For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

conv2d_fprop_with_residual

def conv2d_fprop_with_residual[act_type: DType, filter_type: DType, out_type: DType, *, config: Conv2dConfig[act_type, filter_type, out_type] = Conv2dConfig.default_bf16(), elementwise_lambda_fn: Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None] = None, elementwise_compute_lambda_fn: Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> SIMD[dtype, width]] = None, register_based_epilogue: Bool = True, has_residual: Bool = False](output: TileTensor[out_type, Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type], activation: TileTensor[act_type, Storage=activation.Storage, address_space=activation.address_space, linear_idx_type=activation.linear_idx_type], filter: TileTensor[filter_type, Storage=filter.Storage, address_space=filter.address_space, linear_idx_type=filter.linear_idx_type], source: TileTensor[out_type, Storage=source.Storage, address_space=source.address_space, linear_idx_type=source.linear_idx_type], beta: Float32, problem: Conv2dProblemShape, ctx: DeviceContext)

Launch Conv2D fprop with residual add.

Computes D = Conv(A,B) + beta*C. This function extends conv2d_fprop with residual add support. The epilogue load warp pre-fetches source tensor C via TMA, overlapping with MMA computation for better performance.

The residual add is applied after the optional epilogue lambda: D = lambda(Conv(A,B)) + beta * C

This supports common patterns like:

• Skip connections: D = Conv(A,B) + C (beta=1.0)
• Residual scaling: D = Conv(A,B) + 0.5*C (beta=0.5)
• Fused residual+activation: D = ReLU(Conv(A,B)) + C

Note: The epilogue load warp (warp ID 7) handles C loading when residual is enabled. When has_residual is False or beta is 0, this warp exits early and the kernel behaves identically to conv2d_fprop.

Parameters:

• ​act_type (DType): Data type of the input activation tensor.
• ​filter_type (DType): Data type of the filter weights tensor.
• ​out_type (DType): Data type of the output tensor.
• ​config (Conv2dConfig[act_type, filter_type, out_type]): Kernel configuration (tile sizes, pipeline stages, etc.).
• ​elementwise_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None]): Optional void epilogue lambda applied after output write. Signature: def(IndexList[2], SIMD) -> None.
• ​elementwise_compute_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> SIMD[dtype, width]]): Optional element-wise lambda function for epilogue fusion (bias add, activation). Applied before residual.
• ​register_based_epilogue (Bool): If True, apply lambda in registers (faster).
• ​has_residual (Bool): If True, apply residual add. If False, source is ignored.

Args:

• ​output (TileTensor[out_type, Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type]): Output tensor [N, H_out, W_out, C_out] in NHWC layout (D).
• ​activation (TileTensor[act_type, Storage=activation.Storage, address_space=activation.address_space, linear_idx_type=activation.linear_idx_type]): Input activation [N, H, W, C] in NHWC layout (A).
• ​filter (TileTensor[filter_type, Storage=filter.Storage, address_space=filter.address_space, linear_idx_type=filter.linear_idx_type]): Filter weights [K, R, S, C] in KRSC layout (B).
• ​source (TileTensor[out_type, Storage=source.Storage, address_space=source.address_space, linear_idx_type=source.linear_idx_type]): Source tensor [N, H_out, W_out, C_out] for residual (C).
• ​beta (Float32): Residual scale factor. If 0.0, no residual is applied.
• ​problem (Conv2dProblemShape): Convolution problem shape specification.
• ​ctx (DeviceContext): Device context for kernel launch.

Raises:

Error if kernel launch fails, constraints are violated, or source tensor shape doesn't match output shape.