Function GenerateDwconv2d_Simt

dnn/scripts/cutlass_generator/generator.py:1063–1260

(args, conv_kind)

#
def GenerateDwconv2d_Simt(args, conv_kind):
    ################################################################################
    # warps per threadblock
    ################################################################################
    warpsPerThreadblocks = []
    for warpsPerThreadblock0 in warpsPerThreadblockEdge:
        for warpsPerThreadblock1 in warpsPerThreadblockEdge:
            if (
                warpsPerThreadblock0 / warpsPerThreadblock1 <= warpsPerThreadblockRatio
                and warpsPerThreadblock1 / warpsPerThreadblock0
                <= warpsPerThreadblockRatio
                and warpsPerThreadblock0 * warpsPerThreadblock1
                <= warpsPerThreadblockMax
            ):
                warpsPerThreadblocks.append(
                    [warpsPerThreadblock0, warpsPerThreadblock1]
                )

    ################################################################################
    # warp shapes
    ################################################################################
    warpNumThreads = 32
    warpShapes = []
    for warp0 in warpShapeEdges:
        for warp1 in warpShapeEdges:
            if (
                warp0 / warp1 <= warpShapeRatio
                and warp1 / warp0 <= warpShapeRatio
                and warp0 * warp1 <= warpShapeMax
                and warp0 * warp1 > warpShapeMin
            ):
                warpShapes.append([warp0, warp1])

    # sgemm
    (
        precisionType,
        precisionBits,
        threadblockMaxElements,
        threadblockTilesL0,
    ) = precisions["s"]

    layouts = [(LayoutType.TensorNCHW, LayoutType.TensorNCHW)]

    math_instructions = [
        MathInstruction(
            [1, 1, 1],
            DataType.f32,
            DataType.f32,
            DataType.f32,
            OpcodeClass.Simt,
            MathOperation.multiply_add,
        )
    ]

    min_cc = 50
    max_cc = 1024

    dst_layouts = [LayoutType.TensorNCHW]