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
|
/*
* Copyright (C) 2019-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "shared/source/memory_manager/gfx_partition.h"
#include "shared/source/helpers/aligned_memory.h"
#include "shared/source/helpers/heap_assigner.h"
#include "shared/source/memory_manager/memory_manager.h"
namespace NEO {
const std::array<HeapIndex, 4> GfxPartition::heap32Names{{HeapIndex::HEAP_INTERNAL_DEVICE_MEMORY,
HeapIndex::HEAP_INTERNAL,
HeapIndex::HEAP_EXTERNAL_DEVICE_MEMORY,
HeapIndex::HEAP_EXTERNAL}};
const std::array<HeapIndex, 7> GfxPartition::heapNonSvmNames{{HeapIndex::HEAP_INTERNAL_DEVICE_MEMORY,
HeapIndex::HEAP_INTERNAL,
HeapIndex::HEAP_EXTERNAL_DEVICE_MEMORY,
HeapIndex::HEAP_EXTERNAL,
HeapIndex::HEAP_STANDARD,
HeapIndex::HEAP_STANDARD64KB,
HeapIndex::HEAP_EXTENDED}};
GfxPartition::GfxPartition(OSMemory::ReservedCpuAddressRange &sharedReservedCpuAddressRange) : reservedCpuAddressRange(sharedReservedCpuAddressRange), osMemory(OSMemory::create()) {}
GfxPartition::~GfxPartition() {
osMemory->releaseCpuAddressRange(reservedCpuAddressRange);
reservedCpuAddressRange = {0};
}
void GfxPartition::Heap::init(uint64_t base, uint64_t size) {
this->base = base;
this->size = size;
// Exclude very first and very last 64K from GPU address range allocation
if (size > 2 * GfxPartition::heapGranularity) {
size -= 2 * GfxPartition::heapGranularity;
}
alloc = std::make_unique<HeapAllocator>(base + GfxPartition::heapGranularity, size);
}
void GfxPartition::Heap::initExternalWithFrontWindow(uint64_t base, uint64_t size) {
this->base = base;
this->size = size;
size -= GfxPartition::heapGranularity;
alloc = std::make_unique<HeapAllocator>(base, size, 0u);
}
void GfxPartition::Heap::initWithFrontWindow(uint64_t base, uint64_t size, uint64_t frontWindowSize) {
this->base = base;
this->size = size;
// Exclude very very last 64K from GPU address range allocation
size -= GfxPartition::heapGranularity;
size -= frontWindowSize;
alloc = std::make_unique<HeapAllocator>(base + frontWindowSize, size);
}
void GfxPartition::Heap::initFrontWindow(uint64_t base, uint64_t size) {
this->base = base;
this->size = size;
alloc = std::make_unique<HeapAllocator>(base, size, 0u);
}
void GfxPartition::freeGpuAddressRange(uint64_t ptr, size_t size) {
for (auto heapName : GfxPartition::heapNonSvmNames) {
auto &heap = getHeap(heapName);
if ((ptr > heap.getBase()) && ((ptr + size) < heap.getLimit())) {
heap.free(ptr, size);
break;
}
}
}
bool GfxPartition::init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToReserve, uint32_t rootDeviceIndex, size_t numRootDevices, bool useFrontWindowPool) {
/*
* I. 64-bit builds:
*
* 1) 48-bit Full Range SVM gfx layout:
*
* SVM H0 H1 H2 H3 STANDARD STANDARD64K
* |__________________________________|____|____|____|____|________________|______________|
* | | | | | | | |
* | gfxBase gfxTop
* 0x0 0x0000800000000000 0x0000FFFFFFFFFFFF
*
*
* 2) 47-bit Full Range SVM gfx layout:
*
* gfxSize = 2^47 / 4 = 0x200000000000
* ________________________________________________
* / \
* SVM / H0 H1 H2 H3 STANDARD STANDARD64K \ SVM
* |________________|____|____|____|____|________________|______________|_______________|
* | | | | | | | | |
* | gfxBase gfxTop |
* 0x0 reserveCpuAddressRange(gfxSize) 0x00007FFFFFFFFFFF
* \_____________________________________ SVM _________________________________________/
*
*
*
* 3) Limited Range gfx layout (no SVM):
*
* H0 H1 H2 H3 STANDARD STANDARD64K
* |____|____|____|____|____________________|__________________|
* | | | | | | |
* gfxBase gfxTop
* 0x0 0xFFF...FFF < 47 bit
*
*
* II. 32-bit builds:
*
* 1) 32-bit Full Range SVM gfx layout:
*
* SVM H0 H1 H2 H3 STANDARD STANDARD64K
* |_______|____|____|____|____|________________|______________|
* | | | | | | | |
* | gfxBase gfxTop
* 0x0 0x100000000 gpuAddressSpace
*/
uint64_t gfxTop = gpuAddressSpace + 1;
uint64_t gfxBase = 0x0ull;
const uint64_t gfxHeap32Size = 4 * MemoryConstants::gigaByte;
if (is32bit) {
gfxBase = maxNBitValue(32) + 1;
heapInit(HeapIndex::HEAP_SVM, 0ull, gfxBase);
} else {
if (gpuAddressSpace == maxNBitValue(48)) {
gfxBase = maxNBitValue(48 - 1) + 1;
heapInit(HeapIndex::HEAP_SVM, 0ull, gfxBase);
} else if (gpuAddressSpace == maxNBitValue(47)) {
if (reservedCpuAddressRange.alignedPtr == nullptr) {
if (cpuAddressRangeSizeToReserve == 0) {
return false;
}
reservedCpuAddressRange = osMemory->reserveCpuAddressRange(cpuAddressRangeSizeToReserve, GfxPartition::heapGranularity);
if (reservedCpuAddressRange.originalPtr == nullptr) {
return false;
}
if (!isAligned<GfxPartition::heapGranularity>(reservedCpuAddressRange.alignedPtr)) {
return false;
}
}
gfxBase = reinterpret_cast<uint64_t>(reservedCpuAddressRange.alignedPtr);
gfxTop = gfxBase + cpuAddressRangeSizeToReserve;
heapInit(HeapIndex::HEAP_SVM, 0ull, gpuAddressSpace + 1);
} else if (gpuAddressSpace < maxNBitValue(47)) {
gfxBase = 0ull;
heapInit(HeapIndex::HEAP_SVM, 0ull, 0ull);
} else {
if (!initAdditionalRange(gpuAddressSpace, gfxBase, gfxTop, rootDeviceIndex, numRootDevices)) {
return false;
}
}
}
for (auto heap : GfxPartition::heap32Names) {
if (useFrontWindowPool && HeapAssigner::heapTypeWithFrontWindowPool(heap)) {
heapInitExternalWithFrontWindow(heap, gfxBase, gfxHeap32Size);
size_t externalFrontWindowSize = GfxPartition::externalFrontWindowPoolSize;
heapInitExternalWithFrontWindow(HeapAssigner::mapExternalWindowIndex(heap), heapAllocate(heap, externalFrontWindowSize),
externalFrontWindowSize);
} else if (HeapAssigner::isInternalHeap(heap)) {
heapInitWithFrontWindow(heap, gfxBase, gfxHeap32Size, GfxPartition::internalFrontWindowPoolSize);
heapInitFrontWindow(HeapAssigner::mapInternalWindowIndex(heap), gfxBase, GfxPartition::internalFrontWindowPoolSize);
} else {
heapInit(heap, gfxBase, gfxHeap32Size);
}
gfxBase += gfxHeap32Size;
}
uint64_t gfxStandardSize = alignDown((gfxTop - gfxBase) >> 1, heapGranularity);
heapInit(HeapIndex::HEAP_STANDARD, gfxBase, gfxStandardSize);
gfxBase += gfxStandardSize;
// Split HEAP_STANDARD64K among root devices
auto gfxStandard64KBSize = alignDown(gfxStandardSize / numRootDevices, GfxPartition::heapGranularity);
heapInit(HeapIndex::HEAP_STANDARD64KB, gfxBase + rootDeviceIndex * gfxStandard64KBSize, gfxStandard64KBSize);
return true;
}
} // namespace NEO
|
