void addKernel(const std::string& file_name, const std::string& kernel_name)
{
std::ifstream ifs;
ifs.open(file_name.c_str(), std::ios::in);
if (!ifs.is_open())
{
std::cout << "Impossible to open '" << file_name << "'" << std::endl;
}
std::string kernel_str((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
//If the kernel has not already been added
m_kernels[kernel_name] = kernel_str;
m_program_sources.push_back({m_kernels[kernel_name].c_str(), kernel_str.size()});
}
static void customcl_setup(
std::string cl_program = "blocking-2-v4",
std::string arithmetic = "float") {
err = 0;
// build options for opencl.
std::string cl_build_options =
"-DT=" + arithmetic +
" -DT4=" + arithmetic + "4" +
" -DT8=" + arithmetic + "8" +
" -DT16=" + arithmetic + "16" +
" " + (arithmetic == "double" ? " -DSAMPLE_NEEDS_DOUBLE" : "") +
" " + (arithmetic == "half" ? " -DSAMPLE_NEEDS_HALF" : "");
// clkernel name.
std::string clkernel_path = "clkernel/";
caffe::cl_program = cl_program;
if(cl_program == "blocking-2-v4") {
clkernel_path += "gemm-blocking-2x2-vload4.cl";
}else if(cl_program == "blocking-4-v4") {
clkernel_path += "gemm-blocking-4x4-vload4.cl";
}else if(cl_program == "noblock-v8") {
clkernel_path += "gemm-noblock-vload8.cl";
}
std::ifstream kernel_file(clkernel_path);
std::string kernel_str((std::istreambuf_iterator<char>(kernel_file)),
std::istreambuf_iterator<char>());
viennacl::ocl::current_context().build_options(
"-DT=" + arithmetic +
" -DT4=" + arithmetic + "4" +
" -DT8=" + arithmetic + "8" +
" -DT16=" + arithmetic + "16" +
" " + (arithmetic == "double" ? " -DSAMPLE_NEEDS_DOUBLE" : "") +
" " + (arithmetic == "half" ? " -DSAMPLE_NEEDS_HALF" : ""));
cl_prog = viennacl::ocl::get_context(0).add_program(
kernel_str, "gemm_program");
gemm_exec = cl_prog.get_kernel("gemm");
transpose_exec = cl_prog.get_kernel("transpose");
copy_exec = cl_prog.get_kernel("copy");
transpose_ptr = (void*)clCreateBuffer(
viennacl::ocl::current_context().handle().get(),
CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR,
TRANSPOSE_BUFFER_DIM * TRANSPOSE_BUFFER_DIM * 8,
host_trans_buffer,
&err
);
SAMPLE_CHECK_ERRORS(err);
copy_ptr = (void*)clCreateBuffer(
viennacl::ocl::current_context().handle().get(),
CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR,
TRANSPOSE_BUFFER_DIM * TRANSPOSE_BUFFER_DIM * 8,
host_copy_buffer,
&err
);
SAMPLE_CHECK_ERRORS(err);
result_ptr = (void*)clCreateBuffer(
viennacl::ocl::current_context().handle().get(),
CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR,
TRANSPOSE_BUFFER_DIM * TRANSPOSE_BUFFER_DIM * 8,
host_result_buffer,
&err
);
SAMPLE_CHECK_ERRORS(err);
}
