Skip to content

N
ngraph
Unverified Commit d52473c8 authored by Robert Kimball's avatar Robert Kimball Committed by GitHub
Browse files
Options

Change GPU backend to use op_tbl (#1618) 

* sort op list

* use op_tbl

* throw unsupported_op exception when appropriate

* remove dead code

* Add more use of NGRAPH_OP macro to remove boilerplate definitions/implementations

* revert moving class out of namespace

* change from switch to dispatcher map
parent 3d69bf7a
Hide whitespace changes
Inline Side-by-side
Showing with 33 additions and 139 deletions
src/ngraph/runtime/gpu/gpu_emitter.cpp
View file @ d52473c8
This source diff could not be displayed because it is too large. You can view the blob instead.
src/ngraph/runtime/gpu/gpu_emitter.hpp
View file @ d52473c8
......@@ -24,12 +24,6 @@
#include "ngraph/runtime/gpu/gpu_external_function.hpp"
#include "ngraph/runtime/gpu/gpu_tensor_view_wrapper.hpp"
#define EMITTER_DECL(op_name) \
emit<op_name>(GPU_ExternalFunction * external_function, \
codegen::CodeWriter & writer, \
const ngraph::Node* node, \
const std::vector<GPU_TensorViewWrapper>& args, \
const std::vector<GPU_TensorViewWrapper>& out)
namespace ngraph
{
namespace runtime
......@@ -39,31 +33,17 @@ namespace ngraph
class GPU_Emitter
{
public:
template <typename OP>
static void emit(GPU_ExternalFunction* external_function,
codegen::CodeWriter& writer,
const ngraph::Node* node,
const std::vector<GPU_TensorViewWrapper>& args,
const std::vector<GPU_TensorViewWrapper>& out)
{
throw std::runtime_error("Unimplemented op '" + node->description() +
"' in GPU emitter");
}
static std::function<void(EMIT_ARGS)> get_emit_function(const Node& node);
static void nop(GPU_ExternalFunction* external_function,
codegen::CodeWriter& writer,
const ngraph::Node* node,
const std::vector<GPU_TensorViewWrapper>& args,
const std::vector<GPU_TensorViewWrapper>& out)
{
}
// This defines a collection of function declarations like this
// static void emit_Abs(EMIT_ARGS);
// static void emit_Acos(EMIT_ARGS);
#define NGRAPH_OP(a) static void emit_##a(EMIT_ARGS);
#include "ngraph/op/op_tbl.hpp"
#undef NGRAPH_OP
template <typename T>
static void emit_elementwise(GPU_ExternalFunction* external_function,
codegen::CodeWriter& writer,
const ngraph::Node* node,
const std::vector<GPU_TensorViewWrapper>& args,
const std::vector<GPU_TensorViewWrapper>& out)
static void emit_elementwise(EMIT_ARGS)
{
if (out[0].get_size() == 0)
{
......@@ -104,6 +84,7 @@ namespace ngraph
static std::string node_names(const std::vector<GPU_TensorViewWrapper>& args,
std::initializer_list<int> arg_indexes = {});
};
Shape get_padded_shape(const Shape& input_shape,
const Shape& padding_below,
const Shape& padding_above,
......
src/ngraph/runtime/gpu/gpu_external_function.cpp
View file @ d52473c8
......@@ -24,6 +24,7 @@
#include <string>
#include <tuple>
#include "ngraph/codegen/code_writer.hpp"
#include "ngraph/descriptor/input.hpp"
#include "ngraph/descriptor/layout/dense_tensor_layout.hpp"
#include "ngraph/descriptor/output.hpp"
......@@ -36,6 +37,8 @@
#include "ngraph/op/add.hpp"
#include "ngraph/op/allreduce.hpp"
#include "ngraph/op/and.hpp"
#include "ngraph/op/argmax.hpp"
#include "ngraph/op/argmin.hpp"
#include "ngraph/op/asin.hpp"
#include "ngraph/op/atan.hpp"
#include "ngraph/op/avg_pool.hpp"
......@@ -60,6 +63,7 @@
#include "ngraph/op/less.hpp"
#include "ngraph/op/less_eq.hpp"
#include "ngraph/op/log.hpp"
#include "ngraph/op/lrn.hpp"
#include "ngraph/op/max.hpp"
#include "ngraph/op/max_pool.hpp"
#include "ngraph/op/maximum.hpp"
......@@ -93,10 +97,12 @@
#include "ngraph/op/slice.hpp"
#include "ngraph/op/softmax.hpp"
#include "ngraph/op/sqrt.hpp"
#include "ngraph/op/stop_gradient.hpp"
#include "ngraph/op/subtract.hpp"
#include "ngraph/op/sum.hpp"
#include "ngraph/op/tan.hpp"
#include "ngraph/op/tanh.hpp"
#include "ngraph/op/topk.hpp"
#include "ngraph/pass/common_function_collection.hpp"
#include "ngraph/pass/like_replacement.hpp"
#include "ngraph/runtime/gpu/gpu_backend.hpp"
......@@ -104,6 +110,7 @@
#include "ngraph/runtime/gpu/gpu_external_function.hpp"
#include "ngraph/runtime/gpu/gpu_kernel_emitters.hpp"
#include "ngraph/runtime/gpu/gpu_runtime_context.hpp"
#include "ngraph/runtime/gpu/gpu_tensor_view_wrapper.hpp"
#include "ngraph/runtime/gpu/pass/gpu_layout.hpp"
#include "ngraph/runtime/gpu/pass/tensor_memory_reservation.hpp"
......@@ -157,91 +164,15 @@ static string emit_string_array(const vector<string>& s, size_t max_line_length)
static GPUStaticInitializers s_static_initializers;
#define TI(x) type_index(typeid(x))
static const runtime::gpu::OpMap dispatcher{
{TI(ngraph::op::Add), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Add>},
{TI(ngraph::op::Dot), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Dot>},
{TI(ngraph::op::Multiply), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Multiply>},
{TI(ngraph::op::Parameter), &runtime::gpu::GPU_Emitter::nop},
{TI(ngraph::op::Abs), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Abs>},
{TI(ngraph::op::Concat), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Concat>},
{TI(ngraph::op::Divide), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Divide>},
{TI(ngraph::op::Equal), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Equal>},
{TI(ngraph::op::GetOutputElement),
&runtime::gpu::GPU_Emitter::emit<ngraph::op::GetOutputElement>},
{TI(ngraph::op::Greater), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Greater>},
{TI(ngraph::op::GreaterEq),
&runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::GreaterEq>},
{TI(ngraph::op::Less), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Less>},
{TI(ngraph::op::LessEq), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::LessEq>},
{TI(ngraph::op::Log), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Log>},
{TI(ngraph::op::Maximum), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Maximum>},
{TI(ngraph::op::Minimum), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Minimum>},
{TI(ngraph::op::Negative), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Negative>},
{TI(ngraph::op::NotEqual), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::NotEqual>},
{TI(ngraph::op::Power), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Power>},
{TI(ngraph::op::Select), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Select>},
{TI(ngraph::op::Subtract), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Subtract>},
{TI(ngraph::op::Broadcast), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Broadcast>},
{TI(ngraph::op::Convert), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Convert>},
{TI(ngraph::op::Constant), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Constant>},
{TI(ngraph::op::Reshape), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Reshape>},
{TI(ngraph::op::FunctionCall), &runtime::gpu::GPU_Emitter::emit<ngraph::op::FunctionCall>},
{TI(ngraph::op::Reduce), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Reduce>},
{TI(ngraph::op::Sign), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Sign>},
{TI(ngraph::op::Slice), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Slice>},
{TI(ngraph::op::Sum), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Sum>},
{TI(ngraph::op::Exp), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Exp>},
{TI(ngraph::op::Sin), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Sin>},
{TI(ngraph::op::Sinh), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Sinh>},
{TI(ngraph::op::Cos), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Cos>},
{TI(ngraph::op::Cosh), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Cosh>},
{TI(ngraph::op::Tan), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Tan>},
{TI(ngraph::op::Tanh), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Tanh>},
{TI(ngraph::op::Asin), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Asin>},
{TI(ngraph::op::Acos), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Acos>},
{TI(ngraph::op::Atan), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Atan>},
{TI(ngraph::op::ReplaceSlice), &runtime::gpu::GPU_Emitter::emit<ngraph::op::ReplaceSlice>},
{TI(ngraph::op::OneHot), &runtime::gpu::GPU_Emitter::emit<ngraph::op::OneHot>},
{TI(ngraph::op::Floor), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Floor>},
{TI(ngraph::op::Ceiling), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Ceiling>},
{TI(ngraph::op::Sqrt), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Sqrt>},
{TI(ngraph::op::Convolution), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Convolution>},
{TI(ngraph::op::ConvolutionBackpropFilters),
&runtime::gpu::GPU_Emitter::emit<ngraph::op::ConvolutionBackpropFilters>},
{TI(ngraph::op::ConvolutionBackpropData),
&runtime::gpu::GPU_Emitter::emit<ngraph::op::ConvolutionBackpropData>},
{TI(ngraph::op::Not), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Not>},
{TI(ngraph::op::MaxPool), &runtime::gpu::GPU_Emitter::emit<ngraph::op::MaxPool>},
{TI(ngraph::op::Reverse), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Reverse>},
{TI(ngraph::op::ReverseSequence),
&runtime::gpu::GPU_Emitter::emit<ngraph::op::ReverseSequence>},
{TI(ngraph::op::Result), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Result>},
{TI(ngraph::op::ReduceWindow), &runtime::gpu::GPU_Emitter::emit<ngraph::op::ReduceWindow>},
{TI(ngraph::op::SelectAndScatter),
&runtime::gpu::GPU_Emitter::emit<ngraph::op::SelectAndScatter>},
{TI(ngraph::op::AvgPool), &runtime::gpu::GPU_Emitter::emit<ngraph::op::AvgPool>},
{TI(ngraph::op::AvgPoolBackprop),
&runtime::gpu::GPU_Emitter::emit<ngraph::op::AvgPoolBackprop>},
{TI(ngraph::op::Pad), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Pad>},
{TI(ngraph::op::BatchNorm), &runtime::gpu::GPU_Emitter::emit<ngraph::op::BatchNorm>},
{TI(ngraph::op::BatchNormBackprop),
&runtime::gpu::GPU_Emitter::emit<ngraph::op::BatchNormBackprop>},
{TI(ngraph::op::MaxPoolBackprop),
&runtime::gpu::GPU_Emitter::emit<ngraph::op::MaxPoolBackprop>},
{TI(ngraph::op::Product), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Product>},
{TI(ngraph::op::Max), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Max>},
{TI(ngraph::op::Min), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Min>},
{TI(ngraph::op::Relu), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Relu>},
{TI(ngraph::op::ReluBackprop),
&runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::ReluBackprop>},
{TI(ngraph::op::Softmax), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Softmax>},
{TI(ngraph::op::Sigmoid), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Sigmoid>},
{TI(ngraph::op::SigmoidBackprop),
&runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::SigmoidBackprop>},
{TI(ngraph::op::And), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::And>},
{TI(ngraph::op::Or), &runtime::gpu::GPU_Emitter::emit_elementwise<ngraph::op::Or>}};
void runtime::gpu::GPU_ExternalFunction::emit_op(GPU_ExternalFunction* external_function,
codegen::CodeWriter& writer,
const ngraph::Node* node,
const std::vector<GPU_TensorViewWrapper>& args,
const std::vector<GPU_TensorViewWrapper>& out)
{
auto emit_function = GPU_Emitter::get_emit_function(*node);
emit_function(external_function, writer, node, args, out);
};
const size_t runtime::gpu::GPU_ExternalFunction::GPU_ExternalFunction::s_memory_pool_alignment = 64;
......@@ -548,14 +479,6 @@ void runtime::gpu::GPU_ExternalFunction::emit_functions()
for (shared_ptr<Node> node : m_function_ordered_ops.at(current_function))
{
auto& n = *node;
// Work around a compiler warning (*node inside typeid may have effects
// with shared pointers, which is fine here but clang doesn't like it.)
auto handler = dispatcher.find(type_index(typeid(n)));
if (handler == dispatcher.end())
{
throw ngraph::unsupported_op(node->description());
}
vector<GPU_TensorViewWrapper> in;
vector<string> node_input_names;
vector<string> node_output_names;
......@@ -590,7 +513,7 @@ void runtime::gpu::GPU_ExternalFunction::emit_functions()
auto it = m_node_function_map.find(node.get());
if (it == m_node_function_map.end())
{
handler->second(this, m_writer, node.get(), in, out);
emit_op(this, m_writer, node.get(), in, out);
}
else
{
......@@ -747,13 +670,6 @@ string runtime::gpu::GPU_ExternalFunction::emit_op_as_function(const Node& node,
codegen::CodeWriter writer;
writer << "static void " << function_name << "(";
writer.indent++;
// Work around a compiler warning (*node inside typeid may have effects
// with shared pointers, which is fine here but clang doesn't like it.)
auto handler = dispatcher.find(type_index(typeid(node)));
if (handler == dispatcher.end())
{
throw ngraph::unsupported_op(node.description());
}
vector<GPU_TensorViewWrapper> in;
size_t arg_index = 0;
set<string> arg_names;
......@@ -791,7 +707,7 @@ string runtime::gpu::GPU_ExternalFunction::emit_op_as_function(const Node& node,
writer.indent--;
writer << "\n)\n";
codegen::CodeWriter tmp_writer;
handler->second(this, tmp_writer, &node, in, out);
emit_op(this, tmp_writer, &node, in, out);
string body = tmp_writer.get_code();
if (body.size() > 0 && body[0] == '{')
{
......
src/ngraph/runtime/gpu/gpu_external_function.hpp
View file @ d52473c8
......@@ -36,6 +36,11 @@
#include "ngraph/runtime/gpu/gpu_primitive_emitter.hpp"
#include "ngraph/runtime/gpu/gpu_tensor_view_wrapper.hpp"
#define EMIT_ARGS \
runtime::gpu::GPU_ExternalFunction *external_function, codegen::CodeWriter &writer, \
const Node *node, const std::vector<runtime::gpu::GPU_TensorViewWrapper> &args, \
const std::vector<runtime::gpu::GPU_TensorViewWrapper> &out
namespace ngraph
{
namespace runtime
......@@ -46,15 +51,6 @@ namespace ngraph
class GPU_CallFrame;
struct GPURuntimeContext;
using OpFunction =
std::function<void(GPU_ExternalFunction* external_function,
codegen::CodeWriter&,
const ngraph::Node*,
const std::vector<GPU_TensorViewWrapper>& inputs,
const std::vector<GPU_TensorViewWrapper>& outputs)>;
using OpMap = std::unordered_map<std::type_index, OpFunction>;
class GPU_ExternalFunction : public std::enable_shared_from_this<GPU_ExternalFunction>
{
friend class GPU_CallFrame;
......@@ -97,6 +93,7 @@ namespace ngraph
void emit_debug_function_entry(Node* node);
void emit_debug_function_exit(Node* node);
void emit_temp_mem_pool_allocation(std::shared_ptr<Function> current_function);
void emit_op(EMIT_ARGS);
void release_function() { m_function = nullptr; }
void store_emitted_functions(const std::string& code);
std::string emit_op_as_function(const Node& node, const std::string& function_name);
......
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