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

Renamed tensorcreate and started adding optensorsyncdevice

Browse source
This commit is contained in:
Alejandro Saucedo 2020-09-06 11:38:08 +01:00
parent 236c349aa0
commit ec89fc6d56
17 changed files with 318 additions and 71 deletions
Download patch file Download diff file Expand all files Collapse all files

2
single_include/AggregateHeaders.cpp
View file

@ -6,7 +6,7 @@
#include "kompute/operations/OpAlgoBase.hpp"
#include "kompute/operations/OpAlgoLhsRhsOut.hpp"
#include "kompute/operations/OpMult.hpp"
#include "kompute/operations/OpCreateTensor.hpp"
#include "kompute/operations/OpTensorCreate.hpp"
#include "kompute/operations/OpTensorCopy.hpp"
#include "kompute/Algorithm.hpp"
#include "kompute/Tensor.hpp"

14
single_include/kompute/Kompute.hpp
View file

@ -1509,10 +1509,10 @@ namespace kp {
Operation that creates tensor and manages the memory of the components
created
*/
class OpCreateTensor : public OpBase
class OpTensorCreate : public OpBase
{
public:
OpCreateTensor();
OpTensorCreate();
/**
* Default constructor with parameters that provides the bare minimum
@ -1525,7 +1525,7 @@ class OpCreateTensor : public OpBase
* @param tensors Tensors that will be used to create in operation.
* @param freeTensors Whether operation manages the memory of the Tensors
*/
OpCreateTensor(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
OpTensorCreate(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
std::shared_ptr<vk::Device> device,
std::shared_ptr<vk::CommandBuffer> commandBuffer,
std::vector<std::shared_ptr<Tensor>> tensors);
@ -1534,7 +1534,7 @@ class OpCreateTensor : public OpBase
* Default destructor which in this case expects the parent class to free
* the tensors
*/
~OpCreateTensor() override;
~OpTensorCreate() override;
/**
* In charge of initialising the primary Tensor as well as the staging
@ -1544,8 +1544,10 @@ class OpCreateTensor : public OpBase
void init() override;
/**
* Records the copy command into the GPU memory from the staging or host
* memory depending on the type of tensor.
* Record runs the core actions to create the tensors. For device tensors
* it records a copyCommand to move the data from the staging tensor to the
* device tensor. For staging tensors it performs a mapDataIntoHostMemory
* which would perform immediately as opposed to on sequence eval/submission.
*/
void record() override;