compute-runtime/runtime/command_queue/enqueue_write_buffer_rect.h

112 lines
3.8 KiB
C++

/*
* Copyright (C) 2017-2018 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#pragma once
#include "runtime/command_queue/command_queue_hw.h"
#include "runtime/command_stream/command_stream_receiver.h"
#include "runtime/helpers/kernel_commands.h"
#include "runtime/mem_obj/buffer.h"
#include "runtime/memory_manager/surface.h"
#include "runtime/built_ins/built_ins.h"
#include <new>
namespace OCLRT {
template <typename GfxFamily>
cl_int CommandQueueHw<GfxFamily>::enqueueWriteBufferRect(
Buffer *buffer,
cl_bool blockingWrite,
const size_t *bufferOrigin,
const size_t *hostOrigin,
const size_t *region,
size_t bufferRowPitch,
size_t bufferSlicePitch,
size_t hostRowPitch,
size_t hostSlicePitch,
const void *ptr,
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
cl_event *event) {
MultiDispatchInfo dispatchInfo;
auto isMemTransferNeeded = true;
if (buffer->isMemObjZeroCopy()) {
size_t bufferOffset;
size_t hostOffset;
computeOffsetsValueForRectCommands(&bufferOffset, &hostOffset, bufferOrigin, hostOrigin, region, bufferRowPitch, bufferSlicePitch, hostRowPitch, hostSlicePitch);
isMemTransferNeeded = buffer->checkIfMemoryTransferIsRequired(bufferOffset, hostOffset, ptr, CL_COMMAND_WRITE_BUFFER_RECT);
}
if (!isMemTransferNeeded) {
NullSurface s;
Surface *surfaces[] = {&s};
enqueueHandler<CL_COMMAND_MARKER>(
surfaces,
blockingWrite == CL_TRUE,
dispatchInfo,
numEventsInWaitList,
eventWaitList,
event);
if (event) {
auto pEvent = castToObjectOrAbort<Event>(*event);
pEvent->setCmdType(CL_COMMAND_WRITE_BUFFER_RECT);
}
if (context->isProvidingPerformanceHints()) {
context->providePerformanceHint(CL_CONTEXT_DIAGNOSTICS_LEVEL_GOOD_INTEL, CL_ENQUEUE_WRITE_BUFFER_RECT_DOESNT_REQUIRE_COPY_DATA, static_cast<cl_mem>(buffer), ptr);
}
return CL_SUCCESS;
}
auto &builder = getDevice().getExecutionEnvironment()->getBuiltIns()->getBuiltinDispatchInfoBuilder(EBuiltInOps::CopyBufferRect,
this->getContext(), this->getDevice());
BuiltInOwnershipWrapper builtInLock(builder, this->context);
size_t hostPtrSize = Buffer::calculateHostPtrSize(hostOrigin, region, hostRowPitch, hostSlicePitch);
void *srcPtr = const_cast<void *>(ptr);
MemObjSurface dstBufferSurf(buffer);
HostPtrSurface hostPtrSurf(srcPtr, hostPtrSize, true);
Surface *surfaces[] = {&dstBufferSurf, &hostPtrSurf};
if (region[0] != 0 &&
region[1] != 0 &&
region[2] != 0) {
bool status = createAllocationForHostSurface(hostPtrSurf);
if (!status) {
return CL_OUT_OF_RESOURCES;
}
srcPtr = reinterpret_cast<void *>(hostPtrSurf.getAllocation()->getGpuAddressToPatch());
}
BuiltinDispatchInfoBuilder::BuiltinOpParams dc;
dc.srcPtr = srcPtr;
dc.dstMemObj = buffer;
dc.srcOffset = hostOrigin;
dc.dstOffset = bufferOrigin;
dc.size = region;
dc.srcRowPitch = hostRowPitch;
dc.srcSlicePitch = hostSlicePitch;
dc.dstRowPitch = bufferRowPitch;
dc.dstSlicePitch = bufferSlicePitch;
builder.buildDispatchInfos(dispatchInfo, dc);
enqueueHandler<CL_COMMAND_WRITE_BUFFER_RECT>(
surfaces,
blockingWrite == CL_TRUE,
dispatchInfo,
numEventsInWaitList,
eventWaitList,
event);
if (context->isProvidingPerformanceHints()) {
context->providePerformanceHint(CL_CONTEXT_DIAGNOSTICS_LEVEL_NEUTRAL_INTEL, CL_ENQUEUE_WRITE_BUFFER_RECT_REQUIRES_COPY_DATA, static_cast<cl_mem>(buffer));
}
return CL_SUCCESS;
}
} // namespace OCLRT