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

dispatch_rdna_conv2d

def dispatch_rdna_conv2d[input_type: DType, filter_type: DType, output_type: DType, filter_is_fcrs: Bool, maybe_epilogue_func: Optional[def[dtype: DType, rank: Int, width: SIMDLength, alignment: Int = Int(1)](IndexList[rank], SIMD[dtype, width]) capturing thin -> None] = None, has_residual: Bool = False](input: TileTensor[input_type, Storage=input.Storage, address_space=input.address_space, linear_idx_type=input.linear_idx_type], filter: TileTensor[filter_type, Storage=filter.Storage, address_space=filter.address_space, linear_idx_type=filter.linear_idx_type], output: TileTensor[output_type, Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type], stride: IndexList[Int(2)], dilation: IndexList[Int(2)], symmetric_padding: IndexList[Int(2)], num_groups: Int, ctx: DeviceContext, source_ptr: Pointer[Scalar[output_type], MutAnyOrigin] = Pointer.unsafe_dangling(), beta: Float32 = 0) -> Bool

Try to dispatch Conv2D on RDNA via implicit GEMM (im2col fused into WMMA).

Returns True if the convolution was handled, False if the caller should fall back to another implementation (e.g. MIOpen).

Uses the implicit GEMM kernel when C_in is aligned to BLOCK_K (covers all FLUX VAE shapes), falling back to explicit im2col + matmul otherwise.

When has_residual=True, folds output = conv + beta * source into the conv epilogue (the RDNA implicit-GEMM/im2col kernels have no native residual path). source_ptr is NHWC-contiguous, same shape as output: e.g. ResNet skip connections that the graph compiler fuses into the conv.

Parameters:

• ​input_type (DType): DType of the input tensor elements.
• ​filter_type (DType): DType of the filter tensor elements.
• ​output_type (DType): DType of the output tensor elements.
• ​filter_is_fcrs (Bool): True if filter is laid out as FCRS, False for RSCF.
• ​maybe_epilogue_func (Optional[def[dtype: DType, rank: Int, width: SIMDLength, alignment: Int = Int(1)](IndexList[rank], SIMD[dtype, width]) capturing thin -> None]): Optional elementwise epilogue applied to each output element (defaults to None).
• ​has_residual (Bool): True to fold a scaled residual beta * source_ptr into the conv epilogue (defaults to False).

Args:

• ​input (TileTensor[input_type, Storage=input.Storage, address_space=input.address_space, linear_idx_type=input.linear_idx_type]): Rank-4 NHWC input tensor.
• ​filter (TileTensor[filter_type, Storage=filter.Storage, address_space=filter.address_space, linear_idx_type=filter.linear_idx_type]): Rank-4 filter tensor in FCRS or RSCF layout per filter_is_fcrs.
• ​output (TileTensor[output_type, Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type]): Rank-4 NHWC output tensor written by the convolution.
• ​stride (IndexList[Int(2)]): Spatial stride [stride_h, stride_w]; only [1, 1] is supported.
• ​dilation (IndexList[Int(2)]): Spatial dilation [dilation_h, dilation_w]; only [1, 1] is supported.
• ​symmetric_padding (IndexList[Int(2)]): Symmetric spatial padding [pad_h, pad_w] applied to the input.
• ​num_groups (Int): Number of convolution groups; only 1 is supported.
• ​ctx (DeviceContext): DeviceContext used to enqueue kernels and synchronize.
• ​source_ptr (Pointer[Scalar[output_type], MutAnyOrigin]): NHWC-contiguous residual source with the same shape as output; read only when has_residual=True (defaults to a dangling pointer).
• ​beta (Float32): Scale factor applied to the residual source when has_residual=True (defaults to 0.0).

Returns:

Bool