Module channel_bias_add 

Channel bias add Triton kernels.

Layout: input x and output y are NC-layout, where N = BHW and C = number of channels. Bias is a (C,) vector. Element x[n, c] lives at flat offset n * C + c.

Parallelism: one CTA per channel. Each CTA iterates over all N spatial elements in BLOCK_N-wide tiles, adding the per-channel scalar bias.

Structs

ChannelBiasAddBackward

Backward pass for channel bias add.

ChannelBiasAddForward

Adds a (C,) bias to a tensor in NC layout (N = BHW, C = channels).

ChannelBiasAddRuntimeOp

Combined forward + backward RuntimeOp for channel bias add.

NchwBiasAddBackward

NCHW bias add backward: dx = dy, dbias[c] = sum over (B, H, W) of dy.

NchwBiasAddForward

Adds a (C,) bias to a tensor in NCHW layout.

NchwBiasAddRuntimeOp

RuntimeOp for adding a (C,) bias to an NCHW-layout tensor.

Functions

channel_bias_add_backward

Backward pass for channel bias add.

channel_bias_add_forward

Adds a (C,) bias to a tensor in NC layout (N = BHW, C = channels).

nchw_bias_add_backward

NCHW bias add backward: dx = dy, dbias[c] = sum over (B, H, W) of dy.

nchw_bias_add_forward

Adds a (C,) bias to a tensor in NCHW layout.