thek0tyara/llama-cpp-turboquant
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 6

Amended workgroup to std::array

Browse source
This commit is contained in:
Alejandro Saucedo 2021-02-14 07:18:33 +00:00
parent 7a71ff0751
commit c8370e0a3a
5 changed files with 23 additions and 28 deletions
Download patch file Download diff file Expand all files Collapse all files

26
src/OpAlgoBase.cpp
View file

@ -13,7 +13,7 @@ OpAlgoBase::OpAlgoBase(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,
KomputeWorkgroup komputeWorkgroup,
const std::array<uint32_t, 3>& komputeWorkgroup,
const std::vector<float>& specializationConstants)
: OpBase(physicalDevice, device, commandBuffer, tensors)
{
@ -22,21 +22,21 @@ OpAlgoBase::OpAlgoBase(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
// 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 (komputeWorkgroup.x > 0) {
if (komputeWorkgroup[0] > 0) {
// If at least the x value is provided we use mainly the parameters
// provided
this->mKomputeWorkgroup = {
komputeWorkgroup.x,
komputeWorkgroup.y > 0 ? komputeWorkgroup.y : 1,
komputeWorkgroup.z > 0 ? komputeWorkgroup.z : 1
komputeWorkgroup[0],
komputeWorkgroup[1] > 0 ? komputeWorkgroup[1] : 1,
komputeWorkgroup[2] > 0 ? komputeWorkgroup[2] : 1
};
} else {
this->mKomputeWorkgroup = { tensors[0]->size(), 1, 1 };
}
SPDLOG_INFO("Kompute OpAlgoBase dispatch size X: {}, Y: {}, Z: {}",
this->mKomputeWorkgroup.x,
this->mKomputeWorkgroup.y,
this->mKomputeWorkgroup.z);
this->mKomputeWorkgroup[0],
this->mKomputeWorkgroup[1],
this->mKomputeWorkgroup[2]);
this->mAlgorithm = std::make_shared<Algorithm>(device, commandBuffer, specializationConstants);
}
@ -46,7 +46,7 @@ OpAlgoBase::OpAlgoBase(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
std::shared_ptr<vk::CommandBuffer> commandBuffer,
std::vector<std::shared_ptr<Tensor>>& tensors,
std::string shaderFilePath,
KomputeWorkgroup komputeWorkgroup,
const std::array<uint32_t, 3>& komputeWorkgroup,
const std::vector<float>& specializationConstants)
: OpAlgoBase(physicalDevice, device, commandBuffer, tensors, komputeWorkgroup, specializationConstants)
{
@ -62,7 +62,7 @@ OpAlgoBase::OpAlgoBase(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
std::shared_ptr<vk::CommandBuffer> commandBuffer,
std::vector<std::shared_ptr<Tensor>>& tensors,
const std::vector<char>& shaderDataRaw,
KomputeWorkgroup komputeWorkgroup,
const std::array<uint32_t, 3>& komputeWorkgroup,
const std::vector<float>& specializationConstants)
: OpAlgoBase(physicalDevice, device, commandBuffer, tensors, komputeWorkgroup, specializationConstants)
{
@ -120,9 +120,9 @@ OpAlgoBase::record()
vk::PipelineStageFlagBits::eComputeShader);
}
this->mAlgorithm->recordDispatch(this->mKomputeWorkgroup.x,
this->mKomputeWorkgroup.y,
this->mKomputeWorkgroup.z);
this->mAlgorithm->recordDispatch(this->mKomputeWorkgroup[0],
this->mKomputeWorkgroup[1],
this->mKomputeWorkgroup[2]);
}
void

8
src/OpAlgoLhsRhsOut.cpp
View file

@ -14,7 +14,7 @@ OpAlgoLhsRhsOut::OpAlgoLhsRhsOut(
std::shared_ptr<vk::Device> device,
std::shared_ptr<vk::CommandBuffer> commandBuffer,
std::vector<std::shared_ptr<Tensor>> tensors,
KomputeWorkgroup komputeWorkgroup)
const std::array<uint32_t, 3>& komputeWorkgroup)
// The inheritance is initialised with the copyOutputData to false given that
// this depencendant class handles the transfer of data via staging buffers in
// a granular way.
@ -93,9 +93,9 @@ OpAlgoLhsRhsOut::record()
vk::PipelineStageFlagBits::eHost,
vk::PipelineStageFlagBits::eComputeShader);
this->mAlgorithm->recordDispatch(this->mKomputeWorkgroup.x,
this->mKomputeWorkgroup.y,
this->mKomputeWorkgroup.z);
this->mAlgorithm->recordDispatch(this->mKomputeWorkgroup[0],
this->mKomputeWorkgroup[1],
this->mKomputeWorkgroup[2]);
// Barrier to ensure the shader code is executed before buffer read
this->mTensorOutput->recordBufferMemoryBarrier(

13
src/include/kompute/operations/OpAlgoBase.hpp
View file

@ -22,11 +22,6 @@ namespace kp {
class OpAlgoBase : public OpBase
{
public:
struct KomputeWorkgroup {
uint32_t x;
uint32_t y;
uint32_t z;
};
/**
* Base constructor, should not be used unless explicitly intended.
@ -49,7 +44,7 @@ class OpAlgoBase : public OpBase
std::shared_ptr<vk::Device> device,
std::shared_ptr<vk::CommandBuffer> commandBuffer,
std::vector<std::shared_ptr<Tensor>>& tensors,
KomputeWorkgroup komputeWorkgroup = {},
const std::array<uint32_t, 3>& komputeWorkgroup = {},
const std::vector<float>& specializationConstants = {});
/**
@ -69,7 +64,7 @@ class OpAlgoBase : public OpBase
std::shared_ptr<vk::CommandBuffer> commandBuffer,
std::vector<std::shared_ptr<Tensor>>& tensors,
std::string shaderFilePath,
KomputeWorkgroup komputeWorkgroup = {},
const std::array<uint32_t, 3>& komputeWorkgroup = {},
const std::vector<float>& specializationConstants = {});
/**
@ -88,7 +83,7 @@ class OpAlgoBase : public OpBase
std::shared_ptr<vk::CommandBuffer> commandBuffer,
std::vector<std::shared_ptr<Tensor>>& tensors,
const std::vector<char>& shaderDataRaw,
KomputeWorkgroup komputeWorkgroup = {},
const std::array<uint32_t, 3>& komputeWorkgroup = {},
const std::vector<float>& specializationConstants = {});
/**
@ -137,7 +132,7 @@ class OpAlgoBase : public OpBase
// -------------- ALWAYS OWNED RESOURCES
KomputeWorkgroup mKomputeWorkgroup;
std::array<uint32_t, 3> mKomputeWorkgroup;
std::string mShaderFilePath; ///< Optional member variable which can be provided for the OpAlgoBase to find the data automatically and load for processing
std::vector<char> mShaderDataRaw; ///< Optional member variable which can be provided to contain either the raw shader content or the spirv binary content

2
src/include/kompute/operations/OpAlgoLhsRhsOut.hpp
View file

@ -40,7 +40,7 @@ class OpAlgoLhsRhsOut : public OpAlgoBase
std::shared_ptr<vk::Device> device,
std::shared_ptr<vk::CommandBuffer> commandBuffer,
std::vector<std::shared_ptr<Tensor>> tensors,
KomputeWorkgroup komputeWorkgroup = KomputeWorkgroup());
const std::array<uint32_t, 3>& komputeWorkgroup = {});
/**
* Default destructor, which is in charge of destroying the algorithm

2
src/include/kompute/operations/OpMult.hpp
View file

@ -44,7 +44,7 @@ class OpMult : public OpAlgoBase
std::shared_ptr<vk::Device> device,
std::shared_ptr<vk::CommandBuffer> commandBuffer,
std::vector<std::shared_ptr<Tensor>> tensors,
KomputeWorkgroup komputeWorkgroup = KomputeWorkgroup())
const std::array<uint32_t, 3>& komputeWorkgroup = {})
: OpAlgoBase(physicalDevice, device, commandBuffer, tensors, "", komputeWorkgroup)
{
SPDLOG_DEBUG("Kompute OpMult constructor with params");