/* TODO: there's not much to do here, just build the kernel for HSA.
Perhaps share the same function for all WG sizes in case it's an
SPMD target. */
static void compile (_cl_command_node *cmd)
{
int error;
char bytecode[POCL_FILENAME_LENGTH];
char objfile[POCL_FILENAME_LENGTH];
FILE *file;
char *elf_blob;
size_t file_size, got_size;
hsa_runtime_caller_t caller;
error = snprintf (bytecode, POCL_FILENAME_LENGTH,
"%s/%s", cmd->command.run.tmp_dir,
POCL_PARALLEL_BC_FILENAME);
assert (error >= 0);
error = snprintf (objfile, POCL_FILENAME_LENGTH,
"%s/%s.o", cmd->command.run.tmp_dir,
POCL_PARALLEL_BC_FILENAME);
assert (error >= 0);
error = pocl_llvm_codegen (cmd->command.run.kernel, cmd->device, bytecode, objfile);
assert (error == 0);
/* Load the built AMDGPU ELF file. */
file = fopen (objfile, "rb");
assert (file != NULL);
cmd->command.run.device_data = (void**)malloc (sizeof(void*)*2);
cmd->command.run.device_data[0] = malloc (sizeof(hsa_amd_code_unit_t));
cmd->command.run.device_data[1] = malloc (sizeof(hsa_amd_code_t));
file_size = pocl_file_size (file);
elf_blob = (char*)malloc (file_size);
got_size = fread (elf_blob, 1, file_size, file);
if (file_size != got_size)
POCL_ABORT ("pocl-hsa: could not read the AMD ELF.");
caller.caller = 0;
if (hsa_ext_code_unit_load
(caller, NULL, 0, elf_blob, file_size, NULL, NULL,
(hsa_amd_code_unit_t*)cmd->command.run.device_data[0]) != HSA_STATUS_SUCCESS)
{
POCL_ABORT ("pocl-hsa: error while loading the built AMDGPU ELF binary.");
}
if (hsa_ext_code_unit_get_info
(*(hsa_amd_code_unit_t*)cmd->command.run.device_data[0],
HSA_EXT_CODE_UNIT_INFO_CODE_ENTITY_CODE, 0,
(hsa_amd_code_t*)cmd->command.run.device_data[1]) != HSA_STATUS_SUCCESS)
{
POCL_ABORT ("pocl-hsa: unable to get the code handle to the kernel.");
}
free (elf_blob);
fclose (file);
}
/**
* Generate code from the final bitcode using the LLVM
* tools.
*
* Uses an existing (cached) one, if available.
*
* @param tmpdir The directory of the work-group function bitcode.
* @param return the generated binary filename.
*/
const char*
llvm_codegen (const char* tmpdir, cl_kernel kernel, cl_device_id device) {
const char* pocl_verbose_ptr =
pocl_get_string_option("POCL_VERBOSE", (char*)NULL);
int pocl_verbose = pocl_verbose_ptr && *pocl_verbose_ptr;
char command[COMMAND_LENGTH];
char bytecode[POCL_FILENAME_LENGTH];
char objfile[POCL_FILENAME_LENGTH];
char* module = (char*) malloc(min(POCL_FILENAME_LENGTH,
strlen(tmpdir) + strlen(kernel->function_name) + 5)); // strlen of / .so 4+1
int error;
cl_program program = kernel->program;
error = snprintf
(module, POCL_FILENAME_LENGTH,
"%s/%s.so", tmpdir, kernel->function_name);
assert (error >= 0);
error = snprintf
(objfile, POCL_FILENAME_LENGTH,
"%s/%s.so.o", tmpdir, kernel->function_name);
assert (error >= 0);
if (access (module, F_OK) != 0)
{
error = snprintf (bytecode, POCL_FILENAME_LENGTH,
"%s/%s", tmpdir, POCL_PARALLEL_BC_FILENAME);
assert (error >= 0);
error = pocl_llvm_codegen( kernel, device, bytecode, objfile);
assert (error == 0);
// clang is used as the linker driver in LINK_CMD
error = snprintf (command, COMMAND_LENGTH,
#ifndef ANDROID
LINK_CMD " " HOST_CLANG_FLAGS " " HOST_LD_FLAGS " -o %s %s",
#else
ANDROID_POCL_PREFIX"/bin/ld " HOST_LD_FLAGS " -o %s %s "
" /system/lib/crtend_so.o /system/lib/crtbegin_so.o -ldl -lc ",
#endif
module, objfile);
assert (error >= 0);
if (pocl_verbose) {
fprintf(stderr, "[pocl] executing [%s]\n", command);
fflush(stderr);
}
error = system (command);
assert (error == 0);
/* Save space in kernel cache */
if (!pocl_get_bool_option("POCL_LEAVE_KERNEL_COMPILER_TEMP_FILES", 0))
{
pocl_remove_file(objfile);
pocl_remove_file(bytecode);
}
}
return module;
}
