thek0tyara/llama-cpp-turboquant
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 4

Merge pull request #182 from EthicalML/181_memory_barriers

Browse source 
Updated memory barriers to include staging buffers
This commit is contained in:
Alejandro Saucedo 2021-03-13 10:03:14 +00:00 • committed by GitHub
commit 00f02cb9ad
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
10 changed files with 345 additions and 72 deletions
Download patch file Download diff file Expand all files Collapse all files

1
single_include/AggregateHeaders.cpp
View file

@ -6,6 +6,7 @@
#include "kompute/Tensor.hpp"
#include "kompute/Algorithm.hpp"
#include "kompute/operations/OpBase.hpp"
#include "kompute/operations/OpMemoryBarrier.hpp"
#include "kompute/operations/OpTensorCopy.hpp"
#include "kompute/operations/OpTensorSyncDevice.hpp"
#include "kompute/operations/OpTensorSyncLocal.hpp"

122
single_include/kompute/Kompute.hpp
View file

@ -887,12 +887,9 @@ class Tensor
*
* @param commandBuffer Vulkan Command Buffer to record the commands into
* @param copyFromTensor Tensor to copy the data from
* @param createBarrier Whether to create a barrier that ensures the data is
* copied before further operations. Default is true.
*/
void recordCopyFrom(const vk::CommandBuffer& commandBuffer,
std::shared_ptr<Tensor> copyFromTensor,
bool createBarrier);
std::shared_ptr<Tensor> copyFromTensor);
/**
* Records a copy from the internal staging memory to the device memory
@ -900,11 +897,8 @@ class Tensor
* only be relevant for kp::Tensors of type eDevice.
*
* @param commandBuffer Vulkan Command Buffer to record the commands into
* @param createBarrier Whether to create a barrier that ensures the data is
* copied before further operations. Default is true.
*/
void recordCopyFromStagingToDevice(const vk::CommandBuffer& commandBuffer,
bool createBarrier);
void recordCopyFromStagingToDevice(const vk::CommandBuffer& commandBuffer);
/**
* Records a copy from the internal device memory to the staging memory
@ -912,14 +906,11 @@ class Tensor
* only be relevant for kp::Tensors of type eDevice.
*
* @param commandBuffer Vulkan Command Buffer to record the commands into
* @param createBarrier Whether to create a barrier that ensures the data is
* copied before further operations. Default is true.
*/
void recordCopyFromDeviceToStaging(const vk::CommandBuffer& commandBuffer,
bool createBarrier);
void recordCopyFromDeviceToStaging(const vk::CommandBuffer& commandBuffer);
/**
* Records the buffer memory barrier into the command buffer which
* Records the buffer memory barrier into the primary buffer and command buffer which
* ensures that relevant data transfers are carried out correctly.
*
* @param commandBuffer Vulkan Command Buffer to record the commands into
@ -928,11 +919,26 @@ class Tensor
* @param scrStageMask Pipeline stage flags for source stage mask
* @param dstStageMask Pipeline stage flags for destination stage mask
*/
void recordBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits dstAccessMask,
vk::PipelineStageFlagBits srcStageMask,
vk::PipelineStageFlagBits dstStageMask);
void recordPrimaryBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits dstAccessMask,
vk::PipelineStageFlagBits srcStageMask,
vk::PipelineStageFlagBits dstStageMask);
/**
* Records the buffer memory barrier into the staging buffer and command buffer which
* ensures that relevant data transfers are carried out correctly.
*
* @param commandBuffer Vulkan Command Buffer to record the commands into
* @param srcAccessMask Access flags for source access mask
* @param dstAccessMask Access flags for destination access mask
* @param scrStageMask Pipeline stage flags for source stage mask
* @param dstStageMask Pipeline stage flags for destination stage mask
*/
void recordStagingBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits dstAccessMask,
vk::PipelineStageFlagBits srcStageMask,
vk::PipelineStageFlagBits dstStageMask);
/**
* Constructs a vulkan descriptor buffer info which can be used to specify
@ -1074,8 +1080,13 @@ class Tensor
std::shared_ptr<vk::Buffer> bufferFrom,
std::shared_ptr<vk::Buffer> bufferTo,
vk::DeviceSize bufferSize,
vk::BufferCopy copyRegion,
bool createBarrier);
vk::BufferCopy copyRegion);
void recordBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
const vk::Buffer& buffer,
vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits dstAccessMask,
vk::PipelineStageFlagBits srcStageMask,
vk::PipelineStageFlagBits dstStageMask);
// Private util functions
vk::BufferUsageFlags getPrimaryBufferUsageFlags();
@ -1377,6 +1388,77 @@ class OpBase
namespace kp {
/**
* Operation that provides a general abstraction that simplifies the use of
* algorithm and parameter components which can be used with shaders.
* It exposes the pipeline barrier functionality specifically for memory
* barriers that can be configured through the respective source and destination
* masks
*/
class OpMemoryBarrier : public OpBase
{
public:
/**
* Constructor that stores tensors as well as memory barrier parameters to be
* used to create a pipeline barrier on the respective primary or staging tensor.
*
* @param tensors The tensors to apply the memory barriers on
* @param srcAccessMask The kp::AccessFlagBits for the source access mask
* @param dstAccessMask The kp::AccessFlagBits for the destination access mask
* @param srcStageMask The kp::PipelineStageFlagBits for the source stage mask
* @param dstStageMask The kp::PipelineStageFlagBits for the destination stage mask
* @param barrierOnPrimary Boolean to select primary or secondary buffers on tensors
*/
OpMemoryBarrier(
const std::vector<std::shared_ptr<Tensor>>& tensors,
const vk::AccessFlagBits& srcAccessMask,
const vk::AccessFlagBits& dstAccessMask,
const vk::PipelineStageFlagBits& srcStageMask,
const vk::PipelineStageFlagBits& dstStageMask,
bool barrierOnPrimary = true);
/**
* Default destructor, which is in charge of destroying the reference to the tensors
* and all the relevant access / stage masks created
*/
virtual ~OpMemoryBarrier() override;
/**
* This records the memory barrier with the access and stage masks provided
* across all relevant tensors.
*
* @param commandBuffer The command buffer to record the command into.
*/
virtual void record(const vk::CommandBuffer& commandBuffer) override;
/**
* Does not perform any preEval commands.
*
* @param commandBuffer The command buffer to record the command into.
*/
virtual void preEval(const vk::CommandBuffer& commandBuffer) override;
/**
* Does not perform any postEval commands.
*
* @param commandBuffer The command buffer to record the command into.
*/
virtual void postEval(const vk::CommandBuffer& commandBuffer) override;
private:
const vk::AccessFlagBits mSrcAccessMask;
const vk::AccessFlagBits mDstAccessMask;
const vk::PipelineStageFlagBits mSrcStageMask;
const vk::PipelineStageFlagBits mDstStageMask;
const bool mBarrierOnPrimary;
const std::vector<std::shared_ptr<Tensor>> mTensors;
};
} // End namespace kp
namespace kp {
/**
* Operation that copies the data from the first tensor to the rest of the tensors
* provided, using a record command for all the vectors. This operation does not

6
src/OpAlgoDispatch.cpp
View file

@ -26,11 +26,11 @@ OpAlgoDispatch::record(const vk::CommandBuffer& commandBuffer)
// Barrier to ensure the data is finished writing to buffer memory
for (const std::shared_ptr<Tensor>& tensor :
this->mAlgorithm->getTensors()) {
tensor->recordBufferMemoryBarrier(
tensor->recordPrimaryBufferMemoryBarrier(
commandBuffer,
vk::AccessFlagBits::eHostWrite,
vk::AccessFlagBits::eTransferWrite,
vk::AccessFlagBits::eShaderRead,
vk::PipelineStageFlagBits::eHost,
vk::PipelineStageFlagBits::eTransfer,
vk::PipelineStageFlagBits::eComputeShader);
}

69
src/OpMemoryBarrier.cpp Normal file
View file

@ -0,0 +1,69 @@
#pragma once
#include "kompute/operations/OpMemoryBarrier.hpp"
namespace kp {
OpMemoryBarrier::OpMemoryBarrier(
const std::vector<std::shared_ptr<Tensor>>& tensors,
const vk::AccessFlagBits& srcAccessMask,
const vk::AccessFlagBits& dstAccessMask,
const vk::PipelineStageFlagBits& srcStageMask,
const vk::PipelineStageFlagBits& dstStageMask,
bool barrierOnPrimary)
: mTensors(tensors),
mSrcAccessMask(srcAccessMask),
mDstAccessMask(dstAccessMask),
mSrcStageMask(srcStageMask),
mDstStageMask(dstStageMask),
mBarrierOnPrimary(barrierOnPrimary)
{
KP_LOG_DEBUG("Kompute OpMemoryBarrier constructor");
}
OpMemoryBarrier::~OpMemoryBarrier()
{
KP_LOG_DEBUG("Kompute OpMemoryBarrier destructor started");
}
void
OpMemoryBarrier::record(const vk::CommandBuffer& commandBuffer)
{
KP_LOG_DEBUG("Kompute OpMemoryBarrier record called");
// Barrier to ensure the data is finished writing to buffer memory
if (this->mBarrierOnPrimary) {
for (const std::shared_ptr<Tensor>& tensor : this->mTensors) {
tensor->recordPrimaryBufferMemoryBarrier(
commandBuffer,
this->mSrcAccessMask,
this->mDstAccessMask,
this->mSrcStageMask,
this->mDstStageMask);
}
} else {
for (const std::shared_ptr<Tensor>& tensor : this->mTensors) {
tensor->recordStagingBufferMemoryBarrier(
commandBuffer,
this->mSrcAccessMask,
this->mDstAccessMask,
this->mSrcStageMask,
this->mDstStageMask);
}
}
}
void
OpMemoryBarrier::preEval(const vk::CommandBuffer& commandBuffer)
{
KP_LOG_DEBUG("Kompute OpMemoryBarrier preEval called");
}
void
OpMemoryBarrier::postEval(const vk::CommandBuffer& commandBuffer)
{
KP_LOG_DEBUG("Kompute OpMemoryBarrier postSubmit called");
}
}

2
src/OpTensorCopy.cpp
View file

@ -45,7 +45,7 @@ OpTensorCopy::record(const vk::CommandBuffer& commandBuffer)
// We iterate from the second tensor onwards and record a copy to all
for (size_t i = 1; i < this->mTensors.size(); i++) {
this->mTensors[i]->recordCopyFrom(
commandBuffer, this->mTensors[0], false);
commandBuffer, this->mTensors[0]);
}
}

3
src/OpTensorSyncDevice.cpp
View file

@ -30,8 +30,7 @@ OpTensorSyncDevice::record(const vk::CommandBuffer& commandBuffer)
for (size_t i = 0; i < this->mTensors.size(); i++) {
if (this->mTensors[i]->tensorType() == Tensor::TensorTypes::eDevice) {
this->mTensors[i]->recordCopyFromStagingToDevice(commandBuffer,
false);
this->mTensors[i]->recordCopyFromStagingToDevice(commandBuffer);
}
}
}

16
src/OpTensorSyncLocal.cpp
View file

@ -30,8 +30,20 @@ OpTensorSyncLocal::record(const vk::CommandBuffer& commandBuffer)
for (size_t i = 0; i < this->mTensors.size(); i++) {
if (this->mTensors[i]->tensorType() == Tensor::TensorTypes::eDevice) {
this->mTensors[i]->recordCopyFromDeviceToStaging(commandBuffer,
true);
this->mTensors[i]->recordPrimaryBufferMemoryBarrier(commandBuffer,
vk::AccessFlagBits::eShaderWrite,
vk::AccessFlagBits::eTransferRead,
vk::PipelineStageFlagBits::eComputeShader,
vk::PipelineStageFlagBits::eTransfer);
this->mTensors[i]->recordCopyFromDeviceToStaging(commandBuffer);
this->mTensors[i]->recordPrimaryBufferMemoryBarrier(commandBuffer,
vk::AccessFlagBits::eTransferWrite,
vk::AccessFlagBits::eHostRead,
vk::PipelineStageFlagBits::eTransfer,
vk::PipelineStageFlagBits::eHost);
}
}
}

68
src/Tensor.cpp
View file

@ -72,8 +72,7 @@ Tensor::isInit()
void
Tensor::recordCopyFrom(const vk::CommandBuffer& commandBuffer,
std::shared_ptr<Tensor> copyFromTensor,
bool createBarrier)
std::shared_ptr<Tensor> copyFromTensor)
{
vk::DeviceSize bufferSize(this->memorySize());
@ -85,13 +84,11 @@ Tensor::recordCopyFrom(const vk::CommandBuffer& commandBuffer,
copyFromTensor->mPrimaryBuffer,
this->mPrimaryBuffer,
bufferSize,
copyRegion,
createBarrier);
copyRegion);
}
void
Tensor::recordCopyFromStagingToDevice(const vk::CommandBuffer& commandBuffer,
bool createBarrier)
Tensor::recordCopyFromStagingToDevice(const vk::CommandBuffer& commandBuffer)
{
vk::DeviceSize bufferSize(this->memorySize());
vk::BufferCopy copyRegion(0, 0, bufferSize);
@ -102,13 +99,11 @@ Tensor::recordCopyFromStagingToDevice(const vk::CommandBuffer& commandBuffer,
this->mStagingBuffer,
this->mPrimaryBuffer,
bufferSize,
copyRegion,
createBarrier);
copyRegion);
}
void
Tensor::recordCopyFromDeviceToStaging(const vk::CommandBuffer& commandBuffer,
bool createBarrier)
Tensor::recordCopyFromDeviceToStaging(const vk::CommandBuffer& commandBuffer)
{
vk::DeviceSize bufferSize(this->memorySize());
vk::BufferCopy copyRegion(0, 0, bufferSize);
@ -119,8 +114,7 @@ Tensor::recordCopyFromDeviceToStaging(const vk::CommandBuffer& commandBuffer,
this->mPrimaryBuffer,
this->mStagingBuffer,
bufferSize,
copyRegion,
createBarrier);
copyRegion);
}
void
@ -128,24 +122,49 @@ Tensor::recordCopyBuffer(const vk::CommandBuffer& commandBuffer,
std::shared_ptr<vk::Buffer> bufferFrom,
std::shared_ptr<vk::Buffer> bufferTo,
vk::DeviceSize bufferSize,
vk::BufferCopy copyRegion,
bool createBarrier)
vk::BufferCopy copyRegion)
{
commandBuffer.copyBuffer(*bufferFrom, *bufferTo, copyRegion);
}
if (createBarrier) {
// Buffer to ensure wait until data is copied to staging buffer
this->recordBufferMemoryBarrier(commandBuffer,
vk::AccessFlagBits::eTransferWrite,
vk::AccessFlagBits::eHostRead,
vk::PipelineStageFlagBits::eTransfer,
vk::PipelineStageFlagBits::eHost);
}
void
Tensor::recordPrimaryBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits dstAccessMask,
vk::PipelineStageFlagBits srcStageMask,
vk::PipelineStageFlagBits dstStageMask)
{
KP_LOG_DEBUG("Kompute Tensor recording PRIMARY buffer memory barrier");
this->recordBufferMemoryBarrier(commandBuffer,
*this->mPrimaryBuffer,
srcAccessMask,
dstAccessMask,
srcStageMask,
dstStageMask);
}
void
Tensor::recordStagingBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits dstAccessMask,
vk::PipelineStageFlagBits srcStageMask,
vk::PipelineStageFlagBits dstStageMask)
{
KP_LOG_DEBUG("Kompute Tensor recording PRIMARY buffer memory barrier");
this->recordBufferMemoryBarrier(commandBuffer,
*this->mStagingBuffer,
srcAccessMask,
dstAccessMask,
srcStageMask,
dstStageMask);
}
void
Tensor::recordBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
const vk::Buffer& buffer,
vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits dstAccessMask,
vk::PipelineStageFlagBits srcStageMask,
@ -156,7 +175,7 @@ Tensor::recordBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
vk::DeviceSize bufferSize = this->memorySize();
vk::BufferMemoryBarrier bufferMemoryBarrier;
bufferMemoryBarrier.buffer = *this->mPrimaryBuffer;
bufferMemoryBarrier.buffer = buffer;
bufferMemoryBarrier.size = bufferSize;
bufferMemoryBarrier.srcAccessMask = srcAccessMask;
bufferMemoryBarrier.dstAccessMask = dstAccessMask;
@ -241,7 +260,8 @@ Tensor::getStagingMemoryPropertyFlags()
{
switch (this->mTensorType) {
case TensorTypes::eDevice:
return vk::MemoryPropertyFlagBits::eHostVisible;
return vk::MemoryPropertyFlagBits::eHostVisible |
vk::MemoryPropertyFlagBits::eHostCoherent;
break;
default:
throw std::runtime_error("Kompute Tensor invalid tensor type");

52
src/include/kompute/Tensor.hpp
View file

@ -97,12 +97,9 @@ class Tensor
*
* @param commandBuffer Vulkan Command Buffer to record the commands into
* @param copyFromTensor Tensor to copy the data from
* @param createBarrier Whether to create a barrier that ensures the data is
* copied before further operations. Default is true.
*/
void recordCopyFrom(const vk::CommandBuffer& commandBuffer,
std::shared_ptr<Tensor> copyFromTensor,
bool createBarrier);
std::shared_ptr<Tensor> copyFromTensor);
/**
* Records a copy from the internal staging memory to the device memory
@ -110,11 +107,8 @@ class Tensor
* only be relevant for kp::Tensors of type eDevice.
*
* @param commandBuffer Vulkan Command Buffer to record the commands into
* @param createBarrier Whether to create a barrier that ensures the data is
* copied before further operations. Default is true.
*/
void recordCopyFromStagingToDevice(const vk::CommandBuffer& commandBuffer,
bool createBarrier);
void recordCopyFromStagingToDevice(const vk::CommandBuffer& commandBuffer);
/**
* Records a copy from the internal device memory to the staging memory
@ -122,14 +116,11 @@ class Tensor
* only be relevant for kp::Tensors of type eDevice.
*
* @param commandBuffer Vulkan Command Buffer to record the commands into
* @param createBarrier Whether to create a barrier that ensures the data is
* copied before further operations. Default is true.
*/
void recordCopyFromDeviceToStaging(const vk::CommandBuffer& commandBuffer,
bool createBarrier);
void recordCopyFromDeviceToStaging(const vk::CommandBuffer& commandBuffer);
/**
* Records the buffer memory barrier into the command buffer which
* Records the buffer memory barrier into the primary buffer and command buffer which
* ensures that relevant data transfers are carried out correctly.
*
* @param commandBuffer Vulkan Command Buffer to record the commands into
@ -138,11 +129,27 @@ class Tensor
* @param scrStageMask Pipeline stage flags for source stage mask
* @param dstStageMask Pipeline stage flags for destination stage mask
*/
void recordBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits dstAccessMask,
vk::PipelineStageFlagBits srcStageMask,
vk::PipelineStageFlagBits dstStageMask);
void recordPrimaryBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits dstAccessMask,
vk::PipelineStageFlagBits srcStageMask,
vk::PipelineStageFlagBits dstStageMask);
/**
* Records the buffer memory barrier into the staging buffer and command buffer which
* ensures that relevant data transfers are carried out correctly.
*
* @param commandBuffer Vulkan Command Buffer to record the commands into
* @param srcAccessMask Access flags for source access mask
* @param dstAccessMask Access flags for destination access mask
* @param scrStageMask Pipeline stage flags for source stage mask
* @param dstStageMask Pipeline stage flags for destination stage mask
*/
void recordStagingBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits dstAccessMask,
vk::PipelineStageFlagBits srcStageMask,
vk::PipelineStageFlagBits dstStageMask);
/**
* Constructs a vulkan descriptor buffer info which can be used to specify
@ -284,8 +291,13 @@ class Tensor
std::shared_ptr<vk::Buffer> bufferFrom,
std::shared_ptr<vk::Buffer> bufferTo,
vk::DeviceSize bufferSize,
vk::BufferCopy copyRegion,
bool createBarrier);
vk::BufferCopy copyRegion);
void recordBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
const vk::Buffer& buffer,
vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits dstAccessMask,
vk::PipelineStageFlagBits srcStageMask,
vk::PipelineStageFlagBits dstStageMask);
// Private util functions
vk::BufferUsageFlags getPrimaryBufferUsageFlags();

78
src/include/kompute/operations/OpMemoryBarrier.hpp Normal file
View file

@ -0,0 +1,78 @@
#pragma once
#include "kompute/Core.hpp"
#include "kompute/Algorithm.hpp"
#include "kompute/Tensor.hpp"
#include "kompute/operations/OpBase.hpp"
namespace kp {
/**
* Operation that provides a general abstraction that simplifies the use of
* algorithm and parameter components which can be used with shaders.
* It exposes the pipeline barrier functionality specifically for memory
* barriers that can be configured through the respective source and destination
* masks
*/
class OpMemoryBarrier : public OpBase
{
public:
/**
* Constructor that stores tensors as well as memory barrier parameters to be
* used to create a pipeline barrier on the respective primary or staging tensor.
*
* @param tensors The tensors to apply the memory barriers on
* @param srcAccessMask The kp::AccessFlagBits for the source access mask
* @param dstAccessMask The kp::AccessFlagBits for the destination access mask
* @param srcStageMask The kp::PipelineStageFlagBits for the source stage mask
* @param dstStageMask The kp::PipelineStageFlagBits for the destination stage mask
* @param barrierOnPrimary Boolean to select primary or secondary buffers on tensors
*/
OpMemoryBarrier(
const std::vector<std::shared_ptr<Tensor>>& tensors,
const vk::AccessFlagBits& srcAccessMask,
const vk::AccessFlagBits& dstAccessMask,
const vk::PipelineStageFlagBits& srcStageMask,
const vk::PipelineStageFlagBits& dstStageMask,
bool barrierOnPrimary = true);
/**
* Default destructor, which is in charge of destroying the reference to the tensors
* and all the relevant access / stage masks created
*/
virtual ~OpMemoryBarrier() override;
/**
* This records the memory barrier with the access and stage masks provided
* across all relevant tensors.
*
* @param commandBuffer The command buffer to record the command into.
*/
virtual void record(const vk::CommandBuffer& commandBuffer) override;
/**
* Does not perform any preEval commands.
*
* @param commandBuffer The command buffer to record the command into.
*/
virtual void preEval(const vk::CommandBuffer& commandBuffer) override;
/**
* Does not perform any postEval commands.
*
* @param commandBuffer The command buffer to record the command into.
*/
virtual void postEval(const vk::CommandBuffer& commandBuffer) override;
private:
const vk::AccessFlagBits mSrcAccessMask;
const vk::AccessFlagBits mDstAccessMask;
const vk::PipelineStageFlagBits mSrcStageMask;
const vk::PipelineStageFlagBits mDstStageMask;
const bool mBarrierOnPrimary;
const std::vector<std::shared_ptr<Tensor>> mTensors;
};
} // End namespace kp