Skip to content

N
ngraph
Commit 67248fdb authored by Sergey Shalnov's avatar Sergey Shalnov Committed by Robert Kimball
Browse files
Options

IntelGPU backend: Custom kernels refactoring 3 (#2787)

parent 2b13ae40
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 91 additions and 119 deletions
src/ngraph/runtime/intelgpu/intelgpu_kernels.cpp
View file @ 67248fdb
......@@ -44,3 +44,15 @@ void runtime::intelgpu::CustomKernels::queue_krnl(const krnl_info& krnl_info,
stream.add(kernel_item);
}
}
void runtime::intelgpu::arguments_check(const shared_ptr<Node>& op, size_t input, size_t output)
{
if (op->get_input_size() != input || op->get_output_size() != output)
{
ostringstream os;
os << "Operation \"" << op->description() << "\" input and output sizes mismatch."
<< " Expected input size=" << input << ", provided=" << op->get_input_size()
<< ". Expected output size=" << output << ", provided=" << op->get_output_size();
throw invalid_argument(os.str());
}
}
src/ngraph/runtime/intelgpu/intelgpu_kernels.hpp
View file @ 67248fdb
......@@ -24,11 +24,20 @@
#include "ngraph/node.hpp"
#include "ngraph/op/all.hpp"
#include "ngraph/op/and.hpp"
#include "ngraph/op/any.hpp"
#include "ngraph/op/batch_norm.hpp"
#include "ngraph/op/broadcast.hpp"
#include "ngraph/op/convolution.hpp"
#include "ngraph/op/equal.hpp"
#include "ngraph/op/greater.hpp"
#include "ngraph/op/greater_eq.hpp"
#include "ngraph/op/less.hpp"
#include "ngraph/op/less_eq.hpp"
#include "ngraph/op/max.hpp"
#include "ngraph/op/min.hpp"
#include "ngraph/op/not_equal.hpp"
#include "ngraph/op/or.hpp"
#include "ngraph/op/product.hpp"
#include "ngraph/op/select.hpp"
#include "ngraph/op/slice.hpp"
......@@ -43,6 +52,8 @@ namespace ngraph
{
class CustomKernelInfo;
class CustomKernels;
void arguments_check(const std::shared_ptr<Node>& op, size_t input, size_t output);
}
}
}
......@@ -107,13 +118,24 @@ private:
void queue_krnl(const krnl_info& krn_info, const std::shared_ptr<Node>& op);
krnl_info build_krnl(const std::shared_ptr<op::All>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::And>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Any>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::BatchNormInference>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::BatchNormTraining>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::BatchNormTrainingBackprop>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Broadcast>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Convolution>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::ConvolutionBackpropData>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::ConvolutionBackpropFilters>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Equal>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Greater>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::GreaterEq>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Less>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::LessEq>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Max>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Min>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::NotEqual>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Or>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Product>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Select>& op) const;
krnl_info build_krnl(const std::shared_ptr<op::Slice>& op) const;
......
src/ngraph/runtime/intelgpu/intelgpu_op_batchnorm.hpp deleted 100644 → 0
View file @ 2b13ae40
//*****************************************************************************
// Copyright 2017-2019 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//*****************************************************************************
#pragma once
#include <CPP/topology.hpp>
#include "ngraph/shape.hpp"
#include "ngraph/type/element_type.hpp"
namespace ngraph
{
namespace runtime
{
namespace intelgpu
{
// This implements BatchNorm nGraph operation
// nGraph uses channels in this operation but clDNN uses full input data
void do_batch_norm_operation(cldnn::topology& topology,
const std::string& output_name,
const element::Type& output_type,
double eps,
const std::string& input_name,
const Shape& input_shape,
const std::string& gamma_name,
const std::string& beta_name,
const std::string& mean_name,
const std::string& variance_name);
// This creates mean of the input matrix by Channel axis
void do_create_mean(cldnn::topology& topology,
const std::string& output_name,
const element::Type& output_type,
const std::string& input_name,
const Shape& input_shape,
bool backward);
// This creates variance of the input matrix by Channel axis
void do_create_variance(cldnn::topology& topology,
const std::string& output_name,
const element::Type& output_type,
const std::string& input_name,
const Shape& input_shape,
const std::string& mean_name);
// This creates variance backprop of the input matrix by Channel axis
void do_create_variance_back(cldnn::topology& topology,
const std::string& output_name,
const element::Type& output_type,
double eps,
const std::string& input_name,
const Shape& input_shape,
const std::string& mean_name,
const std::string& variance_name,
const std::string& delta_name);
// This function uses "shape" parameter as input or output Shape
// Shape of all other calculated as first axis from the left
// Example: output[ 4, 3, 2, 8 ] means out_gamma[ 3 ]
void do_batch_norm_backprop_operation(cldnn::topology& topology,
const Shape& shape,
const element::Type& type,
const std::string& gamma_name,
const std::string& beta_name,
const std::string& input_name,
const std::string& mean_name,
const std::string& variance_name,
const std::string& delta_name,
double eps,
const std::string& output_name,
const std::string& output_gamma_name,
const std::string& output_beta_name);
}
}
}
src/ngraph/runtime/intelgpu/intelgpu_op_custom_kernels.cpp
View file @ 67248fdb
......@@ -1277,18 +1277,16 @@ CustomKernels::krnl_info CustomKernels::build_krnl(const shared_ptr<op::Select>&
return {krn_ret};
}
void runtime::intelgpu::do_logic_kernel(cldnn::topology& topology,
const string& input0_name,
const Shape& input0_shape,
const element::Type& input0_type,
const string& input1_name,
const Shape& input1_shape,
const string& output_name,
const Shape& output_shape,
const element::Type& output_type,
const string& operation)
static CustomKernels::krnl_info do_logic_kernel(const shared_ptr<Node>& op, const string& operation)
{
const cldnn::layout layout = IntelGPULayout::create_cldnn_layout(output_type, output_shape);
const string& input0_name = op->get_input_tensor_name(0);
const Shape& input0_shape = op->get_input_shape(0);
const element::Type& input0_type = op->get_input_element_type(0);
const string& input1_name = op->get_input_tensor_name(1);
const Shape& input1_shape = op->get_input_shape(1);
const string& output_name = op->get_output_tensor_name(0);
const Shape& output_shape = op->get_output_shape(0);
const element::Type& output_type = op->get_output_element_type(0);
const string entry_point_name = "logic_" + output_name;
CodeWriter writer;
vector<size_t> gws;
......@@ -1313,15 +1311,14 @@ void runtime::intelgpu::do_logic_kernel(cldnn::topology& topology,
}
writer.block_end();
const cldnn::custom_gpu_primitive op_logical(output_name,
{input0_name, input1_name},
{writer.get_code()},
entry_point_name,
get_kernel_args(2, 1),
"",
layout,
gws);
topology.add(op_logical);
const CustomKernelInfo op_logical(output_name,
output_shape,
output_type,
{input0_name, input1_name},
{writer.get_code()},
entry_point_name,
gws);
return {op_logical};
}
void runtime::intelgpu::do_eltwise_kernel(cldnn::topology& topology,
......@@ -2333,3 +2330,43 @@ size_t runtime::intelgpu::get_max_memory_rss()
return result;
}
CustomKernels::krnl_info CustomKernels::build_krnl(const shared_ptr<op::And>& op) const
{
return do_logic_kernel(op, " && ");
}
CustomKernels::krnl_info CustomKernels::build_krnl(const shared_ptr<op::Equal>& op) const
{
return do_logic_kernel(op, " == ");
}
CustomKernels::krnl_info CustomKernels::build_krnl(const shared_ptr<op::Greater>& op) const
{
return do_logic_kernel(op, " > ");
}
CustomKernels::krnl_info CustomKernels::build_krnl(const shared_ptr<op::GreaterEq>& op) const
{
return do_logic_kernel(op, " >= ");
}
CustomKernels::krnl_info CustomKernels::build_krnl(const shared_ptr<op::Less>& op) const
{
return do_logic_kernel(op, " < ");
}
CustomKernels::krnl_info CustomKernels::build_krnl(const shared_ptr<op::LessEq>& op) const
{
return do_logic_kernel(op, " <= ");
}
CustomKernels::krnl_info CustomKernels::build_krnl(const shared_ptr<op::NotEqual>& op) const
{
return do_logic_kernel(op, " != ");
}
CustomKernels::krnl_info CustomKernels::build_krnl(const shared_ptr<op::Or>& op) const
{
return do_logic_kernel(op, " || ");
}
src/ngraph/runtime/intelgpu/intelgpu_op_custom_kernels.hpp
View file @ 67248fdb
......@@ -100,17 +100,6 @@ namespace ngraph
const element::Type& output_type,
size_t concat_axis);
void do_logic_kernel(cldnn::topology& topology,
const std::string& input0_name,
const Shape& input0_shape,
const element::Type& input0_type,
const std::string& input1_name,
const Shape& input1_shape,
const std::string& output_name,
const Shape& output_shape,
const element::Type& output_type,
const std::string& operation);
void do_eltwise_kernel(cldnn::topology& topology,
const std::string& input0_name,
const Shape& input0_shape,
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment