csullivan commented on code in PR #12895: URL: https://github.com/apache/tvm/pull/12895#discussion_r1002899944
########## tests/python/unittest/test_meta_schedule_vnni_integration.py: ########## @@ -0,0 +1,249 @@ +# 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. +# pylint: disable=missing-docstring +import logging +import tempfile +from typing import Optional + +import numpy as np # type: ignore +import pytest +import tvm +from tvm import meta_schedule as ms +from tvm import relay +from tvm._ffi import register_func +from tvm.tir.schedule import BlockRV, Schedule +from tvm.tir.tensor_intrin.x86 import VNNI_DOT_16x4_INTRIN as VNNI_INTRIN + +logging.basicConfig( + format="%(asctime)s.%(msecs)03d %(levelname)s %(message)s", + datefmt="%Y-%m-%d %H:%M:%S", +) +logging.getLogger("tvm.meta_schedule").setLevel(logging.DEBUG) + + +def _schedule_dense(m: Optional[int], do_tune: bool): + """Manually schedule a dense block, created from TE compute op via CreatePrimFunc, + using VNNI instruction. + """ + + def schedule_fn(sch, dense_block: Optional[BlockRV] = None) -> bool: + if "dense" not in sch.mod.attrs["task_name"]: + return False + if dense_block is None: + dense_block = sch.get_block("compute") + assert "dense_vnni" in sch.get(dense_block).annotations["schedule_rule"] + + post_blocks = sch.get_consumers(dense_block) + if len(post_blocks) > 0: + # Fuse all intermediate post ops into the last op. + # This is equivalent to the traverse_inline function used in TE schedules. + while True: + next_post_blocks = [] + for post_block in post_blocks: + next_consumers = sch.get_consumers(post_block) + if len(next_consumers) > 0: + sch.compute_inline(post_block) + next_post_blocks += next_consumers + if len(next_post_blocks) == 0: + assert len(post_blocks) == 1 + outer_block = post_blocks[0] + a_y, a_x = sch.get_loops(outer_block)[-2:] + break + post_blocks = next_post_blocks + else: + a_y, a_x, _ = sch.get_loops(dense_block)[-3:] + outer_block = dense_block + if do_tune: + y_factors = sch.sample_perfect_tile(a_y, n=2, max_innermost_factor=128) + a_yo, a_yi = sch.split(a_y, factors=y_factors) + else: + a_yo, a_yi = sch.split(a_y, factors=[None, min(m, 64)]) + a_xo, a_xi = sch.split(a_x, factors=[None, 16]) + sch.reorder(a_yo, a_xo, a_yi, a_xi) + fused = sch.fuse(a_yo, a_xo) + if outer_block != dense_block: + # Handle the case when dense is fused with post ops. + sch.vectorize(a_xi) + sch.compute_at(dense_block, a_yi) + a_xi, a_k = sch.get_loops(dense_block)[-2:] + a_ko, a_ki = sch.split(a_k, factors=[None, 4]) + sch.reorder(a_ko, a_xi, a_ki) + # We need to parallelize before decompose_reduction, otherwise the so-called "Compact dataflow" + # condition is violated. + sch.parallel(fused) + dec = sch.decompose_reduction(dense_block, a_ko) + init_loop = sch.get_loops(dec)[-1] + sch.vectorize(init_loop) + sch.tensorize(a_xi, VNNI_INTRIN) + return True + + return schedule_fn + + +def _relay_dense(m, n, k): + data = relay.var("data", shape=(m, k), dtype="uint8") + weight = relay.var("weight", shape=(n, k), dtype="int8") + bias = relay.var("bias", shape=(n,), dtype="int32") + # dense is tuned by the TIR schedule above, bmm is scheduled by TE (topi/x86/batch_matmul.py) + dense = relay.nn.dense(data, weight, out_dtype="int32") + bias_add = relay.nn.bias_add(dense, bias) + relay.const(1, dtype="int32") + out = relay.nn.batch_matmul( + relay.cast(relay.expand_dims(bias_add, 0), "uint8"), + relay.cast(relay.expand_dims(bias_add, 0), "int8"), + out_dtype="int32", + ) + relay_mod = tvm.IRModule.from_expr(out) + data = np.random.uniform(1, 10, size=(m, k)).astype("uint8") + params = { + "weight": np.random.uniform(1, 10, size=(n, k)).astype("int8"), + "bias": np.random.uniform(1, 10, size=(n,)).astype("int32"), + } + + def f_check(lib, dev): + ref = ( + relay.create_executor( + "vm", + mod=relay_mod, + device=dev, + target="llvm", + ) + .evaluate()(data, params["weight"], params["bias"]) + .numpy() + ) + runtime = tvm.contrib.graph_executor.GraphModule(lib["default"](dev)) + runtime.set_input("data", data) + runtime.run() + out = runtime.get_output(0).numpy() + np.testing.assert_equal(out, ref) + + return relay_mod, params, f_check + + [email protected]("Requires cascadelake") +def test_vnni_schedule_fn_database(): + m, n, k = 1024, 1024, 1024 + target = tvm.target.Target("llvm -mcpu=cascadelake -num-cores 4") + dev = tvm.cpu(0) + relay_mod, params, f_check = _relay_dense(m, n, k) + + with ms.database.ScheduleFnDatabase( + _schedule_dense( + m=m, + do_tune=False, + ) + ), tvm.transform.PassContext( + opt_level=3, + config={"relay.backend.use_meta_schedule": True}, + ): + # pylint: disable=W0105 + """The log should say + Warning: Cannot find workload: tvmgen_default_fused_expand_dims + Warning: Cannot find workload: tvmgen_default_fused_cast + Warning: Cannot find workload: tvmgen_default_fused_cast_1 + Warning: Cannot find workload: tvmgen_default_fused_nn_batch_matmul + + This means batch matmul and others are scheduled by TE, and dense (the one not warned) + is found in the meta schedule tuning database during compilation + """ + # pylint: enable=W0105 + lib = relay.build(relay_mod, target=target, params=params) + f_check(lib, dev) + + [email protected]("Requires cascadelake") +def test_vnni_schedule_fn_tune(): + # pylint: disable=W0105 + """ + We can inject and apply a custom TIR scheduling to a TE compute of interest, using + the "schedule_rule" annotation. For example, in topi/x86/dense.py we have the following + declaration for int8 dense targeting the VNNI instruction. + + C = te.compute( + ... + attrs={"schedule_rule": "meta_schedule.dense_vnni"}, + ) + + When the MetaSchedule encounters a TensorIR block with the "schedule_rule" annotation, + it looks up the packed func registry for a function that is associated with the given schedule + rule key ("meta_schedule.dense_vnni" in this example). The signature of such custom schedule + functions must be + + (tir.schedule.Schedule, tir.schedule.BlockRV) -> [tir.schedule.Schedule]. + + The BlockRV argument corresponds to the TE compute annotated with "schedule_rule". + + The relevant code is in meta_schedule/space_generator/post_order_apply.cc. + """ + + def schedule_rule_dense_vnni(sch: Schedule, dense_block: BlockRV): + _schedule_dense(m=None, do_tune=True)(sch, dense_block) + return [sch] + + register_func("meta_schedule.dense_vnni", schedule_rule_dense_vnni) Review Comment: @junrushao @masahi or others, may I ask what the difference is between using the TE annotation as described (e.g. `attrs={"schedule_rule": "meta_schedule.dense_vnni"},` and a corresponding packed func defining the schedule to use, as opposed to just generating the space via ``` space=ms.space_generator.ScheduleFn( _schedule_dense, ... ), ``` ? Is it that in this test case we allow auto scheduling for all ops but apply special manual scheduling for certain ops (dense in this case), whereas if we use the ScheduleFn technique for generating a search space we do not allow other operators to be auto scheduled? Thanks! -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
