Binary spike projections / 二值脉冲投影

普通 Linear/Conv 使用 PyTorch 层，省显存训练通过 memopt 应用压缩和 checkpoint。 旧 SpikeLinear/SpikeConv* 与 spike_linear/spike_conv* 已删除。

Use ordinary PyTorch Linear/Conv layers and memopt compression/checkpointing for memory optimization. The legacy SpikeLinear/SpikeConv* and spike_linear/spike_conv* interfaces have been removed.

if_linear / lif_linear 保留 CUDA FP32、初态梯度和自定义替代梯度语义。 原生扩展可用时执行融合前向并在反向重计算；否则使用 Torch 参考路径，后者不保证 不物化中间脉冲。packed_spike_linear 和 sparse_linear(strategy="sparse") 支持 CUDA FP32/FP16/BF16，以原生 kernel 或 Torch 参考执行。

if_linear / lif_linear preserve CUDA FP32, initial-state gradients and custom surrogates. A built native extension runs fused forward and rematerializes in backward. Otherwise they use Torch reference execution, which does not promise elimination of intermediate spikes. packed_spike_linear and sparse_linear(strategy="sparse") support CUDA FP32/FP16/BF16 using native kernels or Torch reference execution.

普通位压缩是一维、最低位优先的格式；projection 按行打包会逐行补零，不能直接 用全局打包替代。memopt 的压缩器独立保留。普通压缩 CUDA 自动选择 Triton/Torch； CPU 使用 Torch，不依赖 GPU 包。没有 CuPy 依赖。

Flat packing is one-dimensional, least-significant-bit first. Projection packing zero-pads each row and cannot generally be replaced with flat packing. memopt compressors remain independent. Ordinary packing selects Triton/Torch on CUDA and Torch on CPU; no CuPy dependency is needed.

import torch
from spikingjelly.activation_based import functional

x = torch.tensor([[0., 1., 1.], [1., 0., 1.]])
packed = functional.bit_spike_compress(x)
restored = functional.bit_spike_decompress(packed, tuple(x.shape), x.dtype)
assert torch.equal(restored, x)