1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
16
17
xnn_f32_gemm_minmax_ukernel_6x2__neon_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_6x2__neon_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 <= 6);
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 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
66 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
67 if XNN_UNPREDICTABLE(mr != 6) {
68 a5 = a4;
69 c5 = c4;
70 }
71
72 do {
73 float32x2_t vacc0x01 = vld1_f32(w); w += 2;
74 float32x2_t vacc1x01 = vacc0x01;
75 float32x2_t vacc2x01 = vacc0x01;
76 float32x2_t vacc3x01 = vacc0x01;
77 float32x2_t vacc4x01 = vacc0x01;
78 float32x2_t vacc5x01 = vacc0x01;
79
80 size_t k = kc;
81 for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
82 const float32x2_t va0 = vld1_f32(a0); a0 += 2;
83 const float32x2_t va1 = vld1_f32(a1); a1 += 2;
84 const float32x2_t va2 = vld1_f32(a2); a2 += 2;
85 const float32x2_t va3 = vld1_f32(a3); a3 += 2;
86 const float32x2_t va4 = vld1_f32(a4); a4 += 2;
87 const float32x2_t va5 = vld1_f32(a5); a5 += 2;
88
89 const float32x2_t vb01c0 = vld1_f32(w); w += 2;
90
91 vacc0x01 = vmla_lane_f32(vacc0x01, vb01c0, va0, 0);
92 vacc1x01 = vmla_lane_f32(vacc1x01, vb01c0, va1, 0);
93 vacc2x01 = vmla_lane_f32(vacc2x01, vb01c0, va2, 0);
94 vacc3x01 = vmla_lane_f32(vacc3x01, vb01c0, va3, 0);
95 vacc4x01 = vmla_lane_f32(vacc4x01, vb01c0, va4, 0);
96 vacc5x01 = vmla_lane_f32(vacc5x01, vb01c0, va5, 0);
97 const float32x2_t vb01c1 = vld1_f32(w); w += 2;
98
99 vacc0x01 = vmla_lane_f32(vacc0x01, vb01c1, va0, 1);
100 vacc1x01 = vmla_lane_f32(vacc1x01, vb01c1, va1, 1);
101 vacc2x01 = vmla_lane_f32(vacc2x01, vb01c1, va2, 1);
102 vacc3x01 = vmla_lane_f32(vacc3x01, vb01c1, va3, 1);
103 vacc4x01 = vmla_lane_f32(vacc4x01, vb01c1, va4, 1);
104 vacc5x01 = vmla_lane_f32(vacc5x01, vb01c1, va5, 1);
105 }
106 if XNN_UNLIKELY(k != 0) {
107 const float32x2_t va0 = vld1_dup_f32(a0); a0 += 1;
108 const float32x2_t va1 = vld1_dup_f32(a1); a1 += 1;
109 const float32x2_t va2 = vld1_dup_f32(a2); a2 += 1;
110 const float32x2_t va3 = vld1_dup_f32(a3); a3 += 1;
111 const float32x2_t va4 = vld1_dup_f32(a4); a4 += 1;
112 const float32x2_t va5 = vld1_dup_f32(a5); a5 += 1;
113
114 const float32x2_t vb01 = vld1_f32(w); w += 2;
115
116 vacc0x01 = vmla_f32(vacc0x01, va0, vb01);
117 vacc1x01 = vmla_f32(vacc1x01, va1, vb01);
118 vacc2x01 = vmla_f32(vacc2x01, va2, vb01);
119 vacc3x01 = vmla_f32(vacc3x01, va3, vb01);
120 vacc4x01 = vmla_f32(vacc4x01, va4, vb01);
121 vacc5x01 = vmla_f32(vacc5x01, va5, vb01);
122 }
123
124 const float32x2_t vmax = vld1_dup_f32(¶ms->scalar.max);
125 vacc0x01 = vmin_f32(vacc0x01, vmax);
126 vacc1x01 = vmin_f32(vacc1x01, vmax);
127 vacc2x01 = vmin_f32(vacc2x01, vmax);
128 vacc3x01 = vmin_f32(vacc3x01, vmax);
129 vacc4x01 = vmin_f32(vacc4x01, vmax);
130 vacc5x01 = vmin_f32(vacc5x01, vmax);
131
132 const float32x2_t vmin = vld1_dup_f32(¶ms->scalar.min);
133 vacc0x01 = vmax_f32(vacc0x01, vmin);
134 vacc1x01 = vmax_f32(vacc1x01, vmin);
135 vacc2x01 = vmax_f32(vacc2x01, vmin);
136 vacc3x01 = vmax_f32(vacc3x01, vmin);
137 vacc4x01 = vmax_f32(vacc4x01, vmin);
138 vacc5x01 = vmax_f32(vacc5x01, vmin);
139
140 if XNN_LIKELY(nc >= 2) {
141 vst1_f32(c0, vacc0x01);
142 c0 = (float*) ((uintptr_t) c0 + cn_stride);
143 vst1_f32(c1, vacc1x01);
144 c1 = (float*) ((uintptr_t) c1 + cn_stride);
145 vst1_f32(c2, vacc2x01);
146 c2 = (float*) ((uintptr_t) c2 + cn_stride);
147 vst1_f32(c3, vacc3x01);
148 c3 = (float*) ((uintptr_t) c3 + cn_stride);
149 vst1_f32(c4, vacc4x01);
150 c4 = (float*) ((uintptr_t) c4 + cn_stride);
151 vst1_f32(c5, vacc5x01);
152 c5 = (float*) ((uintptr_t) c5 + cn_stride);
153
154 a0 = (const float*) ((uintptr_t) a0 - kc);
155 a1 = (const float*) ((uintptr_t) a1 - kc);
156 a2 = (const float*) ((uintptr_t) a2 - kc);
157 a3 = (const float*) ((uintptr_t) a3 - kc);
158 a4 = (const float*) ((uintptr_t) a4 - kc);
159 a5 = (const float*) ((uintptr_t) a5 - kc);
160
161 nc -= 2;
162 } else {
163 assert(nc == 1);
164 vst1_lane_f32(c0, vacc0x01, 0);
165 vst1_lane_f32(c1, vacc1x01, 0);
166 vst1_lane_f32(c2, vacc2x01, 0);
167 vst1_lane_f32(c3, vacc3x01, 0);
168 vst1_lane_f32(c4, vacc4x01, 0);
169 vst1_lane_f32(c5, vacc5x01, 0);
170
171 nc = 0;
172 }
173 } while (nc != 0);
174 }
175