Add --allow-tf32 perf tuning argument that can be used to enable tf32

Defaults to keeping tf32 disabled.  This is because we haven't fully
verified training results with fp32 enabled.

Author: jannehellsten

docs/train-help.txt

@@ -65,5 +65,6 @@ Options:
   --fp32 BOOL                     Disable mixed-precision training
   --nhwc BOOL                     Use NHWC memory format with FP16
   --nobench BOOL                  Disable cuDNN benchmarking
+  --allow-tf32 BOOL               Allow PyTorch to use TF32 internally
   --workers INT                   Override number of DataLoader workers
   --help                          Show this message and exit.

train.py

@@ -61,6 +61,7 @@ def setup_training_loop_kwargs(
     # Performance options (not included in desc).
     fp32       = None, # Disable mixed-precision training: <bool>, default = False
     nhwc       = None, # Use NHWC memory format with FP16: <bool>, default = False
+    allow_tf32 = None, # Allow PyTorch to use TF32 for matmul and convolutions: <bool>, default = False
     nobench    = None, # Disable cuDNN benchmarking: <bool>, default = False
     workers    = None, # Override number of DataLoader workers: <int>, default = 3
 ):
@@ -343,6 +344,12 @@ def setup_training_loop_kwargs(
     if nobench:
         args.cudnn_benchmark = False
 
+    if allow_tf32 is None:
+        allow_tf32 = False
+    assert isinstance(allow_tf32, bool)
+    if allow_tf32:
+        args.allow_tf32 = True
+
     if workers is not None:
         assert isinstance(workers, int)
         if not workers >= 1:
@@ -425,6 +432,7 @@ def convert(self, value, param, ctx):
 @click.option('--fp32', help='Disable mixed-precision training', type=bool, metavar='BOOL')
 @click.option('--nhwc', help='Use NHWC memory format with FP16', type=bool, metavar='BOOL')
 @click.option('--nobench', help='Disable cuDNN benchmarking', type=bool, metavar='BOOL')
+@click.option('--allow-tf32', help='Allow PyTorch to use TF32 internally', type=bool, metavar='BOOL')
 @click.option('--workers', help='Override number of DataLoader workers', type=int, metavar='INT')
 
 def main(ctx, outdir, dry_run, **config_kwargs):

training/training_loop.py

@@ -115,6 +115,7 @@ def training_loop(
     network_snapshot_ticks  = 50,       # How often to save network snapshots? None = disable.
     resume_pkl              = None,     # Network pickle to resume training from.
     cudnn_benchmark         = True,     # Enable torch.backends.cudnn.benchmark?
+    allow_tf32              = False,    # Enable torch.backends.cuda.matmul.allow_tf32 and torch.backends.cudnn.allow_tf32?
     abort_fn                = None,     # Callback function for determining whether to abort training. Must return consistent results across ranks.
     progress_fn             = None,     # Callback function for updating training progress. Called for all ranks.
 ):
@@ -124,6 +125,8 @@ def training_loop(
     np.random.seed(random_seed * num_gpus + rank)
     torch.manual_seed(random_seed * num_gpus + rank)
     torch.backends.cudnn.benchmark = cudnn_benchmark    # Improves training speed.
+    torch.backends.cuda.matmul.allow_tf32 = allow_tf32  # Allow PyTorch to internally use tf32 for matmul
+    torch.backends.cudnn.allow_tf32 = allow_tf32        # Allow PyTorch to internally use tf32 for convolutions
     conv2d_gradfix.enabled = True                       # Improves training speed.
     grid_sample_gradfix.enabled = True                  # Avoids errors with the augmentation pipe.