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

Updated tensor to copy to memory explicitly, assessing why copy doesn't last through operations

Browse source
This commit is contained in:
Alejandro Saucedo 2020-08-22 18:08:56 +01:00
parent 6f0203b863
commit e047aa3b43
10 changed files with 65 additions and 79 deletions
Download patch file Download diff file Expand all files Collapse all files

45
src/OpMult.cpp
View file

@ -96,8 +96,7 @@ OpMult<tX, tY, tZ>::init(std::vector<std::shared_ptr<Tensor>> tensors)
this->mTensorOutputStaging->init(this->mPhysicalDevice,
this->mDevice,
this->mCommandBuffer,
this->mTensorOutput->data());
this->mCommandBuffer);
// TODO: Make this path configurable
this->mAlgorithm->init("shaders/glsl/opmult.comp.spv", tensors);
@ -110,30 +109,20 @@ 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->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->mTensorRHS->recordBufferMemoryBarrier(
// vk::AccessFlagBits::eShaderWrite,
// vk::AccessFlagBits::eTransferRead,
// vk::PipelineStageFlagBits::eComputeShader,
// vk::PipelineStageFlagBits::eTransfer);
this->mTensorOutput->recordBufferMemoryBarrier(
vk::AccessFlagBits::eShaderWrite,
vk::AccessFlagBits::eTransferRead,
@ -148,16 +137,6 @@ OpMult<tX, tY, tZ>::record()
vk::AccessFlagBits::eHostRead,
vk::PipelineStageFlagBits::eTransfer,
vk::PipelineStageFlagBits::eHost);
//this->mTensorLHS->recordBufferMemoryBarrier(
// vk::AccessFlagBits::eTransferWrite,
// vk::AccessFlagBits::eHostRead,
// vk::PipelineStageFlagBits::eTransfer,
// vk::PipelineStageFlagBits::eHost);
//this->mTensorRHS->recordBufferMemoryBarrier(
// vk::AccessFlagBits::eTransferWrite,
// vk::AccessFlagBits::eHostRead,
// vk::PipelineStageFlagBits::eTransfer,
// vk::PipelineStageFlagBits::eHost);
}
template<uint32_t tX, uint32_t tY, uint32_t tZ>
@ -166,7 +145,7 @@ OpMult<tX, tY, tZ>::postSubmit()
{
SPDLOG_DEBUG("Kompute OpCreateTensor postSubmit called");
this->mTensorOutputStaging->copyDataFromHostBuffer();
this->mTensorOutputStaging->mapDataFromHostMemory();
this->mTensorOutput->setData(this->mTensorOutputStaging->data());
}