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

Initial implementation of tensor working compiling

Browse source
This commit is contained in:
Alejandro Saucedo 2021-03-06 17:25:35 +00:00
parent b81896a780
commit ad18c2e546
12 changed files with 417 additions and 210 deletions
Download patch file Download diff file Expand all files Collapse all files

15
src/Manager.cpp
View file

@ -395,21 +395,6 @@ Manager::createDevice(const std::vector<uint32_t>& familyQueueIndices,
KP_LOG_DEBUG("Kompute Manager compute queue obtained");
}
std::shared_ptr<Tensor>
Manager::tensor(const std::vector<float>& data, Tensor::TensorTypes tensorType)
{
KP_LOG_DEBUG("Kompute Manager tensor creation triggered");
std::shared_ptr<Tensor> tensor{ new kp::Tensor(
this->mPhysicalDevice, this->mDevice, data, tensorType) };
if (this->mManageResources) {
this->mManagedTensors.push_back(tensor);
}
return tensor;
}
std::shared_ptr<Algorithm>
Manager::algorithm(const std::vector<std::shared_ptr<Tensor>>& tensors,
const std::vector<uint32_t>& spirv,

17
src/OpTensorCopy.cpp
View file

@ -13,6 +13,14 @@ OpTensorCopy::OpTensorCopy(const std::vector<std::shared_ptr<Tensor>>& tensors)
throw std::runtime_error(
"Kompute OpTensorCopy called with less than 2 tensor");
}
kp::Tensor::TensorDataTypes dataType = this->mTensors[0]->dataType();
for (const std::shared_ptr<Tensor>& tensor : tensors) {
if (tensor->dataType() != dataType) {
throw std::runtime_error(fmt::format("Attempting to copy tensors of different types from {} to {}",
dataType, tensor->dataType()));
}
}
}
OpTensorCopy::~OpTensorCopy()
@ -43,9 +51,16 @@ OpTensorCopy::postEval(const vk::CommandBuffer& commandBuffer)
{
KP_LOG_DEBUG("Kompute OpTensorCopy postEval called");
// TODO: Simplify with a copyRawData
uint32_t size = this->mTensors[0]->size();
uint32_t dataTypeMemSize = this->mTensors[0]->dataTypeMemorySize();
uint32_t memSize = size * dataTypeMemSize;
void* data = operator new(memSize);
this->mTensors[0]->getRawData(data);
// Copy the data from the first tensor into all the tensors
for (size_t i = 1; i < this->mTensors.size(); i++) {
this->mTensors[i]->setData(this->mTensors[0]->data());
this->mTensors[i]->setRawData(data, size, dataTypeMemSize);
}
}

42
src/Tensor.cpp
View file

@ -17,8 +17,10 @@ Tensor::Tensor(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
this->mPhysicalDevice = physicalDevice;
this->mDevice = device;
this->mDataType = dataType;
this->mTensorType = tensorType;
this->rebuild(data, elementTotalCount, elementMemorySize, dataType, tensorType);
this->rebuild(data, elementTotalCount, elementMemorySize);
}
Tensor::~Tensor()
@ -34,16 +36,12 @@ Tensor::~Tensor()
void
Tensor::rebuild(void* data,
uint32_t elementTotalCount,
uint32_t elementMemorySize,
const TensorDataTypes& dataType,
TensorTypes tensorType)
uint32_t elementMemorySize)
{
KP_LOG_DEBUG("Kompute Tensor rebuilding with size {}", elementTotalCount);
this->mSize = elementTotalCount;
this->mElementMemorySize = elementMemorySize;
this->mDataType = dataType;
this->mTensorType = tensorType;
this->mDataTypeMemorySize = elementMemorySize;
if (this->mPrimaryBuffer || this->mPrimaryMemory) {
KP_LOG_DEBUG(
@ -439,4 +437,34 @@ Tensor::destroy()
KP_LOG_DEBUG("Kompute Tensor successful destroy()");
}
template<>
Tensor::TensorDataTypes
TensorView<bool>::dataType() {
return Tensor::TensorDataTypes::eBool;
}
template<>
Tensor::TensorDataTypes
TensorView<int32_t>::dataType() {
return Tensor::TensorDataTypes::eInt;
}
template<>
Tensor::TensorDataTypes
TensorView<uint32_t>::dataType() {
return Tensor::TensorDataTypes::eUnsignedInt;
}
template<>
Tensor::TensorDataTypes
TensorView<float>::dataType() {
return Tensor::TensorDataTypes::eFloat;
}
template<>
Tensor::TensorDataTypes
TensorView<double>::dataType() {
return Tensor::TensorDataTypes::eDouble;
}
}

19
src/include/kompute/Manager.hpp
View file

@ -74,9 +74,22 @@ class Manager
* @param tensorType The type of tensor to initialize
* @returns Shared pointer with initialised tensor
*/
std::shared_ptr<Tensor> tensor(
const std::vector<float>& data,
Tensor::TensorTypes tensorType = Tensor::TensorTypes::eDevice);
template <typename T>
std::shared_ptr<TensorView<T>> tensor(
const std::vector<T>& data,
Tensor::TensorTypes tensorType = Tensor::TensorTypes::eDevice)
{
KP_LOG_DEBUG("Kompute Manager tensor creation triggered");
std::shared_ptr<TensorView<T>> tensor{ new kp::TensorView<T>(
this->mPhysicalDevice, this->mDevice, data, tensorType) };
if (this->mManageResources) {
this->mManagedTensors.push_back(tensor);
}
return tensor;
}
/**
* Create a managed algorithm that will be destroyed by this manager

183
src/include/kompute/Tensor.hpp
View file

@ -51,14 +51,14 @@ class Tensor
void* data,
uint32_t elementTotalCount,
uint32_t elementMemorySize,
const TensorDataTypes& dataType = TensorDataTypes::eFloat,
const TensorDataTypes& dataType,
const TensorTypes& tensorType = TensorTypes::eDevice);
/**
* Destructor which is in charge of freeing vulkan resources unless they
* have been provided externally.
*/
~Tensor();
virtual ~Tensor();
/**
* Returns the size/magnitude of the Tensor, which will be the total number
@ -68,12 +68,17 @@ class Tensor
*/
// TODO: move to cpp
virtual uint32_t size() {
return this->mElementMemorySize;
return this->mSize;
}
// TODO: move to cpp
virtual uint32_t dataTypeMemorySize() {
return this->mDataTypeMemorySize;
}
// TODO: move to cpp
virtual uint32_t memorySize() {
return this->mSize * this->mElementMemorySize;
return this->mSize * this->mDataTypeMemorySize;
}
/**
@ -98,10 +103,24 @@ class Tensor
// TODO: Decide whether this is one we prefer to have also overriden in the underlying tensorView
// TODO: move to cpp
void getRawData(void* data) {
virtual void getRawData(void* data) {
this->rawMapDataFromHostMemory(data);
}
/**
* Sets / resets the vector data of the tensor. This function does not
* perform any copies into GPU memory and is only performed on the host.
*/
virtual void setRawData(void* data, uint32_t elementTotalCount, uint32_t elementMemorySize) {
if (elementTotalCount * elementMemorySize != this->memorySize()) {
throw std::runtime_error(
"Kompute Tensor Cannot set data of different sizes");
}
this->mSize = elementTotalCount;
this->mDataTypeMemorySize = elementMemorySize;
this->rawMapDataIntoHostMemory(data);
}
/**
* Function to trigger reinitialisation of the tensor buffer and memory with
* new data as well as new potential device type.
@ -111,9 +130,7 @@ class Tensor
*/
void rebuild(void* data,
uint32_t elementTotalCount,
uint32_t elementMemorySize,
const TensorDataTypes& dataType = TensorDataTypes::eFloat,
TensorTypes tensorType = TensorTypes::eDevice);
uint32_t elementMemorySize);
/**
* Destroys and frees the GPU resources which include the buffer and memory.
@ -134,19 +151,6 @@ class Tensor
*/
TensorTypes tensorType();
/**
* Sets / resets the vector data of the tensor. This function does not
* perform any copies into GPU memory and is only performed on the host.
*/
void setRawData(void* data, uint32_t elementTotalCount, uint32_t elementMemorySize) {
if (elementTotalCount * elementMemorySize != this->memorySize()) {
throw std::runtime_error(
"Kompute Tensor Cannot set data of different sizes");
}
this->mSize = elementTotalCount;
this->mElementMemorySize = elementMemorySize;
this->rawMapDataIntoHostMemory(data);
}
/**
* Records a copy from the memory of the tensor provided to the current
@ -211,46 +215,7 @@ class Tensor
*/
vk::DescriptorBufferInfo constructDescriptorBufferInfo();
private:
// -------------- NEVER OWNED RESOURCES
std::shared_ptr<vk::PhysicalDevice> mPhysicalDevice;
std::shared_ptr<vk::Device> mDevice;
// -------------- OPTIONALLY OWNED RESOURCES
std::shared_ptr<vk::Buffer> mPrimaryBuffer;
bool mFreePrimaryBuffer = false;
std::shared_ptr<vk::Buffer> mStagingBuffer;
bool mFreeStagingBuffer = false;
std::shared_ptr<vk::DeviceMemory> mPrimaryMemory;
bool mFreePrimaryMemory = false;
std::shared_ptr<vk::DeviceMemory> mStagingMemory;
bool mFreeStagingMemory = false;
// -------------- ALWAYS OWNED RESOURCES
TensorTypes mTensorType;
TensorDataTypes mDataType;
uint32_t mSize;
uint32_t mElementMemorySize;
void allocateMemoryCreateGPUResources(); // Creates the vulkan buffer
void createBuffer(std::shared_ptr<vk::Buffer> buffer,
vk::BufferUsageFlags bufferUsageFlags);
void allocateBindMemory(std::shared_ptr<vk::Buffer> buffer,
std::shared_ptr<vk::DeviceMemory> memory,
vk::MemoryPropertyFlags memoryPropertyFlags);
void 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);
// Private util functions
vk::BufferUsageFlags getPrimaryBufferUsageFlags();
vk::MemoryPropertyFlags getPrimaryMemoryPropertyFlags();
vk::BufferUsageFlags getStagingBufferUsageFlags();
vk::MemoryPropertyFlags getStagingMemoryPropertyFlags();
protected:
void rawMapDataFromHostMemory(void* data) {
KP_LOG_DEBUG("Kompute Tensor mapping data from host buffer");
@ -300,6 +265,46 @@ class Tensor
this->mDevice->flushMappedMemoryRanges(1, &mappedRange);
this->mDevice->unmapMemory(*hostVisibleMemory);
}
private:
// -------------- NEVER OWNED RESOURCES
std::shared_ptr<vk::PhysicalDevice> mPhysicalDevice;
std::shared_ptr<vk::Device> mDevice;
// -------------- OPTIONALLY OWNED RESOURCES
std::shared_ptr<vk::Buffer> mPrimaryBuffer;
bool mFreePrimaryBuffer = false;
std::shared_ptr<vk::Buffer> mStagingBuffer;
bool mFreeStagingBuffer = false;
std::shared_ptr<vk::DeviceMemory> mPrimaryMemory;
bool mFreePrimaryMemory = false;
std::shared_ptr<vk::DeviceMemory> mStagingMemory;
bool mFreeStagingMemory = false;
// -------------- ALWAYS OWNED RESOURCES
TensorTypes mTensorType;
TensorDataTypes mDataType;
uint32_t mSize;
uint32_t mDataTypeMemorySize;
void allocateMemoryCreateGPUResources(); // Creates the vulkan buffer
void createBuffer(std::shared_ptr<vk::Buffer> buffer,
vk::BufferUsageFlags bufferUsageFlags);
void allocateBindMemory(std::shared_ptr<vk::Buffer> buffer,
std::shared_ptr<vk::DeviceMemory> memory,
vk::MemoryPropertyFlags memoryPropertyFlags);
void 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);
// Private util functions
vk::BufferUsageFlags getPrimaryBufferUsageFlags();
vk::MemoryPropertyFlags getPrimaryMemoryPropertyFlags();
vk::BufferUsageFlags getStagingBufferUsageFlags();
vk::MemoryPropertyFlags getStagingMemoryPropertyFlags();
};
// TODO: Limit T to be only float, bool, double, etc
@ -310,15 +315,21 @@ class TensorView: public Tensor
TensorView(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
std::shared_ptr<vk::Device> device,
const std::vector<T>& data,
const TensorTypes& tensorType = TensorTypes::eDevice);
const TensorTypes& tensorType = TensorTypes::eDevice)
: Tensor(physicalDevice, device, (void*)data.data(), data.size(), sizeof(T), this->dataType())
{
~TensorView();
}
~TensorView() {
}
void rebuild(const std::vector<T>& data,
TensorTypes tensorType = TensorTypes::eDevice) {
this->mData = data;
Tensor::rebuild(data.data(), data.size(), sizeof(T), this->dataType(), tensorType);
Tensor::rebuild(data.data(), data.size(), sizeof(T));
}
std::vector<T>& data() {
@ -338,17 +349,25 @@ class TensorView: public Tensor
this->mData = data;
this->setRawData(this->mData.data(), this->mData.size(), sizeof(T), this->dataType());
Tensor::setRawData(this->mData.data(), this->mData.size(), sizeof(T));
}
void setRawData(void* data, uint32_t elementTotalCount, uint32_t elementMemorySize) override
{
assert(elementMemorySize == sizeof(T));
this->mData = { (T*)data, ((T*)data) + elementTotalCount };
Tensor::setRawData(this->mData.data(), this->mData.size(), sizeof(T));
}
TensorDataTypes dataType() override;
uint32_t size() override {
return this->mData->size();
return this->mData.size();
}
uint32_t memorySize() override {
return this->mData->size() * sizeof(T);
return this->mData.size() * sizeof(T);
}
/**
@ -376,34 +395,4 @@ class TensorView: public Tensor
};
template<>
Tensor::TensorDataTypes
TensorView<bool>::dataType() {
return Tensor::TensorDataTypes::eBool;
}
template<>
Tensor::TensorDataTypes
TensorView<int32_t>::dataType() {
return Tensor::TensorDataTypes::eInt;
}
template<>
Tensor::TensorDataTypes
TensorView<uint32_t>::dataType() {
return Tensor::TensorDataTypes::eUnsignedInt;
}
template<>
Tensor::TensorDataTypes
TensorView<float>::dataType() {
return Tensor::TensorDataTypes::eFloat;
}
template<>
Tensor::TensorDataTypes
TensorView<double>::dataType() {
return Tensor::TensorDataTypes::eDouble;
}
} // End namespace kp