1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
|
/* SPDX-License-Identifier: Apache-2.0
* Copyright 2014-2022 Blender Foundation */
#ifndef __UTIL_ATOMIC_H__
#define __UTIL_ATOMIC_H__
#ifndef __KERNEL_GPU__
/* Using atomic ops header from Blender. */
# include "atomic_ops.h"
# define atomic_add_and_fetch_float(p, x) atomic_add_and_fetch_fl((p), (x))
# define atomic_compare_and_swap_float(p, old_val, new_val) \
atomic_cas_float((p), (old_val), (new_val))
# define atomic_fetch_and_inc_uint32(p) atomic_fetch_and_add_uint32((p), 1)
# define atomic_fetch_and_dec_uint32(p) atomic_fetch_and_add_uint32((p), -1)
# define CCL_LOCAL_MEM_FENCE 0
# define ccl_barrier(flags) ((void)0)
#else /* __KERNEL_GPU__ */
# if defined(__KERNEL_CUDA__) || defined(__KERNEL_HIP__)
# define atomic_add_and_fetch_float(p, x) (atomicAdd((float *)(p), (float)(x)) + (float)(x))
# define atomic_fetch_and_add_uint32(p, x) atomicAdd((unsigned int *)(p), (unsigned int)(x))
# define atomic_fetch_and_sub_uint32(p, x) atomicSub((unsigned int *)(p), (unsigned int)(x))
# define atomic_fetch_and_inc_uint32(p) atomic_fetch_and_add_uint32((p), 1)
# define atomic_fetch_and_dec_uint32(p) atomic_fetch_and_sub_uint32((p), 1)
# define atomic_fetch_and_or_uint32(p, x) atomicOr((unsigned int *)(p), (unsigned int)(x))
ccl_device_inline float atomic_compare_and_swap_float(volatile float *dest,
const float old_val,
const float new_val)
{
union {
unsigned int int_value;
float float_value;
} new_value, prev_value, result;
prev_value.float_value = old_val;
new_value.float_value = new_val;
result.int_value = atomicCAS((unsigned int *)dest, prev_value.int_value, new_value.int_value);
return result.float_value;
}
# define CCL_LOCAL_MEM_FENCE
# define ccl_barrier(flags) __syncthreads()
# endif /* __KERNEL_CUDA__ */
# ifdef __KERNEL_METAL__
// global address space versions
ccl_device_inline float atomic_add_and_fetch_float(volatile ccl_global float *_source,
const float operand)
{
volatile ccl_global atomic_int *source = (ccl_global atomic_int *)_source;
union {
int int_value;
float float_value;
} new_value, prev_value;
prev_value.int_value = atomic_load_explicit(source, memory_order_relaxed);
do {
new_value.float_value = prev_value.float_value + operand;
} while (!atomic_compare_exchange_weak_explicit(source,
&prev_value.int_value,
new_value.int_value,
memory_order_relaxed,
memory_order_relaxed));
return new_value.float_value;
}
# define atomic_fetch_and_add_uint32(p, x) \
atomic_fetch_add_explicit((device atomic_uint *)p, x, memory_order_relaxed)
# define atomic_fetch_and_sub_uint32(p, x) \
atomic_fetch_sub_explicit((device atomic_uint *)p, x, memory_order_relaxed)
# define atomic_fetch_and_inc_uint32(p) \
atomic_fetch_add_explicit((device atomic_uint *)p, 1, memory_order_relaxed)
# define atomic_fetch_and_dec_uint32(p) \
atomic_fetch_sub_explicit((device atomic_uint *)p, 1, memory_order_relaxed)
# define atomic_fetch_and_or_uint32(p, x) \
atomic_fetch_or_explicit((device atomic_uint *)p, x, memory_order_relaxed)
ccl_device_inline float atomic_compare_and_swap_float(volatile ccl_global float *dest,
const float old_val,
const float new_val)
{
int prev_value;
prev_value = __float_as_int(old_val);
atomic_compare_exchange_weak_explicit((ccl_global atomic_int *)dest,
&prev_value,
__float_as_int(new_val),
memory_order_relaxed,
memory_order_relaxed);
return __int_as_float(prev_value);
}
# define atomic_store(p, x) atomic_store_explicit(p, x, memory_order_relaxed)
# define atomic_fetch(p) atomic_load_explicit(p, memory_order_relaxed)
# define CCL_LOCAL_MEM_FENCE mem_flags::mem_threadgroup
# define ccl_barrier(flags) threadgroup_barrier(flags)
# endif /* __KERNEL_METAL__ */
# ifdef __KERNEL_ONEAPI__
ccl_device_inline float atomic_add_and_fetch_float(ccl_global float *p, float x)
{
sycl::atomic_ref<float,
sycl::memory_order::relaxed,
sycl::memory_scope::device,
sycl::access::address_space::ext_intel_global_device_space>
atomic(*p);
return atomic.fetch_add(x);
}
ccl_device_inline float atomic_compare_and_swap_float(ccl_global float *source,
float old_val,
float new_val)
{
sycl::atomic_ref<float,
sycl::memory_order::relaxed,
sycl::memory_scope::device,
sycl::access::address_space::ext_intel_global_device_space>
atomic(*source);
atomic.compare_exchange_weak(old_val, new_val);
return old_val;
}
ccl_device_inline unsigned int atomic_fetch_and_add_uint32(ccl_global unsigned int *p,
unsigned int x)
{
sycl::atomic_ref<unsigned int,
sycl::memory_order::relaxed,
sycl::memory_scope::device,
sycl::access::address_space::ext_intel_global_device_space>
atomic(*p);
return atomic.fetch_add(x);
}
ccl_device_inline int atomic_fetch_and_add_uint32(ccl_global int *p, int x)
{
sycl::atomic_ref<int,
sycl::memory_order::relaxed,
sycl::memory_scope::device,
sycl::access::address_space::ext_intel_global_device_space>
atomic(*p);
return atomic.fetch_add(x);
}
ccl_device_inline unsigned int atomic_fetch_and_sub_uint32(ccl_global unsigned int *p,
unsigned int x)
{
sycl::atomic_ref<unsigned int,
sycl::memory_order::relaxed,
sycl::memory_scope::device,
sycl::access::address_space::ext_intel_global_device_space>
atomic(*p);
return atomic.fetch_sub(x);
}
ccl_device_inline int atomic_fetch_and_sub_uint32(ccl_global int *p, int x)
{
sycl::atomic_ref<int,
sycl::memory_order::relaxed,
sycl::memory_scope::device,
sycl::access::address_space::ext_intel_global_device_space>
atomic(*p);
return atomic.fetch_sub(x);
}
ccl_device_inline unsigned int atomic_fetch_and_inc_uint32(ccl_global unsigned int *p)
{
return atomic_fetch_and_add_uint32(p, 1);
}
ccl_device_inline int atomic_fetch_and_inc_uint32(ccl_global int *p)
{
return atomic_fetch_and_add_uint32(p, 1);
}
ccl_device_inline unsigned int atomic_fetch_and_dec_uint32(ccl_global unsigned int *p)
{
return atomic_fetch_and_sub_uint32(p, 1);
}
ccl_device_inline int atomic_fetch_and_dec_uint32(ccl_global int *p)
{
return atomic_fetch_and_sub_uint32(p, 1);
}
ccl_device_inline unsigned int atomic_fetch_and_or_uint32(ccl_global unsigned int *p,
unsigned int x)
{
sycl::atomic_ref<unsigned int,
sycl::memory_order::relaxed,
sycl::memory_scope::device,
sycl::access::address_space::ext_intel_global_device_space>
atomic(*p);
return atomic.fetch_or(x);
}
ccl_device_inline int atomic_fetch_and_or_uint32(ccl_global int *p, int x)
{
sycl::atomic_ref<int,
sycl::memory_order::relaxed,
sycl::memory_scope::device,
sycl::access::address_space::ext_intel_global_device_space>
atomic(*p);
return atomic.fetch_or(x);
}
# endif /* __KERNEL_ONEAPI__ */
#endif /* __KERNEL_GPU__ */
#endif /* __UTIL_ATOMIC_H__ */
|
