Skip to main content

✅ Tasks

Status:PreliminaryIn progress
Tags:StatusTask Log
Last modified:

Project-wide task log

This page tracks project work that may or may not be connected to a specific experiment. Tasks can describe code, documentation, operations, slides, meetings, server/device setup, or research planning.

The task log is generated from the shared task registry in site/src/data/tasks.js. The same registry now drives the project roadmap on the overview page.

Task lifecycle convention

A task with status todo is unassigned. Once someone starts working on it, set status: 'inprogress', add that person to owners, and fill start_date. Once finished, set status: 'completed' and fill end_date. due_date may be left empty, which renders as UNDEFINED.

Each task is classified by three axes:

  • stage: lifecycle position, such as research, implementation, validation, integration, deployment, or publication.
  • target: the artifact or system being changed, such as enabol, software, hls4ml, cpp_backend, csim, fpga, or paper.
  • action: the kind of work, such as design, development, documentation, experiment, testing, integration, analysis, or writing.

Active Tasks​

Filter tasks68 of 68 shown
HLS4ML-008In progress

HLS4ML - Implement synthesizable controller kernels

CTRL-NONE, CTRL-GT-ORDER-0, and CTRL-GT-ORDER-1 are implemented and wired. GT-0/GT-1 now use raw optimizer update geometry plus division-free candidate search over α²·η²·||ΔG||² ≤ χ²·(||Δθ_raw||²+ε²), with a nonzero alpha candidate floor. CTRL-GT-ORDER-2/2-QA remain.

hls4mlcontrollersglobal-throttleCTRL
Due
UNDEFINED
Start
2026-05-22
End
UNDEFINED
HLS4ML-009In progress

HLS4ML - Add global metric reductions and alpha synchronization

Raw-update controller sensing is separated from controller law (raw_update_sensor_order0 → global_throttle_*_law). The writer emission now runs backpass → SGD proposal → raw-update/dgrad reduction → law → alpha-scaled apply. Per-layer raw update norms and ΔG norms are summed into global accumulators before the law call. CTRL-NONE skips curvature sensing.

hls4mlalpha-syncnorm-reductionglobal-throttle
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned
HLS4ML-037In progress

HLS4ML - Add optional trainable CSIM trace logging

Current slice implemented: generated trainable CSIM writes metadata-bearing comma-delimited .dat traces under tb_data/training/ for loss, alpha, and per-layer epoch-level parameter traces such as dense0/weights.dat and dense0/biases.dat. Repeated epoch/sample/global_step/sample_index columns allow sparse traces with different cadences. Broader trace selection for gradients and controller internals remains open.

hls4mltracedebugcsim
Due
UNDEFINED
Start
2026-05-23
End
UNDEFINED
HLS4ML-035Unassigned

HLS4ML - Add Adam optimizer support

Future Adam support with explicit state buffers and bias-correction policy. The deprecated code mixes double/std::pow paths and large static arrays, so this must be rewritten carefully before synthesis.

hls4mloptimizeradam
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned
HLS4ML-013Unassigned

HLS4ML - Validate one-layer Dense trainable CSIM

Run the minimal one-dense affine regression case through generated trainable firmware in CSIM and compare forward values, loss, gradients, alpha, and updated weights against ENABOL software traces.

hls4mlcsimdenseLIN1
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned
ENB-017Unassigned

Run CSIM validation with global throttle (CTRL-GT-ORDER-0, division-free fix)

GT-0 CSIM failed: two ap_fixed divisions (sqrt + divide) crashed CSIM at step 1. Fixed by replacing α = χ/(ηC+ε) with division-free inequality search α²·η²·||ΔG||² ≤ χ²·(||Δθ_raw||² + ε²). The raw update is the optimizer proposal before alpha/quantized assignment; for SGD it already includes η. Candidate table of binary-fraction α values is searched in descending order with a nonzero alpha floor. Validate: compile 1-Dense model with Kind: CTRL-GT-ORDER-0, run CSIM, confirm α selection, no CSIM errors, loss drops.

csimcontrollerfixed-pointGT0division-free
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
HLS4ML-014Unassigned

HLS4ML - Extend trainable CSIM to two Dense layers with activation

After the one-layer Dense path is correct, validate reverse traversal through Dense -> activation -> Dense before moving to convolution and BatchNorm paths.

hls4mlcsimdense2activation
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned
ENB-019Unassigned

Run FPGA hardware drift-recovery experiment

Collect the real hardware baseline, post-drift initial point, and post-training recovery point for controller-enabled learning.

FPGAdriftrecovery-curve
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned
ENB-018Unassigned

Run HLS synthesis and resource validation

Check resource impact, timing, and generated design viability after CSIM validates behavior.

HLSresourceslatency
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned
ENB-028Done

Document trainable precision schema relationships

Added the first precision relationship document. Unknown trainable precision dependencies are explicitly marked with ?, and the known rule for controller_metric records that dtheta_sq needs fractional bits for squared parameter movement: F_controller_metric >= 2*F_theta + guard bits.

precisionschemacontroller-metrichls4ml
Due
UNDEFINED
Start
2026-05-23
End
2026-05-23
HLS4ML-039Done

HLS4ML - Use raw-update geometry for global throttle control

Replaced actual/throttled parameter-delta controller geometry with raw optimizer update geometry. SGD now runs before the law, raw_update_sensor_order0 sums ||Δθ_raw||² and ||ΔG||² globally, GT-0/GT-1 keep the division-free squared inequality, alpha uses a nonzero candidate floor, and controller.dat logs raw/controlled/actual update norms plus feasibility telemetry.

hls4mlcontrollerraw-updateglobal-throttle
Due
UNDEFINED
Start
2026-05-23
End
2026-05-23
HLS4ML-017Done

HLS4ML - Emit trainable top-level IO for loss and training controls

Done for the minimal path: generated top-level firmware receives ground truth, scalar loss output, alpha output, train_enable, reset_accumulators, and batch_end. The testbench trains from tb_output_predictions.dat, while the Python bridge supplies zero trainable tensors with train_enable=false for inference-safe calls.

hls4mlwritertop-level-ioloss
Due
UNDEFINED
Start
2026-05-22
End
2026-05-22
HLS4ML-021Done

HLS4ML - Implement CTRL-GT-ORDER-0 controller

Implemented CTRL-GT-ORDER-0 as a global law fed by raw_update_sensor_order0. The writer sums per-layer ||Δθ_raw||² and ||ΔG||² contributions into global accumulators, then GT-0 selects the largest candidate alpha satisfying α²·η²·||ΔG||² ≤ χ²·(||Δθ_raw||²+ε²). This avoids sqrt/division and avoids the self-locking actual-update denominator.

hls4mlcontrollerglobal-throttleGT0
Due
UNDEFINED
Start
2026-05-23
End
2026-05-23
HLS4ML-002Done

HLS4ML - Add trainable ModelGraph flow and validation passes

Done in hls4ml-trainable: registered vivado:trainable after template resolution and before writer emission, added vivado:validate_trainable_config for the first supported Dense trainable subset, and added vivado:resolve_trainable_backward_order to attach trainable_forward_path, trainable_forward_order, trainable_backward_order, and trainable_output_layer while rejecting branching graphs. Loss endpoint metadata remains tracked under HLS4ML-006.

hls4mlModelGraphflowsvalidation
Due
UNDEFINED
Start
2026-05-22
End
2026-05-22
HLS4ML-003Done

HLS4ML - Add trainable layer attributes and precision resolution

Done in hls4ml-trainable: Layer initialization now writes a trainable flag and converts configured trainable precision fields into NamedType attributes such as grad_in_t, grad_out_t, raw_update_t, update_t, gradient_accum_t, controller_metric_t, and alpha_t.

hls4mllayer-attributesprecisiongradients
Due
UNDEFINED
Start
2026-05-22
End
2026-05-22
HLS4ML-011Done

HLS4ML - Add minimal Vivado writer hooks for trainable emission

Done in hls4ml-trainable: the Vivado writer copies trainable headers, emits trainable includes/config structs, allocates trainable buffers, emits the minimal loss/backprop/SGD/CTRL-NONE/apply-update call chain, and keeps bridge inference calls safe with train_enable=false stubs.

hls4mlvivado-writercodegenIO
Due
UNDEFINED
Start
2026-05-22
End
2026-05-22
HLS4ML-038Done

HLS4ML - Emit controller metric trace from trainable CSIM

Generated trainable top-level functions now expose controller metric outputs and the trainable CSIM testbench writes tb_data/training/controller.dat. The current schema records raw_update_norm_sq, controlled_update_norm_sq, actual_update_norm_sq, dgrad_norm_sq, dtheta_for_control_sq, stability_lhs_raw, stability_lhs_ctrl, stability_rhs, alpha_feasible, alpha_state, alpha_code, alpha_min, and controller_feasible. TestbenchData loads controller.dat as top-level scalar traces.

hls4mlcontrollertracecurvature
Due
UNDEFINED
Start
2026-05-23
End
2026-05-23
ENB-024Done

Add TestbenchData loader and plots for hls4ml trainable traces

Added TestbenchData.from_dir/from_trainable_dir for recursive tb_data/training/*.dat traces, per-trace metadata parsing, a table-style representation, sparse-cadence trace merging, namespaced nested trace columns, and TestbenchData.plot_training() for metadata, smoothed traces, global-step x-axis, and epoch top axis. FitHistory remains scoped to software training histories.

TestbenchDatahls4mltraining-tracesplots
Due
UNDEFINED
Start
2026-05-23
End
2026-05-23
HLS4ML-020Done

HLS4ML - Validate minimal CTRL-NONE trainable compile path

Done: the one-Dense half_mse -> Dense backpass -> SGD -> CTRL-NONE -> apply-update path runs in CSIM on kona and produces training loss curves. Trainable projects emit config-sized local buffers, static learning rate in trainable_configN, an epoch/shuffle CSIM testbench, compact console progress, and tb_data/training traces.

hls4mlCTRL-NONEcsimminimal
Due
UNDEFINED
Start
2026-05-22
End
2026-05-23
ENB-025Done

Add per-layer parameter trace loading and statistics

Raw weights and biases now load into per-layer TestbenchLayerData objects instead of tb.frame. TestbenchData.from_dir(..., load_weights=True/False/[layers]) controls loading, and tb.stats_frame/tb.scalar_frame expose per-layer parameter summaries for plotting.

TestbenchDataparametersweightsstatistics
Due
UNDEFINED
Start
2026-05-23
End
2026-05-23
ENB-027Done

Standardize TestbenchData training panels for controller diagnostics

TestbenchData.plot_training now promotes expected training panels before generic scalar traces: loss, alpha, and a twin-axis controller-norm panel for ||ΔG|| and ||Δθ_raw||. The norms are derived from controller.dat squared accumulators, and callers can override y-axis scales with scales={metric: scale}.

TestbenchDatacontrollerplotsdiagnostics
Due
UNDEFINED
Start
2026-05-23
End
2026-05-23
ENB-022Done

Add scoped ENABOL toolchain profiles for hls4ml compilation

Added enabol.toolchain with profile loading, host/platform/backend validation, scoped environment application/restoration, required command checks, and tracked example profiles for the kona-ubuntu Vivado/Vitis installs. The default profile is kona-vitis-2024_1 because 2023.2 CSIM is not reliable on Ubuntu 24.04 without a container.

hls4mltoolchainvivadovitis
Due
UNDEFINED
Start
2026-05-22
End
2026-05-22
HLS4ML-012Done

HLS4ML - Build the ENABOL to hls4ml trainable config bridge

Done for the config bridge: enabol.compile.compile builds hls4ml config, emits Model.Training, delegates PrecisionDict to hls4ml precision mapping to enabol.precision, applies toolchain profiles for build=True, exports dataset testbench files, and converts/writes/builds through hls4ml. Reference trace comparison continues under HLS4ML-013 once trainable kernels exist.

hls4mlenabol-bridgePrecisionDictconfig
Due
UNDEFINED
Start
2026-05-22
End
2026-05-22
ENB-026Done

Refactor curvature sensor from controller kernel for multi-layer global reduction

Split global_throttle_order0 into raw_update_sensor_order0 (per-layer raw update and ΔG squared-norm contribution) and global_throttle_order0_law (global law from accumulated norms). Writer emission: backpass, SGD raw proposal, raw-update/dgrad reduction, controller law, then scaled update per layer. GT-1 built on the same architecture.

curvature-sensorglobal-reductionmulti-layerrefactoring
Due
UNDEFINED
Start
2026-05-23
End
2026-05-23

Static Views​

Unassigned Backlog​

HLS4ML-035Unassigned

HLS4ML - Add Adam optimizer support

Future Adam support with explicit state buffers and bias-correction policy. The deprecated code mixes double/std::pow paths and large static arrays, so this must be rewritten carefully before synthesis.

hls4mloptimizeradam
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned
HLS4ML-013Unassigned

HLS4ML - Validate one-layer Dense trainable CSIM

Run the minimal one-dense affine regression case through generated trainable firmware in CSIM and compare forward values, loss, gradients, alpha, and updated weights against ENABOL software traces.

hls4mlcsimdenseLIN1
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned
ENB-017Unassigned

Run CSIM validation with global throttle (CTRL-GT-ORDER-0, division-free fix)

GT-0 CSIM failed: two ap_fixed divisions (sqrt + divide) crashed CSIM at step 1. Fixed by replacing α = χ/(ηC+ε) with division-free inequality search α²·η²·||ΔG||² ≤ χ²·(||Δθ_raw||² + ε²). The raw update is the optimizer proposal before alpha/quantized assignment; for SGD it already includes η. Candidate table of binary-fraction α values is searched in descending order with a nonzero alpha floor. Validate: compile 1-Dense model with Kind: CTRL-GT-ORDER-0, run CSIM, confirm α selection, no CSIM errors, loss drops.

csimcontrollerfixed-pointGT0division-free
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
HLS4ML-014Unassigned

HLS4ML - Extend trainable CSIM to two Dense layers with activation

After the one-layer Dense path is correct, validate reverse traversal through Dense -> activation -> Dense before moving to convolution and BatchNorm paths.

hls4mlcsimdense2activation
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned
ENB-019Unassigned

Run FPGA hardware drift-recovery experiment

Collect the real hardware baseline, post-drift initial point, and post-training recovery point for controller-enabled learning.

FPGAdriftrecovery-curve
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned
ENB-018Unassigned

Run HLS synthesis and resource validation

Check resource impact, timing, and generated design viability after CSIM validates behavior.

HLSresourceslatency
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned

In Progress​

HLS4ML-008In progress

HLS4ML - Implement synthesizable controller kernels

CTRL-NONE, CTRL-GT-ORDER-0, and CTRL-GT-ORDER-1 are implemented and wired. GT-0/GT-1 now use raw optimizer update geometry plus division-free candidate search over α²·η²·||ΔG||² ≤ χ²·(||Δθ_raw||²+ε²), with a nonzero alpha candidate floor. CTRL-GT-ORDER-2/2-QA remain.

hls4mlcontrollersglobal-throttleCTRL
Due
UNDEFINED
Start
2026-05-22
End
UNDEFINED
HLS4ML-009In progress

HLS4ML - Add global metric reductions and alpha synchronization

Raw-update controller sensing is separated from controller law (raw_update_sensor_order0 → global_throttle_*_law). The writer emission now runs backpass → SGD proposal → raw-update/dgrad reduction → law → alpha-scaled apply. Per-layer raw update norms and ΔG norms are summed into global accumulators before the law call. CTRL-NONE skips curvature sensing.

hls4mlalpha-syncnorm-reductionglobal-throttle
Due
UNDEFINED
Start
UNDEFINED
End
UNDEFINED
Unassigned
HLS4ML-037In progress

HLS4ML - Add optional trainable CSIM trace logging

Current slice implemented: generated trainable CSIM writes metadata-bearing comma-delimited .dat traces under tb_data/training/ for loss, alpha, and per-layer epoch-level parameter traces such as dense0/weights.dat and dense0/biases.dat. Repeated epoch/sample/global_step/sample_index columns allow sparse traces with different cadences. Broader trace selection for gradients and controller internals remains open.

hls4mltracedebugcsim
Due
UNDEFINED
Start
2026-05-23
End
UNDEFINED

Blocked​

No tasks match this filter.

Completed Tasks​

IDTaskStatusPriorityStageTargetActionOwnerDueStartEndLinks
ENB-001Build the ENABOL documentation and project registry system
Created the Docusaurus structure, metadata components, people registry, task board, meeting calendar, and implementation/experiment hubs.
docspage
CompletedHighResearch / theoryDocumentation siteDocumentation@Manu B. Valentin2026-05-192026-05-12026-05-16Docs HomeCode Map
ENB-003Study global throttle
Separated static κ-style representational rails from dynamic trajectory control and documented why global throttling preserves update direction.
global-throttle
CompletedHighResearch / theoryTheoryAnalysis@Manu B. Valentin2026-05-202026-05-092026-05-18Global ThrottleHow This Fits
ENB-002Study κ-budgeting failure mode
Separated static κ-style representational rails from dynamic trajectory control and documented why global throttling preserves update direction.
κ-budgeting
CompletedHighResearch / theoryTheoryAnalysis@Manu B. Valentin2026-05-092026-05-012026-05-09Global ThrottleHow This Fits
ENB-028Document trainable precision schema relationships
Added the first precision relationship document. Unknown trainable precision dependencies are explicitly marked with ?, and the known rule for controller_metric records that dtheta_sq needs fractional bits for squared parameter movement: F_controller_metric >= 2*F_theta + guard bits.
precisionschemacontroller-metrichls4ml
DoneHighSystem designprecisionDocumentation@Manu B. ValentinUNDEFINED2026-05-232026-05-23Trainable Precision Schemaprecision.pyPhase 1 Trainable Config
HLS4ML-001HLS4ML - Define the trainable HLSConfig schema
Done in hls4ml-trainable: HLSConfig now has normalized Model.Training defaults and accessors for trainable mode, loss, optimizer, controller, trainable precision, recognized precision fields, and per-layer Training overrides.
hls4mlHLSConfigtrainable-configschema
DoneHighSystem designhls4mlDesign@Manu B. ValentinUNDEFINED2026-05-212026-05-22HLSConfighls4ml IntroPhase 1 Config
ENB-004Define the controller registry and stability stress-test catalog
Established CTRL-* pages and ST-* stress-test pages as stable IDs for experiments and task tracking.
controllersST-catalogmetrics
CompletedHighSystem designController stackDesign@Manu B. Valentin2026-05-202026-05-092026-05-18ControllersAblations
HLS4ML-039HLS4ML - Use raw-update geometry for global throttle control
Replaced actual/throttled parameter-delta controller geometry with raw optimizer update geometry. SGD now runs before the law, raw_update_sensor_order0 sums ||Δθ_raw||² and ||ΔG||² globally, GT-0/GT-1 keep the division-free squared inequality, alpha uses a nonzero candidate floor, and controller.dat logs raw/controlled/actual update norms plus feasibility telemetry.
hls4mlcontrollerraw-updateglobal-throttle
DoneHighImplementationController stackDevelopment@Manu B. ValentinUNDEFINED2026-05-232026-05-23GT-0GT-1Vivado Writertestbench.py
HLS4ML-015HLS4ML - Port minimal trainable static headers
Done: minimal static support files now live under templates/vivado/trainable/{common,losses,backprop,optimizers,controllers}. This includes trainable_trace.h, mse.h, nnet_dense_backprop.h, sgd.h, and global_throttle.h.
hls4mltemplatestrainableheaders
DoneHighImplementationC++ backendDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Template PassesLoss Endpoints
HLS4ML-016HLS4ML - Implement half-MSE loss template and dLdy seed
Done: mse.h implements mse and half_mse using one shift-based core, with half_mse seeding dL/dy = y_hat - y and mse seeding dL/dy = 2 * (y_hat - y).
hls4mllosshalf_msegrad-seed
DoneHighImplementationC++ backendDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Loss Endpoints
HLS4ML-017HLS4ML - Emit trainable top-level IO for loss and training controls
Done for the minimal path: generated top-level firmware receives ground truth, scalar loss output, alpha output, train_enable, reset_accumulators, and batch_end. The testbench trains from tb_output_predictions.dat, while the Python bridge supplies zero trainable tensors with train_enable=false for inference-safe calls.
hls4mlwritertop-level-ioloss
DoneHighImplementationhls4mlDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Vivado WriterLoss Endpoints
HLS4ML-018HLS4ML - Implement minimal Dense backpass without controller
Done: the first Dense backward kernel computes dL/dx, dL/dW, and dL/db accumulation and the writer emits the corresponding call for the resolved backward traversal. Global throttle, kappa, Adam, and delayed-alpha scheduling remain future work.
hls4mldensebackpassminimal
DoneHighImplementationC++ backendDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Template Passes
HLS4ML-019HLS4ML - Implement raw SGD update proposal for Dense
Done: sgd.h computes raw weight/bias update proposals from averaged gradients and learning rate, and writer emission calls SGD before CTRL-NONE alpha application.
hls4mlsgdraw-updatedense
DoneHighImplementationC++ backendDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Optimizer + Controller
HLS4ML-021HLS4ML - Implement CTRL-GT-ORDER-0 controller
Implemented CTRL-GT-ORDER-0 as a global law fed by raw_update_sensor_order0. The writer sums per-layer ||Δθ_raw||² and ||ΔG||² contributions into global accumulators, then GT-0 selects the largest candidate alpha satisfying α²·η²·||ΔG||² ≤ χ²·(||Δθ_raw||²+ε²). This avoids sqrt/division and avoids the self-locking actual-update denominator.
hls4mlcontrollerglobal-throttleGT0
DoneHighImplementationController stackDevelopment@Manu B. ValentinUNDEFINED2026-05-232026-05-23ControllersGlobal Throttle
HLS4ML-002HLS4ML - Add trainable ModelGraph flow and validation passes
Done in hls4ml-trainable: registered vivado:trainable after template resolution and before writer emission, added vivado:validate_trainable_config for the first supported Dense trainable subset, and added vivado:resolve_trainable_backward_order to attach trainable_forward_path, trainable_forward_order, trainable_backward_order, and trainable_output_layer while rejecting branching graphs. Loss endpoint metadata remains tracked under HLS4ML-006.
hls4mlModelGraphflowsvalidation
DoneHighImplementationhls4mlDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22ModelGraph + Flows
HLS4ML-003HLS4ML - Add trainable layer attributes and precision resolution
Done in hls4ml-trainable: Layer initialization now writes a trainable flag and converts configured trainable precision fields into NamedType attributes such as grad_in_t, grad_out_t, raw_update_t, update_t, gradient_accum_t, controller_metric_t, and alpha_t.
hls4mllayer-attributesprecisiongradients
DoneHighImplementationhls4mlDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Layer AttributesPrecisionPhase 1 Config
HLS4ML-004HLS4ML - Port static trainable headers into Vivado templates
Done in hls4ml-trainable: templates/vivado/trainable is the new umbrella for common trace helpers, Dense backprop, MSE/half-MSE losses, SGD raw updates, and CTRL-NONE/global-throttle helpers. Generated firmware now copies this tree into firmware/trainable.
hls4mltrainablelossesvivado-templates
DoneHighImplementationC++ backendDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Template Passes
HLS4ML-005HLS4ML - Implement Dense backpass templates first
Done for the minimal path: nnet_dense_backprop.h computes dL/dx, accumulates dL/dW and dL/db across power-of-two batches, and emits averaged Dense gradients. Writer emission now allocates the required Dense backprop buffers.
hls4mldensebackpasstemplates
DoneHighImplementationC++ backendDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Template Passes
HLS4ML-006HLS4ML - Generate loss endpoints and output-gradient seeds
Done for the first path: vivado:resolve_trainable_loss_endpoints attaches normalized one-output half_mse metadata, the trainable MSE header emits loss and dL/dy with shift-based scaling, and the Vivado writer emits loss config, dL/dy buffers, and loss kernel calls.
hls4mlloss-endpointsfrom-logitsgrad-seed
DoneHighImplementationhls4mlDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Loss Endpoints
HLS4ML-007HLS4ML - Implement two-phase raw update storage for Dense
Done for Dense/CTRL-NONE: backprop emits averaged gradients, SGD emits raw update vectors, and apply_dense_update applies alpha-scaled updates in a separate phase. Global metric reductions remain under HLS4ML-009.
hls4mltwo-phase-updatedenseraw-update
DoneHighImplementationC++ backendDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Optimizer + Controller
HLS4ML-011HLS4ML - Add minimal Vivado writer hooks for trainable emission
Done in hls4ml-trainable: the Vivado writer copies trainable headers, emits trainable includes/config structs, allocates trainable buffers, emits the minimal loss/backprop/SGD/CTRL-NONE/apply-update call chain, and keeps bridge inference calls safe with train_enable=false stubs.
hls4mlvivado-writercodegenIO
DoneHighImplementationhls4mlDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Vivado Writer
ENB-006Implement fake-fixed-point precision and diagnostic summaries
Added layer-indexed precision configuration, rail statistics, quantized update paths, and compact precision summaries.
PrecisionDictap_fixedrailsquantization
CompletedHighImplementationenabol moduleDevelopment@Manu B. Valentin2026-05-202026-05-152026-05-20PrecisionQuantization
ENB-005Refactor enabol neural training into controller-based modules
Split the old training loop into model, controller, optimizer, telemetry, applier, and trainer modules with Controller.from_str and MetricsConfig.
controller.pytraining.pytelemetryMetricsConfig
CompletedHighImplementationenabol moduleDevelopment@Manu B. Valentin2026-05-202026-05-092026-05-20NN ModuleTraining Loop
HLS4ML-038HLS4ML - Emit controller metric trace from trainable CSIM
Generated trainable top-level functions now expose controller metric outputs and the trainable CSIM testbench writes tb_data/training/controller.dat. The current schema records raw_update_norm_sq, controlled_update_norm_sq, actual_update_norm_sq, dgrad_norm_sq, dtheta_for_control_sq, stability_lhs_raw, stability_lhs_ctrl, stability_rhs, alpha_feasible, alpha_state, alpha_code, alpha_min, and controller_feasible. TestbenchData loads controller.dat as top-level scalar traces.
hls4mlcontrollertracecurvature
DoneHighValidation / benchmarkingCSIMDevelopment@Manu B. ValentinUNDEFINED2026-05-232026-05-23Phase 5 Bridge + Validationtestbench.py
ENB-024Add TestbenchData loader and plots for hls4ml trainable traces
Added TestbenchData.from_dir/from_trainable_dir for recursive tb_data/training/*.dat traces, per-trace metadata parsing, a table-style representation, sparse-cadence trace merging, namespaced nested trace columns, and TestbenchData.plot_training() for metadata, smoothed traces, global-step x-axis, and epoch top axis. FitHistory remains scoped to software training histories.
TestbenchDatahls4mltraining-tracesplots
DoneHighValidation / benchmarkingCSIMDevelopment@Manu B. ValentinUNDEFINED2026-05-232026-05-23history.pytestbench.pyPhase 5 Bridge + Validation
HLS4ML-020HLS4ML - Validate minimal CTRL-NONE trainable compile path
Done: the one-Dense half_mse -> Dense backpass -> SGD -> CTRL-NONE -> apply-update path runs in CSIM on kona and produces training loss curves. Trainable projects emit config-sized local buffers, static learning rate in trainable_configN, an epoch/shuffle CSIM testbench, compact console progress, and tb_data/training traces.
hls4mlCTRL-NONEcsimminimal
DoneHighValidation / benchmarkingCSIMTesting@Manu B. ValentinUNDEFINED2026-05-222026-05-23Phase 5 Bridge + Validation
ENB-011Add multi-layer software benchmarks
Move beyond one-layer LIN1 into coupled dense models and nonlinear hidden representations before committing to hardware paths.
dense2relucouplingbenchmarks
DoneHighValidation / benchmarkingSoftware experimentsExperiment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Experiments
ENB-009Run EXP-001 stability phase-diagram sweep
Generated the first software phase diagrams across learning-rate and drift parameters for controlled and uncontrolled training.
EXP-001phase-diagrameta-sweepdrift-sweep
CompletedHighValidation / benchmarkingSoftware experimentsExperiment@Manu B. Valentin2026-05-212026-05-202026-05-21EXP-001
ENB-007Run EXP-000A floating-point global throttle sanity test
Verified the first one-layer software sanity case for global throttling under affine drift.
EXP-000AfloatLIN1drift
CompletedHighValidation / benchmarkingSoftware experimentsExperiment@Manu B. Valentin2026-05-192026-05-152026-05-19EXP-000A
ENB-008Run EXP-000B fake-quantized global throttle sanity test
Validated staged fake-fixed-point paths and showed that throttling can recover in tight quantization regimes where no-controller training oscillates.
EXP-000Bqfxrailsupdate-cosine
CompletedHighValidation / benchmarkingSoftware experimentsExperiment@Manu B. Valentin2026-05-202026-05-152026-05-20EXP-000B
ENB-025Add per-layer parameter trace loading and statistics
Raw weights and biases now load into per-layer TestbenchLayerData objects instead of tb.frame. TestbenchData.from_dir(..., load_weights=True/False/[layers]) controls loading, and tb.stats_frame/tb.scalar_frame expose per-layer parameter summaries for plotting.
TestbenchDataparametersweightsstatistics
DoneMediumValidation / benchmarkingCSIMDevelopment@Manu B. ValentinUNDEFINED2026-05-232026-05-23testbench.pyPhase 5 Bridge + Validation
ENB-027Standardize TestbenchData training panels for controller diagnostics
TestbenchData.plot_training now promotes expected training panels before generic scalar traces: loss, alpha, and a twin-axis controller-norm panel for ||ΔG|| and ||Δθ_raw||. The norms are derived from controller.dat squared accumulators, and callers can override y-axis scales with scales={metric: scale}.
TestbenchDatacontrollerplotsdiagnostics
DoneMediumValidation / benchmarkingplotsDevelopment@Manu B. ValentinUNDEFINED2026-05-232026-05-23testbench.pyPhase 5 Bridge + Validation
ENB-012Compare global throttle against naive safety mechanisms
Use clipping, loose κ projection, and legacy row/column projection as comparisons after the fixed-point path is understood.
ST-006clippingkappa-projectionbaselines
DoneMediumValidation / benchmarkingSoftware experimentsAnalysis@Manu B. ValentinUNDEFINED2026-05-222026-05-22ST-006ST-007
ENB-022Add scoped ENABOL toolchain profiles for hls4ml compilation
Added enabol.toolchain with profile loading, host/platform/backend validation, scoped environment application/restoration, required command checks, and tracked example profiles for the kona-ubuntu Vivado/Vitis installs. The default profile is kona-vitis-2024_1 because 2023.2 CSIM is not reliable on Ubuntu 24.04 without a container.
hls4mltoolchainvivadovitis
DoneHighIntegrationcompileDevelopment@Manu B. ValentinUNDEFINED2026-05-222026-05-22Toolchain ProfilesPhase 5 Bridge + Validation
HLS4ML-012HLS4ML - Build the ENABOL to hls4ml trainable config bridge
Done for the config bridge: enabol.compile.compile builds hls4ml config, emits Model.Training, delegates PrecisionDict to hls4ml precision mapping to enabol.precision, applies toolchain profiles for build=True, exports dataset testbench files, and converts/writes/builds through hls4ml. Reference trace comparison continues under HLS4ML-013 once trainable kernels exist.
hls4mlenabol-bridgePrecisionDictconfig
DoneHighIntegrationenabol moduleIntegration@Manu B. ValentinUNDEFINED2026-05-222026-05-22ENABOL BridgePhase 5 Bridge + Validation
ENB-026Refactor curvature sensor from controller kernel for multi-layer global reduction
Split global_throttle_order0 into raw_update_sensor_order0 (per-layer raw update and ΔG squared-norm contribution) and global_throttle_order0_law (global law from accumulated norms). Writer emission: backpass, SGD raw proposal, raw-update/dgrad reduction, controller law, then scaled update per layer. GT-1 built on the same architecture.
curvature-sensorglobal-reductionmulti-layerrefactoring
DoneHighrefactoringC++ backendDevelopment@Manu B. ValentinUNDEFINED2026-05-232026-05-23ControllersGlobal Throttle

Useful Filters​

<TaskBoard status="todo" />
<TaskBoard owner="__unassigned__" />
<TaskBoard owner="mbvalentin" />
<TaskBoard stage="validation" />
<TaskBoard target="enabol" />
<TaskBoard action="development" />
<TaskBoard tag="global-throttle" />
<TaskBoard statuses={["todo", "inprogress"]} view="labels" />