From b18b5f121fc53dc767b0e1790f5acd34281fb073 Mon Sep 17 00:00:00 2001
From: Yifan Zhao <yifanz16@illinois.edu>
Date: Mon, 1 Feb 2021 04:22:53 -0600
Subject: [PATCH] Finished generating model metadata
---
.../torch2hpvm/torch2hpvm/approxknobs.json | 63 ++++++++--------
.../projects/torch2hpvm/torch2hpvm/compile.py | 74 ++++++++++++++-----
.../torch2hpvm/torch2hpvm/graph_ir.py | 13 +++-
3 files changed, 98 insertions(+), 52 deletions(-)
diff --git a/hpvm/projects/torch2hpvm/torch2hpvm/approxknobs.json b/hpvm/projects/torch2hpvm/torch2hpvm/approxknobs.json
index 2462a512f2..974b536c48 100644
--- a/hpvm/projects/torch2hpvm/torch2hpvm/approxknobs.json
+++ b/hpvm/projects/torch2hpvm/torch2hpvm/approxknobs.json
@@ -2,206 +2,203 @@
{
"name": "11",
"speedup": 1.0,
- "applies_to": [
- "Conv2D",
- "MatMul"
- ]
+ "applies_to": null
},
{
"name": "12",
"speedup": 1.5,
"applies_to": [
- "Conv2D",
- "MatMul"
+ "convolution",
+ "linear"
]
},
{
"name": "151",
"speedup": 3.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "152",
"speedup": 3.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "153",
"speedup": 3.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "154",
"speedup": 3.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "155",
"speedup": 2.25,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "156",
"speedup": 2.25,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "157",
"speedup": 2.25,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "158",
"speedup": 2.25,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "159",
"speedup": 2.25,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "160",
"speedup": 2.25,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "161",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "162",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "163",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "164",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "165",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "166",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "167",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "168",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "261",
"speedup": 3.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "262",
"speedup": 3.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "263",
"speedup": 2.25,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "264",
"speedup": 2.25,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "265",
"speedup": 2.25,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "266",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "267",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "268",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
},
{
"name": "269",
"speedup": 2.0,
"applies_to": [
- "Conv2D"
+ "convolution"
]
}
]
diff --git a/hpvm/projects/torch2hpvm/torch2hpvm/compile.py b/hpvm/projects/torch2hpvm/torch2hpvm/compile.py
index 5b8cb3e1f6..bf1da0270b 100644
--- a/hpvm/projects/torch2hpvm/torch2hpvm/compile.py
+++ b/hpvm/projects/torch2hpvm/torch2hpvm/compile.py
@@ -1,7 +1,7 @@
import os
from pathlib import Path
from tempfile import NamedTemporaryFile
-from typing import IO, NamedTuple, Optional, Sequence, Tuple, Union
+from typing import Dict, IO, List, NamedTuple, Optional, Sequence, Tuple, Union
import onnx
import torch
@@ -33,6 +33,7 @@ class ModelExporter:
testset_name = "test_input.bin", "test_labels.bin"
weight_dir_name = "weights"
source_file_name = "hpvm_c.cpp"
+ metadata_file_name = "ops.json"
def __init__(
self,
@@ -72,34 +73,64 @@ class ModelExporter:
def export_datasets(self):
input_, labels = self.tuneset_name
- self._dump_dataset(self.tune_dataset, self.weight_dir / input_, self.weight_dir / labels)
+ self._dump_dataset(
+ self.tune_dataset, self.weight_dir / input_, self.weight_dir / labels
+ )
input_, labels = self.testset_name
- self._dump_dataset(self.test_dataset, self.weight_dir / input_, self.weight_dir / labels)
+ self._dump_dataset(
+ self.test_dataset, self.weight_dir / input_, self.weight_dir / labels
+ )
def export_metadata(
- self,
- output: PathLike = None, approx_knobs_file: PathLike = def_approx_knobs_file
+ self, output: PathLike, approx_knobs_file: PathLike = def_approx_knobs_file
):
import json
from collections import defaultdict
with Path(approx_knobs_file).open() as f:
knobs = json.load(f)
- ty_knobs = defaultdict(list)
+ KnobInfoT = Tuple[str, float]
+ ty_knobs: Dict[str, List[KnobInfoT]] = defaultdict(list)
+ default_knobs: List[KnobInfoT] = []
for k in knobs:
- for ty in k.pop("applies_to"):
- ty_knobs[ty].append((k["name"], k["speedup"]))
- knobs_used = set()
+ applies_to = k.pop("applies_to")
+ k = k["name"], k["speedup"]
+ if applies_to is None:
+ default_knobs.append(k)
+ continue
+ for ty in applies_to:
+ ty_knobs[ty].append(k)
+ idx = 0
+ op_cost: Dict[str, int] = {}
+ op_knobs: Dict[str, List[str]] = {}
+ knob_speedup: Dict[str, float] = {}
for node in self.dfg.traverse_order:
- knobs = ty_knobs.get(node.op_type, [])
- flops = node.get_flops()
- knobs_used.update(knobs)
- print(f"{node.name} ({node.op_type}) -> {knobs}, {flops}")
+ if not node.hpvm_op_type:
+ continue
+ hpvm_op_name = f"{node.hpvm_op_type}_{idx}"
+ type_knobs = ty_knobs.get(node.hpvm_op_type, [])
+ this_op_knobs = type_knobs + default_knobs
+ knobs_speedup = dict(this_op_knobs)
+ op_cost[hpvm_op_name] = int(node.get_flops()) # May get np.int64
+ op_knobs[hpvm_op_name] = list(knobs_speedup.keys())
+ knob_speedup.update(knobs_speedup)
+ idx += 1
+ with Path(output).open("w") as f:
+ json.dump(
+ {
+ "op_cost": op_cost,
+ "knob_speedup": knob_speedup,
+ "op_knobs": op_knobs,
+ },
+ f,
+ indent=2,
+ )
def export_all(self, output: PathLike = None, batch_size: Optional[int] = None):
default_codefile = self.output_dir / self.source_file_name
self.export_source_code(output or default_codefile, batch_size)
- self.export_metadata()
+ default_metafile = self.output_dir / self.metadata_file_name
+ self.export_metadata(default_metafile)
self.export_weights()
self.export_datasets()
@@ -121,7 +152,9 @@ class ModelExporter:
inputs.tofile(labels_filename)
@classmethod
- def _check_datasets(cls, tune_dataset: DatasetTy, test_dataset: DatasetTy) -> Tuple[int, int, int, int]:
+ def _check_datasets(
+ cls, tune_dataset: DatasetTy, test_dataset: DatasetTy
+ ) -> Tuple[int, int, int, int]:
tune_shape = cls._check_dataset_get_shape(tune_dataset)
test_shape = cls._check_dataset_get_shape(test_dataset)
if tune_shape != test_shape:
@@ -137,9 +170,14 @@ class ModelExporter:
if isinstance(dataset, Dataset):
size = len(dataset)
sample = dataset[0]
- if not isinstance(sample, (np.ndarray, torch.Tensor)) or len(sample.shape) != 4:
- raise ValueError("Dataset must be a 4D tensor due to backend limitation")
- return size, *sample.shape
+ if (
+ not isinstance(sample, (np.ndarray, torch.Tensor))
+ or len(sample.shape) != 4
+ ):
+ raise ValueError(
+ "Dataset must be a 4D tensor due to backend limitation"
+ )
+ return [size, *sample.shape]
if not isinstance(dataset, BinDataset):
raise TypeError("Only BinDataset or PyTorch Dataset are supported")
input_file = Path(dataset.input_file)
diff --git a/hpvm/projects/torch2hpvm/torch2hpvm/graph_ir.py b/hpvm/projects/torch2hpvm/torch2hpvm/graph_ir.py
index 39b6873e19..3309072723 100644
--- a/hpvm/projects/torch2hpvm/torch2hpvm/graph_ir.py
+++ b/hpvm/projects/torch2hpvm/torch2hpvm/graph_ir.py
@@ -15,6 +15,7 @@ class DFGNode(abc.ABC):
"""
op_type = ""
+ hpvm_op_type = ""
def __init__(
self, name: str, input_shapes: Sequence[ShapeT], output_shape: ShapeT, **kwargs
@@ -114,6 +115,7 @@ class WeightTensor(TensorNode):
class Conv2DNode(DFGNode):
op_type = "Conv2D"
+ hpvm_op_type = "convolution"
def __init__(
self,
@@ -205,15 +207,18 @@ class _Pool2DNode(DFGNode, abc.ABC):
class MaxPool2DNode(_Pool2DNode):
pool_type = "0"
op_type = "MaxPool2D"
+ hpvm_op_type = "maxpool"
class AveragePool2DNode(_Pool2DNode):
pool_type = "1"
op_type = "AveragePool2D"
+ hpvm_op_type = "avgpool"
class BiasAddNode(DFGNode):
op_type = "BiasAdd"
+ hpvm_op_type = "add"
def codegen(self):
return "tensorAdd", []
@@ -224,6 +229,7 @@ class BiasAddNode(DFGNode):
class MatMulNode(DFGNode):
op_type = "MatMul"
+ hpvm_op_type = "linear"
def codegen(self):
return "tensorGemmGPU", []
@@ -265,6 +271,7 @@ class MatMulNode(DFGNode):
class SoftMaxNode(DFGNode):
op_type = "SoftMax"
+ hpvm_op_type = "softmax"
def codegen(self):
return "tensorSoftmax", []
@@ -275,6 +282,7 @@ class SoftMaxNode(DFGNode):
class AddNode(DFGNode):
op_type = "Add"
+ hpvm_op_type = "add"
def codegen(self):
return "tensorAdd", []
@@ -285,6 +293,7 @@ class AddNode(DFGNode):
class ReluNode(DFGNode):
op_type = "ReLU"
+ hpvm_op_type = "relu"
def codegen(self):
return "tensorRelu", []
@@ -295,6 +304,7 @@ class ReluNode(DFGNode):
class TanhNode(DFGNode):
op_type = "Tanh"
+ hpvm_op_type = "tanh"
def codegen(self):
return "tensorTanh", []
@@ -304,7 +314,8 @@ class TanhNode(DFGNode):
class BatchNormalizationNode(DFGNode):
- op_type = "BN"
+ op_type = "BatchNorm"
+ hpvm_op_type = "batchnorm"
def __init__(
self,
--
GitLab