def schedule_softmax(outs): """Schedule for softmax op. Parameters ---------- outs: Array of Tensor The computation graph description of softmax in the format of an array of tensors. Returns ------- sch: Schedule The computation schedule for the op. """ outs = [outs] if isinstance(outs, te.tensor.Tensor) else outs s = te.create_schedule([x.op for x in outs]) softmax = outs[0] tgt = Target.current(allow_none=False) op_tag = softmax.op.tag if op_tag == "softmax_output": expsum = softmax.op.input_tensors[1] exp = softmax.op.input_tensors[0] max_elem = s[exp].op.input_tensors[1] elif op_tag == "log_softmax_output": exp = None max_elem = softmax.op.input_tensors[1] expsum = softmax.op.input_tensors[2] else: raise ValueError( "Tag is expected to be softmax_output or log_softmax_output. \ Got {0}".format( op_tag ) ) # The nvptx and rocm backends only supports 32-bits warp shuffle # instructions. # # TODO(tvm-team) Fix nvptx codegen or deprecate nvptx backend. def sched_warp_softmax(): if tgt.kind.name == "nvptx" or tgt.kind.name == "rocm": return softmax.dtype == "float32" or softmax.dtype == "int32" if tgt.kind.name != "cuda": # this is used as the gpu schedule for other arches which may not have warp reductions return False return True if len(softmax.shape) > 2: ops = [max_elem.op, expsum.op, softmax.op] if exp is not None: ops.append(exp.op) for op in ops: s = schedule_injective_from_existing(s, op.output(0)) elif sched_warp_softmax(): # A warp of 32 threads performs a row reduction. num_thread = tgt.thread_warp_size block_x = te.thread_axis("blockIdx.x") thread_x = te.thread_axis((0, num_thread), "threadIdx.x") # (4) softmax xo, xi = s[softmax].split(softmax.op.axis[1], nparts=num_thread) _, xii = s[softmax].split(xi, factor=4) s[softmax].vectorize(xii) s[softmax].bind(xo, thread_x) s[softmax].bind(softmax.op.axis[0], block_x) # (3) expsum k = expsum.op.reduce_axis[0] ko, _ = s[expsum].split(k, nparts=num_thread) s[expsum].bind(ko, thread_x) s[expsum].compute_at(s[softmax], xo) # (2) exp if exp is not None: xo, xi = s[exp].split(exp.op.axis[1], nparts=num_thread) _, xii = s[exp].split(xi, factor=4) s[exp].vectorize(xii) s[exp].bind(xo, thread_x) s[exp].compute_at(s[expsum], expsum.op.axis[0]) s[exp].compute_at(s[softmax], softmax.op.axis[0]) s[exp].set_scope("warp") # (1) max_elem k = max_elem.op.reduce_axis[0] ko, _ = s[max_elem].split(k, nparts=num_thread) s[max_elem].bind(ko, thread_x) if exp is not None: s[max_elem].compute_at(s[exp], xo) else: s[max_elem].bind(ko, thread_x) s[max_elem].bind(max_elem.op.axis[0], block_x) else: num_thread = 64 block_x = te.thread_axis("blockIdx.x") thread_x = te.thread_axis((0, num_thread), "threadIdx.x") if exp is not None: s[exp].bind(exp.op.axis[0], block_x) s[max_elem].bind(max_elem.op.axis[0], block_x) k = expsum.op.reduce_axis[0] ko, ki = s[expsum].split(k, factor=num_thread) EF = s.rfactor(expsum, ki) s[expsum].bind(s[expsum].op.axis[0], block_x) s[expsum].bind(s[expsum].op.reduce_axis[0], thread_x) s[EF].compute_at(s[expsum], s[expsum].op.reduce_axis[0]) s[expsum].set_store_predicate(thread_x.var.equal(0)) tx, xi = s[softmax].split(softmax.op.axis[1], nparts=num_thread) s[softmax].bind(softmax.op.axis[0], block_x) s[softmax].bind(tx, thread_x) return s
def compile_cuda(code, target_format="ptx", arch=None, options=None, path_target=None): """Compile cuda code with NVCC from env. Parameters ---------- code : str The cuda code. target_format : str The target format of nvcc compiler. arch : str The cuda architecture. options : str or list of str The additional options. path_target : str, optional Output file. Return ------ cubin : bytearray The bytearray of the cubin """ if arch is None: # If None, then it will use `tvm.target.Target.current().arch`. # Target arch could be a str like "sm_xx", or a list, such as # [ # "-gencode", "arch=compute_52,code=sm_52", # "-gencode", "arch=compute_70,code=sm_70" # ] compute_version = "".join( get_target_compute_version( Target.current(allow_none=True)).split(".")) arch = [ "-gencode", f"arch=compute_{compute_version},code=sm_{compute_version}" ] temp = utils.tempdir() if target_format not in ["cubin", "ptx", "fatbin"]: raise ValueError("target_format must be in cubin, ptx, fatbin") temp_code = temp.relpath("my_kernel.cu") temp_target = temp.relpath("my_kernel.%s" % target_format) with open(temp_code, "w") as out_file: out_file.write(code) file_target = path_target if path_target else temp_target cmd = ["nvcc"] cmd += ["--%s" % target_format, "-O3"] if isinstance(arch, list): cmd += arch elif isinstance(arch, str): 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
def build( inputs: Union[schedule.Schedule, PrimFunc, IRModule, Mapping[str, IRModule]], args: Optional[List[Union[Buffer, tensor.Tensor, Var]]] = None, target: Optional[Union[str, Target]] = None, target_host: Optional[Union[str, Target]] = None, name: Optional[str] = "default_function", binds: Optional[Mapping[tensor.Tensor, Buffer]] = None, ): """Build a function with arguments as signature. Code will be generated for devices coupled with target information. Parameters ---------- inputs : Union[tvm.te.schedule.Schedule, tvm.tir.PrimFunc, IRModule, Mapping[str, IRModule]] The input to be built args : Optional[List[Union[tvm.tir.Buffer, tensor.Tensor, Var]]] The argument lists to the function. target : Optional[Union[str, Target]] The target and option of the compilation. target_host : Optional[Union[str, Target]] Host compilation target, if target is device. When TVM compiles device specific program such as CUDA, we also need host(CPU) side code to interact with the driver setup the dimensions and parameters correctly. target_host is used to specify the host side codegen target. By default, llvm is used if it is enabled, otherwise a stackvm intepreter is used. name : Optional[str] The name of result function. binds : Optional[Mapping[tensor.Tensor, tvm.tir.Buffer]] Dictionary that maps the binding of symbolic buffer to Tensor. By default, a new buffer is created for each tensor in the argument. Returns ------- ret : tvm.module A module that combines both host and device code. Examples ________ There are two typical example uses of this function depending on the type of the argument `inputs`: 1. it is an IRModule. .. code-block:: python n = 2 A = te.placeholder((n,), name='A') B = te.placeholder((n,), name='B') C = te.compute(A.shape, lambda *i: A(*i) + B(*i), name='C') s = tvm.te.create_schedule(C.op) m = tvm.lower(s, [A, B, C], name="test_add") rt_mod = tvm.build(m, target="llvm") 2. it is a dict of compilation target to IRModule. .. code-block:: python n = 2 A = te.placeholder((n,), name='A') B = te.placeholder((n,), name='B') C = te.compute(A.shape, lambda *i: A(*i) + B(*i), name='C') s1 = tvm.te.create_schedule(C.op) with tvm.target.cuda() as cuda_tgt: s2 = topi.cuda.schedule_injective(cuda_tgt, [C]) m1 = tvm.lower(s1, [A, B, C], name="test_add1") m2 = tvm.lower(s2, [A, B, C], name="test_add2") rt_mod = tvm.build({"llvm": m1, "cuda": m2}, target_host="llvm") Note ---- See the note on :any:`tvm.target` on target string format. """ if isinstance(inputs, schedule.Schedule): if args is None: raise ValueError("args must be given for build from schedule") input_mod = lower(inputs, args, name=name, binds=binds) elif isinstance(inputs, (list, tuple, container.Array)): merged_mod = tvm.IRModule({}) for x in inputs: merged_mod.update(lower(x)) input_mod = merged_mod elif isinstance(inputs, (tvm.IRModule, PrimFunc)): input_mod = lower(inputs) elif not isinstance(inputs, (dict, container.Map)): raise ValueError( f"Inputs must be Schedule, IRModule or dict of target to IRModule, " f"but got {type(inputs)}.") if not isinstance(inputs, (dict, container.Map)): target = Target.current() if target is None else target target = target if target else "llvm" target_input_mod = {target: input_mod} else: target_input_mod = inputs for tar, mod in target_input_mod.items(): if not isinstance(tar, (str, Target)): raise ValueError("The key of inputs must be str or " "Target when inputs is dict.") if not isinstance(mod, tvm.IRModule): raise ValueError("inputs must be Schedule, IRModule," "or dict of str to IRModule.") target_input_mod, target_host = Target.check_and_update_host_consist( target_input_mod, target_host) if not target_host: for tar, mod in target_input_mod.items(): tar = Target(tar) device_type = ndarray.device(tar.kind.name, 0).device_type if device_type == ndarray.cpu(0).device_type: target_host = tar break if not target_host: target_host = "llvm" if tvm.runtime.enabled("llvm") else "stackvm" target_input_mod, target_host = Target.check_and_update_host_consist( target_input_mod, target_host) mod_host_all = tvm.IRModule({}) device_modules = [] for tar, input_mod in target_input_mod.items(): mod_host, mdev = _build_for_device(input_mod, tar, target_host) mod_host_all.update(mod_host) device_modules.append(mdev) # Generate a unified host module. rt_mod_host = codegen.build_module(mod_host_all, target_host) # Import all modules. for mdev in device_modules: if mdev: rt_mod_host.import_module(mdev) if not isinstance(target_host, Target): target_host = Target(target_host) if (target_host.attrs.get("runtime", tvm.runtime.String("c++")) == "c" and target_host.attrs.get("system-lib", 0) == 1): if target_host.kind.name == "c": create_csource_crt_metadata_module = tvm._ffi.get_global_func( "runtime.CreateCSourceCrtMetadataModule") to_return = create_csource_crt_metadata_module([rt_mod_host], target_host) elif target_host.kind.name == "llvm": create_llvm_crt_metadata_module = tvm._ffi.get_global_func( "runtime.CreateLLVMCrtMetadataModule") to_return = create_llvm_crt_metadata_module([rt_mod_host], target_host) else: to_return = rt_mod_host return OperatorModule.from_module(to_return, ir_module_by_target=target_input_mod, name=name)
def build(inputs, args=None, target=None, target_host=None, name="default_function", binds=None): """Build a function with arguments as signature. Code will be generated for devices coupled with target information. Parameters ---------- inputs : tvm.te.Schedule, IRModule, or dict of target to IRModule The schedule to be built args : list of Buffer or Tensor or Var, optional The argument lists to the function. target : str or :any:`tvm.target.Target`, optional The target and option of the compilation. target_host : str or :any:`tvm.target.Target` optional Host compilation target, if target is device. When TVM compiles device specific program such as CUDA, we also need host(CPU) side code to interact with the driver setup the dimensions and parameters correctly. target_host is used to specify the host side codegen target. By default, llvm is used if it is enabled, otherwise a stackvm intepreter is used. name : str, optional The name of result function. binds : dict, optional Dictionary that maps the binding of symbolic buffer to Tensor. By default, a new buffer is created for each tensor in the argument. Returns ------- ret : tvm.module A module that combines both host and device code. Examples ________ There are two typical example uses of this function depending on the type of the argument `inputs`: 1. it is an IRModule. .. code-block:: python n = 2 A = te.placeholder((n,), name='A') B = te.placeholder((n,), name='B') C = te.compute(A.shape, lambda *i: A(*i) + B(*i), name='C') s = tvm.te.create_schedule(C.op) m = tvm.lower(s, [A, B, C], name="test_add") rt_mod = tvm.build(m, target="llvm") 2. it is a dict of compilation target to IRModule. .. code-block:: python n = 2 A = te.placeholder((n,), name='A') B = te.placeholder((n,), name='B') C = te.compute(A.shape, lambda *i: A(*i) + B(*i), name='C') s1 = tvm.te.create_schedule(C.op) with tvm.target.cuda() as cuda_tgt: s2 = topi.cuda.schedule_injective(cuda_tgt, [C]) m1 = tvm.lower(s1, [A, B, C], name="test_add1") m2 = tvm.lower(s2, [A, B, C], name="test_add2") rt_mod = tvm.build({"llvm": m1, "cuda": m2}, target_host="llvm") Note ---- See the note on :any:`tvm.target` on target string format. """ if isinstance(inputs, schedule.Schedule): if args is None: raise ValueError("args must be given for build from schedule") input_mod = lower(inputs, args, name=name, binds=binds) elif isinstance(inputs, (list, tuple, container.Array)): merged_mod = tvm.IRModule({}) for x in inputs: merged_mod.update(x) input_mod = merged_mod elif isinstance(inputs, tvm.IRModule): input_mod = inputs elif not isinstance(inputs, (dict, container.Map)): raise ValueError( "inputs must be Schedule, IRModule or dict of target to IRModule") if not isinstance(inputs, (dict, container.Map)): target = Target.current() if target is None else target target = target if target else "llvm" target_input_mod = {target: input_mod} else: target_input_mod = inputs for tar, mod in target_input_mod.items(): if not isinstance(tar, (str, Target)): raise ValueError("The key of inputs must be str or " "Target when inputs is dict.") if not isinstance(mod, tvm.IRModule): raise ValueError("inputs must be Schedule, IRModule," "or dict of str to IRModule.") if not target_host: for tar, _ in target_input_mod.items(): tar = Target(tar) device_type = ndarray.context(tar.kind.name, 0).device_type if device_type == ndarray.cpu(0).device_type: target_host = tar break if not target_host: target_host = "llvm" if tvm.runtime.enabled("llvm") else "stackvm" mod_host_all = tvm.IRModule({}) device_modules = [] for tar, input_mod in target_input_mod.items(): mod_host, mdev = _build_for_device(input_mod, tar, target_host) mod_host_all.update(mod_host) device_modules.append(mdev) # Generate a unified host module. rt_mod_host = codegen.build_module(mod_host_all, target_host) # Import all modules. for mdev in device_modules: if mdev: rt_mod_host.import_module(mdev) return rt_mod_host
def _get_node_default_compute_dtype(): target = Target.current(True) if target and str(target.kind) == "llvm" and target_has_sse41(target.mcpu): return "float32" return "int64"