def warmup_estimate(ins, outs):
ins = [runtime.array(x, ctx) for x in ins]
outs = [runtime.array(x, ctx) for x in outs]
tensors = ins + outs
func(*tensors)
runtime.gpu(visible_dev_id).sync()
t_start = time.time()
func(*tensors)
runtime.gpu(visible_dev_id).sync()
t_diff = time.time() - t_start
return ins, outs, tensors, t_diff
def compile_cuda(code,
target="ptx",
arch=None,
options=None,
path_target=None):
"""Compile cuda code with NVCC from env.
Parameters
----------
code : str
The cuda code.
target : str
The target format
arch : str
The architecture
options : str or list of str
The additional options
path_target : str, optional
Output file.
Return
------
cubin : bytearray
The bytearray of the cubin
"""
temp = util.tempdir()
if target not in ["cubin", "ptx", "fatbin"]:
raise ValueError("target must be in cubin, ptx, fatbin")
temp_code = temp.relpath("my_kernel.cu")
temp_target = temp.relpath("my_kernel.%s" % target)
with open(temp_code, "w") as out_file:
out_file.write(code)
if arch is None:
if nd.gpu(0).exist:
# auto detect the compute arch argument
arch = "sm_" + "".join(nd.gpu(0).compute_version.split('.'))
else:
raise ValueError(
"arch(sm_xy) is not passed, and we cannot detect it from env")
file_target = path_target if path_target else temp_target
cmd = ["nvcc"]
cmd += ["--%s" % target, "-O3"]
if isinstance(arch, list):
cmd += arch
else:
cmd += ["-arch", arch]
if options:
if isinstance(options, str):
cmd += [options]
elif isinstance(options, list):
cmd += options
else:
raise ValueError("options must be str or list of str")
cmd += ["-o", file_target]
cmd += [temp_code]
proc = subprocess.Popen(cmd,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT)
(out, _) = proc.communicate()
if proc.returncode != 0:
msg = code
msg += "\nCompilation error:\n"
msg += py_str(out)
raise RuntimeError(msg)
data = bytearray(open(file_target, "rb").read())
if not data:
raise RuntimeError("Compilation error: empty result is generated")
return data
def compile_cuda(code, target="ptx", arch=None, options=None, path_target=None):
"""Compile cuda code with NVCC from env.
Parameters
----------
code : str
The cuda code.
target : str
The target format
arch : str
The architecture
options : str or list of str
The additional options
path_target : str, optional
Output file.
Return
------
cubin : bytearray
The bytearray of the cubin
"""
temp = utils.tempdir()
if target not in ["cubin", "ptx", "fatbin"]:
raise ValueError("target must be in cubin, ptx, fatbin")
temp_code = temp.relpath("my_kernel.cu")
temp_target = temp.relpath("my_kernel.%s" % target)
with open(temp_code, "w") as out_file:
out_file.write(code)
if arch is None:
if nd.gpu(0).exist:
# auto detect the compute arch argument
arch = "sm_" + "".join(nd.gpu(0).compute_version.split("."))
else:
raise ValueError("arch(sm_xy) is not passed, and we cannot detect it from env")
file_target = path_target if path_target else temp_target
cmd = ["nvcc"]
cmd += ["--%s" % target, "-O3"]
if isinstance(arch, list):
cmd += arch
else:
cmd += ["-arch", arch]
if options:
if isinstance(options, str):
cmd += [options]
elif isinstance(options, list):
cmd += options
else:
raise ValueError("options must be str or list of str")
cmd += ["-o", file_target]
cmd += [temp_code]
# NOTE: ccbin option can be used to tell nvcc where to find the c++ compiler
# just in case it is not in the path. On Windows it is not in the path by default.
# However, we cannot use TVM_CXX_COMPILER_PATH because the runtime env.
# Because it is hard to do runtime compiler detection, we require nvcc is configured
# correctly by default.
# if cxx_compiler_path != "":
# cmd += ["-ccbin", cxx_compiler_path]
proc = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
(out, _) = proc.communicate()
if proc.returncode != 0:
msg = code
msg += "\nCompilation error:\n"
msg += py_str(out)
raise RuntimeError(msg)
data = bytearray(open(file_target, "rb").read())
if not data:
raise RuntimeError("Compilation error: empty result is generated")
return data