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_ukernel_4x8__neonfma_dup_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_output_params params[restrict static1])18 void xnn_f32_gemminc_ukernel_4x8__neonfma_dup_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_output_params params[restrict static 1])
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
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
62 do {
63 float32x4_t vacc0x0123 = vld1q_f32(acc); acc += 4;
64 float32x4_t vacc0x4567 = vld1q_f32(acc); acc += 4;
65 float32x4_t vacc1x0123 = vld1q_f32(acc); acc += 4;
66 float32x4_t vacc1x4567 = vld1q_f32(acc); acc += 4;
67 float32x4_t vacc2x0123 = vld1q_f32(acc); acc += 4;
68 float32x4_t vacc2x4567 = vld1q_f32(acc); acc += 4;
69 float32x4_t vacc3x0123 = vld1q_f32(acc); acc += 4;
70 float32x4_t vacc3x4567 = vld1q_f32(acc); acc += 4;
71
72 size_t k = kc;
73 for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
74 const float32x2_t va0 = vld1_f32(a0); a0 += 2;
75 const float32x2_t va1 = vld1_f32(a1); a1 += 2;
76 const float32x2_t va2 = vld1_f32(a2); a2 += 2;
77 const float32x2_t va3 = vld1_f32(a3); a3 += 2;
78
79 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
80 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
81
82 const float32x4_t va0c0 = vdupq_lane_f32(va0, 0);
83 const float32x4_t va1c0 = vdupq_lane_f32(va1, 0);
84 const float32x4_t va2c0 = vdupq_lane_f32(va2, 0);
85 const float32x4_t va3c0 = vdupq_lane_f32(va3, 0);
86 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c0, vb0123c0);
87 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c0, vb0123c0);
88 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0);
89 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c0, vb0123c0);
90 vacc0x4567 = vfmaq_f32(vacc0x4567, va0c0, vb4567c0);
91 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c0, vb4567c0);
92 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0);
93 vacc3x4567 = vfmaq_f32(vacc3x4567, va3c0, vb4567c0);
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(va0, 1);
98 const float32x4_t va1c1 = vdupq_lane_f32(va1, 1);
99 const float32x4_t va2c1 = vdupq_lane_f32(va2, 1);
100 const float32x4_t va3c1 = vdupq_lane_f32(va3, 1);
101 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c1, vb0123c1);
102 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1);
103 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1);
104 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c1, vb0123c1);
105 vacc0x4567 = vfmaq_f32(vacc0x4567, va0c1, vb4567c1);
106 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c1, vb4567c1);
107 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c1, vb4567c1);
108 vacc3x4567 = vfmaq_f32(vacc3x4567, va3c1, vb4567c1);
109 }
110 if XNN_UNLIKELY(k != 0) {
111 const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
112 const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
113 const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
114 const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
115
116 const float32x4_t vb0123 = vld1q_f32(w); w += 4;
117 const float32x4_t vb4567 = vld1q_f32(w); w += 4;
118
119 vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123);
120 vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123);
121 vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123);
122 vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123);
123 vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567);
124 vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567);
125 vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567);
126 vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567);
127 }
128 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
129 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
130 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
131 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
132 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
133 vacc0x4567 = vminq_f32(vacc0x4567, vmax);
134 vacc1x4567 = vminq_f32(vacc1x4567, vmax);
135 vacc2x4567 = vminq_f32(vacc2x4567, vmax);
136 vacc3x4567 = vminq_f32(vacc3x4567, vmax);
137
138 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
139 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
140 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
141 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
142 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
143 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
144 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
145 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
146 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
147
148 if XNN_LIKELY(nc >= 8) {
149 vst1q_f32(c3, vacc3x0123);
150 vst1q_f32(c3 + 4, vacc3x4567);
151 c3 = (float*) ((uintptr_t) c3 + cn_stride);
152 vst1q_f32(c2, vacc2x0123);
153 vst1q_f32(c2 + 4, vacc2x4567);
154 c2 = (float*) ((uintptr_t) c2 + cn_stride);
155 vst1q_f32(c1, vacc1x0123);
156 vst1q_f32(c1 + 4, vacc1x4567);
157 c1 = (float*) ((uintptr_t) c1 + cn_stride);
158 vst1q_f32(c0, vacc0x0123);
159 vst1q_f32(c0 + 4, vacc0x4567);
160 c0 = (float*) ((uintptr_t) c0 + cn_stride);
161
162 a3 = (const float*) ((uintptr_t) a3 - kc);
163 a2 = (const float*) ((uintptr_t) a2 - kc);
164 a1 = (const float*) ((uintptr_t) a1 - kc);
165 a0 = (const float*) ((uintptr_t) a0 - kc);
166
167 nc -= 8;
168
169 } else {
170 if (nc & 4) {
171 vst1q_f32(c3, vacc3x0123); c3 += 4;
172 vst1q_f32(c2, vacc2x0123); c2 += 4;
173 vst1q_f32(c1, vacc1x0123); c1 += 4;
174 vst1q_f32(c0, vacc0x0123); c0 += 4;
175
176 vacc3x0123 = vacc3x4567;
177 vacc2x0123 = vacc2x4567;
178 vacc1x0123 = vacc1x4567;
179 vacc0x0123 = vacc0x4567;
180 }
181 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
182 float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
183 float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
184 float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
185 if (nc & 2) {
186 vst1_f32(c3, vacc3x01); c3 += 2;
187 vst1_f32(c2, vacc2x01); c2 += 2;
188 vst1_f32(c1, vacc1x01); c1 += 2;
189 vst1_f32(c0, vacc0x01); c0 += 2;
190
191 vacc3x01 = vget_high_f32(vacc3x0123);
192 vacc2x01 = vget_high_f32(vacc2x0123);
193 vacc1x01 = vget_high_f32(vacc1x0123);
194 vacc0x01 = vget_high_f32(vacc0x0123);
195 }
196 if (nc & 1) {
197 vst1_lane_f32(c3, vacc3x01, 0);
198 vst1_lane_f32(c2, vacc2x01, 0);
199 vst1_lane_f32(c1, vacc1x01, 0);
200 vst1_lane_f32(c0, vacc0x01, 0);
201 }
202
203 nc = 0;
204 }
205 } while (nc != 0);
206 }
207