Some changes in operator BLAS L1, require changing autograd to +.= because the += based on apply2_inline was doing implicit broadcast somehow. But then we have cascading issues that requires changing +.= to use broadcast2 and broadcast2 fixes, and then we have the Stack autograd layer that is failing tests. See 2929cb5 in the laser-iterators-stashed branch

Greptile Summary

This WIP PR replaces the existing applyX_inline / mapX_inline family of higher-order tensor operations with the Laser forEach / forEachSerial / forEachStaged iterators across ~20 files. The motivating benefit — demonstrated clearly in nnp_gru.nim — is the ability to fuse multi-tensor element-wise loops with meaningful variable names rather than cryptic x, y, z placeholders, and to avoid needing bespoke apply4_inline / apply10_inline variants.

Key changes:

• higher_order_applymap.nim now imports and re-exports foreach / foreach_staged, making forEach available throughout the tensor module.
• All in-place scalar and element-wise operators (+=, -=, *=, /=, ^.=, etc.) in operators_broadcasted.nim and operators_blas_l1.nim are ported to forEach.
• Activation functions, optimizer update loops (SGD, SGDMomentum, Adam), error metric functions, and GRU cell forward/backward passes are all migrated.
• nnp_gru.nim gains a clean, readable fused GRU cell loop (7–10 tensors in a single forEach), and the previous n_tmp copy-back workaround is correctly eliminated.
• laser/tensor/initialization.nim disables OpenMP in copyFromRaw and setZero to avoid contention when these are called inside an outer OMP region (referencing issue OpenMP barrier slowness in neural network examples with high core count. #485).
• nnp_sigmoid_cross_entropy.nim reverts its parallel map2_inline + sum: reduction to a sequential for loop pending a working forEachStaged implementation.

Issues to address before merging:

• The acknowledged 13× parallel slowdown in GRU (test_nnp_gru) stems from forEach distributing work element-wise across all batch_size × hidden_size elements of 7–10 tensor arguments, many of which are non-contiguous column slices of the same buffer. Adjacent rows in hidden land on the same cache line, causing severe false sharing. Using forEachSerial for the inner loop (as done in softmax) would recover serial performance until a tiled approach is ready.
• sigmoid_cross_entropy now runs sequentially, a direct performance regression for training workloads.
• copyFromRaw/setZero unconditionally lose their OpenMP path; an omp_in_parallel() guard would preserve performance for standalone use.
• optimizerAdam is changed from func to proc without a comment; the reason is unclear since the function body does not call forEach.

Confidence Score: 2/5

• Not safe to merge: acknowledged 13× GRU parallel regression, reverted parallel sigmoid cross-entropy, and disabled OpenMP in core initialization paths are significant regressions.
• The logical migration of forEach is correct and the code reads much better, but there are three unresolved performance regressions (GRU parallel, sigmoid_cross_entropy sequential, initialization OMP disabled) that the PR itself marks as WIP. Merging in the current state would degrade performance for real training workloads.
• src/arraymancer/nn_primitives/nnp_gru.nim (false sharing / 13× slowdown), src/arraymancer/nn_primitives/nnp_sigmoid_cross_entropy.nim (sequential fallback), src/arraymancer/laser/tensor/initialization.nim (OpenMP removed unconditionally)

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[forEach / forEachSerial / forEachStaged macros] --> B[higher_order_applymap.nim\nexports forEach]
    B --> C[tensor/math_functions.nim\nmelwise_mul, melwise_div, mreciprocal, etc.]
    B --> D[tensor/operators_broadcasted.nim\n+.= -.= *.= /.= ^.=]
    B --> E[tensor/operators_blas_l1.nim\n*= /=]
    B --> F[nn_primitives/nnp_activation.nim\nsigmoid, relu, tanh + backwards]
    B --> G[nn/optimizers/optimizers.nim\nSGD, SGDMomentum, Adam update]
    B --> H[ml/metrics/common_error_functions.nim\nsquared_error, relative_error, etc.]
    B --> I[ml/dimensionality_reduction/pca.nim\nexplained_variance]
    B --> J[stats/kde.nim\nnormalization]

    A --> K[nn_primitives/nnp_gru.nim]
    K --> K1[gru_cell_inference\nforEach over 7 tensor slices]
    K --> K2[gru_cell_forward\nforEach over 10 tensor slices]
    K --> K3[gru_cell_backward\nforEach for dh update]

    A --> L[nn_primitives/nnp_softmax.nim\nforEachSerial inside OMP loop]
    A --> M[nn_primitives/nnp_softmax_cross_entropy.nim\nforEachSerial inside OMP loop]

    N[laser/tensor/initialization.nim] --> N1[copyFromRaw: OMP removed\nserial copyMem]
    N --> N2[setZero: OMP removed\nserial zeroMem]

    O[nnp_sigmoid_cross_entropy.nim] --> O1[parallel sum removed\nsequential for loop]

    style K1 fill:#f99,stroke:#c33
    style K2 fill:#f99,stroke:#c33
    style O1 fill:#f99,stroke:#c33
    style N1 fill:#fc9,stroke:#c63
    style N2 fill:#fc9,stroke:#c63

_{Last reviewed commit: d2d625a}