diff --git a/hpvm/projects/predtuner b/hpvm/projects/predtuner
index 108de0205d1f84082f9722d2e0e7b51dd5f8d997..83e5ca83c83a69b6220ec4344ae31cac16fa8c46 160000
--- a/hpvm/projects/predtuner
+++ b/hpvm/projects/predtuner
@@ -1 +1 @@
-Subproject commit 108de0205d1f84082f9722d2e0e7b51dd5f8d997
+Subproject commit 83e5ca83c83a69b6220ec4344ae31cac16fa8c46
diff --git a/hpvm/projects/torch2hpvm/torch2hpvm/approxknobs.json b/hpvm/projects/torch2hpvm/torch2hpvm/approxknobs.json
index 9d7cb28a8b3fcc2301735c21e99119beb5a89907..64f4e8816b07208d03ccab98624c6c0e477f8e98 100644
--- a/hpvm/projects/torch2hpvm/torch2hpvm/approxknobs.json
+++ b/hpvm/projects/torch2hpvm/torch2hpvm/approxknobs.json
@@ -3,7 +3,8 @@
"name": "11",
"speedup": 1.0,
"applies_to": null,
- "is_baseline": true
+ "is_baseline": true,
+ "devices": null
},
{
"name": "12",
@@ -11,195 +12,439 @@
"applies_to": [
"convolution",
"linear"
- ]
+ ],
+ "devices": ["gpu"]
+ },
+ {
+ "name": "121",
+ "speedup": 2,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "122",
+ "speedup": 2,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "123",
+ "speedup": 2,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "124",
+ "speedup": 2,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "125",
+ "speedup": 1.5,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "126",
+ "speedup": 1.5,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "127",
+ "speedup": 1.5,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "128",
+ "speedup": 1.5,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "129",
+ "speedup": 1.5,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "130",
+ "speedup": 1.5,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "131",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "132",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "133",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "134",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "135",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "136",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "137",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "138",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "231",
+ "speedup": 2,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "232",
+ "speedup": 2,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "233",
+ "speedup": 1.5,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "234",
+ "speedup": 1.5,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "235",
+ "speedup": 1.5,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "236",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "237",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "238",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
+ },
+ {
+ "name": "239",
+ "speedup": 1.33,
+ "applies_to": [
+ "convolution"
+ ],
+ "devices": ["cpu"]
},
{
"name": "151",
"speedup": 3.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "152",
"speedup": 3.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "153",
"speedup": 3.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "154",
"speedup": 3.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "155",
"speedup": 2.25,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "156",
"speedup": 2.25,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "157",
"speedup": 2.25,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "158",
"speedup": 2.25,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "159",
"speedup": 2.25,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "160",
"speedup": 2.25,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "161",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "162",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "163",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "164",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "165",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "166",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "167",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "168",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "261",
"speedup": 3.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "262",
"speedup": 3.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "263",
"speedup": 2.25,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "264",
"speedup": 2.25,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "265",
"speedup": 2.25,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "266",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "267",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "268",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
},
{
"name": "269",
"speedup": 2.0,
"applies_to": [
"convolution"
- ]
+ ],
+ "devices": ["gpu"]
}
]
diff --git a/hpvm/projects/torch2hpvm/torch2hpvm/compile.py b/hpvm/projects/torch2hpvm/torch2hpvm/compile.py
index d53776b363595dd10b8f46f792474b941f444f2b..044eaa10f8632a889dc7183e79b60f409fc21d2d 100644
--- a/hpvm/projects/torch2hpvm/torch2hpvm/compile.py
+++ b/hpvm/projects/torch2hpvm/torch2hpvm/compile.py
@@ -121,45 +121,47 @@ class ModelExporter:
with Path(approx_knobs_file).open() as f:
knobs = json.load(f)
- KnobInfoT = Tuple[str, float]
- ty_knobs: Dict[str, List[KnobInfoT]] = defaultdict(list)
- default_knobs: List[KnobInfoT] = []
+ # Organize knobs into defaults and the ones for certain types
+ ty_knobs: Dict[str, str] = defaultdict(list)
+ default_knobs: List[str] = []
baseline_knob = None
for k in knobs:
- kp = k["name"], k["speedup"]
+ name = k["name"]
if "is_baseline" in k:
if baseline_knob:
raise ValueError("Multiple baseline knobs")
- baseline_knob = k["name"]
+ baseline_knob = name
applies_to = k["applies_to"]
if applies_to is None:
- default_knobs.append(kp)
+ default_knobs.append(name)
continue
for ty in applies_to:
- ty_knobs[ty].append(kp)
+ ty_knobs[ty].append(name)
if not baseline_knob:
raise ValueError("No baseline knob given")
+
+ # Enumerate operators and find knobs for each
idx = 0
op_cost: Dict[str, int] = {}
op_knobs: Dict[str, List[str]] = {}
- knob_speedup: Dict[str, float] = {}
+ used_knobs = set()
for node in self.dfg.traverse_order:
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_knobs[hpvm_op_name] = type_knobs + default_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
+
+ # Write out
with Path(output).open("w") as f:
json.dump(
{
"op_cost": op_cost,
- "knob_speedup": knob_speedup,
"op_knobs": op_knobs,
+ "knob_speedup": {k["name"]: k["speedup"] for k in knobs},
+ "knob_devices": {k["name"]: k["devices"] for k in knobs},
"baseline_knob": baseline_knob,
**self.path_params,
},
diff --git a/hpvm/test/dnn_benchmarks/pytorch/test_tuning/test_tuning.py b/hpvm/test/dnn_benchmarks/pytorch/test_tuning/test_tuning.py
index c589b108d73277977820407f1874cf33053fcc34..9f969a42616804e4b26eb90b3d379a2b5dd08436 100755
--- a/hpvm/test/dnn_benchmarks/pytorch/test_tuning/test_tuning.py
+++ b/hpvm/test/dnn_benchmarks/pytorch/test_tuning/test_tuning.py
@@ -24,7 +24,7 @@ def main():
binary_file, exporter = dnn.export_example_dnn(netname, codegen_dir, True)
metadata_file = codegen_dir / exporter.metadata_file_name
# Tuning
- app = PipedBinaryApp("test", binary_file, metadata_file)
+ app = PipedBinaryApp("test", binary_file, metadata_file, tuning_device="cpu")
tuner = app.get_tuner()
tuner.tune(
5,