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

Updated memory barriers to include staging buffers

Browse source
This commit is contained in:
Alejandro Saucedo 2021-03-09 08:06:27 +00:00
parent 1d1018fa0c
commit 263f392cbb
7 changed files with 230 additions and 175 deletions
Download patch file Download diff file Expand all files Collapse all files

10
src/OpTensorSyncLocal.cpp
View file

@ -30,8 +30,14 @@ OpTensorSyncLocal::record(const vk::CommandBuffer& commandBuffer)
for (size_t i = 0; i < this->mTensors.size(); i++) {
if (this->mTensors[i]->tensorType() == Tensor::TensorTypes::eDevice) {
this->mTensors[i]->recordCopyFromDeviceToStaging(commandBuffer,
true);
this->mTensors[i]->recordCopyFromDeviceToStaging(commandBuffer);
this->mTensors[i]->recordStagingBufferMemoryBarrier(commandBuffer,
vk::AccessFlagBits::eTransferWrite,
vk::AccessFlagBits::eHostRead,
vk::PipelineStageFlagBits::eTransfer,
vk::PipelineStageFlagBits::eHost);
}
}
}