Expand description
Tensor operation kernels and backend dispatch for the Coeus stack.
§Operation families
- Elementwise —
unaryandbinarykernels dispatched viaElementwiseOps::elementwise_unary/ElementwiseOps::elementwise_binary. - Linear algebra —
matmul(),sparseSpMM/SpMV. - Reductions —
reduction:sum,mean,max/min,argmax/argmin,topk, cumulative sum/product scans, plus vector arithmeticdot(flat inner product) andcross(per-channel 3-vector cross alongdim), plus matrix normsfrobenius_norm/frobenius_norm_batched(compose onnormfortorch.linalg.matrix_norm(A, ord='fro')). - Convolution — 1-D/2-D/3-D forward+backward routed through
BackendOps::conv1d/conv2d/conv3d. - Pooling — max and average pooling (2-D/3-D) with backward gradients.
- Resize — nearest/bilinear/bicubic/area spatial resize (1-D/2-D) in
resize, with the matching reverse-mode scatter. - Attention —
attention::scaled_dot_product_attentionwith causal/padding mask support. - Fused expressions —
fuselazy expression DAG evaluated in a single pass on CPU. - Optimizer steps — provider-dispatched
sgd_step,adam_step,adamw_step,rmsprop_step, andadagrad_steponOptimizerOps.
Re-exports§
pub use backend_ops::cross_fold;pub use backend_ops::AttentionOps;pub use backend_ops::AttentionScalar;pub use backend_ops::BackendOps;pub use backend_ops::ConvOps;pub use backend_ops::ConvolutionBackward;pub use backend_ops::ConvolutionForward;pub use backend_ops::CpuBackend;pub use backend_ops::CrossEntropyOps;pub use backend_ops::CrossOps;pub use backend_ops::CtcBatch;pub use backend_ops::CtcOps;pub use backend_ops::ElementwiseOps;pub use backend_ops::FiniteDifference3DOps;pub use backend_ops::MatmulOps;pub use backend_ops::OptimizerOps;pub use backend_ops::OptimizerStateRef;pub use backend_ops::OptimizerStepRule;pub use backend_ops::OptimizerStepValidation;pub use backend_ops::PoolOps;pub use backend_ops::RandomInitOps;pub use backend_ops::ReductionOps;pub use backend_ops::RotateHalfOps;pub use backend_ops::ScalarPowerOps;pub use backend_ops::StaggeredPairOps;pub use backend_ops::UnfoldFoldOps;pub use unary::abs;pub use unary::abs_assign;pub use unary::acos;pub use unary::acosh;pub use unary::asin;pub use unary::asinh;pub use unary::atan;pub use unary::atanh;pub use unary::causal_softmax;pub use unary::ceil;pub use unary::ceil_assign;pub use unary::cos;pub use unary::cos_assign;pub use unary::cosh;pub use unary::elementwise_unary;pub use unary::elementwise_unary_assign;pub use unary::elementwise_unary_to;pub use unary::elu;pub use unary::elu_assign;pub use unary::erf;pub use unary::erfc;pub use unary::exp;pub use unary::exp2;pub use unary::exp_assign;pub use unary::expm1;pub use unary::floor;pub use unary::floor_assign;pub use unary::gelu;pub use unary::gelu_assign;pub use unary::gelu_tanh;pub use unary::gelu_tanh_assign;pub use unary::glu;pub use unary::leaky_relu;pub use unary::leaky_relu_assign;pub use unary::lgamma;pub use unary::log;pub use unary::log10;pub use unary::log1p;pub use unary::log2;pub use unary::log_assign;pub use unary::log_softmax_axis;pub use unary::masked_softmax;pub use unary::mish;pub use unary::mish_assign;pub use unary::neg;pub use unary::neg_assign;pub use unary::pow_scalar;pub use unary::recip;pub use unary::recip_assign;pub use unary::relu;pub use unary::relu_assign;pub use unary::round;pub use unary::round_assign;pub use unary::sigmoid;pub use unary::sigmoid_assign;pub use unary::sign;pub use unary::sign_assign;pub use unary::silu;pub use unary::silu_assign;pub use unary::sin;pub use unary::sin_assign;pub use unary::sinh;pub use unary::softplus;pub use unary::softplus_assign;pub use unary::sqrt;pub use unary::sqrt_assign;pub use unary::tan;pub use unary::tanh;pub use unary::tanh_assign;pub use unary::trunc;pub use unary::trunc_assign;pub use binary::add;pub use binary::add_assign;pub use binary::div;pub use binary::div_assign;pub use binary::elementwise_binary;pub use binary::elementwise_binary_to;pub use binary::eq;pub use binary::ge;pub use binary::gt;pub use binary::le;pub use binary::lt;pub use binary::mul;pub use binary::mul_assign;pub use binary::ne;pub use binary::sub;pub use binary::sub_assign;pub use embedding::embedding;pub use embedding::embedding_backward;pub use embedding::embedding_backward_with_padding_idx;pub use interpolation::linear_interpolation;pub use interpolation::linear_interpolation_backward;pub use interpolation::BoundaryPolicy;pub use interpolation::Dimension;pub use interpolation::InterpolationError;pub use interpolation::InterpolationFailure;pub use interpolation::InterpolationGradients;pub use interpolation::Replicate;pub use interpolation::SupportedDimension;pub use matmul::bmm;pub use matmul::matmul;pub use matmul::matmul_accumulate;pub use matmul::outer;pub use reduction::amax;pub use reduction::amin;pub use reduction::argmax;pub use reduction::argmin;pub use reduction::cross;pub use reduction::cumprod;pub use reduction::cumprod;pub use reduction::cumsum;pub use reduction::cumsum;pub use reduction::dot;pub use reduction::frobenius_norm;pub use reduction::frobenius_norm_batched;pub use reduction::max_axis;pub use reduction::mean;pub use reduction::mean;pub use reduction::mean_axis;pub use reduction::min_axis;pub use reduction::norm;pub use reduction::norm_p;pub use reduction::norm_p_axis;pub use reduction::norm_p_tensor;pub use reduction::prod;pub use reduction::prod_axis;pub use reduction::prod_tensor;pub use reduction::std_dev;pub use reduction::std_dev_axis;pub use reduction::std_mean;pub use reduction::std_mean_axis;pub use reduction::suffix_prod;pub use reduction::suffix_sum;pub use reduction::sum;pub use reduction::sum;pub use reduction::sum_axis;pub use reduction::topk;pub use reduction::topk;pub use reduction::var;pub use reduction::var_axis;pub use reduction::var_mean;pub use reduction::var_mean_axis;pub use resize::resize_1d;pub use resize::resize_1d_backward;pub use resize::resize_1d_backward_with_scale;pub use resize::resize_1d_with_scale;pub use resize::resize_2d;pub use resize::resize_2d_backward;pub use resize::resize_2d_backward_with_scale;pub use resize::resize_2d_with_scale;pub use resize::ResizeMode;pub use rotate_half::rotate_half;pub use shape::broadcast_to;pub use shape::cat;pub use shape::chunk;pub use shape::diag;pub use shape::diagonal;pub use shape::einsum;pub use shape::einsum3;pub use shape::flip;pub use shape::gather;pub use shape::gather;pub use shape::index_put;pub use shape::index_select;pub use shape::masked_fill;pub use shape::masked_select;pub use shape::meshgrid;pub use shape::nonzero;pub use shape::one_hot;pub use shape::pad;pub use shape::repeat_interleave;pub use shape::roll;pub use shape::scatter_add;pub use shape::sort;pub use shape::split;pub use shape::stack;pub use shape::tile;pub use shape::tril;pub use shape::triu;pub use shape::where_cond;pub use sparse::coo_to_csr;pub use sparse::coo_to_dense;pub use sparse::csr_to_dense;pub use sparse::dense_to_coo;pub use sparse::dense_to_csr;pub use sparse::spmm;pub use sparse::spmm_backward_dense;pub use sparse::spmm_backward_values;pub use sparse::spmv;pub use fuse::evaluate_fused_cpu;pub use fuse::evaluate_fused_mean_cpu;pub use fuse::evaluate_fused_reduce_cpu;pub use fuse::scalar;pub use fuse::CpuExprNode;pub use fuse::Expr;pub use fuse::ExprNode;pub use fuse::TensorExprExt;pub use attention::scaled_dot_product_attention;pub use attention::scaled_dot_product_attention_backward;pub use conv_transpose::conv_transpose1d;pub use conv_transpose::conv_transpose2d;pub use conv_transpose::conv_transpose3d;pub use adaptive_pool::adaptive_avg_pool1d;pub use adaptive_pool::adaptive_avg_pool2d;pub use adaptive_pool::adaptive_max_pool1d;pub use adaptive_pool::adaptive_max_pool2d;pub use constructors::geomspace;pub use constructors::linspace;pub use constructors::logspace;pub use unfold_fold::fold1d;pub use unfold_fold::fold2d;pub use unfold_fold::unfold1d;pub use unfold_fold::unfold2d;
Modules§
- adaptive_
pool - Adaptive pooling operations (avg and max in 1D/2D).
- attention
- Scaled dot-product attention with causal and padding mask support.
- backend_
ops - Backend dispatch trait and CPU implementation.
- binary
- Element-wise binary operations (add, sub, mul, div).
- constructors
- Tensor constructors (linspace, logspace, geomspace).
- conv_
transpose - Transposed convolution operations (1-D, 2-D, and 3-D).
- embedding
- Embedding lookup and backward pass.
- fuse
- Device-neutral fused expression DAG with CPU-addressable evaluation support.
- interpolation
- Coordinate-grid interpolation operations. Dimension-generic coordinate-grid interpolation operations.
- matmul
- Matrix multiplication kernels (matmul, bmm, outer).
- reduction
- Reduction operations (sum, mean, max, min, norms, variance). Reduction and selection operators over one or more tensor axes.
- resize
- Spatial resize (nearest / bilinear / bicubic / area) forward and reverse mode. Spatial resize (nearest / bilinear / bicubic / area) with reverse mode.
- rotate_
half - Provider-selected half-vector rotation.
- shape
- Shape manipulation operations (cat, stack, reshape, flip, sort, etc.).
- sparse
- Sparse tensor operations (SpMM, SpMV, format conversions). Sparse tensor operations: COO/CSR format conversions and sparse BLAS routines.
- unary
- Element-wise unary operations (activations, math functions).
- unfold_
fold - Sliding-window extraction and adjoint accumulation. Sliding-window extraction and adjoint accumulation.
Enums§
- Axis
- Arbitrary-even-order staggered gradient/divergence pair for a Yee leapfrog. The spatial axis a gradient or divergence differentiates along.
- Binary
Op - Binary element-wise operation tag.
- Finite
Difference3D Scheme - Generic 3-D finite-difference operator (provider-SSOT for kwavers/CFDrs/helios
first-derivative kernels: central 2nd/4th/6th + Yee staggered forward/backward).
Stencil family + kernel ordering for
FiniteDifference3D. - Reduction
Op - Reduction operation tag.
- UnaryOp
- CPU unary operation dispatch tag.