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
|
/*
* Copyright (C) 2017-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#pragma once
#include "shared/source/built_ins/built_ins.h"
#include "shared/source/command_stream/command_stream_receiver.h"
#include "shared/source/helpers/cache_policy.h"
#include "shared/source/helpers/engine_node_helper.h"
#include "shared/source/memory_manager/unified_memory_manager.h"
#include "shared/source/os_interface/os_context.h"
#include "opencl/source/command_queue/command_queue_hw.h"
#include "opencl/source/command_queue/enqueue_common.h"
#include "opencl/source/helpers/hardware_commands_helper.h"
#include "opencl/source/mem_obj/buffer.h"
#include "opencl/source/memory_manager/mem_obj_surface.h"
#include <new>
namespace NEO {
template <typename GfxFamily>
cl_int CommandQueueHw<GfxFamily>::enqueueReadBuffer(
Buffer *buffer,
cl_bool blockingRead,
size_t offset,
size_t size,
void *ptr,
GraphicsAllocation *mapAllocation,
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
cl_event *event) {
const cl_command_type cmdType = CL_COMMAND_READ_BUFFER;
auto &csr = getCommandStreamReceiverByCommandType(cmdType);
if (nullptr == mapAllocation) {
notifyEnqueueReadBuffer(buffer, !!blockingRead, EngineHelpers::isBcs(csr.getOsContext().getEngineType()));
}
auto rootDeviceIndex = getDevice().getRootDeviceIndex();
bool isMemTransferNeeded = buffer->isMemObjZeroCopy() ? buffer->checkIfMemoryTransferIsRequired(offset, 0, ptr, cmdType) : true;
bool isCpuCopyAllowed = bufferCpuCopyAllowed(buffer, cmdType, blockingRead, size, ptr,
numEventsInWaitList, eventWaitList);
//check if we are dealing with SVM pointer here for which we already have an allocation
if (!mapAllocation && this->getContext().getSVMAllocsManager()) {
auto svmEntry = this->getContext().getSVMAllocsManager()->getSVMAlloc(ptr);
if (svmEntry) {
if ((svmEntry->gpuAllocations.getGraphicsAllocation(rootDeviceIndex)->getGpuAddress() + svmEntry->size) < (castToUint64(ptr) + size)) {
return CL_INVALID_OPERATION;
}
mapAllocation = svmEntry->cpuAllocation ? svmEntry->cpuAllocation : svmEntry->gpuAllocations.getGraphicsAllocation(rootDeviceIndex);
if (isCpuCopyAllowed) {
if (svmEntry->memoryType == DEVICE_UNIFIED_MEMORY) {
isCpuCopyAllowed = false;
}
}
}
}
if (isCpuCopyAllowed) {
if (isMemTransferNeeded) {
return enqueueReadWriteBufferOnCpuWithMemoryTransfer(cmdType, buffer, offset, size, ptr,
numEventsInWaitList, eventWaitList, event);
} else {
return enqueueReadWriteBufferOnCpuWithoutMemoryTransfer(cmdType, buffer, offset, size, ptr,
numEventsInWaitList, eventWaitList, event);
}
} else if (!isMemTransferNeeded) {
return enqueueMarkerForReadWriteOperation(buffer, ptr, cmdType, blockingRead,
numEventsInWaitList, eventWaitList, event);
}
auto eBuiltInOps = EBuiltInOps::CopyBufferToBuffer;
if (forceStateless(buffer->getSize())) {
eBuiltInOps = EBuiltInOps::CopyBufferToBufferStateless;
}
void *dstPtr = ptr;
MemObjSurface bufferSurf(buffer);
HostPtrSurface hostPtrSurf(dstPtr, size);
GeneralSurface mapSurface;
Surface *surfaces[] = {&bufferSurf, nullptr};
if (mapAllocation) {
surfaces[1] = &mapSurface;
mapSurface.setGraphicsAllocation(mapAllocation);
//get offset between base cpu ptr of map allocation and dst ptr
size_t dstOffset = ptrDiff(dstPtr, mapAllocation->getUnderlyingBuffer());
dstPtr = reinterpret_cast<void *>(mapAllocation->getGpuAddress() + dstOffset);
} else {
surfaces[1] = &hostPtrSurf;
if (size != 0) {
bool status = csr.createAllocationForHostSurface(hostPtrSurf, true);
if (!status) {
return CL_OUT_OF_RESOURCES;
}
dstPtr = reinterpret_cast<void *>(hostPtrSurf.getAllocation()->getGpuAddress());
}
}
void *alignedDstPtr = alignDown(dstPtr, 4);
size_t dstPtrOffset = ptrDiff(dstPtr, alignedDstPtr);
BuiltinOpParams dc;
dc.dstPtr = alignedDstPtr;
dc.dstOffset = {dstPtrOffset, 0, 0};
dc.srcMemObj = buffer;
dc.srcOffset = {offset, 0, 0};
dc.size = {size, 0, 0};
dc.transferAllocation = mapAllocation ? mapAllocation : hostPtrSurf.getAllocation();
MultiDispatchInfo dispatchInfo(dc);
if (context->isProvidingPerformanceHints()) {
context->providePerformanceHintForMemoryTransfer(CL_COMMAND_READ_BUFFER, true, static_cast<cl_mem>(buffer), ptr);
if (!isL3Capable(ptr, size)) {
context->providePerformanceHint(CL_CONTEXT_DIAGNOSTICS_LEVEL_BAD_INTEL, CL_ENQUEUE_READ_BUFFER_DOESNT_MEET_ALIGNMENT_RESTRICTIONS, ptr, size, MemoryConstants::pageSize, MemoryConstants::pageSize);
}
}
dispatchBcsOrGpgpuEnqueue<CL_COMMAND_READ_BUFFER>(dispatchInfo, surfaces, eBuiltInOps, numEventsInWaitList, eventWaitList, event, blockingRead);
return CL_SUCCESS;
}
} // namespace NEO
|
