Merge pull request #21 from CentML/dec22-new-devices

Add new devices to Habitat

Author: jimgao1

analyzer/habitat/analysis/__init__.py

@@ -5,6 +5,7 @@
 SPECIAL_OPERATIONS = {
     # Convolution
     'conv2d',
+    'conv_transpose2d',
 
     # Matrix multiply operations
     'linear',

analyzer/habitat/analysis/mlp/dataset_process.py

@@ -35,29 +35,23 @@ def get_dataset(path, features, device_features=None):
         print("Loaded file %s (%d entries)" % (f, len(df.index)))
 
         if device_name not in devices:
-            devices[device_name] = df
-        else:
-            devices[device_name] = devices[device_name].append(df)
+            devices[device_name] = []
+        devices[device_name].append(df)
 
     for device in devices.keys():
+        devices[device] = pd.concat(devices[device])
         print("Device %s contains %d entries" % (device, len(devices[device].index)))
 
     print()
 
-    print("Merging")
-    df_merged = functools.reduce(
-        lambda df1, df2: pd.merge(df1, df2, on=features),
-        devices.values()
-    )
-
     print("Generating dataset")
     # generate vectorized dataset (one entry for each device with device params)
     device_params = get_all_devices(device_features)
 
     x, y = [], []
     for device in devices.keys():
-        df_merged_device = df_merged[features + [device, ]]
-        for row in tqdm(df_merged_device.iterrows(), leave=False, desc=device, total=len(df_merged_device.index)):
+        df_device = devices[device]
+        for row in tqdm(df_device.iterrows(), leave=False, desc=device, total=len(df_device.index)):
             row = row[1]
 
             x.append(list(row[:-1]) + device_params[device])

analyzer/habitat/analysis/mlp/devices.csv

@@ -7,3 +7,6 @@ T4,16,GDDR6,320,40,,8.1,
 V100,16,HBM2,900,80,,14.028,
 A100,40,HBM2,1555,108,9.5,19.0,300.0
 RTX3090,24,GDDR6X,936.2,82,556.0,35.58,35.58
+A40,48,GDDR6,614.9,84,1.168,37.4,299.4
+A4000,16,GDDR6,378.1,48,0.599,19.17,19.17
+RTX4000,8,GDDR6,364.1,36,0.2225,7.119,7.119

analyzer/habitat/analysis/mlp/mlp.py

@@ -89,6 +89,25 @@ def forward(self, x):
 
         return x
 
+class ConvTranspose2DMLP(nn.Module):
+    def __init__(self, layers, layer_size):
+        super().__init__()
+
+        self.features = ['bias', 'batch', 'image_size', 'in_channels', 'out_channels', 'kernel_size', 'stride',
+                         'padding']
+
+        # properly manage device parameters
+        self.fc1 = nn.Linear(len(self.features) + 4, layer_size)
+        self.mlp = MLPBase(layers, layer_size)
+        self.fc2 = nn.Linear(layer_size, 1)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = F.relu(x)
+        x = self.mlp(x)
+        x = self.fc2(x)
+
+        return x
 
 class BMMMLP(nn.Module):
     def __init__(self, layers, layer_size):
@@ -119,6 +138,7 @@ def __init__(self, model_name, layers, layer_size, model_path=None):
             "linear": LinearMLP,
             "lstm": LSTMMLP,
             "conv2d": Conv2DMLP,
+            "conv_transpose2d": ConvTranspose2DMLP,
             "bmm": BMMMLP,
         }[self.model_name](layers, layer_size)
 

analyzer/habitat/analysis/predictor.py

@@ -19,6 +19,10 @@
     'input', 'weight', 'bias', 'stride', 'padding', 'dilation', 'groups',
 ]
 
+CONVTRANSPOSE2D_PARAMS = [
+    'input', 'weight', 'bias', 'stride', 'padding', 'dilation', 'groups',
+]
+
 LINEAR_PARAMS = ['input', 'weight', 'bias']
 
 BMM_PARAMS = ['input', 'mat2', 'out']
@@ -79,6 +83,10 @@ def __init__(
             "bmm", 8, 1024,
             path_to_data("bmm/model.pth"),
         )
+        self.conv_transpose2d_pred = RuntimePredictor(
+            "conv_transpose2d", 8, 1024,
+            path_to_data("conv_transpose2d/model.pth"),
+        )
 
 
     def predict_operation(self, operation, dest_device):
@@ -99,6 +107,8 @@ def predict_operation(self, operation, dest_device):
             return self._special_scale(operation, dest_device, self._linear_scale)
         elif operation.name == 'bmm':
             return self._special_scale(operation, dest_device, self._bmm_scale)
+        elif operation.name == 'conv_transpose2d':
+            return self._special_scale(operation, dest_device, self._conv_transpose2d_scale)
 
         logger.warn('Unhandled special operation: %s', operation.name)
         return PredictedOperation(
@@ -181,6 +191,40 @@ def _conv2d_scale(self, operation, dest_device):
 
         return operation.run_time_ms * pred_dest / pred_orig
 
+    def _conv_transpose2d_scale(self, operation, dest_device):
+        # 1. Merge arguments (give them all names)
+        merged = name_all_arguments(
+            CONVTRANSPOSE2D_PARAMS,
+            operation.arguments.args,
+            operation.arguments.kwargs,
+        )
+
+        # 2. Construct arguments that the predictor expects
+        arguments = dict(
+            batch=merged['input'][0],
+            image_size=merged['input'][2],
+            in_channels=merged['input'][1],
+            out_channels=merged['weight'][0],
+            kernel_size=merged['weight'][2],
+            stride=(
+                merged['stride'][0]
+                if isinstance(merged['stride'], tuple) else merged['stride']
+            ),
+            padding=(
+                merged['padding'][0]
+                if isinstance(merged['padding'], tuple) else merged['padding']
+            ),
+            bias=(1 if merged['bias'] is not None else 0),
+        )
+
+        # 3. Call model to make prediction
+        arguments = [arguments[x] for x in self.conv_transpose2d_pred.model.features]
+
+        pred_dest = self.conv_transpose2d_pred.predict(arguments, dest_device.name)
+        pred_orig = self.conv_transpose2d_pred.predict(arguments, operation.device.name)
+
+        return operation.run_time_ms * pred_dest / pred_orig
+
     def _linear_scale(self, operation, dest_device):
         merged = name_all_arguments(
             LINEAR_PARAMS,

analyzer/habitat/data/bmm/model.pth

@@ -1,5 +1,6 @@
-b2c102539c05fd9169774c9fc32f0732a82b679e  bmm/model.pth
-376bd8d9233c3384be01e638191e842ebc212f5f  conv2d/model.pth
+7e09507e4a97b9b5b2d274b6a5f3bfb7ea1dbd97  bmm/model.pth
+065178ab02df36e7401bd4674a14ee940a1cd8a1  conv2d/model.pth
+d8a583c0b9068cb92276468272935ce5b0b78ca9  convtranspose2d/model.pth
 659a00c6cff529613b40d5166fe7d93f42e8327d  kernels.sqlite
-9f17fb74321fe1a82d9d647cd8c976b724b69a94  linear/model.pth
-ef93bb74178d38660b31ac0bef262c39b1e6ba16  lstm/model.pth
+5f0be970e001ba9b66b7fd9b100f01a2f91ecf8d  linear/model.pth
+a4bd128777ef5c90c3aaa1dd2abfed42e6024f3d  lstm/model.pth

analyzer/habitat/data/checksums

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06c9bb03de9d81ed9d22febcd90a338d5a4e851be4b686e4aa6037fa195354aa
+size 33650493

analyzer/habitat/data/conv2d/model.pth

@@ -161,4 +161,51 @@ RTX3090:
   base_clock_mhz: 1395
   peak_gflops_per_second: 19346
 
+A40:
+  compute_major: 8
+  compute_minor: 6
+  max_threads_per_block: 1024
+  max_threads_per_multiprocessor: 1536
+  regs_per_block: 65536
+  regs_per_multiprocessor: 65536
+  warp_size: 32
+  shared_mem_per_block: 49152
+  shared_mem_per_multiprocessor: 102400
+  num_sms: 84
+  shared_mem_per_block_optin: 101376
+  mem_bandwidth_gb: 614
+  base_clock_mhz: 1305
+  peak_gflops_per_second: 18680
+
+A4000:
+  compute_major: 8
+  compute_minor: 6
+  max_threads_per_block: 1024
+  max_threads_per_multiprocessor: 1536
+  regs_per_block: 65536
+  regs_per_multiprocessor: 65536
+  warp_size: 32
+  shared_mem_per_block: 49152
+  shared_mem_per_multiprocessor: 102400
+  num_sms: 48
+  shared_mem_per_block_optin: 101376
+  mem_bandwidth_gb: 378
+  base_clock_mhz: 735
+  peak_gflops_per_second: 10499
+
+RTX4000:
+  compute_major: 7
+  compute_minor: 5
+  max_threads_per_block: 1024
+  max_threads_per_multiprocessor: 1024
+  regs_per_block: 65536
+  regs_per_multiprocessor: 65536
+  warp_size: 32
+  shared_mem_per_block: 49152
+  shared_mem_per_multiprocessor: 65536
+  num_sms: 36
+  shared_mem_per_block_optin: 65536
+  mem_bandwidth_gb: 364
+  base_clock_mhz: 1005
+  peak_gflops_per_second: 4245
 

analyzer/habitat/data/conv_transpose2d/model.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf50e632cb66f3f8585f65a14d44f36b373ea320f60fb27ab1477d5e6983c249
-size 33647145
+oid sha256:feca09482a1c7f48f9b07fa58abd85b34467107382e33fece92d2fd8af863675
+size 33646397

analyzer/habitat/data/devices.yml

@@ -22,6 +22,13 @@
     habitat.Device.T4,
     habitat.Device.RTX2070,
     habitat.Device.RTX2080Ti,
+
+    habitat.Device.RTX2080Ti,
+    habitat.Device.RTX3090,
+    habitat.Device.A100,
+    habitat.Device.A40,
+    habitat.Device.A4000,
+    habitat.Device.RTX4000,
 ]
 
 RESNET50_BATCHES = [16, 32, 64]
@@ -237,9 +244,9 @@ def main():
     )
 
     run_dcgan_experiments(context)
-    # run_inception_experiments(context)
-    # run_resnet50_experiments(context)
-    # run_gnmt_experiments(context)
+    run_inception_experiments(context)
+    run_resnet50_experiments(context)
+    run_gnmt_experiments(context)
     # run_transformer_experiments(context)
 
 

analyzer/habitat/data/linear/model.pth

@@ -1,11 +1,16 @@
 import logging
+import gc
 import random
 import signal
 
+import torch
+
 import habitat
 from habitat.profiling.operation import OperationProfiler
 from database import Recorder
 
+
+
 logger = logging.getLogger(__name__)
 
 
@@ -111,6 +116,9 @@ def measure_configurations(self, args, num_configs):
             args.rank + 1,
             args.world_size,
         )
+
+        last_count = self._recorder.get_num_recordings()
+
         try:
             for idx, config_id in enumerate(to_record):
                 if idx < num_configs_measured:
@@ -126,6 +134,9 @@ def measure_configurations(self, args, num_configs):
 
                 if idx % 100 == 0:
                     self._recorder.commit()
+                    cur_count = self._recorder.get_num_recordings()
+                    logger.info(f"commit. num_recordings: {cur_count}, new: {cur_count-last_count}")
+                    last_count = cur_count
 
                 if self._shutdown_early:
                     break
@@ -144,10 +155,18 @@ def _measure(self, config):
 
         except RuntimeError as e:
             msg = str(e)
+
+            if "out of memory" in msg:
+                allocated = torch.cuda.memory_allocated()
+                torch.cuda.empty_cache()
+                gc.collect()
+                logger.info(f"Cleared memory: {allocated} -> {torch.cuda.memory_allocated()}")
+
             if ("out of memory" not in msg and
                     "cuDNN error" not in msg and
                     "Calculated padded" not in msg):
                 logger.exception('Unexpected error during measurement.')
+                # logger.info("error: " + msg)
             return None, None
 
     def _record(self, config, forward_result, backward_result):