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
|
/*
* Copyright (C) 2025-2026 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "shared/source/command_stream/host_function.h"
#include "shared/source/command_stream/command_stream_receiver.h"
#include "shared/source/command_stream/host_function_interface.h"
#include "shared/source/command_stream/host_function_scheduler.h"
#include "shared/source/command_stream/host_function_worker_counting_semaphore.h"
#include "shared/source/debug_settings/debug_settings_manager.h"
#include "shared/source/execution_environment/root_device_environment.h"
#include "shared/source/memory_manager/graphics_allocation.h"
namespace NEO {
HostFunctionStreamer::HostFunctionStreamer(CommandStreamReceiver *csr,
GraphicsAllocation *allocation,
void *hostFunctionIdAddress,
const std::function<void(GraphicsAllocation &)> &downloadAllocationImpl,
uint32_t activePartitions,
uint32_t partitionOffset,
bool isTbx,
bool dcFlushRequired)
: hostFunctionIdAddress(reinterpret_cast<uint64_t *>(hostFunctionIdAddress)),
csr(csr),
allocation(allocation),
downloadAllocationImpl(downloadAllocationImpl),
nextHostFunctionId(1), // start from 1 to keep 0 bit for pending/completed status
activePartitions(activePartitions),
partitionOffset(partitionOffset),
isTbx(isTbx),
dcFlushRequired(dcFlushRequired) {
}
uint64_t HostFunctionStreamer::getHostFunctionIdGpuAddress(uint32_t partitionId) const {
auto offset = partitionId * partitionOffset;
return reinterpret_cast<uint64_t>(ptrOffset(hostFunctionIdAddress, offset));
}
uint64_t *HostFunctionStreamer::getHostFunctionIdPtr(uint32_t partitionId) const {
return ptrOffset(hostFunctionIdAddress, partitionId * partitionOffset);
}
uint64_t HostFunctionStreamer::getNextHostFunctionIdAndIncrement() {
// increment by 2 to keep 0 bit for pending/completed status
return nextHostFunctionId.fetch_add(2, std::memory_order_acq_rel);
}
uint64_t HostFunctionStreamer::getHostFunctionId(uint32_t partitionId) const {
auto ptr = ptrOffset(this->hostFunctionIdAddress, partitionId * partitionOffset);
return std::atomic_ref<uint64_t>(*ptr).load(std::memory_order_acquire);
}
void HostFunctionStreamer::signalHostFunctionCompletion(const HostFunction &hostFunction) {
setHostFunctionIdAsCompleted();
endInOrderExecution();
}
void HostFunctionStreamer::prepareForExecution(const HostFunction &hostFunction) {
startInOrderExecution();
pendingHostFunctions.fetch_sub(1, std::memory_order_acq_rel);
}
uint32_t HostFunctionStreamer::getActivePartitions() const {
return activePartitions;
}
bool HostFunctionStreamer::getDcFlushRequired() const { return dcFlushRequired; }
void HostFunctionStreamer::updateTbxData() {
constexpr uint32_t allBanks = std::numeric_limits<uint32_t>::max();
allocation->setTbxWritable(true, allBanks);
for (auto partitionId = 0u; partitionId < activePartitions; partitionId++) {
auto offset = ptrDiff(getHostFunctionIdGpuAddress(partitionId), allocation->getGpuAddress());
csr->writeMemory(*allocation, true, offset, sizeof(uint64_t));
}
allocation->setTbxWritable(false, allBanks);
}
void HostFunctionStreamer::setHostFunctionIdAsCompleted() {
auto setAsCompleted = [this]() {
for (auto partitionId = 0u; partitionId < activePartitions; partitionId++) {
auto ptr = getHostFunctionIdPtr(partitionId);
std::atomic_ref<uint64_t>(*ptr).store(HostFunctionStatus::completed, std::memory_order_release);
}
};
if (isTbx) {
auto lock = csr->obtainTagAllocationDownloadLock();
setAsCompleted();
updateTbxData();
} else {
setAsCompleted();
}
}
void HostFunctionStreamer::endInOrderExecution() {
inOrderExecutionInProgress.store(false, std::memory_order_release);
}
void HostFunctionStreamer::startInOrderExecution() {
inOrderExecutionInProgress.store(true, std::memory_order_release);
}
bool HostFunctionStreamer::isInOrderExecutionInProgress() const {
return inOrderExecutionInProgress.load(std::memory_order_acquire);
}
HostFunction HostFunctionStreamer::getHostFunction(uint64_t hostFunctionId) {
std::unique_lock lock(hostFunctionsMutex);
auto node = hostFunctions.extract(hostFunctionId);
if (!node) {
UNRECOVERABLE_IF(true);
return HostFunction{};
}
return std::move(node.mapped());
}
void HostFunctionStreamer::addHostFunction(uint64_t hostFunctionId, HostFunction &&hostFunction) {
{
std::unique_lock lock(hostFunctionsMutex);
hostFunctions.emplace(hostFunctionId, std::move(hostFunction));
}
pendingHostFunctions.fetch_add(1, std::memory_order_acq_rel);
}
GraphicsAllocation *HostFunctionStreamer::getHostFunctionIdAllocation() const {
return allocation;
}
void HostFunctionStreamer::downloadHostFunctionAllocation() const {
if (isTbx) {
downloadAllocationImpl(*allocation);
}
}
uint64_t HostFunctionStreamer::getHostFunctionReadyToExecute() const {
if (pendingHostFunctions.load(std::memory_order_acquire) == 0) {
return HostFunctionStatus::completed;
}
if (isInOrderExecutionInProgress()) {
return HostFunctionStatus::completed;
}
downloadHostFunctionAllocation();
uint64_t hostFunctionId = HostFunctionStatus::completed;
for (auto partitionId = 0u; partitionId < activePartitions; partitionId++) {
hostFunctionId = getHostFunctionId(partitionId);
bool hostFunctionNotReady = hostFunctionId == HostFunctionStatus::completed;
if (hostFunctionNotReady) {
return HostFunctionStatus::completed;
}
}
return hostFunctionId;
}
namespace HostFunctionFactory {
void createAndSetHostFunctionWorker(HostFunctionWorkerMode hostFunctionWorkerMode,
bool skipHostFunctionExecution,
CommandStreamReceiver *csr,
RootDeviceEnvironment *rootDeviceEnvironment) {
if (csr->getHostFunctionWorker() != nullptr) {
return;
}
switch (hostFunctionWorkerMode) {
default:
case HostFunctionWorkerMode::defaultMode:
case HostFunctionWorkerMode::countingSemaphore:
csr->setHostFunctionWorker(new HostFunctionWorkerCountingSemaphore(skipHostFunctionExecution));
break;
case HostFunctionWorkerMode::schedulerWithThreadPool: {
auto scheduler = rootDeviceEnvironment->getHostFunctionScheduler();
if (scheduler == nullptr) {
int32_t nWorkers = (debugManager.flags.HostFunctionThreadPoolSize.get() > 0)
? debugManager.flags.HostFunctionThreadPoolSize.get()
: HostFunctionThreadPoolHelper::unlimitedThreads;
auto createdScheduler = std::make_unique<HostFunctionScheduler>(skipHostFunctionExecution,
nWorkers);
rootDeviceEnvironment->setHostFunctionScheduler(std::move(createdScheduler));
}
scheduler = rootDeviceEnvironment->getHostFunctionScheduler();
csr->setHostFunctionWorker(scheduler);
break;
}
}
}
} // namespace HostFunctionFactory
} // namespace NEO
|
