Enable Asymmetric DistConv Kernels

ci_test/unit_tests/test_unit_layer_convolution_asymmetric_distconv.py

@@ -0,0 +1,73 @@
+import lbann
+import numpy as np
+import test_util
+import pytest
+import os
+import sys
+import lbann.contrib.launcher
+import lbann.contrib.args
+
+# Bamboo utilities
+current_file = os.path.realpath(__file__)
+current_dir = os.path.dirname(current_file)
+sys.path.insert(0, os.path.join(os.path.dirname(current_dir), 'common_python'))
+import tools
+
+@pytest.mark.parametrize('num_dims', [2, 3])
+@test_util.lbann_test(check_gradients=True,
+                      environment=lbann.contrib.args.get_distconv_environment(),
+                      time_limit=10)
+def test_simple(num_dims):
+    try:
+        import torch
+        import torch.nn as nn
+    except:
+        pytest.skip('PyTorch is required to run this test.')
+
+    torch.manual_seed(20240216)
+    # Two samples of 4x16x16 or 4x16x16x16 tensors
+    shape = [2, 4] + [16] * num_dims
+    x = torch.randn(shape)
+    if num_dims == 2:
+        ConvClass = nn.Conv2d
+        kerenel_size = (3, 1)
+        padding = (1, 0)
+        group_name = 'height_groups'
+    else:
+        ConvClass = nn.Conv3d
+        kerenel_size = (5, 3, 1)
+        padding = (2, 1, 0)
+        group_name = 'depth_groups'
+
+    conv = ConvClass(4, 8, kerenel_size, padding=padding, bias=False)
+    with torch.no_grad():
+        ref = conv(x)
+
+    tester = test_util.ModelTester()
+    x = tester.inputs(x.numpy())
+    ref = tester.make_reference(ref.numpy())
+
+    # Test layer
+    kernel = conv.weight.detach().numpy()
+    kernel_weights = lbann.Weights(
+        initializer=lbann.ValueInitializer(values=np.nditer(kernel)),
+        name=f'kernel_{num_dims}d'
+    )
+    ps = {group_name: tools.gpus_per_node(lbann)}
+    y = lbann.Convolution(
+        x,
+        weights=(kernel_weights,),
+        num_dims=num_dims,
+        out_channels=8,
+        kernel_size=kerenel_size,
+        stride=1,
+        padding=padding,
+        dilation=1,
+        has_bias=False,
+        parallel_strategy=ps,
+        name=f'conv_{num_dims}d'
+    )
+    y = lbann.Identity(y)
+    tester.set_loss(lbann.MeanSquaredError(y, ref))
+    tester.set_check_gradients_tensor(lbann.Square(y))
+    return tester

src/layers/learning/convolution.cpp

@@ -264,10 +264,6 @@ bool convolution_layer<TensorDataType, Layout, Device>::is_distconv_supported()
 {
   const auto& kernel_dims = get_kernel_dims();
   for (int i = 0; i < dc::get_num_spatial_dims(*this); i++) {
-    if (kernel_dims[2 + i] != kernel_dims[2]) {
-      dc::MPIRootPrintStreamDebug() << "Nonsymmetric kernel not supported";
-      return false;
-    }
     if (kernel_dims[2 + i] != this->m_pads[i] / this->m_dilations[i] * 2 + 1) {
       dc::MPIRootPrintStreamDebug()
         << "Unsupported as padding does not match the kernel size";