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
|
/* SPDX-License-Identifier: Apache-2.0
* Copyright 2011-2022 Blender Foundation */
#ifdef WITH_HIP
# include "device/hip/queue.h"
# include "device/hip/device_impl.h"
# include "device/hip/graphics_interop.h"
# include "device/hip/kernel.h"
CCL_NAMESPACE_BEGIN
/* HIPDeviceQueue */
HIPDeviceQueue::HIPDeviceQueue(HIPDevice *device)
: DeviceQueue(device), hip_device_(device), hip_stream_(nullptr)
{
const HIPContextScope scope(hip_device_);
hip_device_assert(hip_device_, hipStreamCreateWithFlags(&hip_stream_, hipStreamNonBlocking));
}
HIPDeviceQueue::~HIPDeviceQueue()
{
const HIPContextScope scope(hip_device_);
hipStreamDestroy(hip_stream_);
}
int HIPDeviceQueue::num_concurrent_states(const size_t state_size) const
{
const int max_num_threads = hip_device_->get_num_multiprocessors() *
hip_device_->get_max_num_threads_per_multiprocessor();
int num_states = ((max_num_threads == 0) ? 65536 : max_num_threads) * 16;
const char *factor_str = getenv("CYCLES_CONCURRENT_STATES_FACTOR");
if (factor_str) {
const float factor = (float)atof(factor_str);
if (factor != 0.0f) {
num_states = max((int)(num_states * factor), 1024);
}
else {
VLOG(3) << "CYCLES_CONCURRENT_STATES_FACTOR evaluated to 0";
}
}
VLOG(3) << "GPU queue concurrent states: " << num_states << ", using up to "
<< string_human_readable_size(num_states * state_size);
return num_states;
}
int HIPDeviceQueue::num_concurrent_busy_states() const
{
const int max_num_threads = hip_device_->get_num_multiprocessors() *
hip_device_->get_max_num_threads_per_multiprocessor();
if (max_num_threads == 0) {
return 65536;
}
return 4 * max_num_threads;
}
void HIPDeviceQueue::init_execution()
{
/* Synchronize all textures and memory copies before executing task. */
HIPContextScope scope(hip_device_);
hip_device_->load_texture_info();
hip_device_assert(hip_device_, hipDeviceSynchronize());
debug_init_execution();
}
bool HIPDeviceQueue::enqueue(DeviceKernel kernel,
const int work_size,
DeviceKernelArguments const &args)
{
if (hip_device_->have_error()) {
return false;
}
debug_enqueue(kernel, work_size);
const HIPContextScope scope(hip_device_);
const HIPDeviceKernel &hip_kernel = hip_device_->kernels.get(kernel);
/* Compute kernel launch parameters. */
const int num_threads_per_block = hip_kernel.num_threads_per_block;
const int num_blocks = divide_up(work_size, num_threads_per_block);
int shared_mem_bytes = 0;
switch (kernel) {
case DEVICE_KERNEL_INTEGRATOR_QUEUED_PATHS_ARRAY:
case DEVICE_KERNEL_INTEGRATOR_QUEUED_SHADOW_PATHS_ARRAY:
case DEVICE_KERNEL_INTEGRATOR_ACTIVE_PATHS_ARRAY:
case DEVICE_KERNEL_INTEGRATOR_TERMINATED_PATHS_ARRAY:
case DEVICE_KERNEL_INTEGRATOR_SORTED_PATHS_ARRAY:
case DEVICE_KERNEL_INTEGRATOR_COMPACT_PATHS_ARRAY:
case DEVICE_KERNEL_INTEGRATOR_TERMINATED_SHADOW_PATHS_ARRAY:
case DEVICE_KERNEL_INTEGRATOR_COMPACT_SHADOW_PATHS_ARRAY:
/* See parall_active_index.h for why this amount of shared memory is needed. */
shared_mem_bytes = (num_threads_per_block + 1) * sizeof(int);
break;
default:
break;
}
/* Launch kernel. */
assert_success(hipModuleLaunchKernel(hip_kernel.function,
num_blocks,
1,
1,
num_threads_per_block,
1,
1,
shared_mem_bytes,
hip_stream_,
const_cast<void **>(args.values),
0),
"enqueue");
return !(hip_device_->have_error());
}
bool HIPDeviceQueue::synchronize()
{
if (hip_device_->have_error()) {
return false;
}
const HIPContextScope scope(hip_device_);
assert_success(hipStreamSynchronize(hip_stream_), "synchronize");
debug_synchronize();
return !(hip_device_->have_error());
}
void HIPDeviceQueue::zero_to_device(device_memory &mem)
{
assert(mem.type != MEM_GLOBAL && mem.type != MEM_TEXTURE);
if (mem.memory_size() == 0) {
return;
}
/* Allocate on demand. */
if (mem.device_pointer == 0) {
hip_device_->mem_alloc(mem);
}
/* Zero memory on device. */
assert(mem.device_pointer != 0);
const HIPContextScope scope(hip_device_);
assert_success(
hipMemsetD8Async((hipDeviceptr_t)mem.device_pointer, 0, mem.memory_size(), hip_stream_),
"zero_to_device");
}
void HIPDeviceQueue::copy_to_device(device_memory &mem)
{
assert(mem.type != MEM_GLOBAL && mem.type != MEM_TEXTURE);
if (mem.memory_size() == 0) {
return;
}
/* Allocate on demand. */
if (mem.device_pointer == 0) {
hip_device_->mem_alloc(mem);
}
assert(mem.device_pointer != 0);
assert(mem.host_pointer != nullptr);
/* Copy memory to device. */
const HIPContextScope scope(hip_device_);
assert_success(
hipMemcpyHtoDAsync(
(hipDeviceptr_t)mem.device_pointer, mem.host_pointer, mem.memory_size(), hip_stream_),
"copy_to_device");
}
void HIPDeviceQueue::copy_from_device(device_memory &mem)
{
assert(mem.type != MEM_GLOBAL && mem.type != MEM_TEXTURE);
if (mem.memory_size() == 0) {
return;
}
assert(mem.device_pointer != 0);
assert(mem.host_pointer != nullptr);
/* Copy memory from device. */
const HIPContextScope scope(hip_device_);
assert_success(
hipMemcpyDtoHAsync(
mem.host_pointer, (hipDeviceptr_t)mem.device_pointer, mem.memory_size(), hip_stream_),
"copy_from_device");
}
void HIPDeviceQueue::assert_success(hipError_t result, const char *operation)
{
if (result != hipSuccess) {
const char *name = hipewErrorString(result);
hip_device_->set_error(
string_printf("%s in HIP queue %s (%s)", name, operation, debug_active_kernels().c_str()));
}
}
unique_ptr<DeviceGraphicsInterop> HIPDeviceQueue::graphics_interop_create()
{
return make_unique<HIPDeviceGraphicsInterop>(this);
}
CCL_NAMESPACE_END
#endif /* WITH_HIP */
|
