Module conv 

Convolution requests for the cuDNN actor (Phase 2 frontend graph API).

Three op families:

• ConvFwdRequest<T> — y = conv(x, w) [+ optional bias + optional activation, when fused via epilogue].
• ConvBwdDataRequest<T> — dx = conv_bwd_data(w, dy).
• ConvBwdFilterRequest<T> — dw = conv_bwd_filter(x, dy).

Each supports 1D / 2D / 3D, NCHW + NHWC packed layouts (or fully strided), arbitrary group count, and arbitrary dilation.

Structs

ConvBwdDataRequest

Backward-data convolution: dx = alpha * conv_bwd_data(w, dy) + beta * dx.

ConvBwdFilterRequest

Backward-filter convolution: dw = alpha * conv_bwd_filter(x, dy) + beta * dw.

ConvDescParams

Convolution descriptor parameters, dimension-generic.

ConvFwdRequest

Forward convolution: y = alpha * conv(x, w) + beta * y, optionally with a fused bias / activation tail.

Enums

EpilogueKind

Optional fused epilogue tail attached to conv-fwd. The bias is represented as an opaque marker on the spec layer (the graph builder records “there is a bias of this dtype + shape”); the concrete GpuRef<T> lives on ConvFwdRequest proper.

Functions

build_conv_bwd_data_graph

Build the spec-side conv-bwd-data op graph.

build_conv_bwd_filter_graph

Build the spec-side conv-bwd-filter op graph.

build_conv_fwd_graph

Build the spec-side conv-fwd op graph, parameterised by dtype + dims. Independent of GpuRef so callers (tests, plan-cache lookup) can build it without owning device buffers.