Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[ GPU ] split kernel registration from forwarding function in rmsnorm_layer_cl #2804

Merged
merged 3 commits into from
Dec 19, 2024
+120 −98
Merged

[ GPU ] split kernel registration from forwarding function in rmsnorm_layer_cl #2804

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
bbda540
[ GPU/OpenCL ] change reshape_cl to inherit LayerImplCl
EunjuYang Nov 5, 2024
8702596
[ GPU/OpenCL ] change rmsnorm_layer_cl to inherit LayerImplCl
EunjuYang Nov 6, 2024
78907d7
[ Prop ] Update bn layer properties to follow naming convention
EunjuYang Dec 2, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 5 additions & 3 deletions nntrainer/cl_context.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,9 +53,11 @@ static void add_default_object(ClContext &cc) {
ml::train::LayerType::LAYER_RESHAPE);
}

// @todo rmsnormlayercl also needs to be updated.
cc.registerFactory(nntrainer::createLayer<RMSNormLayerCl>,
RMSNormLayerCl::type, ml::train::LayerType::LAYER_RMSNORM);
if (RMSNormLayerCl::registerClKernels()) {
cc.registerFactory(nntrainer::createLayer<RMSNormLayerCl>,
RMSNormLayerCl::type,
ml::train::LayerType::LAYER_RMSNORM);
}

if (ConcatLayerCl::registerClKernels()) {
cc.registerFactory(nntrainer::createLayer<ConcatLayerCl>,
Expand Down
69 changes: 52 additions & 17 deletions nntrainer/layers/cl_layers/rmsnorm_layer_cl.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -91,13 +91,12 @@ static constexpr size_t SINGLE_INOUT_IDX = 0;

enum RMSParams { gamma };

RMSNormLayerCl::RMSNormLayerCl() : LayerImpl() { wt_idx.fill(0); }
RMSNormLayerCl::RMSNormLayerCl() : LayerImplCl() { wt_idx.fill(0); }

void RMSNormLayerCl::finalize(InitLayerContext &context) {
std::vector<TensorDim> dim = context.getInputDimensions();
context.setOutputDimensions(dim);
auto &rmsparams_gamma =
std::get<props::RMS_NORM_GAMMA_INIT_GPU>(rmsnorm_props);
auto &rmsparams_gamma = std::get<props::RMS_NORM_GAMMA_INIT>(rmsnorm_props);

TensorDim gamma_dim(
1, 1, 1, dim[0].width(),
Expand All @@ -123,9 +122,6 @@ void RMSNormLayerCl::forwarding(RunLayerContext &context, bool training) {
}
}

opencl::Kernel RMSNormLayerCl::kernel_rmsnorm;
opencl::Kernel RMSNormLayerCl::kernel_rmsnorm_fp16;

void RMSNormLayerCl::rmsnormProcess(Tensor const &input, Tensor &result,
Tensor const &gamma, const float epsilon) {
bool ret = false;
Expand All @@ -138,11 +134,8 @@ void RMSNormLayerCl::rmsnormProcess(Tensor const &input, Tensor &result,
int w = input.width();

do {
ClContext::SharedPtrClKernel kernel_rmsnorm_ptr =
cl_context_ref.registerClKernel(rmsnorm_cl_kernel_, "rmsnorm_cl");
if (!kernel_rmsnorm_ptr) {
break;
}

auto kernel_rmsnorm_ptr = layer_kernel_ptrs[Kernels::RMSNORM_CL];

opencl::Buffer inputbuf(cl_context_ref.context_inst_, dim1 * sizeof(float),
true, nullptr);
Expand Down Expand Up @@ -219,6 +212,7 @@ void RMSNormLayerCl::rmsnormProcess(Tensor const &input, Tensor &result,
} while (false);
}

#ifdef ENABLE_FP16
void RMSNormLayerCl::rmsnormProcess_fp16(Tensor const &input, Tensor &result,
Tensor const &gamma,
const float epsilon) {
Expand All @@ -232,12 +226,8 @@ void RMSNormLayerCl::rmsnormProcess_fp16(Tensor const &input, Tensor &result,
int h = input.height();
int w = input.width();
do {
ClContext::SharedPtrClKernel kernel_rmsnorm_ptr =
cl_context_ref.registerClKernel(rmsnorm_cl_kernel_fp16_,
"rmsnorm_cl_fp16");
if (!kernel_rmsnorm_ptr) {
break;
}
auto kernel_rmsnorm_ptr = layer_kernel_ptrs[Kernels::RMSNORM_CL_FP16];

opencl::Buffer inputbuf(cl_context_ref.context_inst_,
dim1 * sizeof(cl_half), true, nullptr);

Expand Down Expand Up @@ -308,6 +298,7 @@ void RMSNormLayerCl::rmsnormProcess_fp16(Tensor const &input, Tensor &result,
}
} while (false);
}
#endif

void RMSNormLayerCl::incremental_forwarding(nntrainer::RunLayerContext &context,
unsigned int from, unsigned int to,
Expand Down Expand Up @@ -339,7 +330,11 @@ void RMSNormLayerCl::incremental_forwarding(nntrainer::RunLayerContext &context,
if (in_step.getDataType() == ml::train::TensorDim::DataType::FP32) {
rmsnormProcess(in, out, gamma, epsilon);
} else {
#ifdef ENABLE_FP16
rmsnormProcess_fp16(in, out, gamma, epsilon);
#else
throw std::runtime_error("enable-fp16 is not enabled");
#endif
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

it might be good to throw an error when fp16 is not enabled!

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you for the comment. I updated it :)

}
}

Expand All @@ -362,4 +357,44 @@ void RMSNormLayerCl::setProperty(const std::vector<std::string> &values) {
LayerImpl::setProperty(remain_props);
}

bool RMSNormLayerCl::registerClKernels() {

// check if already registered
if (!layer_kernel_ptrs.empty()) {
ml_loge("kernels for concat layer are already registered.");
return false;
}

do {

ClContext::SharedPtrClKernel kernel_rmsnorm_ptr = nullptr;

kernel_rmsnorm_ptr =
cl_context_ref.registerClKernel(rmsnorm_cl_kernel_, "rmsnorm_cl");
if (!kernel_rmsnorm_ptr) {
ml_loge("OpenCL Error: Fail to register rmsnorm_cl kernel");
break;
}
layer_kernel_ptrs.emplace_back(kernel_rmsnorm_ptr);

#ifdef ENABLE_FP16
kernel_rmsnorm_ptr = cl_context_ref.registerClKernel(
rmsnorm_cl_kernel_fp16_, "rmsnorm_cl_fp16");
if (!kernel_rmsnorm_ptr) {
ml_loge("OpenCL Error: Fail to register rmsnorm_cl_fp16 kernel");
break;
}
layer_kernel_ptrs.emplace_back(kernel_rmsnorm_ptr);
#endif

return true;

} while (false);

// clear all registered kernels if any error occurs during registration
layer_kernel_ptrs.clear();

return false;
}

} // namespace nntrainer
48 changes: 17 additions & 31 deletions nntrainer/layers/cl_layers/rmsnorm_layer_cl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@
#ifdef __cplusplus

#include <common_properties.h>
#include <layer_impl.h>
#include <layer_impl_cl.h>
#include <nntrainer_log.h>

#include <cl_context.h>
Expand All @@ -25,36 +25,11 @@

namespace nntrainer {

namespace props {

/**
* @brief RMS_NORM_GAMMA_INIT_GPU Initialization Enumeration Information
*
*/
class RMS_NORM_GAMMA_INIT_GPU final
: public ::nntrainer::EnumProperty<::nntrainer::props::InitializerInfo> {
public:
/**
* @brief Construct a RMS_NORM_GAMMA_INIT object
*/
RMS_NORM_GAMMA_INIT_GPU(
::nntrainer::Initializer value = ::nntrainer::Initializer::ONES) {
set(value);
};
using prop_tag = enum_class_prop_tag;
static constexpr const char *key = "gamma_initializer";
};
}; // namespace props

/**
* @class RMSNormLayer
* @brief RMS Norm layer
*/

class RMSNormLayerCl : public LayerImpl {

private:
inline static ClContext cl_context_ref;
class RMSNormLayerCl : public LayerImplCl {

public:
/**
Expand Down Expand Up @@ -118,9 +93,6 @@ class RMSNormLayerCl : public LayerImpl {
*/
const std::string getType() const override { return RMSNormLayerCl::type; };

static opencl::Kernel kernel_rmsnorm;
static opencl::Kernel kernel_rmsnorm_fp16;

/**
* @brief Process data and dimensions for rms norm operation
* @param[in] input Tensor
Expand Down Expand Up @@ -153,12 +125,26 @@ class RMSNormLayerCl : public LayerImpl {
*/
void setProperty(const std::vector<std::string> &values) override;

/**
* @brief registerClKernels
*/
static bool registerClKernels();

inline static const std::string type = "rmsnorm";

private:
std::array<unsigned int, 1> wt_idx;
std::tuple<props::RMS_NORM_GAMMA_INIT_GPU, props::Epsilon>

std::tuple<props::RMS_NORM_GAMMA_INIT, props::Epsilon>
rmsnorm_props; /**< rmsnorm layer properties */

inline static std::vector<ClContext::SharedPtrClKernel>
layer_kernel_ptrs; /**< kernel list relevant with this layer */

enum Kernels {
RMSNORM_CL,
RMSNORM_CL_FP16,
};
};
} // namespace nntrainer

Expand Down
13 changes: 13 additions & 0 deletions nntrainer/layers/common_properties.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1075,6 +1075,19 @@ class BNPARAMS_BETA_INIT final : public EnumProperty<InitializerInfo> {
static constexpr const char *key = "beta_initializer";
};

/**
* @brief RMS_NORM_GAMMA_INIT Initialization Enumeration Information
*/
class RMS_NORM_GAMMA_INIT final : public EnumProperty<InitializerInfo> {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It is quite strange to use capital letters here.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Oh I checked BNPPARAMS_*_INIT only.
As you suggested, it seems more natural to update it to follow our convention :) I will apply it

public:
/**
* @brief Construct a RMS_NORM_GAMMA_INIT object
*/
RMS_NORM_GAMMA_INIT(Initializer value = Initializer::ONES) { set(value); };
using prop_tag = enum_class_prop_tag;
static constexpr const char *key = "gamma_initializer";
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It seems that there is a duplication of key value ("gamma_initializer") for BNPARAMS_GAMMA_INIT.
How about changing the BNPARAMS_GAMMA_INIT class to GammaInitializer and using it in common?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you! That's my bad. It must be critical if it wasn't fixed. I will update it.

};

/**
* @brief Enumeration of tensor regularization type
*/
Expand Down