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 struct

Conv2DKernelConfig

struct Conv2DKernelConfig

Conv-specific geometry for AMD4WaveMatmul's conv2d entry point.

Companion to MatmulKernelConfig: the matmul shape config drives the 4-wave block/warp/MMA tiling, while this struct carries the extra conv2d parameters the im2col loader needs to materialize the A operand from a 4D NHWC input. Pass both at the kernel callsite.

Fields

• ​H (Int): Input spatial height (NHWC dim 1). Used as H_eff when use_runtime_hw=False; ignored otherwise (loader reads runtime H).
• ​W (Int): Input spatial width (NHWC dim 2). Same semantics as H.
• ​H_out (Int): Output spatial height = (H + 2*pad_h - dilation_h*(R-1) - 1) // stride_h + 1. Comptime-only on the static-HW path.
• ​W_out (Int): Output spatial width = (W + 2*pad_w - dilation_w*(S-1) - 1) // stride_w + 1.
• ​R (Int): Filter spatial height.
• ​S (Int): Filter spatial width.
• ​stride_h (Int): Conv vertical stride (>= 1).
• ​stride_w (Int): Conv horizontal stride (>= 1).
• ​dilation_h (Int): Conv vertical dilation (>= 1).
• ​dilation_w (Int): Conv horizontal dilation (>= 1).
• ​pad_h (Int): Vertical pad (>= 0). When > 0, halo lanes route to the SRD-OOB sentinel for zero-clamp behavior.
• ​pad_w (Int): Horizontal pad (>= 0).
• ​C_in (Int): Real input channel count. When > 0, lets the caller K-pad the filter (allocate filter as [Cout, K_padded] where K_padded = round_up(RSC_in, 2*BK) and zero-fill the trailing K rows). When 0, the kernel derives C_in = K_filter // (R*S) and asserts exact divisibility. In 3D mode the divisor is Q*R*S.
• ​use_runtime_hw (Bool): When True, H/W/H_out/W_out (and D/D_out in 3D mode) are runtime values read from the input tensor (typically via graph-compiler symbolic resolution). When False, all conv geometry is comptime.
• ​Q (Int): Filter temporal extent (3D only). Q == 1 keeps the kernel in 2D mode (4D NHWC input, K = RSC). Q > 1 activates 3D mode (5D NDHWC input, K = QRS*C).
• ​D (Int): Input temporal depth (NDHWC dim 1). 2D mode: unused.
• ​D_out (Int): Output temporal depth. 2D mode: unused.
• ​stride_d (Int): Conv temporal stride (>= 1). 2D mode: unused.
• ​dilation_d (Int): Conv temporal dilation (>= 1). 2D mode: unused.
• ​pad_d (Int): Temporal pad (>= 0). 2D mode: unused. When > 0, halo lanes route to the SRD-OOB sentinel for zero-clamp behavior.

Implemented traits

AnyType, Copyable, Deinitable, ImplicitlyCopyable, Movable, Writable

Methods

__init__

def __init__(out self, *, H: Int = Int(1), W: Int = Int(1), H_out: Int = Int(1), W_out: Int = Int(1), R: Int = Int(1), S: Int = Int(1), stride_h: Int = Int(1), stride_w: Int = Int(1), dilation_h: Int = Int(1), dilation_w: Int = Int(1), pad_h: Int = Int(0), pad_w: Int = Int(0), C_in: Int = Int(0), use_runtime_hw: Bool = False, Q: Int = Int(1), D: Int = Int(1), D_out: Int = Int(1), stride_d: Int = Int(1), dilation_d: Int = Int(1), pad_d: Int = Int(0))

Constructs a Conv2DKernelConfig from the conv geometry.

Args:

• ​H (Int): Input spatial height.
• ​W (Int): Input spatial width.
• ​H_out (Int): Output spatial height.
• ​W_out (Int): Output spatial width.
• ​R (Int): Filter spatial height.
• ​S (Int): Filter spatial width.
• ​stride_h (Int): Conv vertical stride.
• ​stride_w (Int): Conv horizontal stride.
• ​dilation_h (Int): Conv vertical dilation.
• ​dilation_w (Int): Conv horizontal dilation.
• ​pad_h (Int): Vertical pad.
• ​pad_w (Int): Horizontal pad.
• ​C_in (Int): Real input channel count.
• ​use_runtime_hw (Bool): When True, treat H/W/H_out/W_out (and D/D_out in 3D mode) as placeholders to be replaced at kernel-launch time with values read from the input tensor.
• ​Q (Int): Filter temporal extent. Q == 1 (default) keeps the kernel in 2D mode; Q > 1 activates 3D mode.
• ​D (Int): Input temporal depth (3D mode).
• ​D_out (Int): Output temporal depth (3D mode).
• ​stride_d (Int): Conv temporal stride (3D mode).
• ​dilation_d (Int): Conv temporal dilation (3D mode).
• ​pad_d (Int): Temporal pad (3D mode).

write_to

def write_to(self, mut writer: T)

Writes a compact conv geometry tag to writer.

Args:

• ​writer (T): Sink for the rendered tag.

write_repr_to

def write_repr_to(self, mut writer: T)

Writes a debug representation of this conv config.

Args:

• ​writer (T): Sink for the rendered tag.