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

294
single_include/kompute/Kompute.hpp
View file

@ -762,7 +762,7 @@ class Shader
* GLSL compiler
* @return The compiled SPIR-V binary in unsigned int32 format
*/
static std::vector<uint32_t> compileSources(
static std::vector<uint32_t> compile_sources(
const std::vector<std::string>& sources,
const std::vector<std::string>& files = {},
const std::string& entryPoint = "main",
@ -783,7 +783,7 @@ class Shader
* GLSL compiler
* @return The compiled SPIR-V binary in unsigned int32 format
*/
static std::vector<uint32_t> compileSource(
static std::vector<uint32_t> compile_source(
const std::string& source,
const std::string& entryPoint = "main",
std::vector<std::pair<std::string, std::string>> definitions = {},
@ -818,6 +818,14 @@ class Tensor
eHost = 1, ///< Type is host memory, source and destination
eStorage = 2, ///< Type is Device memory (only)
};
enum class TensorDataTypes
{
eBool = 0,
eInt = 1,
eUnsignedInt = 2,
eFloat = 3,
eDouble = 4,
};
/**
* Constructor with data provided which would be used to create the
@ -831,14 +839,78 @@ class Tensor
*/
Tensor(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
std::shared_ptr<vk::Device> device,
const std::vector<float>& data,
void* data,
uint32_t elementTotalCount,
uint32_t elementMemorySize,
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
* of elements across all dimensions
*
* @return Unsigned integer representing the total number of elements
*/
// TODO: move to cpp
virtual uint32_t size() {
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->mDataTypeMemorySize;
}
/**
* Retrieve the underlying data type of the Tensor
*
* @return Data type of tensor of type kp::Tensor::TensorDataTypes
*/
virtual TensorDataTypes dataType() {
return this->mDataType;
}
/**
* Maps data from the Host Visible GPU memory into the data vector. It
* requires the Tensor to be of staging type for it to work.
*/
virtual void mapDataFromHostMemory();
/**
* Maps data from the data vector into the Host Visible GPU memory. It
* requires the tensor to be of staging type for it to work.
*/
virtual void mapDataIntoHostMemory();
// TODO: Decide whether this is one we prefer to have also overriden in the underlying tensorView
// TODO: move to cpp
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
@ -847,8 +919,9 @@ class Tensor
* @param data Vector of data to use to initialise vector from
* @param tensorType The type to use for the tensor
*/
void rebuild(const std::vector<float>& data,
TensorTypes tensorType = TensorTypes::eDevice);
void rebuild(void* data,
uint32_t elementTotalCount,
uint32_t elementMemorySize);
/**
* Destroys and frees the GPU resources which include the buffer and memory.
@ -862,32 +935,6 @@ class Tensor
*/
bool isInit();
/**
* Returns the vector of data currently contained by the Tensor. It is
* important to ensure that there is no out-of-sync data with the GPU
* memory.
*
* @return Reference to vector of elements representing the data in the
* tensor.
*/
std::vector<float>& data();
/**
* Overrides the subscript operator to expose the underlying data's
* subscript operator which in this case would be its underlying
* vector's.
*
* @param i The index where the element will be returned from.
* @return Returns the element in the position requested.
*/
float& operator[](int index);
/**
* Returns the size/magnitude of the Tensor, which will be the total number
* of elements across all dimensions
*
* @return Unsigned integer representing the total number of elements
*/
uint32_t size();
/**
* Retrieve the tensor type of the Tensor
*
@ -895,12 +942,6 @@ 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 setData(const std::vector<float>& data);
/**
* Records a copy from the memory of the tensor provided to the current
* thensor. This is intended to pass memory into a processing, to perform
@ -963,17 +1004,57 @@ class Tensor
* @return Descriptor buffer info with own buffer
*/
vk::DescriptorBufferInfo constructDescriptorBufferInfo();
/**
* Maps data from the Host Visible GPU memory into the data vector. It
* requires the Tensor to be of staging type for it to work.
*/
void mapDataFromHostMemory();
/**
* Maps data from the data vector into the Host Visible GPU memory. It
* requires the tensor to be of staging type for it to work.
*/
void mapDataIntoHostMemory();
protected:
void rawMapDataFromHostMemory(void* data) {
KP_LOG_DEBUG("Kompute Tensor mapping data from host buffer");
std::shared_ptr<vk::DeviceMemory> hostVisibleMemory = nullptr;
if (this->mTensorType == TensorTypes::eHost) {
hostVisibleMemory = this->mPrimaryMemory;
} else if (this->mTensorType == TensorTypes::eDevice) {
hostVisibleMemory = this->mStagingMemory;
} else {
KP_LOG_WARN(
"Kompute Tensor mapping data not supported on storage tensor");
return;
}
vk::DeviceSize bufferSize = this->memorySize();
void* mapped = this->mDevice->mapMemory(
*hostVisibleMemory, 0, bufferSize, vk::MemoryMapFlags());
vk::MappedMemoryRange mappedMemoryRange(*hostVisibleMemory, 0, bufferSize);
this->mDevice->invalidateMappedMemoryRanges(mappedMemoryRange);
memcpy(data, mapped, bufferSize);
this->mDevice->unmapMemory(*hostVisibleMemory);
}
void rawMapDataIntoHostMemory(void* data) {
KP_LOG_DEBUG("Kompute Tensor mapping data into host buffer");
std::shared_ptr<vk::DeviceMemory> hostVisibleMemory = nullptr;
if (this->mTensorType == TensorTypes::eHost) {
hostVisibleMemory = this->mPrimaryMemory;
} else if (this->mTensorType == TensorTypes::eDevice) {
hostVisibleMemory = this->mStagingMemory;
} else {
KP_LOG_WARN(
"Kompute Tensor mapping data not supported on storage tensor");
return;
}
vk::DeviceSize bufferSize = this->memorySize();
void* mapped = this->mDevice->mapMemory(
*hostVisibleMemory, 0, bufferSize, vk::MemoryMapFlags());
memcpy(mapped, data, bufferSize);
vk::MappedMemoryRange mappedRange(*hostVisibleMemory, 0, bufferSize);
this->mDevice->flushMappedMemoryRanges(1, &mappedRange);
this->mDevice->unmapMemory(*hostVisibleMemory);
}
private:
// -------------- NEVER OWNED RESOURCES
std::shared_ptr<vk::PhysicalDevice> mPhysicalDevice;
@ -990,9 +1071,10 @@ class Tensor
bool mFreeStagingMemory = false;
// -------------- ALWAYS OWNED RESOURCES
std::vector<float> mData;
TensorTypes mTensorType = TensorTypes::eDevice;
TensorTypes mTensorType;
TensorDataTypes mDataType;
uint32_t mSize;
uint32_t mDataTypeMemorySize;
void allocateMemoryCreateGPUResources(); // Creates the vulkan buffer
void createBuffer(std::shared_ptr<vk::Buffer> buffer,
@ -1012,9 +1094,98 @@ class Tensor
vk::MemoryPropertyFlags getPrimaryMemoryPropertyFlags();
vk::BufferUsageFlags getStagingBufferUsageFlags();
vk::MemoryPropertyFlags getStagingMemoryPropertyFlags();
uint64_t memorySize();
};
// TODO: Limit T to be only float, bool, double, etc
template <typename T>
class TensorView: public Tensor
{
public:
TensorView(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
std::shared_ptr<vk::Device> device,
const std::vector<T>& data,
const TensorTypes& tensorType = TensorTypes::eDevice)
: Tensor(physicalDevice, device, (void*)data.data(), data.size(), sizeof(T), this->dataType())
{
}
~TensorView() {
}
void rebuild(const std::vector<T>& data,
TensorTypes tensorType = TensorTypes::eDevice) {
this->mData = data;
Tensor::rebuild(data.data(), data.size(), sizeof(T));
}
std::vector<T>& data() {
return this->mData;
}
T& operator[](int index) {
return this->mData[index];
}
void setData(const std::vector<T>& data) {
if (data.size() != this->mData.size()) {
throw std::runtime_error(
"Kompute TensorView Cannot set data of different sizes");
}
this->mData = data;
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();
}
uint32_t memorySize() override {
return this->mData.size() * sizeof(T);
}
/**
* Maps data from the Host Visible GPU memory into the data vector. It
* requires the Tensor to be of staging type for it to work.
*/
void mapDataFromHostMemory() override {
KP_LOG_DEBUG("Kompute TensorView mapDataFromHostMemory copying data");
this->rawMapDataFromHostMemory(this->mData.data());
}
/**
* Maps data from the data vector into the Host Visible GPU memory. It
* requires the tensor to be of staging type for it to work.
*/
void mapDataIntoHostMemory() override {
KP_LOG_DEBUG("Kompute TensorView mapDataIntoHostMemory copying data");
this->rawMapDataIntoHostMemory(this->mData.data());
}
private:
// -------------- ALWAYS OWNED RESOURCES
std::vector<T> mData;
};
} // End namespace kp
namespace kp {
@ -1883,9 +2054,22 @@ 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