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c2, vget_high_f32(va1), 0);
107 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c2, vget_high_f32(va2), 0);
108 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c2, vget_high_f32(va3), 0);
109 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c2, vget_high_f32(va0), 0);
110 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c2, vget_high_f32(va1), 0);
111 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c2, vget_high_f32(va2), 0);
112 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c2, vget_high_f32(va3), 0);
113
114 const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
115 const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
116
117 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c3, vget_high_f32(va0), 1);
118 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c3, vget_high_f32(va1), 1);
119 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c3, vget_high_f32(va2), 1);
120 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c3, vget_high_f32(va3), 1);
121 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c3, vget_high_f32(va0), 1);
122 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c3, vget_high_f32(va1), 1);
123 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c3, vget_high_f32(va2), 1);
124 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c3, vget_high_f32(va3), 1);
125 }
126 if XNN_UNLIKELY(k != 0) {
127 do {
128 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
129 const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
130 const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
131 const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
132
133 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
134 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
135
136 vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);
137 vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123);
138 vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123);
139 vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123);
140 vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);
141 vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567);
142 vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567);
143 vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567);
144
145 k -= sizeof(float);
146 } while (k != 0);
147 }
148 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
149 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
150 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
151 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
152 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
153 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
154 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
155 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
156 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
157
158 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
159 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
160 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
161 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
162 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
163 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
164 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
165 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
166 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
167
168 if XNN_LIKELY(nc >= 8) {
169 vst1q_f32(c3, vacc3x0123);
170 vst1q_f32(c3 + 4, vacc3x4567);
171 c3 = (float*) ((uintptr_t) c3 + cn_stride);
172 vst1q_f32(c2, vacc2x0123);
173 vst1q_f32(c2 + 4, vacc2x4567);
174 c2 = (float*) ((uintptr_t) c2 + cn_stride);
175 vst1q_f32(c1, vacc1x0123);
176 vst1q_f32(c1 + 4, vacc1x4567);
177 c1 = (float*) ((uintptr_t) c1 + cn_stride);
178 vst1q_f32(c0, vacc0x0123);
179 vst1q_f32(c0 + 4, vacc0x4567);
180 c0 = (float*) ((uintptr_t) c0 + cn_stride);
181
182 a3 = (const float*) ((uintptr_t) a3 - kc);
183 a2 = (const float*) ((uintptr_t) a2 - kc);
184 a1 = (const float*) ((uintptr_t) a1 - kc);
185 a0 = (const float*) ((uintptr_t) a0 - kc);
186
187 nc -= 8;
188
189 } else {
190 if (nc & 4) {
191 vst1q_f32(c3, vacc3x0123); c3 += 4;
192 vst1q_f32(c2, vacc2x0123); c2 += 4;
193 vst1q_f32(c1, vacc1x0123); c1 += 4;
194 vst1q_f32(c0, vacc0x0123); c0 += 4;
195
196 vacc3x0123 = vacc3x4567;
197 vacc2x0123 = vacc2x4567;
198 vacc1x0123 = vacc1x4567;
199 vacc0x0123 = vacc0x4567;
200 }
201 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
202 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
203 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
204 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
205 if (nc & 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 vacc3x01 = vget_high_f32(vacc3x0123);
212 vacc2x01 = vget_high_f32(vacc2x0123);
213 vacc1x01 = vget_high_f32(vacc1x0123);
214 vacc0x01 = vget_high_f32(vacc0x0123);
215 }
216 if (nc & 1) {
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