1 // Auto-generated file. Do not edit!
2 // Template: src/f16-gemm/neonfp16arith-ld64.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/common.h>
16
17 #include <xnnpack/gemm.h>
18
19
xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64(size_t mr,size_t nc,size_t kc,const void * restrict a,size_t a_stride,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,const void * restrict acc,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f16_gemminc_minmax_ukernel_1x16__neonfp16arith_ld64(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const void* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 void* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const void*restrict acc,
31 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
32 {
33 assert(mr != 0);
34 assert(mr <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(__fp16) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41 assert(acc != NULL);
42
43 const __fp16* a0 = (const __fp16*) a;
44 __fp16* c0 = (__fp16*) c;
45
46 do {
47 float16x8_t vacc0x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
48 float16x8_t vacc0x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
49
50 size_t k = kc;
51 while (k >= 4 * sizeof(__fp16)) {
52 const float16x4_t va0 = vld1_f16(a0); a0 += 4;
53
54 const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
55 const float16x8_t vb89ABCDEFc0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
56
57 #if XNN_ARCH_ARM64
58 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
59 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);
60 #else
61 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
62
63 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
64 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c0, vb89ABCDEFc0);
65 #endif
66 const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
67 const float16x8_t vb89ABCDEFc1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
68
69 #if XNN_ARCH_ARM64
70 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
71 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);
72 #else
73 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
74
75 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
76 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c1, vb89ABCDEFc1);
77 #endif
78 const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
79 const float16x8_t vb89ABCDEFc2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
80
81 #if XNN_ARCH_ARM64
82 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
83 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);
84 #else
85 const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
86
87 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
88 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c2, vb89ABCDEFc2);
89 #endif
90 const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
91 const float16x8_t vb89ABCDEFc3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
92
93 #if XNN_ARCH_ARM64
94 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
95 vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);
96 #else
97 const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
98
99 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
100 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c3, vb89ABCDEFc3);
101 #endif
102
103 k -= 4 * sizeof(__fp16);
104 }
105 if XNN_UNLIKELY(k != 0) {
106 do {
107 const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
108
109 const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
110 const float16x8_t vb89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
111
112 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
113 vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0, vb89ABCDEF);
114
115 k -= sizeof(__fp16);
116 } while (k != 0);
117 }
118
119
120 const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
121 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
122 vacc0x89ABCDEF = vminq_f16(vacc0x89ABCDEF, vmax);
123
124 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min));
125 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
126 vacc0x89ABCDEF = vmaxq_f16(vacc0x89ABCDEF, vmin);
127
128 if XNN_LIKELY(nc >= 16) {
129 vst1q_f16(c0, vacc0x01234567);
130 vst1q_f16(c0 + 8, vacc0x89ABCDEF);
131 c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
132
133 a0 = (const __fp16*) ((uintptr_t) a0 - kc);
134
135 nc -= 16;
136 } else {
137 if (nc & 8) {
138 vst1q_f16(c0, vacc0x01234567); c0 += 8;
139
140 vacc0x01234567 = vacc0x89ABCDEF;
141 }
142 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
143 if (nc & 4) {
144 vst1_f16(c0, vacc0x0123); c0 += 4;
145
146 vacc0x0123 = vget_high_f16(vacc0x01234567);
147 }
148 if (nc & 2) {
149 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
150
151 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
152 }
153 if (nc & 1) {
154 vst1_lane_f16(c0, vacc0x0123, 0);
155 }
156
157 nc = 0;
158 }
159 } while (nc != 0);
160 }
161