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

Updated examples

Browse source
This commit is contained in:
Alejandro Saucedo 2021-02-28 17:07:17 +00:00
parent 63e220a8a4
commit 4fddf74ca7
11 changed files with 408 additions and 405 deletions
Download patch file Download diff file Expand all files Collapse all files

95
examples/android/android-simple/app/src/main/cpp/KomputeModelML.cpp
View file

@ -20,61 +20,62 @@ void KomputeModelML::train(std::vector<float> yData, std::vector<float> xIData,
uint32_t ITERATIONS = 100; uint32_t ITERATIONS = 100;
float learningRate = 0.1; float learningRate = 0.1;
std::shared_ptr<kp::Tensor> xI{ new kp::Tensor(xIData) };
std::shared_ptr<kp::Tensor> xJ{ new kp::Tensor(xJData) };
std::shared_ptr<kp::Tensor> y{ new kp::Tensor(yData) };
std::shared_ptr<kp::Tensor> wIn{ new kp::Tensor({ 0.001, 0.001 }) };
std::shared_ptr<kp::Tensor> wOutI{ new kp::Tensor(zerosData) };
std::shared_ptr<kp::Tensor> wOutJ{ new kp::Tensor(zerosData) };
std::shared_ptr<kp::Tensor> bIn{ new kp::Tensor({ 0 }) };
std::shared_ptr<kp::Tensor> bOut{ new kp::Tensor(zerosData) };
std::shared_ptr<kp::Tensor> lOut{ new kp::Tensor(zerosData) };
std::vector<std::shared_ptr<kp::Tensor>> params = { xI, xJ, y,
wIn, wOutI, wOutJ,
bIn, bOut, lOut };
{ {
kp::Manager mgr; kp::Manager mgr;
{ std::shared_ptr<kp::Tensor> xI = mgr.tensor(xIData);
mgr.rebuild(params); std::shared_ptr<kp::Tensor> xJ = mgr.tensor(xJData);
std::shared_ptr<kp::Sequence> sq = mgr.sequence(); std::shared_ptr<kp::Tensor> y = mgr.tensor(yData);
// Record op algo base std::shared_ptr<kp::Tensor> wIn = mgr.tensor({ 0.001, 0.001 });
sq->begin(); std::shared_ptr<kp::Tensor> wOutI = mgr.tensor(zerosData);
std::shared_ptr<kp::Tensor> wOutJ = mgr.tensor(zerosData);
sq->record<kp::OpTensorSyncDevice>({ wIn, bIn }); std::shared_ptr<kp::Tensor> bIn = mgr.tensor({ 0 });
std::shared_ptr<kp::Tensor> bOut = mgr.tensor(zerosData);
// Newer versions of Android are able to use shaderc to read raw string std::shared_ptr<kp::Tensor> lOut = mgr.tensor(zerosData);
sq->record<kp::OpAlgoCreate>(
params, kp::Shader::compile_source(LR_SHADER));
sq->record<kp::OpTensorSyncLocal>({ wOutI, wOutJ, bOut, lOut }); std::vector<std::shared_ptr<kp::Tensor>> params = { xI, xJ, y,
wIn, wOutI, wOutJ,
bIn, bOut, lOut };
sq->end(); std::vector<uint32_t> spirv(
(uint32_t*)kp::shader_data::shaders_glsl_logisticregression_comp_spv,
(uint32_t*)(kp::shader_data::shaders_glsl_logisticregression_comp_spv
+ kp::shader_data::shaders_glsl_logisticregression_comp_spv_len));
// Iterate across all expected iterations std::shared_ptr<kp::Algorithm> algo =
for (size_t i = 0; i < ITERATIONS; i++) { mgr.algorithm(params, spirv, kp::Workgroup({ 5 }), kp::Constants({ 5.0 }));
sq->eval(); mgr.sequence()->eval<kp::OpTensorSyncDevice>(params);
for (size_t j = 0; j < bOut->size(); j++) { std::shared_ptr<kp::Sequence> sq = mgr.sequence()
wIn->data()[0] -= learningRate * wOutI->data()[j]; ->record<kp::OpTensorSyncDevice>({ wIn, bIn })
wIn->data()[1] -= learningRate * wOutJ->data()[j]; ->record<kp::OpAlgoDispatch>(algo)
bIn->data()[0] -= learningRate * bOut->data()[j]; ->record<kp::OpTensorSyncLocal>({ wOutI, wOutJ, bOut, lOut });
}
// Iterate across all expected iterations
for (size_t i = 0; i < ITERATIONS; i++) {
sq->eval();
for (size_t j = 0; j < bOut->size(); j++) {
wIn->data()[0] -= learningRate * wOutI->data()[j];
wIn->data()[1] -= learningRate * wOutJ->data()[j];
bIn->data()[0] -= learningRate * bOut->data()[j];
} }
} }
}
this->mWeights = kp::Tensor(wIn->data()); KP_LOG_INFO("RESULT: <<<<<<<<<<<<<<<<<<<");
this->mBias = kp::Tensor(bIn->data()); KP_LOG_INFO("{}", wIn->data()[0]);
KP_LOG_INFO("{}", wIn->data()[1]);
KP_LOG_INFO("{}", bIn->data()[0]);
this->mWeights = wIn;
this->mBias = bIn;
}
} }
std::vector<float> KomputeModelML::predict(std::vector<float> xI, std::vector<float> xJ) { std::vector<float> KomputeModelML::predict(std::vector<float> xI, std::vector<float> xJ) {
@ -88,9 +89,9 @@ std::vector<float> KomputeModelML::predict(std::vector<float> xI, std::vector<fl
for (size_t i = 0; i < xI.size(); i++) { for (size_t i = 0; i < xI.size(); i++) {
float xIVal = xI[i]; float xIVal = xI[i];
float xJVal = xJ[i]; float xJVal = xJ[i];
float result = (xIVal * this->mWeights.data()[0] float result = (xIVal * this->mWeights->data()[0]
+ xJVal * this->mWeights.data()[1] + xJVal * this->mWeights->data()[1]
+ this->mBias.data()[0]); + this->mBias->data()[0]);
// Instead of using sigmoid we'll just return full numbers // Instead of using sigmoid we'll just return full numbers
float var = result > 0 ? 1 : 0; float var = result > 0 ? 1 : 0;
@ -103,13 +104,13 @@ std::vector<float> KomputeModelML::predict(std::vector<float> xI, std::vector<fl
std::vector<float> KomputeModelML::get_params() { std::vector<float> KomputeModelML::get_params() {
std::vector<float> retVector; std::vector<float> retVector;
if(this->mWeights.size() + this->mBias.size() == 0) { if(this->mWeights->size() + this->mBias->size() == 0) {
return retVector; return retVector;
} }
retVector.push_back(this->mWeights.data()[0]); retVector.push_back(this->mWeights->data()[0]);
retVector.push_back(this->mWeights.data()[1]); retVector.push_back(this->mWeights->data()[1]);
retVector.push_back(this->mBias.data()[0]); retVector.push_back(this->mBias->data()[0]);
retVector.push_back(99.0); retVector.push_back(99.0);
return retVector; return retVector;

5
examples/android/android-simple/app/src/main/cpp/KomputeModelML.hpp
View file

@ -4,6 +4,7 @@
#include <vector> #include <vector>
#include <string> #include <string>
#include <memory>
#include "kompute/Kompute.hpp" #include "kompute/Kompute.hpp"
@ -20,8 +21,8 @@ public:
std::vector<float> get_params(); std::vector<float> get_params();
private: private:
kp::Tensor mWeights; std::shared_ptr<kp::Tensor> mWeights;
kp::Tensor mBias; std::shared_ptr<kp::Tensor> mBias;
}; };

11
examples/array_multiplication/src/Main.cpp
View file

@ -37,11 +37,14 @@ int main()
} }
)"); )");
mgr.evalOpDefault<kp::OpAlgoCreate>( std::vector<std::shared_ptr<kp::Tensor>> params = { tensorInA, tensorInB, tensorOut };
{ tensorInA, tensorInB, tensorOut },
kp::Shader::compile_source(shader));
mgr.evalOpDefault<kp::OpTensorSyncLocal>({tensorOut}); std::shared_ptr<kp::Algorithm> algo = mgr.algorithm(params, kp::Shader::compile_source(shader));
mgr.sequence()
->record<kp::OpTensorSyncDevice>(params)
->record<kp::OpAlgoDispatch>(algo)
->record<kp::OpTensorSyncLocal>(params);
// prints "Output { 0 4 12 }" // prints "Output { 0 4 12 }"
std::cout<< "Output: { "; std::cout<< "Output: { ";

13
examples/godot_examples/custom_module/kompute_summator/KomputeSummatorNode.cpp
View file

@ -31,7 +31,7 @@ void KomputeSummatorNode::_init() {
std::cout << "CALLING INIT" << std::endl; std::cout << "CALLING INIT" << std::endl;
this->mPrimaryTensor = this->mManager.tensor({ 0.0 }); this->mPrimaryTensor = this->mManager.tensor({ 0.0 });
this->mSecondaryTensor = this->mManager.tensor({ 0.0 }); this->mSecondaryTensor = this->mManager.tensor({ 0.0 });
this->mSequence = this->mManager.sequence("AdditionSeq"); this->mSequence = this->mManager.sequence();
// We now record the steps in the sequence // We now record the steps in the sequence
if (std::shared_ptr<kp::Sequence> sq = this->mSequence) if (std::shared_ptr<kp::Sequence> sq = this->mSequence)
@ -51,7 +51,11 @@ void KomputeSummatorNode::_init() {
} }
)"); )");
sq->begin(); std::shared_ptr<kp::Algorithm> algo =
mgr.algorithm(
{ this->mPrimaryTensor, this->mSecondaryTensor },
kp::Shader::compile_source(shader));
// First we ensure secondary tensor loads to GPU // First we ensure secondary tensor loads to GPU
// No need to sync the primary tensor as it should not be changed // No need to sync the primary tensor as it should not be changed
@ -59,15 +63,12 @@ void KomputeSummatorNode::_init() {
{ this->mSecondaryTensor }); { this->mSecondaryTensor });
// Then we run the operation with both tensors // Then we run the operation with both tensors
sq->record<kp::OpAlgoCreate>( sq->record<kp::OpAlgoDispatch>(algo)
{ this->mPrimaryTensor, this->mSecondaryTensor },
kp::Shader::compile_source(shader));
// We map the result back to local // We map the result back to local
sq->record<kp::OpTensorSyncLocal>( sq->record<kp::OpTensorSyncLocal>(
{ this->mPrimaryTensor }); { this->mPrimaryTensor });
sq->end();
} }
else { else {
throw std::runtime_error("Sequence pointer no longer available"); throw std::runtime_error("Sequence pointer no longer available");

83
examples/godot_logistic_regression/custom_module/kompute_model_ml/KomputeModelMLNode.cpp
View file

@ -29,54 +29,41 @@ void KomputeModelMLNode::train(Array yArr, Array xIArr, Array xJArr) {
uint32_t ITERATIONS = 100; uint32_t ITERATIONS = 100;
float learningRate = 0.1; float learningRate = 0.1;
std::shared_ptr<kp::Tensor> xI{ new kp::Tensor(xIData) };
std::shared_ptr<kp::Tensor> xJ{ new kp::Tensor(xJData) };
std::shared_ptr<kp::Tensor> y{ new kp::Tensor(yData) };
std::shared_ptr<kp::Tensor> wIn{ new kp::Tensor({ 0.001, 0.001 }) };
std::shared_ptr<kp::Tensor> wOutI{ new kp::Tensor(zerosData) };
std::shared_ptr<kp::Tensor> wOutJ{ new kp::Tensor(zerosData) };
std::shared_ptr<kp::Tensor> bIn{ new kp::Tensor({ 0 }) };
std::shared_ptr<kp::Tensor> bOut{ new kp::Tensor(zerosData) };
std::shared_ptr<kp::Tensor> lOut{ new kp::Tensor(zerosData) };
std::vector<std::shared_ptr<kp::Tensor>> params = { xI, xJ, y,
wIn, wOutI, wOutJ,
bIn, bOut, lOut };
{ {
kp::Manager mgr; kp::Manager mgr;
mgr.rebuild(params); std::shared_ptr<kp::Tensor> xI = mgr.tensor(xIData);
std::shared_ptr<kp::Tensor> xJ = mgr.tensor(xJData);
std::shared_ptr<kp::Tensor> y = mgr.tensor(yData);
std::shared_ptr<kp::Tensor> wIn = mgr.tensor({ 0.001, 0.001 });
std::shared_ptr<kp::Tensor> wOutI = mgr.tensor(zerosData);
std::shared_ptr<kp::Tensor> wOutJ = mgr.tensor(zerosData);
std::shared_ptr<kp::Tensor> bIn = mgr.tensor({ 0 });
std::shared_ptr<kp::Tensor> bOut = mgr.tensor(zerosData);
std::shared_ptr<kp::Tensor> lOut = mgr.tensor(zerosData);
std::vector<std::shared_ptr<kp::Tensor>> params = { xI, xJ, y,
wIn, wOutI, wOutJ,
bIn, bOut, lOut };
{ {
std::shared_ptr<kp::Sequence> sq = mgr.sequence(); std::vector<uint32_t> spirv(
(uint32_t*)kp::shader_data::shaders_glsl_logisticregression_comp_spv,
(uint32_t*)(kp::shader_data::shaders_glsl_logisticregression_comp_spv
+ kp::shader_data::shaders_glsl_logisticregression_comp_spv_len));
// Record op algo base std::shared_ptr<kp::Algorithm> algo = mgr.algorithm(params, spirv);
sq->begin();
sq->record<kp::OpTensorSyncDevice>({ wIn, bIn }); mgr.sequence()->eval<kp::OpTensorSyncDevice>(params);
#ifdef KOMPUTE_ANDROID_SHADER_FROM_STRING std::shared_ptr<kp::Sequence> sq = mgr.sequence()
// Newer versions of Android are able to use shaderc to read raw string ->record<kp::OpTensorSyncDevice>({ wIn, bIn })
sq->record<kp::OpAlgoCreate>( ->record<kp::OpAlgoDispatch>(algo)
params, std::vector<char>(LR_SHADER.begin(), LR_SHADER.end())); ->record<kp::OpTensorSyncLocal>({ wOutI, wOutJ, bOut, lOut });
#else
// Older versions of Android require the SPIRV binary directly
sq->record<kp::OpAlgoCreate>(
params, std::vector<char>(
kp::shader_data::shaders_glsl_logisticregression_comp_spv,
kp::shader_data::shaders_glsl_logisticregression_comp_spv
+ kp::shader_data::shaders_glsl_logisticregression_comp_spv_len
));
#endif
sq->record<kp::OpTensorSyncLocal>({ wOutI, wOutJ, bOut, lOut });
sq->end();
// Iterate across all expected iterations // Iterate across all expected iterations
for (size_t i = 0; i < ITERATIONS; i++) { for (size_t i = 0; i < ITERATIONS; i++) {
@ -90,15 +77,15 @@ void KomputeModelMLNode::train(Array yArr, Array xIArr, Array xJArr) {
} }
} }
} }
KP_LOG_INFO("RESULT: <<<<<<<<<<<<<<<<<<<");
KP_LOG_INFO(wIn->data()[0]);
KP_LOG_INFO(wIn->data()[1]);
KP_LOG_INFO(bIn->data()[0]);
this->mWeights = kp::Tensor(wIn->data());
this->mBias = kp::Tensor(bIn->data());
} }
KP_LOG_INFO("RESULT: <<<<<<<<<<<<<<<<<<<");
KP_LOG_INFO(wIn->data()[0]);
KP_LOG_INFO(wIn->data()[1]);
KP_LOG_INFO(bIn->data()[0]);
this->mWeights = kp::Tensor(wIn->data());
this->mBias = kp::Tensor(bIn->data());
} }
Array KomputeModelMLNode::predict(Array xI, Array xJ) { Array KomputeModelMLNode::predict(Array xI, Array xJ) {

101
examples/godot_logistic_regression/gdnative_shared/src/KomputeModelML.cpp
View file

@ -33,54 +33,41 @@ void KomputeModelML::train(Array yArr, Array xIArr, Array xJArr) {
uint32_t ITERATIONS = 100; uint32_t ITERATIONS = 100;
float learningRate = 0.1; float learningRate = 0.1;
std::shared_ptr<kp::Tensor> xI{ new kp::Tensor(xIData) };
std::shared_ptr<kp::Tensor> xJ{ new kp::Tensor(xJData) };
std::shared_ptr<kp::Tensor> y{ new kp::Tensor(yData) };
std::shared_ptr<kp::Tensor> wIn{ new kp::Tensor({ 0.001, 0.001 }) };
std::shared_ptr<kp::Tensor> wOutI{ new kp::Tensor(zerosData) };
std::shared_ptr<kp::Tensor> wOutJ{ new kp::Tensor(zerosData) };
std::shared_ptr<kp::Tensor> bIn{ new kp::Tensor({ 0 }) };
std::shared_ptr<kp::Tensor> bOut{ new kp::Tensor(zerosData) };
std::shared_ptr<kp::Tensor> lOut{ new kp::Tensor(zerosData) };
std::vector<std::shared_ptr<kp::Tensor>> params = { xI, xJ, y,
wIn, wOutI, wOutJ,
bIn, bOut, lOut };
{ {
kp::Manager mgr; kp::Manager mgr;
std::shared_ptr<kp::Tensor> xI = mgr.tensor(xIData);
std::shared_ptr<kp::Tensor> xJ = mgr.tensor(xJData);
std::shared_ptr<kp::Tensor> y = mgr.tensor(yData);
std::shared_ptr<kp::Tensor> wIn = mgr.tensor({ 0.001, 0.001 });
std::shared_ptr<kp::Tensor> wOutI = mgr.tensor(zerosData);
std::shared_ptr<kp::Tensor> wOutJ = mgr.tensor(zerosData);
std::shared_ptr<kp::Tensor> bIn = mgr.tensor({ 0 });
std::shared_ptr<kp::Tensor> bOut = mgr.tensor(zerosData);
std::shared_ptr<kp::Tensor> lOut = mgr.tensor(zerosData);
std::vector<std::shared_ptr<kp::Tensor>> params = { xI, xJ, y,
wIn, wOutI, wOutJ,
bIn, bOut, lOut };
{ {
mgr.rebuild(params); std::vector<uint32_t> spirv(
(uint32_t*)kp::shader_data::shaders_glsl_logisticregression_comp_spv,
(uint32_t*)(kp::shader_data::shaders_glsl_logisticregression_comp_spv
+ kp::shader_data::shaders_glsl_logisticregression_comp_spv_len));
std::shared_ptr<kp::Sequence> sq = mgr.sequence(); std::shared_ptr<kp::Algorithm> algo = mgr.algorithm(params, spirv);
// Record op algo base mgr.sequence()->eval<kp::OpTensorSyncDevice>(params);
sq->begin();
sq->record<kp::OpTensorSyncDevice>({ wIn, bIn }); std::shared_ptr<kp::Sequence> sq = mgr.sequence()
->record<kp::OpTensorSyncDevice>({ wIn, bIn })
#ifdef KOMPUTE_ANDROID_SHADER_FROM_STRING ->record<kp::OpAlgoDispatch>(algo)
// Newer versions of Android are able to use shaderc to read raw string ->record<kp::OpTensorSyncLocal>({ wOutI, wOutJ, bOut, lOut });
sq->record<kp::OpAlgoCreate>(
params, std::vector<char>(LR_SHADER.begin(), LR_SHADER.end()));
#else
// Older versions of Android require the SPIRV binary directly
sq->record<kp::OpAlgoCreate>(
params, std::vector<char>(
kp::shader_data::shaders_glsl_logisticregression_comp_spv,
kp::shader_data::shaders_glsl_logisticregression_comp_spv
+ kp::shader_data::shaders_glsl_logisticregression_comp_spv_len
));
#endif
sq->record<kp::OpTensorSyncLocal>({ wOutI, wOutJ, bOut, lOut });
sq->end();
// Iterate across all expected iterations // Iterate across all expected iterations
for (size_t i = 0; i < ITERATIONS; i++) { for (size_t i = 0; i < ITERATIONS; i++) {
@ -94,15 +81,15 @@ void KomputeModelML::train(Array yArr, Array xIArr, Array xJArr) {
} }
} }
} }
KP_LOG_INFO("RESULT: <<<<<<<<<<<<<<<<<<<");
KP_LOG_INFO(wIn->data()[0]);
KP_LOG_INFO(wIn->data()[1]);
KP_LOG_INFO(bIn->data()[0]);
this->mWeights = wIn;
this->mBias = bIn;
} }
KP_LOG_INFO("RESULT: <<<<<<<<<<<<<<<<<<<");
KP_LOG_INFO(wIn->data()[0]);
KP_LOG_INFO(wIn->data()[1]);
KP_LOG_INFO(bIn->data()[0]);
this->mWeights = kp::Tensor(wIn->data());
this->mBias = kp::Tensor(bIn->data());
} }
Array KomputeModelML::predict(Array xI, Array xJ) { Array KomputeModelML::predict(Array xI, Array xJ) {
@ -116,9 +103,9 @@ Array KomputeModelML::predict(Array xI, Array xJ) {
for (size_t i = 0; i < xI.size(); i++) { for (size_t i = 0; i < xI.size(); i++) {
float xIVal = xI[i]; float xIVal = xI[i];
float xJVal = xJ[i]; float xJVal = xJ[i];
float result = (xIVal * this->mWeights.data()[0] float result = (xIVal * this->mWeights->data()[0]
+ xJVal * this->mWeights.data()[1] + xJVal * this->mWeights->data()[1]
+ this->mBias.data()[0]); + this->mBias->data()[0]);
// Instead of using sigmoid we'll just return full numbers // Instead of using sigmoid we'll just return full numbers
Variant var = result > 0 ? 1 : 0; Variant var = result > 0 ? 1 : 0;
@ -131,15 +118,15 @@ Array KomputeModelML::predict(Array xI, Array xJ) {
Array KomputeModelML::get_params() { Array KomputeModelML::get_params() {
Array retArray; Array retArray;
KP_LOG_INFO(this->mWeights.size() + this->mBias.size()); KP_LOG_INFO(this->mWeights->size() + this->mBias->size());
if(this->mWeights.size() + this->mBias.size() == 0) { if(this->mWeights->size() + this->mBias->size() == 0) {
return retArray; return retArray;
} }
retArray.push_back(this->mWeights.data()[0]); retArray.push_back(this->mWeights->data()[0]);
retArray.push_back(this->mWeights.data()[1]); retArray.push_back(this->mWeights->data()[1]);
retArray.push_back(this->mBias.data()[0]); retArray.push_back(this->mBias->data()[0]);
retArray.push_back(99.0); retArray.push_back(99.0);
return retArray; return retArray;

4
examples/godot_logistic_regression/gdnative_shared/src/KomputeModelML.hpp
View file

@ -28,8 +28,8 @@ public:
static void _register_methods(); static void _register_methods();
private: private:
kp::Tensor mWeights; std::shared_ptr<kp::Tensor> mWeights;
kp::Tensor mBias; std::shared_ptr<kp::Tensor> mBias;
}; };
static std::string LR_SHADER = R"( static std::string LR_SHADER = R"(

45
examples/logistic_regression/src/Main.cpp
View file

@ -15,44 +15,39 @@ int main()
uint32_t ITERATIONS = 100; uint32_t ITERATIONS = 100;
float learningRate = 0.1; float learningRate = 0.1;
std::shared_ptr<kp::Tensor> xI{ new kp::Tensor({ 0, 1, 1, 1, 1 }) }; kp::Manager mgr;
std::shared_ptr<kp::Tensor> xJ{ new kp::Tensor({ 0, 0, 0, 1, 1 }) };
std::shared_ptr<kp::Tensor> y{ new kp::Tensor({ 0, 0, 0, 1, 1 }) }; std::shared_ptr<kp::Tensor> xI = mgr.tensor({ 0, 1, 1, 1, 1 });
std::shared_ptr<kp::Tensor> xJ = mgr.tensor({ 0, 0, 0, 1, 1 });
std::shared_ptr<kp::Tensor> wIn{ new kp::Tensor({ 0.001, 0.001 }) }; std::shared_ptr<kp::Tensor> y = mgr.tensor({ 0, 0, 0, 1, 1 });
std::shared_ptr<kp::Tensor> wOutI{ new kp::Tensor({ 0, 0, 0, 0, 0 }) };
std::shared_ptr<kp::Tensor> wOutJ{ new kp::Tensor({ 0, 0, 0, 0, 0 }) };
std::shared_ptr<kp::Tensor> bIn{ new kp::Tensor({ 0 }) }; std::shared_ptr<kp::Tensor> wIn = mgr.tensor({ 0.001, 0.001 });
std::shared_ptr<kp::Tensor> bOut{ new kp::Tensor({ 0, 0, 0, 0, 0 }) }; std::shared_ptr<kp::Tensor> wOutI = mgr.tensor({ 0, 0, 0, 0, 0 });
std::shared_ptr<kp::Tensor> wOutJ = mgr.tensor({ 0, 0, 0, 0, 0 });
std::shared_ptr<kp::Tensor> lOut{ new kp::Tensor({ 0, 0, 0, 0, 0 }) }; std::shared_ptr<kp::Tensor> bIn = mgr.tensor({ 0 });
std::shared_ptr<kp::Tensor> bOut = mgr.tensor({ 0, 0, 0, 0, 0 });
std::shared_ptr<kp::Tensor> lOut = mgr.tensor({ 0, 0, 0, 0, 0 });
std::vector<std::shared_ptr<kp::Tensor>> params = { xI, xJ, y, std::vector<std::shared_ptr<kp::Tensor>> params = { xI, xJ, y,
wIn, wOutI, wOutJ, wIn, wOutI, wOutJ,
bIn, bOut, lOut }; bIn, bOut, lOut };
kp::Manager mgr; std::vector<uint32_t> spirv(
mgr.rebuild(params);
std::shared_ptr<kp::Sequence> sq = mgr.sequence();
// Record op algo base
sq->begin();
sq->record<kp::OpTensorSyncDevice>({ wIn, bIn });
sq->record<kp::OpAlgoCreate>(
params, std::vector<uint32_t>(
(uint32_t*)kp::shader_data::shaders_glsl_logisticregression_comp_spv, (uint32_t*)kp::shader_data::shaders_glsl_logisticregression_comp_spv,
(uint32_t*)(kp::shader_data::shaders_glsl_logisticregression_comp_spv (uint32_t*)(kp::shader_data::shaders_glsl_logisticregression_comp_spv
+ kp::shader_data::shaders_glsl_logisticregression_comp_spv_len))); + kp::shader_data::shaders_glsl_logisticregression_comp_spv_len));
sq->record<kp::OpTensorSyncLocal>({ wOutI, wOutJ, bOut, lOut }); std::shared_ptr<kp::Algorithm> algo = mgr.algorithm(params, spirv);
sq->end(); mgr.sequence()->eval<kp::OpTensorSyncDevice>(params);
std::shared_ptr<kp::Sequence> sq = mgr.sequence()
->record<kp::OpTensorSyncDevice>({ wIn, bIn })
->record<kp::OpAlgoDispatch>(algo)
->record<kp::OpTensorSyncLocal>({ wOutI, wOutJ, bOut, lOut });
// Iterate across all expected iterations // Iterate across all expected iterations
for (size_t i = 0; i < ITERATIONS; i++) { for (size_t i = 0; i < ITERATIONS; i++) {

435
single_include/kompute/Kompute.hpp
View file

@ -647,12 +647,19 @@ extern py::object kp_debug, kp_info, kp_warning, kp_error;
#define KP_LOG_DEBUG(...) #define KP_LOG_DEBUG(...)
#else #else
#if defined(VK_USE_PLATFORM_ANDROID_KHR) #if defined(VK_USE_PLATFORM_ANDROID_KHR)
#define KP_LOG_DEBUG(...) \ #define KP_LOG_DEBUG(...) \
((void)__android_log_print(ANDROID_LOG_DEBUG, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__))) ((void)__android_log_write( \
ANDROID_LOG_DEBUG, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__).c_str()))
#elif defined(KOMPUTE_BUILD_PYTHON) #elif defined(KOMPUTE_BUILD_PYTHON)
#define KP_LOG_DEBUG(...) kp_debug(fmt::format(__VA_ARGS__)) #define KP_LOG_DEBUG(...) kp_debug(fmt::format(__VA_ARGS__))
#else #else
#define KP_LOG_DEBUG(...) fmt::print("[{} {}] [debug] [{}:{}] {}\n", __DATE__, __TIME__, __FILE__, __LINE__, fmt::format(__VA_ARGS__)) #define KP_LOG_DEBUG(...) \
fmt::print("[{} {}] [debug] [{}:{}] {}\n", \
__DATE__, \
__TIME__, \
__FILE__, \
__LINE__, \
fmt::format(__VA_ARGS__))
#endif // VK_USE_PLATFORM_ANDROID_KHR #endif // VK_USE_PLATFORM_ANDROID_KHR
#endif // SPDLOG_ACTIVE_LEVEL > 1 #endif // SPDLOG_ACTIVE_LEVEL > 1
@ -660,12 +667,19 @@ extern py::object kp_debug, kp_info, kp_warning, kp_error;
#define KP_LOG_INFO(...) #define KP_LOG_INFO(...)
#else #else
#if defined(VK_USE_PLATFORM_ANDROID_KHR) #if defined(VK_USE_PLATFORM_ANDROID_KHR)
#define KP_LOG_INFO(...) \ #define KP_LOG_INFO(...) \
((void)__android_log_print(ANDROID_LOG_INFO, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__))) ((void)__android_log_write( \
ANDROID_LOG_INFO, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__).c_str()))
#elif defined(KOMPUTE_BUILD_PYTHON) #elif defined(KOMPUTE_BUILD_PYTHON)
#define KP_LOG_INFO(...) kp_info(fmt::format(__VA_ARGS__)) #define KP_LOG_INFO(...) kp_info(fmt::format(__VA_ARGS__))
#else #else
#define KP_LOG_INFO(...) fmt::print("[{} {}] [debug] [{}:{}] {}\n", __DATE__, __TIME__, __FILE__, __LINE__, fmt::format(__VA_ARGS__)) #define KP_LOG_INFO(...) \
fmt::print("[{} {}] [debug] [{}:{}] {}\n", \
__DATE__, \
__TIME__, \
__FILE__, \
__LINE__, \
fmt::format(__VA_ARGS__))
#endif // VK_USE_PLATFORM_ANDROID_KHR #endif // VK_USE_PLATFORM_ANDROID_KHR
#endif // SPDLOG_ACTIVE_LEVEL > 2 #endif // SPDLOG_ACTIVE_LEVEL > 2
@ -673,12 +687,19 @@ extern py::object kp_debug, kp_info, kp_warning, kp_error;
#define KP_LOG_WARN(...) #define KP_LOG_WARN(...)
#else #else
#if defined(VK_USE_PLATFORM_ANDROID_KHR) #if defined(VK_USE_PLATFORM_ANDROID_KHR)
#define KP_LOG_WARN(...) \ #define KP_LOG_WARN(...) \
((void)__android_log_print(ANDROID_LOG_WARN, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__))) ((void)__android_log_write( \
ANDROID_LOG_WARN, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__).c_str()))
#elif defined(KOMPUTE_BUILD_PYTHON) #elif defined(KOMPUTE_BUILD_PYTHON)
#define KP_LOG_WARN(...) kp_warning(fmt::format(__VA_ARGS__)) #define KP_LOG_WARN(...) kp_warning(fmt::format(__VA_ARGS__))
#else #else
#define KP_LOG_WARN(...) fmt::print("[{} {}] [debug] [{}:{}] {}\n", __DATE__, __TIME__, __FILE__, __LINE__, fmt::format(__VA_ARGS__)) #define KP_LOG_WARN(...) \
fmt::print("[{} {}] [debug] [{}:{}] {}\n", \
__DATE__, \
__TIME__, \
__FILE__, \
__LINE__, \
fmt::format(__VA_ARGS__))
#endif // VK_USE_PLATFORM_ANDROID_KHR #endif // VK_USE_PLATFORM_ANDROID_KHR
#endif // SPDLOG_ACTIVE_LEVEL > 3 #endif // SPDLOG_ACTIVE_LEVEL > 3
@ -686,12 +707,19 @@ extern py::object kp_debug, kp_info, kp_warning, kp_error;
#define KP_LOG_ERROR(...) #define KP_LOG_ERROR(...)
#else #else
#if defined(VK_USE_PLATFORM_ANDROID_KHR) #if defined(VK_USE_PLATFORM_ANDROID_KHR)
#define KP_LOG_ERROR(...) \ #define KP_LOG_ERROR(...) \
((void)__android_log_print(ANDROID_LOG_ERROR, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__))) ((void)__android_log_write( \
ANDROID_LOG_ERROR, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__).c_str()))
#elif defined(KOMPUTE_BUILD_PYTHON) #elif defined(KOMPUTE_BUILD_PYTHON)
#define KP_LOG_ERROR(...) kp_error(fmt::format(__VA_ARGS__)) #define KP_LOG_ERROR(...) kp_error(fmt::format(__VA_ARGS__))
#else #else
#define KP_LOG_ERROR(...) fmt::print("[{} {}] [debug] [{}:{}] {}\n", __DATE__, __TIME__, __FILE__, __LINE__, fmt::format(__VA_ARGS__)) #define KP_LOG_ERROR(...) \
fmt::print("[{} {}] [debug] [{}:{}] {}\n", \
__DATE__, \
__TIME__, \
__FILE__, \
__LINE__, \
fmt::format(__VA_ARGS__))
#endif // VK_USE_PLATFORM_ANDROID_KHR #endif // VK_USE_PLATFORM_ANDROID_KHR
#endif // SPDLOG_ACTIVE_LEVEL > 4 #endif // SPDLOG_ACTIVE_LEVEL > 4
#endif // KOMPUTE_SPDLOG_ENABLED #endif // KOMPUTE_SPDLOG_ENABLED
@ -701,9 +729,9 @@ extern py::object kp_debug, kp_info, kp_warning, kp_error;
#include <iostream> #include <iostream>
#include <vector> #include <vector>
#include <SPIRV/GlslangToSpv.h>
#include <glslang/Include/ResourceLimits.h> #include <glslang/Include/ResourceLimits.h>
#include <glslang/Public/ShaderLang.h> #include <glslang/Public/ShaderLang.h>
#include <SPIRV/GlslangToSpv.h>
namespace kp { namespace kp {
@ -711,117 +739,120 @@ namespace kp {
// Has been adobted by: // Has been adobted by:
// https://github.com/KhronosGroup/glslang/blob/master/StandAlone/ResourceLimits.cpp // https://github.com/KhronosGroup/glslang/blob/master/StandAlone/ResourceLimits.cpp
const TBuiltInResource defaultResource = { const TBuiltInResource defaultResource = {
/* .MaxLights = */ 0, /* .MaxLights = */ 0,
/* .MaxClipPlanes = */ 0, /* .MaxClipPlanes = */ 0,
/* .MaxTextureUnits = */ 0, /* .MaxTextureUnits = */ 0,
/* .MaxTextureCoords = */ 0, /* .MaxTextureCoords = */ 0,
/* .MaxVertexAttribs = */ 64, /* .MaxVertexAttribs = */ 64,
/* .MaxVertexUniformComponents = */ 4096, /* .MaxVertexUniformComponents = */ 4096,
/* .MaxVaryingFloats = */ 64, /* .MaxVaryingFloats = */ 64,
/* .MaxVertexTextureImageUnits = */ 0, /* .MaxVertexTextureImageUnits = */ 0,
/* .MaxCombinedTextureImageUnits = */ 0, /* .MaxCombinedTextureImageUnits = */ 0,
/* .MaxTextureImageUnits = */ 0, /* .MaxTextureImageUnits = */ 0,
/* .MaxFragmentUniformComponents = */ 0, /* .MaxFragmentUniformComponents = */ 0,
/* .MaxDrawBuffers = */ 0, /* .MaxDrawBuffers = */ 0,
/* .MaxVertexUniformVectors = */ 128, /* .MaxVertexUniformVectors = */ 128,
/* .MaxVaryingVectors = */ 8, /* .MaxVaryingVectors = */ 8,
/* .MaxFragmentUniformVectors = */ 0, /* .MaxFragmentUniformVectors = */ 0,
/* .MaxVertexOutputVectors = */ 16, /* .MaxVertexOutputVectors = */ 16,
/* .MaxFragmentInputVectors = */ 0, /* .MaxFragmentInputVectors = */ 0,
/* .MinProgramTexelOffset = */ -8, /* .MinProgramTexelOffset = */ -8,
/* .MaxProgramTexelOffset = */ 7, /* .MaxProgramTexelOffset = */ 7,
/* .MaxClipDistances = */ 8, /* .MaxClipDistances = */ 8,
/* .MaxComputeWorkGroupCountX = */ 65535, /* .MaxComputeWorkGroupCountX = */ 65535,
/* .MaxComputeWorkGroupCountY = */ 65535, /* .MaxComputeWorkGroupCountY = */ 65535,
/* .MaxComputeWorkGroupCountZ = */ 65535, /* .MaxComputeWorkGroupCountZ = */ 65535,
/* .MaxComputeWorkGroupSizeX = */ 1024, /* .MaxComputeWorkGroupSizeX = */ 1024,
/* .MaxComputeWorkGroupSizeY = */ 1024, /* .MaxComputeWorkGroupSizeY = */ 1024,
/* .MaxComputeWorkGroupSizeZ = */ 64, /* .MaxComputeWorkGroupSizeZ = */ 64,
/* .MaxComputeUniformComponents = */ 1024, /* .MaxComputeUniformComponents = */ 1024,
/* .MaxComputeTextureImageUnits = */ 16, /* .MaxComputeTextureImageUnits = */ 16,
/* .MaxComputeImageUniforms = */ 8, /* .MaxComputeImageUniforms = */ 8,
/* .MaxComputeAtomicCounters = */ 8, /* .MaxComputeAtomicCounters = */ 8,
/* .MaxComputeAtomicCounterBuffers = */ 1, /* .MaxComputeAtomicCounterBuffers = */ 1,
/* .MaxVaryingComponents = */ 60, /* .MaxVaryingComponents = */ 60,
/* .MaxVertexOutputComponents = */ 64, /* .MaxVertexOutputComponents = */ 64,
/* .MaxGeometryInputComponents = */ 64, /* .MaxGeometryInputComponents = */ 64,
/* .MaxGeometryOutputComponents = */ 128, /* .MaxGeometryOutputComponents = */ 128,
/* .MaxFragmentInputComponents = */ 0, /* .MaxFragmentInputComponents = */ 0,
/* .MaxImageUnits = */ 0, /* .MaxImageUnits = */ 0,
/* .MaxCombinedImageUnitsAndFragmentOutputs = */ 0, /* .MaxCombinedImageUnitsAndFragmentOutputs = */ 0,
/* .MaxCombinedShaderOutputResources = */ 8, /* .MaxCombinedShaderOutputResources = */ 8,
/* .MaxImageSamples = */ 0, /* .MaxImageSamples = */ 0,
/* .MaxVertexImageUniforms = */ 0, /* .MaxVertexImageUniforms = */ 0,
/* .MaxTessControlImageUniforms = */ 0, /* .MaxTessControlImageUniforms = */ 0,
/* .MaxTessEvaluationImageUniforms = */ 0, /* .MaxTessEvaluationImageUniforms = */ 0,
/* .MaxGeometryImageUniforms = */ 0, /* .MaxGeometryImageUniforms = */ 0,
/* .MaxFragmentImageUniforms = */ 0, /* .MaxFragmentImageUniforms = */ 0,
/* .MaxCombinedImageUniforms = */ 0, /* .MaxCombinedImageUniforms = */ 0,
/* .MaxGeometryTextureImageUnits = */ 0, /* .MaxGeometryTextureImageUnits = */ 0,
/* .MaxGeometryOutputVertices = */ 256, /* .MaxGeometryOutputVertices = */ 256,
/* .MaxGeometryTotalOutputComponents = */ 1024, /* .MaxGeometryTotalOutputComponents = */ 1024,
/* .MaxGeometryUniformComponents = */ 1024, /* .MaxGeometryUniformComponents = */ 1024,
/* .MaxGeometryVaryingComponents = */ 64, /* .MaxGeometryVaryingComponents = */ 64,
/* .MaxTessControlInputComponents = */ 128, /* .MaxTessControlInputComponents = */ 128,
/* .MaxTessControlOutputComponents = */ 128, /* .MaxTessControlOutputComponents = */ 128,
/* .MaxTessControlTextureImageUnits = */ 0, /* .MaxTessControlTextureImageUnits = */ 0,
/* .MaxTessControlUniformComponents = */ 1024, /* .MaxTessControlUniformComponents = */ 1024,
/* .MaxTessControlTotalOutputComponents = */ 4096, /* .MaxTessControlTotalOutputComponents = */ 4096,
/* .MaxTessEvaluationInputComponents = */ 128, /* .MaxTessEvaluationInputComponents = */ 128,
/* .MaxTessEvaluationOutputComponents = */ 128, /* .MaxTessEvaluationOutputComponents = */ 128,
/* .MaxTessEvaluationTextureImageUnits = */ 16, /* .MaxTessEvaluationTextureImageUnits = */ 16,
/* .MaxTessEvaluationUniformComponents = */ 1024, /* .MaxTessEvaluationUniformComponents = */ 1024,
/* .MaxTessPatchComponents = */ 120, /* .MaxTessPatchComponents = */ 120,
/* .MaxPatchVertices = */ 32, /* .MaxPatchVertices = */ 32,
/* .MaxTessGenLevel = */ 64, /* .MaxTessGenLevel = */ 64,
/* .MaxViewports = */ 16, /* .MaxViewports = */ 16,
/* .MaxVertexAtomicCounters = */ 0, /* .MaxVertexAtomicCounters = */ 0,
/* .MaxTessControlAtomicCounters = */ 0, /* .MaxTessControlAtomicCounters = */ 0,
/* .MaxTessEvaluationAtomicCounters = */ 0, /* .MaxTessEvaluationAtomicCounters = */ 0,
/* .MaxGeometryAtomicCounters = */ 0, /* .MaxGeometryAtomicCounters = */ 0,
/* .MaxFragmentAtomicCounters = */ 0, /* .MaxFragmentAtomicCounters = */ 0,
/* .MaxCombinedAtomicCounters = */ 8, /* .MaxCombinedAtomicCounters = */ 8,
/* .MaxAtomicCounterBindings = */ 1, /* .MaxAtomicCounterBindings = */ 1,
/* .MaxVertexAtomicCounterBuffers = */ 0, /* .MaxVertexAtomicCounterBuffers = */ 0,
/* .MaxTessControlAtomicCounterBuffers = */ 0, /* .MaxTessControlAtomicCounterBuffers = */ 0,
/* .MaxTessEvaluationAtomicCounterBuffers = */ 0, /* .MaxTessEvaluationAtomicCounterBuffers = */ 0,
/* .MaxGeometryAtomicCounterBuffers = */ 0, /* .MaxGeometryAtomicCounterBuffers = */ 0,
/* .MaxFragmentAtomicCounterBuffers = */ 0, /* .MaxFragmentAtomicCounterBuffers = */ 0,
/* .MaxCombinedAtomicCounterBuffers = */ 1, /* .MaxCombinedAtomicCounterBuffers = */ 1,
/* .MaxAtomicCounterBufferSize = */ 16384, /* .MaxAtomicCounterBufferSize = */ 16384,
/* .MaxTransformFeedbackBuffers = */ 4, /* .MaxTransformFeedbackBuffers = */ 4,
/* .MaxTransformFeedbackInterleavedComponents = */ 64, /* .MaxTransformFeedbackInterleavedComponents = */ 64,
/* .MaxCullDistances = */ 8, /* .MaxCullDistances = */ 8,
/* .MaxCombinedClipAndCullDistances = */ 8, /* .MaxCombinedClipAndCullDistances = */ 8,
/* .MaxSamples = */ 4, /* .MaxSamples = */ 4,
/* .maxMeshOutputVerticesNV = */ 256, /* .maxMeshOutputVerticesNV = */ 256,
/* .maxMeshOutputPrimitivesNV = */ 512, /* .maxMeshOutputPrimitivesNV = */ 512,
/* .maxMeshWorkGroupSizeX_NV = */ 32, /* .maxMeshWorkGroupSizeX_NV = */ 32,
/* .maxMeshWorkGroupSizeY_NV = */ 1, /* .maxMeshWorkGroupSizeY_NV = */ 1,
/* .maxMeshWorkGroupSizeZ_NV = */ 1, /* .maxMeshWorkGroupSizeZ_NV = */ 1,
/* .maxTaskWorkGroupSizeX_NV = */ 32, /* .maxTaskWorkGroupSizeX_NV = */ 32,
/* .maxTaskWorkGroupSizeY_NV = */ 1, /* .maxTaskWorkGroupSizeY_NV = */ 1,
/* .maxTaskWorkGroupSizeZ_NV = */ 1, /* .maxTaskWorkGroupSizeZ_NV = */ 1,
/* .maxMeshViewCountNV = */ 4, /* .maxMeshViewCountNV = */ 4,
/* .maxDualSourceDrawBuffersEXT = */ 1, /* .maxDualSourceDrawBuffersEXT = */ 1,
/* .limits = */ { /* .limits = */
/* .nonInductiveForLoops = */ 1, {
/* .whileLoops = */ 1, /* .nonInductiveForLoops = */ 1,
/* .doWhileLoops = */ 1, /* .whileLoops = */ 1,
/* .generalUniformIndexing = */ 1, /* .doWhileLoops = */ 1,
/* .generalAttributeMatrixVectorIndexing = */ 1, /* .generalUniformIndexing = */ 1,
/* .generalVaryingIndexing = */ 1, /* .generalAttributeMatrixVectorIndexing = */ 1,
/* .generalSamplerIndexing = */ 1, /* .generalVaryingIndexing = */ 1,
/* .generalVariableIndexing = */ 1, /* .generalSamplerIndexing = */ 1,
/* .generalConstantMatrixVectorIndexing = */ 1, /* .generalVariableIndexing = */ 1,
}}; /* .generalConstantMatrixVectorIndexing = */ 1,
}
};
/** /**
Shader utily class with functions to compile and process glsl files. Shader utily class with functions to compile and process glsl files.
*/ */
class Shader { class Shader
public: {
public:
/** /**
* Compile multiple sources with optional filenames. Currently this function * Compile multiple sources with optional filenames. Currently this function
* uses the glslang C++ interface which is not thread safe so this funciton * uses the glslang C++ interface which is not thread safe so this funciton
@ -833,35 +864,36 @@ public:
* @param files A list of file names respective to each of the sources * @param files A list of file names respective to each of the sources
* @param entryPoint The function name to use as entry point * @param entryPoint The function name to use as entry point
* @param definitions List of pairs containing key value definitions * @param definitions List of pairs containing key value definitions
* @param resourcesLimit A list that contains the resource limits for the GLSL compiler * @param resourcesLimit A list that contains the resource limits for the
* GLSL compiler
* @return The compiled SPIR-V binary in unsigned int32 format * @return The compiled SPIR-V binary in unsigned int32 format
*/ */
static std::vector<uint32_t> compile_sources( static std::vector<uint32_t> compile_sources(
const std::vector<std::string>& sources, const std::vector<std::string>& sources,
const std::vector<std::string>& files = {}, const std::vector<std::string>& files = {},
const std::string& entryPoint = "main", const std::string& entryPoint = "main",
std::vector<std::pair<std::string,std::string>> definitions = {}, std::vector<std::pair<std::string, std::string>> definitions = {},
const TBuiltInResource& resources = defaultResource); const TBuiltInResource& resources = defaultResource);
/** /**
* Compile a single glslang source from string value. Currently this function * Compile a single glslang source from string value. Currently this
* uses the glslang C++ interface which is not thread safe so this funciton * function uses the glslang C++ interface which is not thread safe so this
* should not be called from multiple threads concurrently. If you have a * funciton should not be called from multiple threads concurrently. If you
* online shader processing multithreading use-case that can't use offline * have a online shader processing multithreading use-case that can't use
* compilation please open an issue. * offline compilation please open an issue.
* *
* @param source An individual raw glsl shader in string format * @param source An individual raw glsl shader in string format
* @param entryPoint The function name to use as entry point * @param entryPoint The function name to use as entry point
* @param definitions List of pairs containing key value definitions * @param definitions List of pairs containing key value definitions
* @param resourcesLimit A list that contains the resource limits for the GLSL compiler * @param resourcesLimit A list that contains the resource limits for the
* GLSL compiler
* @return The compiled SPIR-V binary in unsigned int32 format * @return The compiled SPIR-V binary in unsigned int32 format
*/ */
static std::vector<uint32_t> compile_source( static std::vector<uint32_t> compile_source(
const std::string& source, const std::string& source,
const std::string& entryPoint = "main", const std::string& entryPoint = "main",
std::vector<std::pair<std::string,std::string>> definitions = {}, std::vector<std::pair<std::string, std::string>> definitions = {},
const TBuiltInResource& resources = defaultResource); const TBuiltInResource& resources = defaultResource);
}; };
} }
@ -919,7 +951,7 @@ class Tensor
* otherwise there is no need to copy from host memory. * otherwise there is no need to copy from host memory.
*/ */
void rebuild(const std::vector<float>& data, void rebuild(const std::vector<float>& data,
TensorTypes tensorType = TensorTypes::eDevice); TensorTypes tensorType = TensorTypes::eDevice);
/** /**
* Destroys and frees the GPU resources which include the buffer and memory. * Destroys and frees the GPU resources which include the buffer and memory.
@ -990,9 +1022,8 @@ class Tensor
* @param createBarrier Whether to create a barrier that ensures the data is * @param createBarrier Whether to create a barrier that ensures the data is
* copied before further operations. Default is true. * copied before further operations. Default is true.
*/ */
void recordCopyFromStagingToDevice( void recordCopyFromStagingToDevice(const vk::CommandBuffer& commandBuffer,
const vk::CommandBuffer& commandBuffer, bool createBarrier);
bool createBarrier);
/** /**
* Records a copy from the internal device memory to the staging memory * Records a copy from the internal device memory to the staging memory
@ -1003,9 +1034,8 @@ class Tensor
* @param createBarrier Whether to create a barrier that ensures the data is * @param createBarrier Whether to create a barrier that ensures the data is
* copied before further operations. Default is true. * copied before further operations. Default is true.
*/ */
void recordCopyFromDeviceToStaging( void recordCopyFromDeviceToStaging(const vk::CommandBuffer& commandBuffer,
const vk::CommandBuffer& commandBuffer, bool createBarrier);
bool createBarrier);
/** /**
* Records the buffer memory barrier into the command buffer which * Records the buffer memory barrier into the command buffer which
@ -1017,12 +1047,11 @@ class Tensor
* @param scrStageMask Pipeline stage flags for source stage mask * @param scrStageMask Pipeline stage flags for source stage mask
* @param dstStageMask Pipeline stage flags for destination stage mask * @param dstStageMask Pipeline stage flags for destination stage mask
*/ */
void recordBufferMemoryBarrier( void recordBufferMemoryBarrier(const vk::CommandBuffer& commandBuffer,
const vk::CommandBuffer& commandBuffer, vk::AccessFlagBits srcAccessMask,
vk::AccessFlagBits srcAccessMask, vk::AccessFlagBits dstAccessMask,
vk::AccessFlagBits dstAccessMask, vk::PipelineStageFlagBits srcStageMask,
vk::PipelineStageFlagBits srcStageMask, vk::PipelineStageFlagBits dstStageMask);
vk::PipelineStageFlagBits dstStageMask);
/** /**
* Constructs a vulkan descriptor buffer info which can be used to specify * Constructs a vulkan descriptor buffer info which can be used to specify
@ -1070,11 +1099,11 @@ class Tensor
std::shared_ptr<vk::DeviceMemory> memory, std::shared_ptr<vk::DeviceMemory> memory,
vk::MemoryPropertyFlags memoryPropertyFlags); vk::MemoryPropertyFlags memoryPropertyFlags);
void recordCopyBuffer(const vk::CommandBuffer& commandBuffer, void recordCopyBuffer(const vk::CommandBuffer& commandBuffer,
std::shared_ptr<vk::Buffer> bufferFrom, std::shared_ptr<vk::Buffer> bufferFrom,
std::shared_ptr<vk::Buffer> bufferTo, std::shared_ptr<vk::Buffer> bufferTo,
vk::DeviceSize bufferSize, vk::DeviceSize bufferSize,
vk::BufferCopy copyRegion, vk::BufferCopy copyRegion,
bool createBarrier); bool createBarrier);
// Private util functions // Private util functions
vk::BufferUsageFlags getPrimaryBufferUsageFlags(); vk::BufferUsageFlags getPrimaryBufferUsageFlags();
@ -1094,8 +1123,7 @@ namespace kp {
*/ */
class Algorithm class Algorithm
{ {
public: public:
/** /**
* Default constructor for Algorithm * Default constructor for Algorithm
* *
@ -1103,12 +1131,11 @@ public:
* @param commandBuffer The vulkan command buffer to bind the pipeline and * @param commandBuffer The vulkan command buffer to bind the pipeline and
* shaders * shaders
*/ */
Algorithm( Algorithm(std::shared_ptr<vk::Device> device,
std::shared_ptr<vk::Device> device, const std::vector<std::shared_ptr<Tensor>>& tensors = {},
const std::vector<std::shared_ptr<Tensor>>& tensors = {}, const std::vector<uint32_t>& spirv = {},
const std::vector<uint32_t>& spirv = {}, const Workgroup& workgroup = {},
const Workgroup& workgroup = {}, const Constants& specializationConstants = {});
const Constants& specializationConstants = {});
/** /**
* Initialiser for the shader data provided to the algorithm as well as * Initialiser for the shader data provided to the algorithm as well as
@ -1116,14 +1143,13 @@ public:
* *
* @param shaderFileData The bytes in spir-v format of the shader * @param shaderFileData The bytes in spir-v format of the shader
* @tensorParams The Tensors to be used in the Algorithm / shader for * @tensorParams The Tensors to be used in the Algorithm / shader for
* @specalizationInstalces The specialization parameters to pass to the function * @specalizationInstalces The specialization parameters to pass to the
* processing * function processing
*/ */
void rebuild( void rebuild(const std::vector<std::shared_ptr<Tensor>>& tensors = {},
const std::vector<std::shared_ptr<Tensor>>& tensors = {}, const std::vector<uint32_t>& spirv = {},
const std::vector<uint32_t>& spirv = {}, const Workgroup& workgroup = {},
const Workgroup& workgroup = {}, const Constants& specializationConstants = {});
const Constants& specializationConstants = {});
/** /**
* Destructor for Algorithm which is responsible for freeing and desroying * Destructor for Algorithm which is responsible for freeing and desroying
@ -1143,7 +1169,8 @@ public:
void bindCore(const vk::CommandBuffer& commandBuffer); void bindCore(const vk::CommandBuffer& commandBuffer);
void bindPush(const vk::CommandBuffer& commandBuffer, const Constants& pushConstants); void bindPush(const vk::CommandBuffer& commandBuffer,
const Constants& pushConstants);
bool isInit(); bool isInit();
@ -1155,7 +1182,7 @@ public:
void destroy(); void destroy();
private: private:
// -------------- NEVER OWNED RESOURCES // -------------- NEVER OWNED RESOURCES
std::shared_ptr<vk::Device> mDevice; std::shared_ptr<vk::Device> mDevice;
std::vector<std::shared_ptr<Tensor>> mTensors; std::vector<std::shared_ptr<Tensor>> mTensors;
@ -1489,7 +1516,7 @@ namespace kp {
/** /**
* Container of operations that can be sent to GPU as batch * Container of operations that can be sent to GPU as batch
*/ */
class Sequence: public std::enable_shared_from_this<Sequence> class Sequence : public std::enable_shared_from_this<Sequence>
{ {
public: public:
/** /**
@ -1526,8 +1553,9 @@ class Sequence: public std::enable_shared_from_this<Sequence>
* which allows for extensible configurations on initialisation. * which allows for extensible configurations on initialisation.
*/ */
template<typename T, typename... TArgs> template<typename T, typename... TArgs>
std::shared_ptr<Sequence> std::shared_ptr<Sequence> record(
record(std::vector<std::shared_ptr<Tensor>> tensors, TArgs&&... params) std::vector<std::shared_ptr<Tensor>> tensors,
TArgs&&... params)
{ {
KP_LOG_DEBUG("Kompute Sequence record function started"); KP_LOG_DEBUG("Kompute Sequence record function started");
@ -1536,14 +1564,13 @@ class Sequence: public std::enable_shared_from_this<Sequence>
"OpBase derived classes"); "OpBase derived classes");
KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance"); KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance");
std::shared_ptr<T> op{ std::shared_ptr<T> op{ new T(tensors, std::forward<TArgs>(params)...) };
new T(tensors, std::forward<TArgs>(params)...) };
return this->record(op); return this->record(op);
} }
template<typename T, typename... TArgs> template<typename T, typename... TArgs>
std::shared_ptr<Sequence> std::shared_ptr<Sequence> record(std::shared_ptr<Algorithm> algorithm,
record(std::shared_ptr<Algorithm> algorithm, TArgs&&... params) TArgs&&... params)
{ {
KP_LOG_DEBUG("Kompute Sequence record function started"); KP_LOG_DEBUG("Kompute Sequence record function started");
@ -1552,8 +1579,8 @@ class Sequence: public std::enable_shared_from_this<Sequence>
"OpBase derived classes"); "OpBase derived classes");
KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance"); KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance");
std::shared_ptr<T> op{ std::shared_ptr<T> op{ new T(algorithm,
new T(algorithm, std::forward<TArgs>(params)...) }; std::forward<TArgs>(params)...) };
return this->record(op); return this->record(op);
} }
@ -1576,8 +1603,8 @@ class Sequence: public std::enable_shared_from_this<Sequence>
*/ */
// TODO: Aim to have only a single function with tensors/algorithm // TODO: Aim to have only a single function with tensors/algorithm
template<typename T, typename... TArgs> template<typename T, typename... TArgs>
std::shared_ptr<Sequence> std::shared_ptr<Sequence> eval(std::vector<std::shared_ptr<Tensor>> tensors,
eval(std::vector<std::shared_ptr<Tensor>> tensors, TArgs&&... params) TArgs&&... params)
{ {
KP_LOG_DEBUG("Kompute Sequence record function started"); KP_LOG_DEBUG("Kompute Sequence record function started");
@ -1586,16 +1613,16 @@ class Sequence: public std::enable_shared_from_this<Sequence>
"OpBase derived classes"); "OpBase derived classes");
KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance"); KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance");
std::shared_ptr<T> op{ std::shared_ptr<T> op{ new T(tensors, std::forward<TArgs>(params)...) };
new T(tensors, std::forward<TArgs>(params)...) };
// TODO: Aim to be able to handle errors when returning without throw except // TODO: Aim to be able to handle errors when returning without throw
// except
return this->eval(op); return this->eval(op);
} }
// Needded as otherise can't use initialiser list // Needded as otherise can't use initialiser list
template<typename T, typename... TArgs> template<typename T, typename... TArgs>
std::shared_ptr<Sequence> std::shared_ptr<Sequence> eval(std::shared_ptr<Algorithm> algorithm,
eval(std::shared_ptr<Algorithm> algorithm, TArgs&&... params) TArgs&&... params)
{ {
KP_LOG_DEBUG("Kompute Sequence record function started"); KP_LOG_DEBUG("Kompute Sequence record function started");
@ -1604,8 +1631,8 @@ class Sequence: public std::enable_shared_from_this<Sequence>
"OpBase derived classes"); "OpBase derived classes");
KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance"); KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance");
std::shared_ptr<T> op{ std::shared_ptr<T> op{ new T(algorithm,
new T(algorithm, std::forward<TArgs>(params)...) }; std::forward<TArgs>(params)...) };
return this->eval(op); return this->eval(op);
} }
@ -1627,8 +1654,9 @@ class Sequence: public std::enable_shared_from_this<Sequence>
* @return shared_ptr<Sequence> of the Sequence class itself * @return shared_ptr<Sequence> of the Sequence class itself
*/ */
template<typename T, typename... TArgs> template<typename T, typename... TArgs>
std::shared_ptr<Sequence> std::shared_ptr<Sequence> evalAsync(
evalAsync(std::vector<std::shared_ptr<Tensor>> tensors, TArgs&&... params) std::vector<std::shared_ptr<Tensor>> tensors,
TArgs&&... params)
{ {
KP_LOG_DEBUG("Kompute Sequence record function started"); KP_LOG_DEBUG("Kompute Sequence record function started");
@ -1637,15 +1665,14 @@ class Sequence: public std::enable_shared_from_this<Sequence>
"OpBase derived classes"); "OpBase derived classes");
KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance"); KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance");
std::shared_ptr<T> op{ std::shared_ptr<T> op{ new T(tensors, std::forward<TArgs>(params)...) };
new T(tensors, std::forward<TArgs>(params)...) };
return this->evalAsync(op); return this->evalAsync(op);
} }
// Needed as otherwise it's not possible to use initializer lists // Needed as otherwise it's not possible to use initializer lists
template<typename T, typename... TArgs> template<typename T, typename... TArgs>
std::shared_ptr<Sequence> std::shared_ptr<Sequence> evalAsync(std::shared_ptr<Algorithm> algorithm,
evalAsync(std::shared_ptr<Algorithm> algorithm, TArgs&&... params) TArgs&&... params)
{ {
KP_LOG_DEBUG("Kompute Sequence record function started"); KP_LOG_DEBUG("Kompute Sequence record function started");
@ -1654,8 +1681,8 @@ class Sequence: public std::enable_shared_from_this<Sequence>
"OpBase derived classes"); "OpBase derived classes");
KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance"); KP_LOG_DEBUG("Kompute Sequence creating OpBase derived class instance");
std::shared_ptr<T> op{ std::shared_ptr<T> op{ new T(algorithm,
new T(algorithm, std::forward<TArgs>(params)...) }; std::forward<TArgs>(params)...) };
return this->evalAsync(op); return this->evalAsync(op);
} }
@ -1670,7 +1697,8 @@ class Sequence: public std::enable_shared_from_this<Sequence>
std::shared_ptr<Sequence> evalAwait(uint64_t waitFor = UINT64_MAX); std::shared_ptr<Sequence> evalAwait(uint64_t waitFor = UINT64_MAX);
/** /**
* Clear function clears all operations currently recorded and starts recording again. * Clear function clears all operations currently recorded and starts
* recording again.
*/ */
void clear(); void clear();
@ -1821,10 +1849,10 @@ class Manager
Tensor::TensorTypes tensorType = Tensor::TensorTypes::eDevice); Tensor::TensorTypes tensorType = Tensor::TensorTypes::eDevice);
std::shared_ptr<Algorithm> algorithm( std::shared_ptr<Algorithm> algorithm(
const std::vector<std::shared_ptr<Tensor>>& tensors = {}, const std::vector<std::shared_ptr<Tensor>>& tensors = {},
const std::vector<uint32_t>& spirv = {}, const std::vector<uint32_t>& spirv = {},
const Workgroup& workgroup = {}, const Workgroup& workgroup = {},
const Constants& specializationConstants = {}); const Constants& specializationConstants = {});
void destroy(); void destroy();
void clear(); void clear();
@ -1856,7 +1884,8 @@ class Manager
// Create functions // Create functions
void createInstance(); void createInstance();
void createDevice(const std::vector<uint32_t>& familyQueueIndices = {}, uint32_t hysicalDeviceIndex = 0); void createDevice(const std::vector<uint32_t>& familyQueueIndices = {},
uint32_t hysicalDeviceIndex = 0);
}; };
} // End namespace kp } // End namespace kp

16
src/include/kompute/Core.hpp
View file

@ -61,8 +61,8 @@ extern py::object kp_debug, kp_info, kp_warning, kp_error;
#else #else
#if defined(VK_USE_PLATFORM_ANDROID_KHR) #if defined(VK_USE_PLATFORM_ANDROID_KHR)
#define KP_LOG_DEBUG(...) \ #define KP_LOG_DEBUG(...) \
((void)__android_log_print( \ ((void)__android_log_write( \
ANDROID_LOG_DEBUG, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__))) ANDROID_LOG_DEBUG, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__).c_str()))
#elif defined(KOMPUTE_BUILD_PYTHON) #elif defined(KOMPUTE_BUILD_PYTHON)
#define KP_LOG_DEBUG(...) kp_debug(fmt::format(__VA_ARGS__)) #define KP_LOG_DEBUG(...) kp_debug(fmt::format(__VA_ARGS__))
#else #else
@ -81,8 +81,8 @@ extern py::object kp_debug, kp_info, kp_warning, kp_error;
#else #else
#if defined(VK_USE_PLATFORM_ANDROID_KHR) #if defined(VK_USE_PLATFORM_ANDROID_KHR)
#define KP_LOG_INFO(...) \ #define KP_LOG_INFO(...) \
((void)__android_log_print( \ ((void)__android_log_write( \
ANDROID_LOG_INFO, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__))) ANDROID_LOG_INFO, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__).c_str()))
#elif defined(KOMPUTE_BUILD_PYTHON) #elif defined(KOMPUTE_BUILD_PYTHON)
#define KP_LOG_INFO(...) kp_info(fmt::format(__VA_ARGS__)) #define KP_LOG_INFO(...) kp_info(fmt::format(__VA_ARGS__))
#else #else
@ -101,8 +101,8 @@ extern py::object kp_debug, kp_info, kp_warning, kp_error;
#else #else
#if defined(VK_USE_PLATFORM_ANDROID_KHR) #if defined(VK_USE_PLATFORM_ANDROID_KHR)
#define KP_LOG_WARN(...) \ #define KP_LOG_WARN(...) \
((void)__android_log_print( \ ((void)__android_log_write( \
ANDROID_LOG_WARN, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__))) ANDROID_LOG_WARN, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__).c_str()))
#elif defined(KOMPUTE_BUILD_PYTHON) #elif defined(KOMPUTE_BUILD_PYTHON)
#define KP_LOG_WARN(...) kp_warning(fmt::format(__VA_ARGS__)) #define KP_LOG_WARN(...) kp_warning(fmt::format(__VA_ARGS__))
#else #else
@ -121,8 +121,8 @@ extern py::object kp_debug, kp_info, kp_warning, kp_error;
#else #else
#if defined(VK_USE_PLATFORM_ANDROID_KHR) #if defined(VK_USE_PLATFORM_ANDROID_KHR)
#define KP_LOG_ERROR(...) \ #define KP_LOG_ERROR(...) \
((void)__android_log_print( \ ((void)__android_log_write( \
ANDROID_LOG_ERROR, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__))) ANDROID_LOG_ERROR, KOMPUTE_LOG_TAG, fmt::format(__VA_ARGS__).c_str()))
#elif defined(KOMPUTE_BUILD_PYTHON) #elif defined(KOMPUTE_BUILD_PYTHON)
#define KP_LOG_ERROR(...) kp_error(fmt::format(__VA_ARGS__)) #define KP_LOG_ERROR(...) kp_error(fmt::format(__VA_ARGS__))
#else #else

1
src/include/kompute/operations/OpBase.hpp
View file

@ -1,7 +1,6 @@
#pragma once #pragma once
#include "kompute/Core.hpp" #include "kompute/Core.hpp"
#include "kompute/Tensor.hpp" #include "kompute/Tensor.hpp"
#include "kompute/Algorithm.hpp" #include "kompute/Algorithm.hpp"