bbfdab79d9
## Summary - Keep the Python test launcher close to plain `pytest -n auto`, move nightly tests under `tests/nightly/python`, remove obsolete launchers and collection bookkeeping, and partition CPU/GPU jobs with explicit `gpu` marker expressions. - Repair exact-pointer regressions at their owning boundaries: packed raw-string ABI values, CUDA/Metal matrix intrinsic pointers, internal TE extern offsets, MetaSchedule scalar annotations, localized auto-tensorization scope matching, and typed DLTensor fixture fields. - Preserve typed workspace calls in TIR and cast pointer-returning external calls in CodeGenC, covered by a plain-TIRx 1024-byte global workspace that is compiled as C++. - Finish phasing out value-bearing Relax `R.Prim` annotations by requiring an explicit dtype, removing obsolete value-based contracts, and expressing the DISCO rank-dependent slices as explicit scalar `call_tir` inputs. - Gate the distributed callback on the optional DISCO runtime, NCCL, and at least two GPUs so capability-limited jobs skip instead of failing. - Remove the non-demonstrating pointer probe, use direct TVMScript comparison for packed strings, and remove the four designated legacy testing modules. The seven repaired CPU categories cover packed raw strings (7 failures), CUDA/Metal matrix access-pointer types (7), internal TE extern offsets (1), a typed DLTensor fixture (1), MetaSchedule scalar annotations (1), CodeGenC workspace return casts (12), and localized auto-tensorization storage-scope matching (19). ## Validation - Base: `ded6ad8dd212869c881efb5590f8a33fc972728e` - Head: `a7277e86dbcfe0638c8c252d36760859c4ab4297` - All 35 locally available original failing node IDs pass across the focused runs. - The full focused TE, TIR builtin-lowering, and CodeGenC files pass: 61 tests. - The complete touched Relax/TVMScript set plus PlanAndUpdateBufferAllocationLocation passes with 784 passed, 20 skipped, and 1 expected failure. - The DISCO callback collects and skips when its runtime or two-GPU environment is unavailable. - Six direct mapping tests, twelve tensor-core sketches, and the dp4a sketch pass unchanged. - The compiler rebuild, branch-wide pre-commit hooks, and full-range whitespace checks pass. - The 13 broad CBLAS/TFLite nodes remain dependency-gated; their owning TE and generated-C regressions compile. No merge is included in this change.
135 lines
4.3 KiB
Python
135 lines
4.3 KiB
Python
# Licensed to the Apache Software Foundation (ASF) under one
|
|
# or more contributor license agreements. See the NOTICE file
|
|
# distributed with this work for additional information
|
|
# regarding copyright ownership. The ASF licenses this file
|
|
# to you under the Apache License, Version 2.0 (the
|
|
# "License"); you may not use this file except in compliance
|
|
# with the License. You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing,
|
|
# software distributed under the License is distributed on an
|
|
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
|
# KIND, either express or implied. See the License for the
|
|
# specific language governing permissions and limitations
|
|
# under the License.
|
|
# ruff: noqa: RUF005
|
|
|
|
import numpy as np
|
|
|
|
import tvm
|
|
import tvm.script.relax as R
|
|
from tvm.relax.backend.contrib.nnapi import partition_for_nnapi
|
|
from tvm.support import ndk, utils
|
|
|
|
|
|
# pylint: disable=import-outside-toplevel,missing-function-docstring
|
|
def reshape_matmul(mod: tvm.IRModule):
|
|
from tvm.relax import Expr
|
|
from tvm.relax.dpl import DFPattern, rewrite_call
|
|
from tvm.relax.dpl.pattern import is_op, wildcard
|
|
|
|
input0 = wildcard()
|
|
input1 = wildcard()
|
|
pattern = is_op("relax.matmul")(input0, input1)
|
|
|
|
def _rewriter(expr: Expr, matches: dict[DFPattern, Expr]):
|
|
i0 = matches[input0]
|
|
i1 = matches[input1]
|
|
if len(i0.ty.shape) == 2 and len(i1.ty.shape) == 2:
|
|
i0_shape = [1] + [*i0.ty.shape.values]
|
|
i1_shape = [1] + [*i1.ty.shape.values]
|
|
oshape = matches[pattern].ty.shape
|
|
return R.reshape(R.matmul(R.reshape(i0, i0_shape), R.reshape(i1, i1_shape)), oshape)
|
|
return expr
|
|
|
|
mod["main"] = rewrite_call(pattern, _rewriter, mod["main"])
|
|
return mod
|
|
|
|
|
|
def decompose_clip(mod: tvm.IRModule) -> tvm.IRModule:
|
|
from tvm.relax import Expr
|
|
from tvm.relax.dpl import DFPattern, rewrite_call
|
|
from tvm.relax.dpl.pattern import is_op, wildcard
|
|
|
|
input_pattern = wildcard()
|
|
min_pattern = wildcard()
|
|
max_pattern = wildcard()
|
|
pattern = is_op("relax.clip")(input_pattern, min_pattern, max_pattern)
|
|
|
|
def _rewriter(expr: Expr, matches: dict[DFPattern, Expr]) -> Expr: # pylint: disable=unused-argument
|
|
dtype = matches[input_pattern].ty.dtype
|
|
return R.minimum(
|
|
R.maximum(
|
|
matches[input_pattern],
|
|
R.const(np.array(matches[min_pattern].value.value).astype(dtype), dtype),
|
|
),
|
|
R.const(np.array(matches[max_pattern].value.value).astype(dtype), dtype),
|
|
)
|
|
|
|
mod["main"] = rewrite_call(pattern, _rewriter, mod["main"])
|
|
return mod
|
|
|
|
|
|
def _build(mod, enable_nnapi):
|
|
if isinstance(mod, tvm.ir.Call):
|
|
mod = tvm.IRModule.from_expr(mod)
|
|
|
|
if enable_nnapi:
|
|
mod = tvm.relax.transform.FoldConstant()(mod)
|
|
mod = reshape_matmul(mod)
|
|
mod = decompose_clip(mod)
|
|
mod = partition_for_nnapi(mod)
|
|
|
|
mod = tvm.relax.transform.RunCodegen()(mod)
|
|
ex = tvm.compile(mod, target={"kind": "llvm", "mtriple": "aarch64-linux-android"})
|
|
|
|
return ex
|
|
|
|
|
|
def _run(remote, tracker, ex, inputs):
|
|
tmp = utils.tempdir()
|
|
so_name = "test_mod.so"
|
|
so_path = tmp / so_name
|
|
ex.export_library(str(so_path), fcompile=ndk.create_shared, options=["-shared", "-fPIC", "-lm"])
|
|
|
|
remote.upload(so_path)
|
|
dev = remote.cpu(0)
|
|
|
|
try:
|
|
# Execute the model on the remote.
|
|
remote_ex = remote.load_module(so_name)
|
|
vm = tvm.relax.VirtualMachine(remote_ex, device=dev)
|
|
|
|
inputs = [x.copyto(dev) for x in inputs]
|
|
|
|
vm.set_input("main", *inputs)
|
|
vm.invoke_stateful("main")
|
|
output = vm.get_outputs("main")
|
|
output = output.numpy()
|
|
except Exception as e:
|
|
# Re-raise all exceptions
|
|
raise e
|
|
finally:
|
|
# Manually close the connection.
|
|
# See https://discuss.tvm.apache.org/t/trouble-with-rpc-session/14008/.
|
|
#
|
|
# TODO: Remove if it does not happen on Python 3.11.
|
|
remote._sess.get_function("CloseRPCConnection")()
|
|
tracker.close()
|
|
pass
|
|
|
|
return output
|
|
|
|
|
|
def build_and_run(
|
|
remote,
|
|
tracker,
|
|
mod,
|
|
inputs,
|
|
enable_nnapi=False,
|
|
):
|
|
ex = _build(mod, enable_nnapi)
|
|
return _run(remote, tracker, ex, inputs)
|