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

Reformatted

Browse source
This commit is contained in:
Alejandro Saucedo 2020-08-26 05:47:21 +01:00
parent c22247534a
commit 1748694b3a
8 changed files with 82 additions and 113 deletions
Download patch file Download diff file Expand all files Collapse all files

22
src/Algorithm.cpp
View file

@ -41,9 +41,9 @@ Algorithm::init(const std::vector<char>& shaderFileData,
this->createPipeline();
}
void Algorithm::createDescriptorPool() {
}
void
Algorithm::createDescriptorPool()
{}
void
Algorithm::createParameters(std::vector<std::shared_ptr<Tensor>>& tensorParams)
@ -52,9 +52,10 @@ Algorithm::createParameters(std::vector<std::shared_ptr<Tensor>>& tensorParams)
// TODO: Explore design for having multiple descriptor pool sizes
std::vector<vk::DescriptorPoolSize> descriptorPoolSizes = {
vk::DescriptorPoolSize(vk::DescriptorType::eStorageBuffer,
static_cast<uint32_t>(tensorParams.size()) // Descriptor count
)
vk::DescriptorPoolSize(
vk::DescriptorType::eStorageBuffer,
static_cast<uint32_t>(tensorParams.size()) // Descriptor count
)
};
// TODO: Explore design for having more than 1 set configurable
@ -120,11 +121,12 @@ Algorithm::createParameters(std::vector<std::shared_ptr<Tensor>>& tensorParams)
nullptr, // Descriptor image info
&descriptorBufferInfo));
this->mDevice->updateDescriptorSets(computeWriteDescriptorSets, nullptr);
this->mDevice->updateDescriptorSets(computeWriteDescriptorSets,
nullptr);
}
SPDLOG_DEBUG("Kompute Algorithm updating descriptor sets");
//this->mDevice->updateDescriptorSets(computeWriteDescriptorSets, nullptr);
// this->mDevice->updateDescriptorSets(computeWriteDescriptorSets, nullptr);
SPDLOG_DEBUG("Kompue Algorithm successfully run init");
}
@ -135,7 +137,9 @@ Algorithm::createShaderModule(const std::vector<char>& shaderFileData)
SPDLOG_DEBUG("Kompute Algorithm createShaderModule started");
vk::ShaderModuleCreateInfo shaderModuleInfo(
vk::ShaderModuleCreateFlags(), shaderFileData.size(), (uint32_t*)shaderFileData.data());
vk::ShaderModuleCreateFlags(),
shaderFileData.size(),
(uint32_t*)shaderFileData.data());
SPDLOG_DEBUG("Kompute Algorithm Creating shader module. ShaderFileSize: {}",
shaderFileData.size());

8
src/Manager.cpp
View file

@ -66,9 +66,13 @@ Manager::~Manager()
}
Sequence
Manager::constructSequence() {
Manager::constructSequence()
{
SPDLOG_DEBUG("Kompute Manager creating Sequence object");
return Sequence(this->mPhysicalDevice, this->mDevice, this->mComputeQueue, this->mComputeQueueFamilyIndex);
return Sequence(this->mPhysicalDevice,
this->mDevice,
this->mComputeQueue,
this->mComputeQueueFamilyIndex);
}
void

21
src/Tensor.cpp Executable file → Normal file
View file

@ -15,7 +15,8 @@ Tensor::Tensor()
Tensor::Tensor(std::vector<uint32_t> data, TensorTypes tensorType)
{
SPDLOG_DEBUG("Kompute Tensor constructor data: {}, and type: {}", data, tensorType);
SPDLOG_DEBUG(
"Kompute Tensor constructor data: {}, and type: {}", data, tensorType);
this->mData = data;
this->mShape = { static_cast<uint32_t>(data.size()) };
@ -60,8 +61,9 @@ Tensor::init(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
std::shared_ptr<vk::Device> device,
std::shared_ptr<vk::CommandBuffer> commandBuffer)
{
SPDLOG_DEBUG(
"Kompute Tensor running init with Vulkan params and num data elementS: {}", this->mData.size());
SPDLOG_DEBUG("Kompute Tensor running init with Vulkan params and num data "
"elementS: {}",
this->mData.size());
this->mPhysicalDevice = physicalDevice;
this->mDevice = device;
@ -178,8 +180,9 @@ Tensor::mapDataFromHostMemory()
SPDLOG_DEBUG("Kompute Tensor mapping data from host buffer");
if (this->mTensorType != TensorTypes::eStaging) {
spdlog::error("Mapping tensor data manually from DEVICE buffer instead of "
"using record GPU command with staging buffer");
spdlog::error(
"Mapping tensor data manually from DEVICE buffer instead of "
"using record GPU command with staging buffer");
return;
}
@ -198,10 +201,12 @@ Tensor::mapDataIntoHostMemory()
SPDLOG_DEBUG("Kompute Tensor local mapping tensor data to host buffer");
// TODO: Verify if there are situations where we want to copy to device memory
// TODO: Verify if there are situations where we want to copy to device
// memory
if (this->mTensorType != TensorTypes::eStaging) {
spdlog::error("Mapping tensor data manually to DEVICE memory instead of "
"using record GPU command with staging buffer");
spdlog::error(
"Mapping tensor data manually to DEVICE memory instead of "
"using record GPU command with staging buffer");
return;
}

1
src/include/kompute/Algorithm.hpp
View file

@ -52,7 +52,6 @@ class Algorithm
// Parameters
void createParameters(std::vector<std::shared_ptr<Tensor>>& tensorParams);
void createDescriptorPool();
};
} // End namespace kp

2
src/include/kompute/Core.hpp
View file

@ -8,5 +8,3 @@
#endif
#include <spdlog/spdlog.h>

85
src/include/kompute/OpMult.hpp
View file

@ -66,8 +66,8 @@ OpMult<tX, tY, tZ>::OpMult()
// 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)
std::shared_ptr<vk::Device> device,
std::shared_ptr<vk::CommandBuffer> commandBuffer)
: OpBase(physicalDevice, device, commandBuffer)
{
SPDLOG_DEBUG("Kompute OpMult constructor with params");
@ -79,7 +79,6 @@ 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>
@ -99,20 +98,24 @@ OpMult<tX, tY, tZ>::init(std::vector<std::shared_ptr<Tensor>> tensors)
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
// 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
// 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 {
} 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);
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() &&
@ -138,16 +141,17 @@ OpMult<tX, tY, tZ>::init(std::vector<std::shared_ptr<Tensor>> tensors)
this->mTensorOutputStaging = std::make_shared<Tensor>(
this->mTensorOutput->data(), Tensor::TensorTypes::eStaging);
this->mTensorOutputStaging->init(this->mPhysicalDevice,
this->mDevice,
this->mCommandBuffer);
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);
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");
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";
@ -160,7 +164,8 @@ OpMult<tX, tY, tZ>::init(std::vector<std::shared_ptr<Tensor>> tensors)
fileStream.read(shaderDataRaw, shaderFileSize);
fileStream.close();
std::vector<char> shaderFileData(shaderDataRaw, shaderDataRaw + shaderFileSize);
std::vector<char> shaderFileData(shaderDataRaw,
shaderDataRaw + shaderFileSize);
#endif
SPDLOG_DEBUG("Kompute OpMult Initialising algorithm component");
@ -176,43 +181,43 @@ OpMult<tX, tY, tZ>::record()
// 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);
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);
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);
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);
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);
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);
vk::AccessFlagBits::eTransferWrite,
vk::AccessFlagBits::eHostRead,
vk::PipelineStageFlagBits::eTransfer,
vk::PipelineStageFlagBits::eHost);
}
template<uint32_t tX, uint32_t tY, uint32_t tZ>
@ -229,5 +234,3 @@ OpMult<tX, tY, tZ>::postSubmit()
}
#endif // #ifndef OPMULT_CPP

6
src/include/kompute/Sequence.hpp
View file

@ -37,10 +37,12 @@ class Sequence
std::unique_ptr<OpBase> baseOpPtr{ baseOp };
SPDLOG_DEBUG("Kompute Sequence running init on OpBase derived class instance");
SPDLOG_DEBUG(
"Kompute Sequence running init on OpBase derived class instance");
baseOpPtr->init(tensors);
SPDLOG_DEBUG("Kompute Sequence running record on OpBase derived class instance");
SPDLOG_DEBUG(
"Kompute Sequence running record on OpBase derived class instance");
baseOpPtr->record();
mOperations.push_back(std::move(baseOpPtr));