mirror of
https://github.com/intel/compute-runtime.git
synced 2025-12-22 18:25:05 +08:00
d4c1631ac7
Related-To: NEO-12197 Currently for new resources user thread must wait before submitting actual workload. With this commit, instead of waiting on user thread, request is sent to background ULLS controller thread and additional semaphore is programmed. ULLS controller will perform actual wait and signal semaphore when paging fence reaches required value. Signed-off-by: Szymon Morek <szymon.morek@intel.com>
101 lines
3.3 KiB
C++
101 lines
3.3 KiB
C++
/*
|
|
* Copyright (C) 2018-2024 Intel Corporation
|
|
*
|
|
* SPDX-License-Identifier: MIT
|
|
*
|
|
*/
|
|
|
|
#pragma once
|
|
#include "shared/source/command_container/cmdcontainer.h"
|
|
#include "shared/source/command_stream/csr_definitions.h"
|
|
#include "shared/source/command_stream/task_count_helper.h"
|
|
#include "shared/source/helpers/pipe_control_args.h"
|
|
#include "shared/source/utilities/idlist.h"
|
|
|
|
#include <vector>
|
|
namespace NEO {
|
|
class Device;
|
|
class Event;
|
|
class FlushStampTracker;
|
|
class GraphicsAllocation;
|
|
|
|
struct PagingFenceSemaphoreInfo {
|
|
bool requiresBlockingResidencyHandling = true;
|
|
uint64_t pagingFenceValue = 0u;
|
|
};
|
|
|
|
struct BatchBuffer {
|
|
BatchBuffer(GraphicsAllocation *commandBufferAllocation,
|
|
size_t startOffset,
|
|
size_t chainedBatchBufferStartOffset,
|
|
uint64_t taskStartAddress,
|
|
GraphicsAllocation *chainedBatchBuffer,
|
|
bool lowPriority,
|
|
QueueThrottle throttle,
|
|
uint64_t sliceCount,
|
|
size_t usedSize,
|
|
LinearStream *stream,
|
|
void *endCmdPtr,
|
|
uint32_t numCsrClients,
|
|
bool hasStallingCmds,
|
|
bool hasRelaxedOrderingDependencies,
|
|
bool dispatchMonitorFence);
|
|
BatchBuffer() {}
|
|
GraphicsAllocation *commandBufferAllocation = nullptr;
|
|
ResidencyContainer *allocationsForResidency = nullptr;
|
|
size_t startOffset = 0u;
|
|
size_t chainedBatchBufferStartOffset = 0u;
|
|
uint64_t taskStartAddress = 0; // if task not available, use CSR stream
|
|
|
|
GraphicsAllocation *chainedBatchBuffer = nullptr;
|
|
bool lowPriority = false;
|
|
QueueThrottle throttle = QueueThrottle::MEDIUM;
|
|
uint64_t sliceCount = QueueSliceCount::defaultSliceCount;
|
|
size_t usedSize = 0u;
|
|
|
|
// only used in drm csr in gem close worker active mode
|
|
LinearStream *stream = nullptr;
|
|
void *endCmdPtr = nullptr;
|
|
uint32_t numCsrClients = 0;
|
|
|
|
PagingFenceSemaphoreInfo pagingFenceSemInfo{};
|
|
|
|
bool hasStallingCmds = false;
|
|
bool hasRelaxedOrderingDependencies = false;
|
|
bool disableFlatRingBuffer = false;
|
|
bool dispatchMonitorFence = false;
|
|
};
|
|
|
|
struct CommandBuffer : public IDNode<CommandBuffer> {
|
|
CommandBuffer(Device &device);
|
|
CommandBuffer &operator=(const CommandBuffer &) = delete;
|
|
CommandBuffer &operator=(CommandBuffer &&) = delete;
|
|
~CommandBuffer() override;
|
|
ResidencyContainer surfaces;
|
|
BatchBuffer batchBuffer;
|
|
void *batchBufferEndLocation = nullptr;
|
|
uint32_t inspectionId = 0;
|
|
TaskCountType taskCount = 0u;
|
|
void *pipeControlThatMayBeErasedLocation = nullptr;
|
|
void *epiloguePipeControlLocation = nullptr;
|
|
PipeControlArgs epiloguePipeControlArgs;
|
|
std::unique_ptr<FlushStampTracker> flushStamp;
|
|
Device &device;
|
|
};
|
|
|
|
struct CommandBufferList : public IDList<CommandBuffer, false, true, false> {};
|
|
|
|
using ResourcePackage = StackVec<GraphicsAllocation *, 128>;
|
|
|
|
class SubmissionAggregator {
|
|
public:
|
|
void recordCommandBuffer(CommandBuffer *commandBuffer);
|
|
void aggregateCommandBuffers(ResourcePackage &resourcePackage, size_t &totalUsedSize, size_t totalMemoryBudget, uint32_t osContextId);
|
|
CommandBufferList &peekCmdBufferList() { return cmdBuffers; }
|
|
|
|
protected:
|
|
CommandBufferList cmdBuffers;
|
|
uint32_t inspectionId = 1;
|
|
};
|
|
} // namespace NEO
|