Skip to content

N
ngraph
Commit 2e295d27 authored by fenglei.tian's avatar fenglei.tian
Browse files
Options

fix merge bug and apply clang format

parent 809dda4f
Show whitespace changes
Inline Side-by-side
Showing with 68 additions and 105 deletions
src/ngraph/runtime/gpu/gpu_emitter.cpp
View file @ 2e295d27
......@@ -91,20 +91,23 @@
#include "ngraph/ops/sum.hpp"
#include "ngraph/ops/tan.hpp"
#include "ngraph/ops/tanh.hpp"
#include "ngraph/runtime/gpu/gpu_emitter.hpp"
#include "ngraph/runtime/gpu/gpu_cuda_kernel_emitters.hpp"
#include "ngraph/runtime/gpu/gpu_emitter.hpp"
#include "ngraph/runtime/gpu/gpu_kernel_emitters.hpp"
#include "ngraph/util.hpp"
using namespace std;
using namespace ngraph;
namespace ngraph
{
}
namespace runtime
{
namespace gpu
{
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Abs)
{
{
if (out[0].get_size() == 0)
{
return;
......@@ -112,15 +115,15 @@ namespace ngraph
writer << "{ // " << node->get_name() << "\n";
writer.indent++;
writer << "int count = " << out[0].get_size() << ";\n";
writer << "ngraph::runtime::gpu::emit_abs((void*) " << args[0].get_name() << ", (void*) "
<< out[0].get_name() << ", count);\n";
writer << "ngraph::runtime::gpu::emit_abs((void*) " << args[0].get_name()
<< ", (void*) " << out[0].get_name() << ", count);\n";
writer.indent--;
writer << "}\n";
}
}
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Add)
{
{
if (out[0].get_size() == 0)
{
return;
......@@ -158,17 +161,17 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
<< "descriptor," << out[0].get_name() << ");\n";
writer.indent--;
writer << "}\n";
}
}
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Dot)
{
{
if (out[0].get_size() == 0)
{
return;
}
const ngraph::op::Dot* dot = static_cast<const ngraph::op::Dot*>(n);
const ngraph::op::Dot* dot = static_cast<const ngraph::op::Dot*>(node);
const Shape& arg0_shape = args[0].get_shape();
const Shape& arg1_shape = args[1].get_shape();
if (arg0_shape.empty() || arg1_shape.empty())
......@@ -184,7 +187,8 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
<< "1," << out[0].get_name() << ", 1);\n";
writer << "cublasSscal("
<< "cublas_handle,"
<< "count ," << first.get_name() << "," << out[0].get_name() << ", 1);\n";
<< "count ," << first.get_name() << "," << out[0].get_name()
<< ", 1);\n";
writer.indent--;
writer << "}\n";
return;
......@@ -195,8 +199,8 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
{
writer << "{ // " << node->get_name() << "\n";
writer.indent++;
writer << "runtime::gpu::cuda_memset(" << out[0].get_name() << ", 0, " << out[0].get_size()
<< " * sizeof(float));\n";
writer << "runtime::gpu::cuda_memset(" << out[0].get_name() << ", 0, "
<< out[0].get_size() << " * sizeof(float));\n";
writer.indent--;
writer << "}\n";
return;
......@@ -224,7 +228,8 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
<< "cublas_handle,"
<< "CUBLAS_OP_T," << arg0_shape[0] << "," << arg0_shape[1] << ","
<< "&alpha," // Alpha
<< args[0].get_name() << "," << arg0_shape[1] << "," << args[1].get_name() << ","
<< args[0].get_name() << "," << arg0_shape[1] << ","
<< args[1].get_name() << ","
<< "1,"
<< "&beta," // beta
<< out[0].get_name() << ","
......@@ -270,15 +275,14 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
}
else
{
throw std::runtime_error(node->get_name() + " with more then 2D is not implemented.");
throw std::runtime_error(node->get_name() +
" with more then 2D is not implemented.");
}
}
}
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Maximum)
{
{
if (out[0].get_size() == 0)
{
return;
......@@ -320,7 +324,7 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Minimum)
{
{
if (out[0].get_size() == 0)
{
return;
......@@ -362,7 +366,7 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Negative)
{
{
if (out[0].get_size() == 0)
{
return;
......@@ -400,35 +404,15 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
<< "descriptor," << out[0].get_name() << ");\n";
writer.indent--;
writer << "}\n";
}
}
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Broadcast)
{
if (out[0].get_size() == 0)
{
return;
}
auto broadcast = static_cast<const ngraph::op::Broadcast*>(n);
auto arg_shape = args[0].get_shape();
auto result_shape = out[0].get_shape();
auto& axes = broadcast->get_broadcast_axes();
//broadcast axes is empty, do a copy
if (axes.empty())
if (out[0].get_size() == 0)
{
writer << "{ // " << node->get_name() << " \n";
writer.indent++;
writer << "runtime::gpu::cuda_memcpyDtD(" << out[0].get_name() << ", " << args[0].get_name()
<< ", " << out[0].get_size() << " * " << out[0].get_element_type().size() << ");\n";
writer.indent--;
writer << "}\n";
return;
}
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Broadcast)
{
auto broadcast = static_cast<const ngraph::op::Broadcast*>(node);
auto arg_shape = args[0].get_shape();
auto result_shape = out[0].get_shape();
......@@ -490,50 +474,20 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
throw std::runtime_error(node->get_name() + " is not implemented.");
}
}
}
}
if (is_one_axes)
{
int repeat_times = 1;
for (int i = 0; i < axes_v.size(); i++)
{
repeat_times *= result_shape[axes_v[i]];
}
int repeat_size = 1;
for (int i = *axes_v.rbegin() + 1; i < result_shape.size(); i++)
{
repeat_size *= result_shape[i];
}
writer << "{ // " << node->get_name() << " \n";
writer.indent++;
writer << "runtime::gpu::emit_broadcast(" << args[0].get_name() << ", " << out[0].get_name()
<< ", " << repeat_size << ", " << repeat_times << ", " << out[0].get_size()
<< ");\n";
writer.indent--;
writer << "}\n";
}
else
{
throw std::runtime_error(node->get_name() + " is not implemented.");
}
}
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Constant)
{
}
{
}
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Reshape)
{
{
if (out[0].get_size() == 0)
{
return;
}
auto reshape = static_cast<const op::Reshape*>(n);
auto reshape = static_cast<const op::Reshape*>(node);
writer << "{ // " << node->get_name() << "\n";
writer.indent++;
auto arg_shape = args[0].get_shape();
......@@ -557,8 +511,9 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
{
writer << "{ // " << node->get_name() << " 1\n";
writer.indent++;
writer << "runtime::gpu::cuda_memcpyDtD(" << out[0].get_name() << ", " << args[0].get_name()
<< ", " << out[0].get_size() << " * " << out[0].get_element_type().size() << ");\n";
writer << "runtime::gpu::cuda_memcpyDtD(" << out[0].get_name() << ", "
<< args[0].get_name() << ", " << out[0].get_size() << " * "
<< out[0].get_element_type().size() << ");\n";
writer.indent--;
writer << "}\n";
}
......@@ -578,8 +533,8 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
<< "&alpha," // Alpha
<< args[0].get_name() << "," << arg_shape[1] << ","
<< "&beta," // beta
<< args[0].get_name() << "," << arg_shape[1] << "," << out[0].get_name() << ","
<< result_shape[1] << ");\n";
<< args[0].get_name() << "," << arg_shape[1] << "," << out[0].get_name()
<< "," << result_shape[1] << ");\n";
writer << "cublasSetPointerMode(cublas_handle, CUBLAS_POINTER_MODE_DEVICE);\n";
writer.indent--;
writer << "}\n";
......@@ -588,11 +543,12 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
else
{
throw runtime_error(
"Axis permutation in reshape is not implemented yet for tensors with rank>2");
"Axis permutation in reshape is not implemented yet for tensors with "
"rank>2");
}
writer.indent--;
writer << "}\n";
}
}
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::FunctionCall)
......@@ -601,7 +557,7 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Multiply)
{
{
if (out[0].get_size() == 0)
{
return;
......@@ -639,11 +595,11 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
<< "descriptor," << out[0].get_name() << ");\n";
writer.indent--;
writer << "}\n";
}
}
v template <>
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Sqrt)
{
{
if (out[0].get_size() == 0)
{
return;
......@@ -682,18 +638,19 @@ cudnnSetOpTensorDescriptor(opTensorDesc,
writer.indent--;
writer << "}\n";
}
}
}
}
template <>
void GPU_Emitter::EMITTER_DECL(ngraph::op::Result)
{
{
writer << "{ //" << node->get_name() << "\n";
writer.indent++;
writer << "runtime::gpu::cuda_memcpyDtD(" << out[0].get_name() << ", " << args[0].get_name()
<< ", " << out[0].get_size() << " * " << out[0].get_element_type().size() << ");\n";
writer << "runtime::gpu::cuda_memcpyDtD(" << out[0].get_name() << ", "
<< args[0].get_name() << ", " << out[0].get_size() << " * "
<< out[0].get_element_type().size() << ");\n";
writer.indent--;
writer << "}\n";
return;
}
}
}
}
src/ngraph/runtime/gpu/gpu_emitter.hpp
View file @ 2e295d27
......@@ -46,7 +46,8 @@ namespace ngraph
const std::vector<GPU_TensorViewWrapper>& args,
const std::vector<GPU_TensorViewWrapper>& out)
{
throw std::runtime_error("Unimplemented op in GPU emitter for " + node->get_name());
throw std::runtime_error("Unimplemented op in GPU emitter for " +
node->get_name());
}
static void nop(GPU_ExternalFunction* external_function,
......
src/ngraph/runtime/gpu/gpu_external_function.cpp
View file @ 2e295d27
......@@ -187,7 +187,8 @@ static const ngraph::runtime::gpu::OpMap dispatcher{
{TI(ngraph::op::Convert), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Convert>},
{TI(ngraph::op::Constant), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Constant>},
{TI(ngraph::op::Reshape), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Reshape>},
{TI(ngraph::op::FunctionCall), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::FunctionCall>},
{TI(ngraph::op::FunctionCall),
&ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::FunctionCall>},
{TI(ngraph::op::Reduce), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Reduce>},
{TI(ngraph::op::Sign), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Sign>},
{TI(ngraph::op::Slice), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Slice>},
......@@ -202,12 +203,14 @@ static const ngraph::runtime::gpu::OpMap dispatcher{
{TI(ngraph::op::Asin), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Asin>},
{TI(ngraph::op::Acos), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Acos>},
{TI(ngraph::op::Atan), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Atan>},
{TI(ngraph::op::ReplaceSlice), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::ReplaceSlice>},
{TI(ngraph::op::ReplaceSlice),
&ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::ReplaceSlice>},
{TI(ngraph::op::OneHot), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::OneHot>},
{TI(ngraph::op::Floor), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Floor>},
{TI(ngraph::op::Ceiling), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Ceiling>},
{TI(ngraph::op::Sqrt), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Sqrt>},
{TI(ngraph::op::Convolution), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Convolution>},
{TI(ngraph::op::Convolution),
&ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Convolution>},
{TI(ngraph::op::ConvolutionBackpropFilters),
&ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::ConvolutionBackpropFilters>},
{TI(ngraph::op::ConvolutionBackpropData),
......@@ -216,7 +219,8 @@ static const ngraph::runtime::gpu::OpMap dispatcher{
{TI(ngraph::op::MaxPool), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::MaxPool>},
{TI(ngraph::op::Reverse), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Reverse>},
{TI(ngraph::op::Result), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Result>},
{TI(ngraph::op::ReduceWindow), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::ReduceWindow>},
{TI(ngraph::op::ReduceWindow),
&ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::ReduceWindow>},
{TI(ngraph::op::SelectAndScatter),
&ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::SelectAndScatter>},
{TI(ngraph::op::AvgPool), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::AvgPool>},
......@@ -232,7 +236,8 @@ static const ngraph::runtime::gpu::OpMap dispatcher{
{TI(ngraph::op::Max), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Max>},
{TI(ngraph::op::Min), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Min>},
{TI(ngraph::op::Relu), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Relu>},
{TI(ngraph::op::ReluBackprop), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::ReluBackprop>},
{TI(ngraph::op::ReluBackprop),
&ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::ReluBackprop>},
{TI(ngraph::op::Softmax), &ngraph::runtime::gpu::GPU_Emitter::emit<ngraph::op::Softmax>},
};
......@@ -564,8 +569,8 @@ using namespace std;
size_t temp_pool_size = current_function->get_temporary_pool_size();
writer << "// Allocate the memory pool\n";
// TODO memory pool malloc.
writer << "void* pool_base_ptr = ngraph::runtime::gpu::create_gpu_buffer(" << temp_pool_size
<< ");\n";
writer << "void* pool_base_ptr = ngraph::runtime::gpu::create_gpu_buffer("
<< temp_pool_size << ");\n";
// Add temporaries to the variable name map
for (shared_ptr<Node> node : current_function->get_ordered_ops())
......@@ -641,9 +646,9 @@ using namespace std;
{
if (contains(constants, tv.get()))
{
writer << "ngraph::runtime::gpu::cuda_memcpyHtD(outputs[" << output_index << "], "
<< tv->get_tensor().get_name() << ", " << tv->get_tensor().size()
<< ");\n";
writer << "ngraph::runtime::gpu::cuda_memcpyHtD(outputs[" << output_index
<< "], " << tv->get_tensor().get_name() << ", "
<< tv->get_tensor().size() << ");\n";
}
else
{
......
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