Skip to content

N
ngraph
Unverified Commit b5467550 authored by Chris Sullivan's avatar Chris Sullivan Committed by GitHub
Browse files
Options

Updated gpu cpp files with consistent use of namespaces (cosmetic) (#629) 

* Updated namespace use in cpp files.
parent a32fdab5
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 150 additions and 187 deletions
src/ngraph/runtime/gpu/gpu_cuda_context_manager.cpp
View file @ b5467550
......@@ -19,25 +19,18 @@
#include "ngraph/runtime/gpu/gpu_cuda_context_manager.hpp"
namespace ngraph
using namespace ngraph;
runtime::gpu::CudaContextManager& runtime::gpu::CudaContextManager::instance()
{
namespace runtime
{
namespace gpu
{
CudaContextManager& CudaContextManager::instance()
{
static CudaContextManager manager;
return manager;
}
static CudaContextManager manager;
return manager;
}
CudaContextManager::CudaContextManager()
{
CUDA_SAFE_CALL(cuInit(0));
CUDA_SAFE_CALL(cuDeviceGet(&m_device, 0));
CUDA_SAFE_CALL(cuCtxCreate(&m_context, 0, m_device));
m_context_ptr = std::make_shared<CUcontext>(m_context);
}
}
}
runtime::gpu::CudaContextManager::CudaContextManager()
{
CUDA_SAFE_CALL(cuInit(0));
CUDA_SAFE_CALL(cuDeviceGet(&m_device, 0));
CUDA_SAFE_CALL(cuCtxCreate(&m_context, 0, m_device));
m_context_ptr = std::make_shared<CUcontext>(m_context);
}
src/ngraph/runtime/gpu/gpu_cuda_function_builder.cpp
View file @ b5467550
......@@ -20,46 +20,39 @@
#include "ngraph/runtime/gpu/gpu_cuda_function_builder.hpp"
#include "ngraph/runtime/gpu/gpu_util.hpp"
namespace ngraph
using namespace ngraph;
std::shared_ptr<CUfunction> runtime::gpu::CudaFunctionBuilder::get(const std::string& name,
const std::string& kernel,
int number_of_options,
const char** options)
{
namespace runtime
{
namespace gpu
{
std::shared_ptr<CUfunction> CudaFunctionBuilder::get(const std::string& name,
const std::string& kernel,
int number_of_options,
const char** options)
{
nvrtcProgram prog;
NVRTC_SAFE_CALL(nvrtcCreateProgram(&prog,
kernel.c_str(),
"op.cu",
0, // numHeaders
NULL, // headers
NULL)); // includeNames
nvrtcProgram prog;
NVRTC_SAFE_CALL(nvrtcCreateProgram(&prog,
kernel.c_str(),
"op.cu",
0, // numHeaders
NULL, // headers
NULL)); // includeNames
nvrtcResult compile_result = nvrtcCompileProgram(prog, number_of_options, options);
nvrtcResult compile_result = nvrtcCompileProgram(prog, number_of_options, options);
if (compile_result != NVRTC_SUCCESS)
{
throw std::runtime_error("compile error: \n" + kernel + "\n options");
}
if (compile_result != NVRTC_SUCCESS)
{
throw std::runtime_error("compile error: \n" + kernel + "\n options");
}
size_t ptx_size;
NVRTC_SAFE_CALL(nvrtcGetPTXSize(prog, &ptx_size));
char* ptx = new char[ptx_size];
NVRTC_SAFE_CALL(nvrtcGetPTX(
prog,
size_t ptx_size;
NVRTC_SAFE_CALL(nvrtcGetPTXSize(prog, &ptx_size));
char* ptx = new char[ptx_size];
NVRTC_SAFE_CALL(
nvrtcGetPTX(prog,
ptx)); // Load the generated PTX and get a handle to the parent kernel.
NVRTC_SAFE_CALL(nvrtcDestroyProgram(&prog)); // Destroy the program.
NVRTC_SAFE_CALL(nvrtcDestroyProgram(&prog)); // Destroy the program.
CUmodule module;
CUfunction function;
CUDA_SAFE_CALL(cuModuleLoadDataEx(&module, ptx, 0, 0, 0));
CUDA_SAFE_CALL(cuModuleGetFunction(&function, module, name.c_str()));
return std::make_shared<CUfunction>(function);
}
}
}
CUmodule module;
CUfunction function;
CUDA_SAFE_CALL(cuModuleLoadDataEx(&module, ptx, 0, 0, 0));
CUDA_SAFE_CALL(cuModuleGetFunction(&function, module, name.c_str()));
return std::make_shared<CUfunction>(function);
}
src/ngraph/runtime/gpu/gpu_cuda_function_pool.cpp
View file @ b5467550
......@@ -26,40 +26,31 @@
static const std::string s_output_dir = "gpu_codegen";
namespace ngraph
using namespace ngraph;
runtime::gpu::CudaFunctionPool& runtime::gpu::CudaFunctionPool::instance()
{
namespace runtime
{
namespace gpu
{
CudaFunctionPool& CudaFunctionPool::instance()
{
static CudaFunctionPool pool;
return pool;
}
static CudaFunctionPool pool;
return pool;
}
void CudaFunctionPool::set(const std::string& name, const std::string& kernel)
{
const char* opts[] = {"--gpu-architecture=compute_35",
"--relocatable-device-code=true"};
std::string filename =
file_util::path_join(s_output_dir, "cuda_kernel_" + name + "_codegen.cu");
std::ofstream out(filename);
out << kernel;
out.close();
m_function_map.insert(
{name, CudaFunctionBuilder::get("cuda_" + name, kernel, 2, opts)});
}
void runtime::gpu::CudaFunctionPool::set(const std::string& name, const std::string& kernel)
{
const char* opts[] = {"--gpu-architecture=compute_35", "--relocatable-device-code=true"};
std::string filename =
file_util::path_join(s_output_dir, "cuda_kernel_" + name + "_codegen.cu");
std::ofstream out(filename);
out << kernel;
out.close();
m_function_map.insert({name, CudaFunctionBuilder::get("cuda_" + name, kernel, 2, opts)});
}
std::shared_ptr<CUfunction> CudaFunctionPool::get(const std::string& name)
{
auto it = m_function_map.find(name);
if (it != m_function_map.end())
{
return (*it).second;
}
return nullptr;
}
}
std::shared_ptr<CUfunction> runtime::gpu::CudaFunctionPool::get(const std::string& name)
{
auto it = m_function_map.find(name);
if (it != m_function_map.end())
{
return (*it).second;
}
return nullptr;
}
src/ngraph/runtime/gpu/gpu_cuda_kernel_builder.cpp
View file @ b5467550
......@@ -16,74 +16,67 @@
#include "ngraph/runtime/gpu/gpu_cuda_kernel_builder.hpp"
#include "ngraph/codegen/code_writer.hpp"
namespace ngraph
using namespace ngraph;
void runtime::gpu::CudaKernelBuilder::get_elementwise_op(codegen::CodeWriter& writer,
const std::string& name,
const std::string& data_type,
const std::string& op,
const size_t& num_inputs)
{
namespace runtime
writer << "extern \"C\" __global__ void cuda_" << name << "(";
for (size_t i = 0; i < num_inputs; i++)
{
writer << data_type << "* in" << i << ", ";
}
writer << data_type << "* out,"
<< "size_t n)\n";
writer << "{\n";
writer.indent++;
{
namespace gpu
writer << "size_t tid = blockIdx.x * blockDim.x + threadIdx.x; \n";
writer << "if (tid < n)\n";
writer << "{\n";
writer.indent++;
{
void CudaKernelBuilder::get_elementwise_op(codegen::CodeWriter& writer,
const std::string& name,
const std::string& data_type,
const std::string& op,
const size_t& num_inputs)
writer << "out[tid] = " << op << "(";
for (size_t i = 0; i < num_inputs - 1; i++)
{
writer << "extern \"C\" __global__ void cuda_" << name << "(";
for (size_t i = 0; i < num_inputs; i++)
{
writer << data_type << "* in" << i << ", ";
}
writer << data_type << "* out,"
<< "size_t n)\n";
writer << "{\n";
writer.indent++;
{
writer << "size_t tid = blockIdx.x * blockDim.x + threadIdx.x; \n";
writer << "if (tid < n)\n";
writer << "{\n";
writer.indent++;
{
writer << "out[tid] = " << op << "(";
for (size_t i = 0; i < num_inputs - 1; i++)
{
writer << "in" << i << "[tid], ";
}
writer << "in" << num_inputs - 1 << "[tid]);\n";
}
writer.indent--;
writer << "}\n";
}
writer.indent--;
writer << "}\n";
return;
writer << "in" << i << "[tid], ";
}
writer << "in" << num_inputs - 1 << "[tid]);\n";
}
writer.indent--;
writer << "}\n";
}
writer.indent--;
writer << "}\n";
void CudaKernelBuilder::get_device_helper(codegen::CodeWriter& writer,
const std::string& name,
const std::string& data_type,
const std::string& math_kernel,
const size_t& num_inputs)
{
if (math_kernel.size())
{
writer << "__device__ " << data_type << " " << name << "(";
for (size_t i = 0; i < num_inputs - 1; i++)
{
writer << data_type << " x" << i << ", ";
}
writer << data_type << " x" << num_inputs - 1;
writer << ")\n";
writer << "{\n";
writer.indent++;
{
writer << "return " + math_kernel << ";\n";
}
writer.indent--;
writer << "}\n";
}
return;
}
return;
}
void runtime::gpu::CudaKernelBuilder::get_device_helper(codegen::CodeWriter& writer,
const std::string& name,
const std::string& data_type,
const std::string& math_kernel,
const size_t& num_inputs)
{
if (math_kernel.size())
{
writer << "__device__ " << data_type << " " << name << "(";
for (size_t i = 0; i < num_inputs - 1; i++)
{
writer << data_type << " x" << i << ", ";
}
writer << data_type << " x" << num_inputs - 1;
writer << ")\n";
writer << "{\n";
writer.indent++;
{
writer << "return " + math_kernel << ";\n";
}
writer.indent--;
writer << "}\n";
}
return;
}
src/ngraph/runtime/gpu/gpu_cuda_kernel_emitters.cpp
View file @ b5467550
......@@ -20,26 +20,22 @@
#include "ngraph/runtime/gpu/gpu_cuda_kernel_emitters.hpp"
#include "ngraph/runtime/gpu/gpu_cuda_kernel_ops.hpp"
namespace ngraph
using namespace ngraph;
void runtime::gpu::emit_broadcast(
void* in, void* out, size_t repeat_size, size_t repeat_times, size_t count)
{
namespace runtime
std::string name = "broadcast";
// Create an instance of nvrtcProgram with the code string.
if (CudaFunctionPool::instance().get(name) == nullptr)
{
namespace gpu
{
void emit_broadcast(
void* in, void* out, size_t repeat_size, size_t repeat_times, size_t count)
{
std::string name = "broadcast";
// Create an instance of nvrtcProgram with the code string.
if (CudaFunctionPool::instance().get(name) == nullptr)
{
std::string kernel;
std::string data_type("float");
std::string kernel;
std::string data_type("float");
kernel = R"(
kernel = R"(
extern "C" __global__
void cuda_)" + name + "(" + data_type +
"* in, " + data_type + "* out, size_t m, size_t k, size_t n)\n" + R"(
void cuda_)" + name +
"(" + data_type + "* in, " + data_type + "* out, size_t m, size_t k, size_t n)\n" +
R"(
{
size_t tid = blockIdx.x * blockDim.x + threadIdx.x;
if(tid < n)
......@@ -48,28 +44,25 @@ void cuda_)" + name + "(" + data_type +
out[tid] = in[idx];
}
})";
CudaFunctionPool::instance().set(name, kernel);
}
CudaFunctionPool::instance().set(name, kernel);
}
//convert runtime ptr to driver api ptr
CUdeviceptr d_ptr_in, d_ptr_out;
d_ptr_in = CUdeviceptr(in);
d_ptr_out = CUdeviceptr(out);
//convert runtime ptr to driver api ptr
CUdeviceptr d_ptr_in, d_ptr_out;
d_ptr_in = CUdeviceptr(in);
d_ptr_out = CUdeviceptr(out);
void* args_list[] = {&d_ptr_in, &d_ptr_out, &repeat_size, &repeat_times, &count};
CUDA_SAFE_CALL(cuLaunchKernel(*CudaFunctionPool::instance().get(name).get(),
static_cast<unsigned int>(count),
1,
1, // grid dim
1,
1,
1, // block dim
0,
NULL, // shared mem and stream
args_list,
0)); // arguments
CUDA_SAFE_CALL(cuCtxSynchronize()); // Retrieve and print output.
}
}
}
void* args_list[] = {&d_ptr_in, &d_ptr_out, &repeat_size, &repeat_times, &count};
CUDA_SAFE_CALL(cuLaunchKernel(*CudaFunctionPool::instance().get(name).get(),
static_cast<unsigned int>(count),
1,
1, // grid dim
1,
1,
1, // block dim
0,
NULL, // shared mem and stream
args_list,
0)); // arguments
CUDA_SAFE_CALL(cuCtxSynchronize()); // Retrieve and print output.
}
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