Moved ops files to operations subfolder
This commit is contained in:
Alejandro Saucedo
parent
6fb99c089b
commit
99daaf16cb
7 changed files with 8 additions and 9 deletions
|
|
@ -1,282 +0,0 @@
|
|||
#pragma once
|
||||
|
||||
#include <fstream>
|
||||
|
||||
#include "kompute/Core.hpp"
|
||||
|
||||
#include "kompute/shaders/shaderopmult.hpp"
|
||||
|
||||
#include "kompute/Algorithm.hpp"
|
||||
#include "kompute/Tensor.hpp"
|
||||
|
||||
#include "kompute/OpBase.hpp"
|
||||
|
||||
namespace kp {
|
||||
|
||||
/**
|
||||
* Operation that performs multiplication on two tensors and outpus on third
|
||||
* tensor. The template parameters specify the processing GPU layout number of
|
||||
* iterations for each x, y, z parameter. More specifically, this will be the
|
||||
* input to ".dispatch(uint32_t tX, uint32_t tY, uint32_t, tZ)"
|
||||
*/
|
||||
template<uint32_t tX = 0, uint32_t tY = 0, uint32_t tZ = 0>
|
||||
class OpMult : public OpBase
|
||||
{
|
||||
public:
|
||||
/**
|
||||
* Base constructor, should not be used unless explicitly intended.
|
||||
*/
|
||||
OpMult();
|
||||
|
||||
/**
|
||||
* Default constructor with parameters that provides the bare minimum
|
||||
* requirements for the operations to be able to create and manage their
|
||||
* sub-components.
|
||||
*
|
||||
* @param physicalDevice Vulkan physical device used to find device queues
|
||||
* @param device Vulkan logical device for passing to Algorithm
|
||||
* @param commandBuffer Vulkan Command Buffer to record commands into
|
||||
* @param tensors Tensors that are to be used in this operation
|
||||
* @param freeTensors Whether operation manages the memory of the Tensors
|
||||
*/
|
||||
OpMult(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,
|
||||
bool freeTensors = false);
|
||||
|
||||
/**
|
||||
* Default destructor, which is in charge of destroying the algorithm
|
||||
* components but does not destroy the underlying tensors
|
||||
*/
|
||||
~OpMult();
|
||||
|
||||
/**
|
||||
* The init function is responsible for ensuring that all of the tensors
|
||||
* provided are aligned with requirements such as LHS, RHS and Output
|
||||
* tensors, and creates the algorithm component which processes the
|
||||
* computation.
|
||||
*/
|
||||
void init() override;
|
||||
|
||||
/**
|
||||
* This records the commands that are to be sent to the GPU. This includes
|
||||
* the barriers that ensure the memory has been copied before going in and
|
||||
* out of the shader, as well as the dispatch operation that sends the
|
||||
* shader processing to the gpu. This function also records the GPU memory
|
||||
* copy of the output data for the staging bufffer so it can be read by the
|
||||
* host.
|
||||
*/
|
||||
void record() override;
|
||||
|
||||
/**
|
||||
* Executes after the recorded commands are submitted, and performs a copy
|
||||
* of the GPU Device memory into the staging buffer so the output data can
|
||||
* be retrieved.
|
||||
*/
|
||||
void postSubmit() override;
|
||||
|
||||
private:
|
||||
// Always owned resources
|
||||
std::shared_ptr<Tensor> mTensorOutputStaging;
|
||||
|
||||
// Optionally owned resources
|
||||
std::shared_ptr<Algorithm> mAlgorithm;
|
||||
bool mFreeAlgorithm = false;
|
||||
|
||||
// Never owned resources
|
||||
std::shared_ptr<Tensor> mTensorLHS;
|
||||
std::shared_ptr<Tensor> mTensorRHS;
|
||||
std::shared_ptr<Tensor> mTensorOutput;
|
||||
|
||||
uint32_t mX;
|
||||
uint32_t mY;
|
||||
uint32_t mZ;
|
||||
};
|
||||
|
||||
} // End namespace kp
|
||||
|
||||
// Including implemenation for template class
|
||||
#ifndef OPMULT_CPP
|
||||
#define OPMULT_CPP
|
||||
|
||||
namespace kp {
|
||||
|
||||
template<uint32_t tX, uint32_t tY, uint32_t tZ>
|
||||
OpMult<tX, tY, tZ>::OpMult()
|
||||
{
|
||||
SPDLOG_DEBUG("Kompute OpMult constructor base");
|
||||
}
|
||||
|
||||
template<uint32_t tX, uint32_t tY, uint32_t tZ>
|
||||
OpMult<tX, tY, tZ>::OpMult(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,
|
||||
bool freeTensors)
|
||||
: OpBase(physicalDevice, device, commandBuffer, tensors, freeTensors)
|
||||
{
|
||||
SPDLOG_DEBUG("Kompute OpMult constructor with params");
|
||||
|
||||
this->mAlgorithm = std::make_shared<Algorithm>(device, commandBuffer);
|
||||
}
|
||||
|
||||
template<uint32_t tX, uint32_t tY, uint32_t tZ>
|
||||
OpMult<tX, tY, tZ>::~OpMult()
|
||||
{
|
||||
SPDLOG_DEBUG("Kompute OpMult destructor started");
|
||||
}
|
||||
|
||||
template<uint32_t tX, uint32_t tY, uint32_t tZ>
|
||||
void
|
||||
OpMult<tX, tY, tZ>::init()
|
||||
{
|
||||
SPDLOG_DEBUG("Kompute OpMult init called");
|
||||
|
||||
if (this->mTensors.size() < 3) {
|
||||
throw std::runtime_error(
|
||||
"Kompute OpMult called with less than 1 tensor");
|
||||
} else if (this->mTensors.size() > 3) {
|
||||
spdlog::warn("Kompute OpMult called with more than 3 this->mTensors");
|
||||
}
|
||||
|
||||
this->mTensorLHS = this->mTensors[0];
|
||||
this->mTensorRHS = this->mTensors[1];
|
||||
this->mTensorOutput = this->mTensors[2];
|
||||
|
||||
// The dispatch size is set up based on either explicitly provided template
|
||||
// parameters or by default it would take the shape and size of the tensors
|
||||
if (tX > 0) {
|
||||
// If at least the x value is provided we use mainly the parameters
|
||||
// provided
|
||||
this->mX = tX;
|
||||
this->mY = tY > 0 ? tY : 1;
|
||||
this->mZ = tZ > 0 ? tZ : 1;
|
||||
} else {
|
||||
// TODO: Fully support the full size dispatch using size for the shape
|
||||
this->mX = this->mTensorLHS->size();
|
||||
this->mY = 1;
|
||||
this->mZ = 1;
|
||||
}
|
||||
spdlog::info("Kompute OpMult dispatch size X: {}, Y: {}, Z: {}",
|
||||
this->mX,
|
||||
this->mY,
|
||||
this->mZ);
|
||||
|
||||
// TODO: Explore adding a validate function
|
||||
if (!(this->mTensorLHS->isInit() && this->mTensorRHS->isInit() &&
|
||||
this->mTensorOutput->isInit())) {
|
||||
throw std::runtime_error(
|
||||
"Kompute OpMult all tensor parameters must be initialised. LHS: " +
|
||||
std::to_string(this->mTensorLHS->isInit()) +
|
||||
" RHS: " + std::to_string(this->mTensorRHS->isInit()) +
|
||||
" Output: " + std::to_string(this->mTensorOutput->isInit()));
|
||||
}
|
||||
|
||||
// TODO: Explore use-cases where tensors shouldn't be the same size, and how
|
||||
// to deal with those situations
|
||||
if (!(this->mTensorLHS->size() == this->mTensorRHS->size() &&
|
||||
this->mTensorRHS->size() == this->mTensorOutput->size())) {
|
||||
throw std::runtime_error(
|
||||
"Kompute OpMult all tensor parameters must be the same size LHS: " +
|
||||
std::to_string(this->mTensorLHS->size()) +
|
||||
" RHS: " + std::to_string(this->mTensorRHS->size()) +
|
||||
" Output: " + std::to_string(this->mTensorOutput->size()));
|
||||
}
|
||||
|
||||
this->mTensorOutputStaging = std::make_shared<Tensor>(
|
||||
this->mTensorOutput->data(), Tensor::TensorTypes::eStaging);
|
||||
|
||||
this->mTensorOutputStaging->init(
|
||||
this->mPhysicalDevice, this->mDevice, this->mCommandBuffer);
|
||||
|
||||
#if RELEASE
|
||||
std::vector<char> shaderFileData(
|
||||
shader_data::shaders_glsl_opmult_comp_spv,
|
||||
shader_data::shaders_glsl_opmult_comp_spv +
|
||||
kp::shader_data::shaders_glsl_opmult_comp_spv_len);
|
||||
#else
|
||||
SPDLOG_DEBUG(
|
||||
"Kompute OpMult Running debug loading shaders directly from spirv file");
|
||||
|
||||
// TODO: Move to utility function
|
||||
std::string shaderFilePath = "shaders/glsl/opmult.comp.spv";
|
||||
std::ifstream fileStream(shaderFilePath,
|
||||
std::ios::binary | std::ios::in | std::ios::ate);
|
||||
|
||||
size_t shaderFileSize = fileStream.tellg();
|
||||
fileStream.seekg(0, std::ios::beg);
|
||||
char* shaderDataRaw = new char[shaderFileSize];
|
||||
fileStream.read(shaderDataRaw, shaderFileSize);
|
||||
fileStream.close();
|
||||
|
||||
std::vector<char> shaderFileData(shaderDataRaw,
|
||||
shaderDataRaw + shaderFileSize);
|
||||
#endif
|
||||
|
||||
SPDLOG_DEBUG("Kompute OpMult Initialising algorithm component");
|
||||
|
||||
this->mAlgorithm->init(shaderFileData, this->mTensors);
|
||||
}
|
||||
|
||||
template<uint32_t tX, uint32_t tY, uint32_t tZ>
|
||||
void
|
||||
OpMult<tX, tY, tZ>::record()
|
||||
{
|
||||
SPDLOG_DEBUG("Kompute OpMult record called");
|
||||
|
||||
// Barrier to ensure the data is finished writing to buffer memory
|
||||
this->mTensorLHS->recordBufferMemoryBarrier(
|
||||
vk::AccessFlagBits::eHostWrite,
|
||||
vk::AccessFlagBits::eShaderRead,
|
||||
vk::PipelineStageFlagBits::eHost,
|
||||
vk::PipelineStageFlagBits::eComputeShader);
|
||||
this->mTensorRHS->recordBufferMemoryBarrier(
|
||||
vk::AccessFlagBits::eHostWrite,
|
||||
vk::AccessFlagBits::eShaderRead,
|
||||
vk::PipelineStageFlagBits::eHost,
|
||||
vk::PipelineStageFlagBits::eComputeShader);
|
||||
|
||||
this->mAlgorithm->recordDispatch(this->mX, this->mY, this->mZ);
|
||||
|
||||
// Barrier to ensure the shader code is executed before buffer read
|
||||
this->mTensorLHS->recordBufferMemoryBarrier(
|
||||
vk::AccessFlagBits::eShaderWrite,
|
||||
vk::AccessFlagBits::eTransferRead,
|
||||
vk::PipelineStageFlagBits::eComputeShader,
|
||||
vk::PipelineStageFlagBits::eTransfer);
|
||||
this->mTensorOutput->recordBufferMemoryBarrier(
|
||||
vk::AccessFlagBits::eShaderWrite,
|
||||
vk::AccessFlagBits::eTransferRead,
|
||||
vk::PipelineStageFlagBits::eComputeShader,
|
||||
vk::PipelineStageFlagBits::eTransfer);
|
||||
|
||||
this->mTensorOutputStaging->recordCopyFrom(this->mTensorOutput);
|
||||
|
||||
// Buffer to ensure wait until data is copied to staging buffer
|
||||
this->mTensorLHS->recordBufferMemoryBarrier(
|
||||
vk::AccessFlagBits::eTransferWrite,
|
||||
vk::AccessFlagBits::eHostRead,
|
||||
vk::PipelineStageFlagBits::eTransfer,
|
||||
vk::PipelineStageFlagBits::eHost);
|
||||
this->mTensorOutput->recordBufferMemoryBarrier(
|
||||
vk::AccessFlagBits::eTransferWrite,
|
||||
vk::AccessFlagBits::eHostRead,
|
||||
vk::PipelineStageFlagBits::eTransfer,
|
||||
vk::PipelineStageFlagBits::eHost);
|
||||
}
|
||||
|
||||
template<uint32_t tX, uint32_t tY, uint32_t tZ>
|
||||
void
|
||||
OpMult<tX, tY, tZ>::postSubmit()
|
||||
{
|
||||
SPDLOG_DEBUG("Kompute OpMult postSubmit called");
|
||||
|
||||
this->mTensorOutputStaging->mapDataFromHostMemory();
|
||||
|
||||
this->mTensorOutput->setData(this->mTensorOutputStaging->data());
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
#endif // #ifndef OPMULT_CPP
|
||||
Loading…
Add table
Add a link
Reference in a new issue