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

conv3d_gpu_naive_ndhwc_qrscf

def conv3d_gpu_naive_ndhwc_qrscf[input_layout: Layout, filter_layout: Layout, output_layout: Layout, input_type: DType, filter_type: DType, output_type: DType, block_size: Int, maybe_epilogue_func: Optional[def[dtype: DType, rank: Int, width: SIMDLength, alignment: Int = Int(1)](IndexList[rank], SIMD[dtype, width]) capturing thin -> None]](input: LayoutTensor[input_type, input_layout, MutAnyOrigin], filter: LayoutTensor[filter_type, filter_layout, MutAnyOrigin], output: LayoutTensor[output_type, output_layout, MutAnyOrigin], stride: IndexList[Int(3)], dilation: IndexList[Int(3)], padding: IndexList[Int(3)], num_groups: Int)

Naive GPU kernel for 3D NDHWC convolution with QRSCF filter layout.

Each thread computes one output voxel across all output channels, iterating over the Q x R x S filter window with vectorized input loads and scalar filter accumulation.

Arguments: input: Input activation tensor in NDHWC layout. filter: Convolution weight tensor in QRSCF layout. output: Output activation tensor in NDHWC layout. stride: Per-spatial-axis convolution stride as (depth, height, width). dilation: Per-spatial-axis filter dilation as (depth, height, width). padding: Per-spatial-axis symmetric padding as (depth, height, width). num_groups: Number of convolution groups for grouped convolution.

Parameters:

input_layout (Layout): Memory layout of the input tensor (NDHWC).
filter_layout (Layout): Memory layout of the filter tensor (QRSCF).
output_layout (Layout): Memory layout of the output tensor (NDHWC).
input_type (DType): Element type of the input tensor.
filter_type (DType): Element type of the filter tensor.
output_type (DType): Element type of the output tensor.
block_size (Int): Thread block size used for both x and y block dimensions.
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 SIMD epilogue applied to each computed output value in place of a direct store.