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

Mojo function

dispatch_amd_4wave_conv2d

def dispatch_amd_4wave_conv2d[input_type: DType, filter_type: DType, output_type: DType, filter_is_fcrs: Bool, has_residual: Bool = False, elementwise_lambda_fn: Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None] = None](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: Optional[Pointer[Scalar[output_type], MutAnyOrigin, _safe=False]] = None, beta: Float32 = 0) -> Bool

Try to dispatch a Conv2D to amd_4wave_conv on MI355X.

Returns True if the convolution was handled; False if the caller should fall back (typically to MIOpen). See module docstring for the full acceptance criteria.

When has_residual=True and source_ptr is set, computes D = Conv(A, B) + beta * source via the in-kernel fused residual path (amd_4wave_conv[has_residual=True]). The source pointer is expected to point to an NHWC-contiguous buffer with the same shape as output. When has_residual=False (default), the call is identical to the no-residual variant: no extra ABI overhead beyond the launch packet's 16 bytes (DCE'd source_ptr / stride / beta).

Parameters:

input_type (DType): DType of the input activation tensor; must be float8_e4m3fn, bfloat16, or float16.
filter_type (DType): DType of the filter tensor; must equal input_type.
output_type (DType): DType of the output tensor; bfloat16 for FP8 inputs, otherwise tracks input_type.
filter_is_fcrs (Bool): True when filter is in FCRS layout; False for RSCF. Selects the transpose path and the R/S dim indices.
has_residual (Bool): True to fuse D = Conv + beta * source in-kernel via source_ptr (defaults to False).
elementwise_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None]): Optional fused 2D-coord epilogue lambda wrapping a 4D NHWC elementwise_simd_epilogue_type; None fuses no epilogue (defaults to None).

Args:

input (TileTensor[input_type, Storage=input.Storage, address_space=input.address_space, linear_idx_type=input.linear_idx_type]): Input activation TileTensor, rank-4 NHWC with static C_in; N, H, W may be dynamic.
filter (TileTensor[filter_type, Storage=filter.Storage, address_space=filter.address_space, linear_idx_type=filter.linear_idx_type]): Filter TileTensor, rank-4 FCRS or RSCF per filter_is_fcrs, with static R, S, C_in, C_out.
output (TileTensor[output_type, Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type]): Output activation TileTensor, rank-4 NHWC, mutable, with static C_out.
stride (IndexList[Int(2)]): Convolution stride [stride_h, stride_w]; must be (1, 1) or (2, 2) with stride_h == stride_w.
dilation (IndexList[Int(2)]): Dilation [dh, dw]; must be (1, 1).
symmetric_padding (IndexList[Int(2)]): Symmetric padding [ph, pw]; must be (0, 0), (1, 1), or (2, 2) with ph == pw.
num_groups (Int): Number of convolution groups; must be 1.
ctx (DeviceContext): DeviceContext for kernel launches and buffer allocation.
source_ptr (Optional[Pointer[Scalar[output_type], MutAnyOrigin, _safe=False]]): Optional pointer to an NHWC-contiguous residual source buffer with the same shape as output; read only when has_residual=True (defaults to None).
beta (Float32): Scale factor for the residual term in D = Conv + beta * source (defaults to 0.0).

Returns:

Bool