1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/MRx2-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 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/igemm.h>
16
17
xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64(
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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 6);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(float) == 0);
37 assert(ks != 0);
38 assert(ks % (6 * 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 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 4) {
59 c4 = c3;
60 }
61 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
62 if XNN_UNPREDICTABLE(mr != 6) {
63 c5 = c4;
64 }
65
66 do {
67 float32x2_t vacc0x01 = vld1_f32(w); w += 2;
68 float32x2_t vacc1x01 = vacc0x01;
69 float32x2_t vacc2x01 = vacc0x01;
70 float32x2_t vacc3x01 = vacc0x01;
71 float32x2_t vacc4x01 = vacc0x01;
72 float32x2_t vacc5x01 = vacc0x01;
73
74 size_t p = ks;
75 do {
76 const float* restrict a0 = a[0];
77 assert(a0 != NULL);
78 if XNN_UNPREDICTABLE(a0 != zero) {
79 a0 = (const float*) ((uintptr_t) a0 + a_offset);
80 }
81 const float* restrict a1 = a[1];
82 assert(a1 != NULL);
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const float*) ((uintptr_t) a1 + a_offset);
85 }
86 const float* restrict a2 = a[2];
87 assert(a2 != NULL);
88 if XNN_UNPREDICTABLE(a2 != zero) {
89 a2 = (const float*) ((uintptr_t) a2 + a_offset);
90 }
91 const float* restrict a3 = a[3];
92 assert(a3 != NULL);
93 if XNN_UNPREDICTABLE(a3 != zero) {
94 a3 = (const float*) ((uintptr_t) a3 + a_offset);
95 }
96 const float* restrict a4 = a[4];
97 assert(a4 != NULL);
98 if XNN_UNPREDICTABLE(a4 != zero) {
99 a4 = (const float*) ((uintptr_t) a4 + a_offset);
100 }
101 const float* restrict a5 = a[5];
102 assert(a5 != NULL);
103 if XNN_UNPREDICTABLE(a5 != zero) {
104 a5 = (const float*) ((uintptr_t) a5 + a_offset);
105 }
106 a += 6;
107
108 size_t k = kc;
109 for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
110 const float32x2_t va0 = vld1_f32(a0); a0 += 2;
111 const float32x2_t va1 = vld1_f32(a1); a1 += 2;
112 const float32x2_t va2 = vld1_f32(a2); a2 += 2;
113 const float32x2_t va3 = vld1_f32(a3); a3 += 2;
114 const float32x2_t va4 = vld1_f32(a4); a4 += 2;
115 const float32x2_t va5 = vld1_f32(a5); a5 += 2;
116
117 const float32x2_t vb01c0 = vld1_f32(w); w += 2;
118
119 #if XNN_ARCH_ARM64
120 vacc0x01 = vfma_lane_f32(vacc0x01, vb01c0, va0, 0);
121 vacc1x01 = vfma_lane_f32(vacc1x01, vb01c0, va1, 0);
122 vacc2x01 = vfma_lane_f32(vacc2x01, vb01c0, va2, 0);
123 vacc3x01 = vfma_lane_f32(vacc3x01, vb01c0, va3, 0);
124 vacc4x01 = vfma_lane_f32(vacc4x01, vb01c0, va4, 0);
125 vacc5x01 = vfma_lane_f32(vacc5x01, vb01c0, va5, 0);
126 #else
127 const float32x2_t va0c0 = vdup_lane_f32(va0, 0);
128 const float32x2_t va1c0 = vdup_lane_f32(va1, 0);
129 const float32x2_t va2c0 = vdup_lane_f32(va2, 0);
130 const float32x2_t va3c0 = vdup_lane_f32(va3, 0);
131 const float32x2_t va4c0 = vdup_lane_f32(va4, 0);
132 const float32x2_t va5c0 = vdup_lane_f32(va5, 0);
133 vacc0x01 = vfma_f32(vacc0x01, va0c0, vb01c0);
134 vacc1x01 = vfma_f32(vacc1x01, va1c0, vb01c0);
135 vacc2x01 = vfma_f32(vacc2x01, va2c0, vb01c0);
136 vacc3x01 = vfma_f32(vacc3x01, va3c0, vb01c0);
137 vacc4x01 = vfma_f32(vacc4x01, va4c0, vb01c0);
138 vacc5x01 = vfma_f32(vacc5x01, va5c0, vb01c0);
139 #endif
140 const float32x2_t vb01c1 = vld1_f32(w); w += 2;
141
142 #if XNN_ARCH_ARM64
143 vacc0x01 = vfma_lane_f32(vacc0x01, vb01c1, va0, 1);
144 vacc1x01 = vfma_lane_f32(vacc1x01, vb01c1, va1, 1);
145 vacc2x01 = vfma_lane_f32(vacc2x01, vb01c1, va2, 1);
146 vacc3x01 = vfma_lane_f32(vacc3x01, vb01c1, va3, 1);
147 vacc4x01 = vfma_lane_f32(vacc4x01, vb01c1, va4, 1);
148 vacc5x01 = vfma_lane_f32(vacc5x01, vb01c1, va5, 1);
149 #else
150 const float32x2_t va0c1 = vdup_lane_f32(va0, 1);
151 const float32x2_t va1c1 = vdup_lane_f32(va1, 1);
152 const float32x2_t va2c1 = vdup_lane_f32(va2, 1);
153 const float32x2_t va3c1 = vdup_lane_f32(va3, 1);
154 const float32x2_t va4c1 = vdup_lane_f32(va4, 1);
155 const float32x2_t va5c1 = vdup_lane_f32(va5, 1);
156 vacc0x01 = vfma_f32(vacc0x01, va0c1, vb01c1);
157 vacc1x01 = vfma_f32(vacc1x01, va1c1, vb01c1);
158 vacc2x01 = vfma_f32(vacc2x01, va2c1, vb01c1);
159 vacc3x01 = vfma_f32(vacc3x01, va3c1, vb01c1);
160 vacc4x01 = vfma_f32(vacc4x01, va4c1, vb01c1);
161 vacc5x01 = vfma_f32(vacc5x01, va5c1, vb01c1);
162 #endif
163 }
164 if XNN_UNLIKELY(k != 0) {
165 const float32x2_t va0 = vld1_dup_f32(a0);
166 const float32x2_t va1 = vld1_dup_f32(a1);
167 const float32x2_t va2 = vld1_dup_f32(a2);
168 const float32x2_t va3 = vld1_dup_f32(a3);
169 const float32x2_t va4 = vld1_dup_f32(a4);
170 const float32x2_t va5 = vld1_dup_f32(a5);
171
172 const float32x2_t vb01 = vld1_f32(w); w += 2;
173
174 vacc0x01 = vfma_f32(vacc0x01, va0, vb01);
175 vacc1x01 = vfma_f32(vacc1x01, va1, vb01);
176 vacc2x01 = vfma_f32(vacc2x01, va2, vb01);
177 vacc3x01 = vfma_f32(vacc3x01, va3, vb01);
178 vacc4x01 = vfma_f32(vacc4x01, va4, vb01);
179 vacc5x01 = vfma_f32(vacc5x01, va5, vb01);
180 }
181 p -= 6 * sizeof(void*);
182 } while (p != 0);
183
184 const float32x2_t vmax = vld1_dup_f32(¶ms->scalar.max);
185 vacc0x01 = vmin_f32(vacc0x01, vmax);
186 vacc1x01 = vmin_f32(vacc1x01, vmax);
187 vacc2x01 = vmin_f32(vacc2x01, vmax);
188 vacc3x01 = vmin_f32(vacc3x01, vmax);
189 vacc4x01 = vmin_f32(vacc4x01, vmax);
190 vacc5x01 = vmin_f32(vacc5x01, vmax);
191
192 const float32x2_t vmin = vld1_dup_f32(¶ms->scalar.min);
193 vacc0x01 = vmax_f32(vacc0x01, vmin);
194 vacc1x01 = vmax_f32(vacc1x01, vmin);
195 vacc2x01 = vmax_f32(vacc2x01, vmin);
196 vacc3x01 = vmax_f32(vacc3x01, vmin);
197 vacc4x01 = vmax_f32(vacc4x01, vmin);
198 vacc5x01 = vmax_f32(vacc5x01, vmin);
199
200 if XNN_LIKELY(nc >= 2) {
201 vst1_f32(c5, vacc5x01);
202 c5 = (float*) ((uintptr_t) c5 + cn_stride);
203 vst1_f32(c4, vacc4x01);
204 c4 = (float*) ((uintptr_t) c4 + cn_stride);
205 vst1_f32(c3, vacc3x01);
206 c3 = (float*) ((uintptr_t) c3 + cn_stride);
207 vst1_f32(c2, vacc2x01);
208 c2 = (float*) ((uintptr_t) c2 + cn_stride);
209 vst1_f32(c1, vacc1x01);
210 c1 = (float*) ((uintptr_t) c1 + cn_stride);
211 vst1_f32(c0, vacc0x01);
212 c0 = (float*) ((uintptr_t) c0 + cn_stride);
213
214 a = (const float**restrict) ((uintptr_t) a - ks);
215 nc -= 2;
216 } else {
217 assert(nc == 1);
218 vst1_lane_f32(c5, vacc5x01, 0);
219 vst1_lane_f32(c4, vacc4x01, 0);
220 vst1_lane_f32(c3, vacc3x01, 0);
221 vst1_lane_f32(c2, vacc2x01, 0);
222 vst1_lane_f32(c1, vacc1x01, 0);
223 vst1_lane_f32(c0, vacc0x01, 0);
224
225 nc = 0;
226 }
227 } while (nc != 0);
228 }
229