Speed up Conv1D/Conv2D with large dilation_rate (closes #132)

[Dobiasd]

Summary

Fixes the long-standing slowdown for Conv1D/Conv2D with large dilation_rate reported in #132.

The previous implementation inflated the filter with zeros at construction time (dilate_filter → dilate_tensor). For a 3-tap kernel with dilation_rate=64, this produced a 129-tap kernel where 126 of the 129 entries are zero — but the Eigen GEMM still multiplied through every one of them.

This PR keeps filters un-dilated and threads dilation_rate through convolve(). In the strides=(1,1) fast path, the loop now iterates over the original (y_filt, x_filt) positions and runs one GEMM per position, indexing the padded input at (y_filt * dil_y, x_filt * dil_x). The existing output_temp wraparound trick keeps working as long as out_width_temp = out_width + (f_width - 1) * dil_x (which equals the padded input width).

Benchmark

Stacked-WaveNet style net (3 levels × 7 branches at dilations 1, 2, 4, 8, 16, 32, 64; 16 channels; input shape (18000, 3)):

	Forward pass
master	159 ms
this PR	49 ms
same model with all dilation_rate=1	46 ms

So the dilation tax drops from ~3.5× to ~6 %.

Scope

• conv_2d_layer (covers Conv1D + Conv2D) is updated.
• conv_2d_transpose_layer, depthwise_conv_2d_layer, separable_conv_2d_layer are unchanged — they continue to pre-dilate filters where applicable. Those paths are not the bottleneck reported in the issue.
• Conv3D's dilate_filter_3d could get the same treatment in a follow-up; left out of this PR to keep it focused.

Test plan

• test_model_exhaustive.json — passes within 1e-5 tolerance (covers Conv1D dilation_rate=2 causal, Conv2D dilation_rate=(2,3) same, Conv3D dilation_rate=(2,2,3) same).
• WaveNet model from the benchmark — passes within 1e-5 against the Keras reference.
• Mini 2D dilated model with padding='same' and asymmetric dilation (2, 3) — passes.
• Forward pass on VGG19 (no dilation, exercises the unchanged s1x1 and strided paths) — runs.

🤖 Generated with Claude Code