llama-cpp-turboquant/src/OpTensorSyncDevice.cpp
2020-09-06 12:52:45 +01:00

91 lines
2.8 KiB
C++

#include "kompute/Tensor.hpp"
#include "kompute/operations/OpTensorSyncDevice.hpp"
namespace kp {
OpTensorSyncDevice::OpTensorSyncDevice()
{
SPDLOG_DEBUG("Kompute OpTensorSyncDevice constructor base");
}
OpTensorSyncDevice::OpTensorSyncDevice(
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)
: OpBase(physicalDevice, device, commandBuffer, tensors, false)
{
SPDLOG_DEBUG("Kompute OpTensorSyncDevice constructor with params");
}
OpTensorSyncDevice::~OpTensorSyncDevice()
{
SPDLOG_DEBUG("Kompute OpTensorSyncDevice destructor started");
}
void
OpTensorSyncDevice::init()
{
SPDLOG_DEBUG("Kompute OpTensorSyncDevice init called");
if (this->mTensors.size() < 1) {
throw std::runtime_error(
"Kompute OpTensorSyncDevice called with less than 1 tensor");
}
for (std::shared_ptr<Tensor> tensor: this->mTensors) {
if (tensor->isInit()) {
throw std::runtime_error("Kompute OpTensorSyncDevice: Tensor has already been initialized");
}
if (tensor->tensorType() == Tensor::TensorTypes::eStorage) {
throw std::runtime_error("Kompute OpTensorSyncLocal tensor parameter is of type TensorTypes::eStorage and hence cannot be used to receive or pass data.");
}
if (tensor->tensorType() == Tensor::TensorTypes::eDevice) {
std::shared_ptr<Tensor> stagingTensor = std::make_shared<Tensor>(
tensor->data(), Tensor::TensorTypes::eStaging);
stagingTensor->init(
this->mPhysicalDevice, this->mDevice);
stagingTensor->mapDataIntoHostMemory();
this->mStagingTensors.push_back(stagingTensor);
} else {
tensor->mapDataIntoHostMemory();
// We push a nullptr when no staging tensor is needed to match
// index number in array to have one to one mapping with tensors
this->mStagingTensors.push_back(nullptr);
}
}
}
void
OpTensorSyncDevice::record()
{
SPDLOG_DEBUG("Kompute OpTensorSyncDevice record called");
for (size_t i = 0; i < this->mTensors.size(); i++) {
if (this->mTensors[i]->tensorType() == Tensor::TensorTypes::eDevice) {
this->mTensors[i]->recordCopyFrom(this->mCommandBuffer, this->mStagingTensors[i], false);
}
}
}
void
OpTensorSyncDevice::postSubmit()
{
SPDLOG_DEBUG("Kompute OpTensorSyncDevice postSubmit called");
// Remove all staging tensors as they are not required after operation
SPDLOG_DEBUG("Kompute OpTensorSyncDevice destroying staging tensors");
// TODO: This would cause issues if there is no CPU barrier
this->mStagingTensors.clear();
}
}