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
VarlenConvIO
struct VarlenConvIO[weight_origin: Origin[mut=weight_origin.mut], bias_origin: Origin[mut=bias_origin.mut], out_origin: MutOrigin, dtype: DType, weight_layout: TensorLayout, bias_layout: TensorLayout, out_layout: TensorLayout, weight_engine: TensorEngine, bias_engine: TensorEngine, out_engine: TensorEngine, weight_addr: AddressSpace, bias_addr: AddressSpace, out_addr: AddressSpace, weight_idx: DType, bias_idx: DType, out_idx: DType, XFn: def[width: Int, alignment: Int](Int, Int) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable, //, channels_last: Bool]
Owner of the forward-path DRAM reads/writes for varlen causal conv1d.
Holds the input reader, the (dim, width) weight, (dim,) bias, and
output TileTensor views and exposes one method per access verb.
x and output are the caller's physical tensors: (dim, seqlen), or
(seqlen, dim) when channels_last. The methods take the logical
(d, s) and order the coordinates by channels_last, so the axis order stays
out of the kernels.
Every view parameter is inferred from the constructor arguments, so the
owner adapts to whatever storage the caller's views carry. Only the store
target output pins a mutable origin.
The CPU reference is the owner's user. The GPU forward kernel indexes its tensors directly: on Metal, that kernel's accesses through these fields read zero and drop stores.
Parameters
- weight_origin (
Origin[mut=mut]): Inferred origin of the weight view. - bias_origin (
Origin[mut=mut]): Inferred origin of the bias view. - out_origin (
MutOrigin): Inferred mutable origin of the output view. - dtype (
DType): Element type ofx,weight,biasandoutput. - weight_layout (
TensorLayout): Layout type of the(dim, width)weight view. - bias_layout (
TensorLayout): Layout type of the(dim,)bias view. - out_layout (
TensorLayout): Layout type of the output view. - weight_engine (
TensorEngine): Inferred engine of the weight view. - bias_engine (
TensorEngine): Inferred engine of the bias view. - out_engine (
TensorEngine): Inferred engine of the output view. - weight_addr (
AddressSpace): Inferred address space of the weight view. - bias_addr (
AddressSpace): Inferred address space of the bias view. - out_addr (
AddressSpace): Inferred address space of the output view. - weight_idx (
DType): Inferred linear-index type of the weight view. - bias_idx (
DType): Inferred linear-index type of the bias view. - out_idx (
DType): Inferred linear-index type of the output view. - XFn (
def[width: Int, alignment: Int](Int, Int) -> SIMD[dtype, width]&RegisterPassable&ImplicitlyCopyable): Inferred reader ofxat its physical(row, column)index. - channels_last (
Bool): Whetherxandoutputare(seqlen, dim)rather than(dim, seqlen).
Fields
- x_fn (
XFn): - weight (
TileTensor[dtype, weight_layout, weight_origin, Engine=weight_engine, address_space=weight_addr, linear_idx_type=weight_idx]): - bias (
TileTensor[dtype, bias_layout, bias_origin, Engine=bias_engine, address_space=bias_addr, linear_idx_type=bias_idx]): - output (
TileTensor[dtype, out_layout, out_origin, Engine=out_engine, address_space=out_addr, linear_idx_type=out_idx]):
Implemented traits
AnyType,
Copyable,
Deinitable,
ImplicitlyCopyable,
Movable
Methods
dim
load_x
def load_x[width: Int = Int(1)](self, d: Int, s: Int) -> SIMD[dtype, width]
Load x[d:d + width, s]: windowed history read of the input.
Parameters:
- width (
Int): The number of consecutive channels; above 1 needschannels_last, where they are contiguous.
Args:
Returns:
load_weight
def load_weight(self, d: Int, w: Int) -> Scalar[dtype]
Load weight[d, w]: per-channel conv tap.
Args:
- d (
Int): The channel index into the(dim, width)weight view. - w (
Int): The convolution tap index in[0, width).
Returns:
load_bias
def load_bias[width: Int = Int(1)](self, d: Int) -> SIMD[dtype, width]
Load bias[d:d + width], or zero when bias is empty.
Parameters:
- width (
Int): The number of consecutive channels.
Args:
- d (
Int): The first channel index into the(dim,)bias view.
Returns:
store_out
def store_out[width: Int = Int(1)](self, d: Int, s: Int, val: SIMD[dtype, width])
Store output[d:d + width, s] = val: convolution output.
Parameters:
- width (
Int): The number of consecutive channels, as inload_x.
Args:
- d (
Int): The first channel index. - s (
Int): The packed sequence position. - val (
SIMD[dtype, width]): The convolution result to store.