llama-cpp-turboquant/src/include/kompute/OpMult.hpp

238 lines
7.4 KiB
C++

#pragma once
#include <fstream>
#include <vulkan/vulkan.h>
#include <vulkan/vulkan.hpp>
// SPDLOG_ACTIVE_LEVEL must be defined before spdlog.h import
#if DEBUG
#define SPDLOG_ACTIVE_LEVEL SPDLOG_LEVEL_DEBUG
#endif
#include <spdlog/spdlog.h>
#include "kompute/shaders/shaderopmult.hpp"
#include "kompute/Algorithm.hpp"
#include "kompute/Tensor.hpp"
#include "kompute/OpBase.hpp"
namespace kp {
template<uint32_t tX = 0, uint32_t tY = 0, uint32_t tZ = 0>
class OpMult : public OpBase
{
public:
OpMult();
OpMult(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
std::shared_ptr<vk::Device> device,
std::shared_ptr<vk::CommandBuffer> commandBuffer);
~OpMult();
void init(std::vector<std::shared_ptr<Tensor>> tensors) override;
void record() override;
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");
}
// TODO: Remove physicalDevice from main initialiser
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)
: OpBase(physicalDevice, device, commandBuffer)
{
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(std::vector<std::shared_ptr<Tensor>> tensors)
{
SPDLOG_DEBUG("Kompute OpMult init called");
if (tensors.size() < 3) {
throw std::runtime_error(
"Kompute OpMult called with less than 1 tensor");
} else if (tensors.size() > 3) {
spdlog::warn("Kompute OpMult called with more than 3 tensors");
}
this->mTensorLHS = tensors[0];
this->mTensorRHS = tensors[1];
this->mTensorOutput = tensors[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
// 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
this->mAlgorithm->init(shaderFileData, tensors);
}
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