Remove memory-aware node bindings #123
Conversation
📝 WalkthroughSummary by CodeRabbit
WalkthroughRemoved memory-aware node parsing/binding extensions and all Neureka "Wmem" binding/tiling variants; introduced requantization helper utilities and refactored Neureka tile-constraint classes to condition on weight SRAM vs non‑SRAM placement and to integrate requant load scheduling. Changes
Sequence Diagram(s)sequenceDiagram
participant Parser as Parser
participant Engine as Engine/Mapper
participant Tiler as Tiler
participant Binding as TilingReadyBinding
Note over Parser,Engine: Before: parser → (WmemBinding + Binding)
Parser->>Engine: map to (WmemBinding + Binding)
Engine->>Tiler: provide Wmem + non‑Wmem tiling bindings
Tiler->>Binding: use Wmem-aware constraints
Note over Parser,Engine: After: parser → Binding (direct)
Parser->>Engine: map to Binding
Engine->>Tiler: provide non‑Wmem tiling bindings
Tiler->>Binding: use unified / requant-aware constraints
sequenceDiagram
participant Constraint as TileConstraint
participant Ctxt as NetworkContext
participant Requant as RequantHelpers
participant Schedule as TilingSchedule
Constraint->>Ctxt: lookup weight buffer (VariableBuffer)
alt weight memory == SRAM
Ctxt-->>Constraint: SRAM
Constraint->>Schedule: compute per-cube weight_addr_offset (SRAM path)
else weight memory != SRAM
Ctxt-->>Constraint: non‑SRAM
Constraint->>Requant: requantLoadSchedule(absoluteOutputCubes, ctxt, op)
Requant-->>Schedule: mul/add load schedule per cube
Constraint->>Schedule: merge weight info into input/output loads
end
Schedule-->>Constraint: final tiling load schedule
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Possibly related PRs
Suggested labels
Suggested reviewers
WalkthroughRemoved memory-aware parsing/binding extensions and Wmem-prefixed binding/tiling variants across the deployer and Neureka target; introduced requantization helper utilities and refactored tile-constraint classes to handle SRAM vs non‑SRAM weight placement and to integrate requant load scheduling. Changes
Sequence Diagram(s)sequenceDiagram
participant Parser as Parser (e.g., NeurekaRQSPWParser)
participant Engine as Engine Mapper
participant Tiler as Tiler
participant Binding as TilingReadyBinding
Note over Parser,Binding: Before: parser mapped to concatenated Wmem + non‑Wmem bindings
Parser->>Engine: map to (WmemBinding + Binding)
Engine->>Tiler: produce Wmem tiling-ready bindings
Tiler->>Binding: use Wmem-aware tile constraints
Note over Parser,Binding: After: direct one-to-one mapping
Parser->>Engine: map to Binding
Engine->>Tiler: produce tiling-ready binding (non-Wmem)
Tiler->>Binding: use unified / requant-aware tile constraints
sequenceDiagram
participant Constraint as TileConstraint
participant Mem as MemoryStore
participant Requant as RequantHelpers
participant Schedule as TilingSchedule
Constraint->>Mem: lookup weight buffer memory level
alt Weight in SRAM
Mem-->>Constraint: SRAM
Constraint->>Schedule: compute per-cube weight_addr_offset
else Weight in non‑SRAM
Mem-->>Constraint: non-SRAM
Constraint->>Requant: call requantLoadSchedule
Requant-->>Schedule: return requantized load schedule (mul/add)
Constraint->>Schedule: merge weight into input/output loads
end
Schedule-->>Constraint: final tiling load schedule
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Possibly related PRs
Suggested labels
Suggested reviewers
Pre-merge checks and finishing touches❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
✨ Finishing touches
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
Deeploy/CommonExtensions/NetworkDeployers/NetworkDeployerWrapper.py (1)
5-64: RestoreNetworkContextimport to avoid runtime failure
NetworkContextis still used in the type hints (e.g., Line 51) but is no longer imported. Because these annotations are evaluated at module import time (nofrom __future__ import annotationssafeguard), the module will now raise aNameErroras soon as it loads. Please reintroduce the import (or switch to a quoted forward reference).
🧹 Nitpick comments (12)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py (2)
227-227: Addstrict=Truetozip()for safety.Python 3.10+ supports
strict=Trueto catch length mismatches between iterables. This prevents silent bugs ifoutputCubesandinputLoadSchedulediverge in length.Apply this diff:
- for cube, load in zip(outputCubes, inputLoadSchedule): + for cube, load in zip(outputCubes, inputLoadSchedule, strict=True):
257-261: Addstrict=Truetozip()for safety.Similar to line 227, this
zip()should includestrict=Trueto catch length mismatches.Apply this diff:
newInputLoadSchedule = [{ **load, **rqLoad - } for load, rqLoad in zip(tilingSchedule.inputLoadSchedule, requantSchedule)] + } for load, rqLoad in zip(tilingSchedule.inputLoadSchedule, requantSchedule, strict=True)]Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py (1)
64-68: Addstrict=Truetozip()for safety.For consistency and to catch potential length mismatches, add
strict=True.Apply this diff:
newInputLoadSchedule = [{ **schedule, "mul": mul, "add": add, - } for schedule, mul, add in zip(tilingSchedule.inputLoadSchedule, inputMulCubes, inputAddCubes)] + } for schedule, mul, add in zip(tilingSchedule.inputLoadSchedule, inputMulCubes, inputAddCubes, strict=True)]Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py (4)
161-162: Consider using_prefix for unused unpacked variables.Static analysis flags
BatchOffset,HOffset,WOffset, andBatchSizeas unused. While the current code works, prefixing with_makes the intent explicit.Apply this diff:
- (BatchOffset, HOffset, WOffset, COffset) = cube.offset - (BatchSize, HSize, WSize, CSize) = cube.dims + (_BatchOffset, _HOffset, _WOffset, COffset) = cube.offset + (_BatchSize, HSize, WSize, CSize) = cube.dims
212-232: Code duplication with NeurekaDepthwiseConstraint.This weight handling logic is nearly identical to lines 210-230 in NeurekaDepthwiseConstraint.py. The only difference is the weight cube construction (different dimension ordering).
Consider extracting this pattern into a helper function in RequantHelpers.py to reduce duplication across Dense, Depthwise, and Pointwise constraint classes.
229-229: Addstrict=Truetozip()for safety.For consistency with best practices in Python 3.10+, add
strict=True.Apply this diff:
- for cube, load in zip(outputCubes, inputLoadSchedule): + for cube, load in zip(outputCubes, inputLoadSchedule, strict=True):
259-263: Addstrict=Truetozip()for safety.Apply this diff:
newInputLoadSchedule = [{ **load, **rqLoad - } for load, rqLoad in zip(tilingSchedule.inputLoadSchedule, requantSchedule)] + } for load, rqLoad in zip(tilingSchedule.inputLoadSchedule, requantSchedule, strict=True)]Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py (5)
48-52: Document why SRAM weights are constrained to Max().Unlike the Dense and Depthwise variants, the Pointwise constraint explicitly constrains
weightOutChannelVar == weightOutChannelVar.Max()for SRAM weights. This difference suggests Pointwise has specific requirements for SRAM weight tiling.Consider adding a comment explaining why Pointwise forces full channel tiling for SRAM weights while Dense/Depthwise don't.
191-192: Consider using_prefix for unused unpacked variables.Static analysis flags
BatchOffset,HOffset,WOffset, andBatchSizeas unused.Apply this diff:
- (BatchOffset, HOffset, WOffset, COffset) = cube.offset - (BatchSize, HSize, WSize, CSize) = cube.dims + (_BatchOffset, _HOffset, _WOffset, COffset) = cube.offset + (_BatchSize, HSize, WSize, CSize) = cube.dims
242-262: Code duplication with Dense and Depthwise constraints.This weight handling logic is duplicated across NeurekaDenseConstraint.py (lines 212-232), NeurekaDepthwiseConstraint.py (lines 210-230), and here. Consider consolidating this pattern into a shared helper function.
259-259: Addstrict=Truetozip()for safety.Apply this diff:
- for cube, load in zip(outputCubes, inputLoadSchedule): + for cube, load in zip(outputCubes, inputLoadSchedule, strict=True):
289-293: Addstrict=Truetozip()for safety.Apply this diff:
newInputLoadSchedule = [{ **load, **rqLoad - } for load, rqLoad in zip(tilingSchedule.inputLoadSchedule, requantSchedule)] + } for load, rqLoad in zip(tilingSchedule.inputLoadSchedule, requantSchedule, strict=True)]
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (10)
Deeploy/CommonExtensions/NetworkDeployers/NetworkDeployerWrapper.py(1 hunks)Deeploy/MemoryLevelExtension/MemoryLevels.py(1 hunks)Deeploy/MemoryLevelExtension/NetworkDeployers/MemoryLevelDeployer.py(1 hunks)Deeploy/Targets/Neureka/Bindings.py(0 hunks)Deeploy/Targets/Neureka/Engine.py(1 hunks)Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py(7 hunks)Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py(6 hunks)Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py(5 hunks)Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py(1 hunks)Deeploy/Targets/Neureka/Tiler.py(1 hunks)
💤 Files with no reviewable changes (1)
- Deeploy/Targets/Neureka/Bindings.py
🧰 Additional context used
🧬 Code graph analysis (8)
Deeploy/Targets/Neureka/Engine.py (2)
Deeploy/DeeployTypes.py (1)
NodeMapper(1716-1872)Deeploy/Targets/Neureka/Parsers.py (6)
NeurekaRQSPWConv2DParser(140-162)NeurekaPWConv2DParser(125-137)NeurekaRQSDWConv2DParser(99-122)NeurekaDWConv2DParser(80-96)NeurekaRQSDenseConv2DParser(180-202)NeurekaDenseConv2DParser(165-177)
Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py (6)
Deeploy/DeeployTypes.py (2)
NetworkContext(564-1076)add(737-774)Deeploy/TilingExtension/MemoryConstraints.py (1)
NodeMemoryConstraint(95-167)Deeploy/TilingExtension/TilerModel.py (3)
TilerModel(34-402)addTensorDimToModel(143-157)getTensorDimVar(131-135)Deeploy/TilingExtension/TilingCodegen.py (4)
AbsoluteHyperRectangle(39-49)HyperRectangle(24-35)TilingSchedule(53-122)VariableReplacementScheme(126-158)Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py (2)
serializeTilingSolution(92-236)serializeTilingSolution(247-268)Deeploy/TilingExtension/TileConstraint.py (1)
extractBaseAddr(56-74)
Deeploy/CommonExtensions/NetworkDeployers/NetworkDeployerWrapper.py (1)
Deeploy/DeeployTypes.py (3)
CodeGenVerbosity(49-55)NetworkDeployer(3231-3596)ONNXLayer(1875-2203)
Deeploy/MemoryLevelExtension/NetworkDeployers/MemoryLevelDeployer.py (1)
Deeploy/DeeployTypes.py (3)
NetworkContext(564-1076)NetworkOptimizationPass(2263-2283)NetworkOptimizer(2286-2309)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py (5)
Deeploy/DeeployTypes.py (3)
NetworkContext(564-1076)VariableBuffer(232-391)lookup(776-808)Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py (2)
requantAddGeometricalConstraint(14-34)requantLoadSchedule(76-94)Deeploy/TilingExtension/TilerModel.py (2)
getTensorDimVar(131-135)TilerModel(34-402)Deeploy/TilingExtension/TilingCodegen.py (4)
HyperRectangle(24-35)calculateFlatOffsetInBytes(239-242)TilingSchedule(53-122)VariableReplacementScheme(126-158)Deeploy/TilingExtension/TileConstraint.py (1)
extractBaseAddr(56-74)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py (5)
Deeploy/DeeployTypes.py (3)
NetworkContext(564-1076)VariableBuffer(232-391)lookup(776-808)Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py (2)
requantAddGeometricalConstraint(14-34)requantLoadSchedule(76-94)Deeploy/TilingExtension/TilerModel.py (2)
getTensorDimVar(131-135)TilerModel(34-402)Deeploy/TilingExtension/TilingCodegen.py (4)
HyperRectangle(24-35)calculateFlatOffsetInBytes(239-242)TilingSchedule(53-122)VariableReplacementScheme(126-158)Deeploy/TilingExtension/TileConstraint.py (1)
extractBaseAddr(56-74)
Deeploy/Targets/Neureka/Tiler.py (3)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py (2)
NeurekaDenseConv2DTileConstraint(21-236)NeurekaRQSDenseConv2DTileConstraint(239-268)Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py (2)
NeurekaDWConv2DTileConstraint(21-234)NeurekaRQSDWConv2DTileConstraint(237-266)Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py (2)
NeurekaPWConv2DTileConstraint(21-266)NeurekaRQSPWConv2DTileConstraint(269-298)
Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py (5)
Deeploy/DeeployTypes.py (3)
NetworkContext(564-1076)VariableBuffer(232-391)lookup(776-808)Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py (2)
requantAddGeometricalConstraint(14-34)requantLoadSchedule(76-94)Deeploy/Targets/PULPOpen/TileConstraints/ConvTileConstraint.py (1)
computeInputCube(320-347)Deeploy/TilingExtension/TilingCodegen.py (4)
HyperRectangle(24-35)calculateFlatOffsetInBytes(239-242)TilingSchedule(53-122)VariableReplacementScheme(126-158)Deeploy/TilingExtension/TileConstraint.py (1)
extractBaseAddr(56-74)
🪛 Ruff (0.14.0)
Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py
48-48: Undefined name cls
(F821)
68-68: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py
227-227: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
261-261: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py
161-161: Unpacked variable BatchOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
161-161: Unpacked variable HOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
161-161: Unpacked variable WOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
162-162: Unpacked variable BatchSize is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
229-229: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
263-263: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py
191-191: Unpacked variable BatchOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
191-191: Unpacked variable HOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
191-191: Unpacked variable WOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
192-192: Unpacked variable BatchSize is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
259-259: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
293-293: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
🔇 Additional comments (7)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py (2)
210-230: LGTM! Clean SRAM-aware weight handling.The conditional weight handling properly separates the SRAM path (computing offsets per tile) from the non-SRAM path (merging base offsets and adding weight to load schedule). This aligns with the PR's goal to simplify by removing memory-aware bindings while still supporting SRAM weights.
237-266: LGTM! Requantization subclass correctly delegates.The
NeurekaRQSDWConv2DTileConstraintclass properly delegates to the base class and then applies requantization-specific constraints viarequantAddGeometricalConstraintandrequantLoadSchedule. This composition pattern is clean and maintainable.Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py (2)
14-34: LGTM! Clean constraint addition for requantization.This helper correctly adds dimension variables for mul/add buffers and constrains their channel dimensions to match the output channel. The logic is straightforward and well-commented.
76-94: LGTM! Clean load schedule generation.This helper correctly constructs HyperRectangles for mul/add buffers based on output channel offsets and sizes. The logic matches the pattern used in the tile constraint classes.
Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py (2)
54-58: LGTM! Clean conditional weight constraint.Unlike the depthwise variant, this implementation doesn't have a TODO and cleanly handles the SRAM case with a
passstatement, making the intent clear.
239-268: LGTM! Requantization integration follows established pattern.The
NeurekaRQSDenseConv2DTileConstraintclass correctly implements the same pattern as the depthwise variant, delegating to the base class and then applying requantization constraints.Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py (1)
269-298: LGTM! Consistent requantization pattern.The
NeurekaRQSPWConv2DTileConstraintclass follows the same clean pattern as Dense and Depthwise variants.
lukamac
force-pushed
the
remove-memory-aware-binding
branch
from
6f72b86 to
f0d4de7
Compare
…ve then unused memory-aware bindings
…de bindings anymore
lukamac
force-pushed
the
remove-memory-aware-binding
branch
from
f0d4de7 to
28faade
Compare
lukamac
changed the title
There was a problem hiding this comment.
Actionable comments posted: 0
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py (1)
86-88: Bug: kernel dims swapped (height vs width).Height should compare to dim_kernel_y, width to dim_kernel_x.
- tilerModel.addConstraint(inputHeightVar >= parseDict['dim_kernel_x']) - tilerModel.addConstraint(inputWidthVar >= parseDict['dim_kernel_y']) + tilerModel.addConstraint(inputHeightVar >= parseDict['dim_kernel_y']) + tilerModel.addConstraint(inputWidthVar >= parseDict['dim_kernel_x'])
♻️ Duplicate comments (1)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py (1)
52-58: Clarify the TODO comment or remove it.The TODO at line 54 suggests uncertainty about whether to uncomment the constraint
weightOutChannelVar == weightOutChannelVar.Max()for SRAM weights. This conditional logic is critical for handling weights in dedicated weight memory.If skipping the constraint is intentional for SRAM weights, document why. Otherwise, either uncomment it or remove the TODO.
Could you clarify the intended behavior for SRAM weights? Should the constraint be:
- Skipped entirely (current behavior)
- Uncommented to enforce full channel loading
- Something else?
🧹 Nitpick comments (7)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py (2)
210-230: Consider addingstrict=Trueto zip() call.The conditional weight handling logic correctly separates SRAM and non-SRAM paths:
- SRAM weights: pre-calculate offsets for direct access
- Non-SRAM weights: include in load schedule like other inputs
However, static analysis flags the
zip()call at line 227 for missingstrict=parameter. While the lists should have matching lengths by construction, addingstrict=Truemakes this assumption explicit and will fail fast if violated.Consider applying this change:
- for cube, load in zip(outputCubes, inputLoadSchedule): + for cube, load in zip(outputCubes, inputLoadSchedule, strict=True):
237-266: LGTM: Requantized DW conv constraint correctly extends base class.The
NeurekaRQSDWConv2DTileConstraintclass properly layers requantization support onto the base DW convolution constraints by:
- Calling base
addGeometricalConstraintthen applyingrequantAddGeometricalConstraint- Calling base
serializeTilingSolutionthen merging in requant load schedulesThis design allows requantization to be added without duplicating the base convolution tiling logic.
Consider adding
strict=Trueto thezip()call at line 261 for the same reason as mentioned in the previous comment:- } for load, rqLoad in zip(tilingSchedule.inputLoadSchedule, requantSchedule)] + } for load, rqLoad in zip(tilingSchedule.inputLoadSchedule, requantSchedule, strict=True)]Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py (2)
190-195: Silence unused tuple-unpack vars.Prefix unused items with underscores to satisfy linters and reduce noise.
- (BatchOffset, HOffset, WOffset, COffset) = cube.offset - (BatchSize, HSize, WSize, CSize) = cube.dims + (_BatchOffset, _HOffset, _WOffset, _COffset) = cube.offset + (_BatchSize, HSize, WSize, CSize) = cube.dims
259-262: Use zip(..., strict=True) to guard against length mismatches
Since the project requires Python >=3.10, update your loops as follows:- for cube, load in zip(outputCubes, inputLoadSchedule): + for cube, load in zip(outputCubes, inputLoadSchedule, strict=True):- } for load, rqLoad in zip(tilingSchedule.inputLoadSchedule, requantSchedule)] + } for load, rqLoad in zip(tilingSchedule.inputLoadSchedule, requantSchedule, strict=True)]Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py (3)
65-66: Avoid hard-coded 3x3; use parsed kernel dims (optional).Use dim_kernel_y/x to generalize and reduce magic numbers.
- tilerModel.addConstraint((outputHeightVar == (effectiveHeight - (3 - 1) - 1) // strides[0] + 1)) - tilerModel.addConstraint((outputWidthVar == (effectiveWidth - (3 - 1) - 1) // strides[1] + 1)) + kH, kW = parseDict["dim_kernel_y"], parseDict["dim_kernel_x"] + tilerModel.addConstraint((outputHeightVar == (effectiveHeight - (kH - 1) - 1) // strides[0] + 1)) + tilerModel.addConstraint((outputWidthVar == (effectiveWidth - (kW - 1) - 1) // strides[1] + 1))
161-163: Silence unused tuple-unpack vars.Prefix unused items to satisfy linters.
- (BatchOffset, HOffset, WOffset, COffset) = cube.offset - (BatchSize, HSize, WSize, CSize) = cube.dims + (_BatchOffset, _HOffset, _WOffset, _COffset) = cube.offset + (_BatchSize, HSize, WSize, CSize) = cube.dims
229-231: Addstrict=Trueto zip() calls (optional defensive programming).Since the project requires Python >= 3.10 (per
pyproject.toml), usestrict=Trueunconditionally rather than conditional logic. This ensures the iterables always have matching lengths:
- Line 229:
for cube, load in zip(outputCubes, inputLoadSchedule, strict=True):- Line 263:
for load, rqLoad in zip(tilingSchedule.inputLoadSchedule, requantSchedule, strict=True)]
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (11)
CHANGELOG.md(4 hunks)Deeploy/CommonExtensions/NetworkDeployers/NetworkDeployerWrapper.py(1 hunks)Deeploy/MemoryLevelExtension/MemoryLevels.py(1 hunks)Deeploy/MemoryLevelExtension/NetworkDeployers/MemoryLevelDeployer.py(1 hunks)Deeploy/Targets/Neureka/Bindings.py(0 hunks)Deeploy/Targets/Neureka/Engine.py(1 hunks)Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py(7 hunks)Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py(6 hunks)Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py(5 hunks)Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py(1 hunks)Deeploy/Targets/Neureka/Tiler.py(1 hunks)
💤 Files with no reviewable changes (1)
- Deeploy/Targets/Neureka/Bindings.py
🚧 Files skipped from review as they are similar to previous changes (2)
- Deeploy/MemoryLevelExtension/MemoryLevels.py
- Deeploy/Targets/Neureka/Engine.py
🧰 Additional context used
🧬 Code graph analysis (6)
Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py (3)
Deeploy/DeeployTypes.py (1)
NetworkContext(564-1076)Deeploy/TilingExtension/TilerModel.py (3)
TilerModel(34-402)addTensorDimToModel(143-157)getTensorDimVar(131-135)Deeploy/TilingExtension/TilingCodegen.py (2)
AbsoluteHyperRectangle(39-49)HyperRectangle(24-35)
Deeploy/MemoryLevelExtension/NetworkDeployers/MemoryLevelDeployer.py (1)
Deeploy/DeeployTypes.py (4)
NetworkContext(564-1076)NetworkOptimizationPass(2263-2283)NetworkOptimizer(2286-2309)StructBuffer(489-509)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py (4)
Deeploy/DeeployTypes.py (3)
NetworkContext(564-1076)VariableBuffer(232-391)lookup(776-808)Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py (2)
requantAddGeometricalConstraint(12-32)requantLoadSchedule(35-53)Deeploy/TilingExtension/TilerModel.py (2)
getTensorDimVar(131-135)TilerModel(34-402)Deeploy/TilingExtension/TilingCodegen.py (2)
HyperRectangle(24-35)calculateFlatOffsetInBytes(239-242)
Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py (3)
Deeploy/DeeployTypes.py (3)
NetworkContext(564-1076)VariableBuffer(232-391)lookup(776-808)Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py (2)
requantAddGeometricalConstraint(12-32)requantLoadSchedule(35-53)Deeploy/TilingExtension/TilingCodegen.py (2)
HyperRectangle(24-35)calculateFlatOffsetInBytes(239-242)
Deeploy/Targets/Neureka/Tiler.py (3)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py (1)
NeurekaDenseConv2DTileConstraint(21-236)Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py (1)
NeurekaDWConv2DTileConstraint(21-234)Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py (1)
NeurekaPWConv2DTileConstraint(21-266)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py (6)
Deeploy/DeeployTypes.py (3)
NetworkContext(564-1076)VariableBuffer(232-391)lookup(776-808)Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py (2)
requantAddGeometricalConstraint(12-32)requantLoadSchedule(35-53)Deeploy/TilingExtension/TilerModel.py (2)
getTensorDimVar(131-135)TilerModel(34-402)Deeploy/Targets/PULPOpen/TileConstraints/ConvTileConstraint.py (1)
computeInputCube(320-347)Deeploy/TilingExtension/TilingCodegen.py (4)
HyperRectangle(24-35)calculateFlatOffsetInBytes(239-242)TilingSchedule(53-122)VariableReplacementScheme(126-158)Deeploy/TilingExtension/TileConstraint.py (1)
extractBaseAddr(56-74)
🪛 Ruff (0.14.0)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py
227-227: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
261-261: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py
191-191: Unpacked variable BatchOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
191-191: Unpacked variable HOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
191-191: Unpacked variable WOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
192-192: Unpacked variable BatchSize is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
259-259: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
293-293: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
Deeploy/Targets/Neureka/TileConstraints/NeurekaDenseConstraint.py
161-161: Unpacked variable BatchOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
161-161: Unpacked variable HOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
161-161: Unpacked variable WOffset is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
162-162: Unpacked variable BatchSize is never used
Prefix it with an underscore or any other dummy variable pattern
(RUF059)
229-229: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
263-263: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (150)
- GitHub Check: mempool-models / test-runner-mempool
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (simpleRegression, 45000, 30000, 16000, 8, L3) / test-runner-siracusa-tiled (30000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (6000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (5000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (12000)
- GitHub Check: siracusa-kernels-tiled-doublebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-kernels / test-runner-siracusa
- GitHub Check: siracusa-kernels-tiled-singlebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (Transformer, 15000, 8, L3) / test-runner-siracusa-neureka-tiled (15000)
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (microLlama/microLlama1, 10000, 8, L3) / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: generic-kernels / test-runner-generic
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (miniMobileNet, 2000, 8, L3) / test-runner-siracusa-neureka-tiled (2000)
- GitHub Check: siracusa-neureka-models-tiled-doublebuffer-L3-wmem (microLlama/microLlama1, 10000, 8, true, L3, t... / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: mempool-models / test-runner-mempool
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (simpleRegression, 45000, 30000, 16000, 8, L3) / test-runner-siracusa-tiled (30000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (6000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (5000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (12000)
- GitHub Check: siracusa-kernels-tiled-doublebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-kernels / test-runner-siracusa
- GitHub Check: siracusa-kernels-tiled-singlebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (Transformer, 15000, 8, L3) / test-runner-siracusa-neureka-tiled (15000)
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (microLlama/microLlama1, 10000, 8, L3) / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: generic-kernels / test-runner-generic
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (miniMobileNet, 2000, 8, L3) / test-runner-siracusa-neureka-tiled (2000)
- GitHub Check: siracusa-neureka-models-tiled-doublebuffer-L3-wmem (microLlama/microLlama1, 10000, 8, true, L3, t... / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: mempool-models / test-runner-mempool
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (simpleRegression, 45000, 30000, 16000, 8, L3) / test-runner-siracusa-tiled (30000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (6000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (5000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (12000)
- GitHub Check: siracusa-kernels-tiled-doublebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-kernels / test-runner-siracusa
- GitHub Check: siracusa-kernels-tiled-singlebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (Transformer, 15000, 8, L3) / test-runner-siracusa-neureka-tiled (15000)
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (microLlama/microLlama1, 10000, 8, L3) / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: generic-kernels / test-runner-generic
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (miniMobileNet, 2000, 8, L3) / test-runner-siracusa-neureka-tiled (2000)
- GitHub Check: siracusa-neureka-models-tiled-doublebuffer-L3-wmem (microLlama/microLlama1, 10000, 8, true, L3, t... / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: mempool-models / test-runner-mempool
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (simpleRegression, 45000, 30000, 16000, 8, L3) / test-runner-siracusa-tiled (30000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (6000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (5000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (12000)
- GitHub Check: siracusa-kernels-tiled-doublebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-kernels / test-runner-siracusa
- GitHub Check: siracusa-kernels-tiled-singlebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (Transformer, 15000, 8, L3) / test-runner-siracusa-neureka-tiled (15000)
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (microLlama/microLlama1, 10000, 8, L3) / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: generic-kernels / test-runner-generic
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (miniMobileNet, 2000, 8, L3) / test-runner-siracusa-neureka-tiled (2000)
- GitHub Check: siracusa-neureka-models-tiled-doublebuffer-L3-wmem (microLlama/microLlama1, 10000, 8, true, L3, t... / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: mempool-models / test-runner-mempool
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (simpleRegression, 45000, 30000, 16000, 8, L3) / test-runner-siracusa-tiled (30000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (6000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (5000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (12000)
- GitHub Check: siracusa-kernels-tiled-doublebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-kernels / test-runner-siracusa
- GitHub Check: siracusa-kernels-tiled-singlebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (Transformer, 15000, 8, L3) / test-runner-siracusa-neureka-tiled (15000)
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (microLlama/microLlama1, 10000, 8, L3) / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: generic-kernels / test-runner-generic
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (miniMobileNet, 2000, 8, L3) / test-runner-siracusa-neureka-tiled (2000)
- GitHub Check: siracusa-neureka-models-tiled-doublebuffer-L3-wmem (microLlama/microLlama1, 10000, 8, true, L3, t... / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: mempool-models / test-runner-mempool
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (simpleRegression, 45000, 30000, 16000, 8, L3) / test-runner-siracusa-tiled (30000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (6000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (5000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (12000)
- GitHub Check: siracusa-kernels-tiled-doublebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-kernels / test-runner-siracusa
- GitHub Check: siracusa-kernels-tiled-singlebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (Transformer, 15000, 8, L3) / test-runner-siracusa-neureka-tiled (15000)
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (microLlama/microLlama1, 10000, 8, L3) / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: generic-kernels / test-runner-generic
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (miniMobileNet, 2000, 8, L3) / test-runner-siracusa-neureka-tiled (2000)
- GitHub Check: siracusa-neureka-models-tiled-doublebuffer-L3-wmem (microLlama/microLlama1, 10000, 8, true, L3, t... / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: mempool-models / test-runner-mempool
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (simpleRegression, 45000, 30000, 16000, 8, L3) / test-runner-siracusa-tiled (30000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (6000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (5000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (12000)
- GitHub Check: siracusa-kernels-tiled-doublebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-kernels / test-runner-siracusa
- GitHub Check: siracusa-kernels-tiled-singlebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (Transformer, 15000, 8, L3) / test-runner-siracusa-neureka-tiled (15000)
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (microLlama/microLlama1, 10000, 8, L3) / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: generic-kernels / test-runner-generic
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (miniMobileNet, 2000, 8, L3) / test-runner-siracusa-neureka-tiled (2000)
- GitHub Check: siracusa-neureka-models-tiled-doublebuffer-L3-wmem (microLlama/microLlama1, 10000, 8, true, L3, t... / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: mempool-models / test-runner-mempool
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (simpleRegression, 45000, 30000, 16000, 8, L3) / test-runner-siracusa-tiled (30000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (6000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (5000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (12000)
- GitHub Check: siracusa-kernels-tiled-doublebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-kernels / test-runner-siracusa
- GitHub Check: siracusa-kernels-tiled-singlebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (Transformer, 15000, 8, L3) / test-runner-siracusa-neureka-tiled (15000)
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (microLlama/microLlama1, 10000, 8, L3) / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: generic-kernels / test-runner-generic
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (miniMobileNet, 2000, 8, L3) / test-runner-siracusa-neureka-tiled (2000)
- GitHub Check: siracusa-neureka-models-tiled-doublebuffer-L3-wmem (microLlama/microLlama1, 10000, 8, true, L3, t... / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: mempool-models / test-runner-mempool
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (simpleRegression, 45000, 30000, 16000, 8, L3) / test-runner-siracusa-tiled (30000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (6000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (5000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (12000)
- GitHub Check: siracusa-kernels-tiled-doublebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-kernels / test-runner-siracusa
- GitHub Check: siracusa-kernels-tiled-singlebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (Transformer, 15000, 8, L3) / test-runner-siracusa-neureka-tiled (15000)
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (microLlama/microLlama1, 10000, 8, L3) / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: generic-kernels / test-runner-generic
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (miniMobileNet, 2000, 8, L3) / test-runner-siracusa-neureka-tiled (2000)
- GitHub Check: siracusa-neureka-models-tiled-doublebuffer-L3-wmem (microLlama/microLlama1, 10000, 8, true, L3, t... / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: mempool-models / test-runner-mempool
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (simpleRegression, 45000, 30000, 16000, 8, L3) / test-runner-siracusa-tiled (30000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (60000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (6000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (Attention, 60000, 10000, 5000, 2500, 8, L3) / test-runner-siracusa-tiled (5000)
- GitHub Check: siracusa-models-tiled-singlebuffer-L3 (miniMobileNet, 60000, 12000, 6000, 8, L3) / test-runner-siracusa-tiled (12000)
- GitHub Check: siracusa-kernels-tiled-doublebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-kernels / test-runner-siracusa
- GitHub Check: siracusa-kernels-tiled-singlebuffer-L2 / test-runner-siracusa-tiled
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (Transformer, 15000, 8, L3) / test-runner-siracusa-neureka-tiled (15000)
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (microLlama/microLlama1, 10000, 8, L3) / test-runner-siracusa-neureka-tiled (10000)
- GitHub Check: generic-kernels / test-runner-generic
- GitHub Check: siracusa-neureka-models-tiled-singlebuffer-L3 (miniMobileNet, 2000, 8, L3) / test-runner-siracusa-neureka-tiled (2000)
- GitHub Check: siracusa-neureka-models-tiled-doublebuffer-L3-wmem (microLlama/microLlama1, 10000, 8, true, L3, t... / test-runner-siracusa-neureka-tiled (10000)
🔇 Additional comments (5)
Deeploy/Targets/Neureka/TileConstraints/RequantHelpers.py (2)
12-32: LGTM: Requantization constraint logic is correct.The function properly adds mul and add tensor dimensions to the tiler model and enforces channel equality constraints. The implementation correctly uses the TilerModel API to retrieve dimension variables and add constraints.
35-53: LGTM: Load schedule generation is correct.The function correctly generates per-cube load schedules for mul and add buffers by:
- Extracting channel offset and size from each output cube
- Creating HyperRectangles with zeros for all dimensions except the last (channel) dimension
- Properly preserving the rank of the mul/add tensors as mentioned in the PR objectives
This implementation aligns with the NHWC layout convention used by Neureka.
Deeploy/Targets/Neureka/TileConstraints/NeurekaDepthwiseConstraint.py (1)
9-9: LGTM: Required imports added.The imports for
VariableBufferand requantization helpers (requantAddGeometricalConstraint,requantLoadSchedule) are necessary for the new memory-level checking and requantization functionality introduced in this PR.Also applies to: 12-12
Deeploy/Targets/Neureka/TileConstraints/NeurekaPointwiseConstraint.py (1)
48-52: SRAM-aware weight constraint is correct.Using weightBuffer._memoryLevel to switch channel constraint is the right replacement for Wmem variants.
Deeploy/Targets/Neureka/Tiler.py (1)
6-13: Wmem bindings removal verified as complete.Ripgrep confirms zero stale references to removed Wmem variants, MemoryAwareNodeBinding, or NodeMemoryLevelChecker across the codebase. The CHANGELOG documents the intentional removal at lines 78–80, and the import cleanup in Tiler.py is consistent with the broader cleanup. Code changes are sound.
…eady expects annotations
lukamac
force-pushed
the
remove-memory-aware-binding
branch
from
e9f8bef to
b27c5ac
Compare
lukamac
changed the title
1 participant
Remove memory-aware node bindings because it makes the great parser refactor easier.
The memory-aware node bindings exist only for Neureka to be able to separate bindings that use the dedicated weight memory vs. those who don't.
But those bindings can simply be rewritten to check whether the weights reside in weight memory and change behavior accordingly.
By removing the memory-aware node bindings, we remove another dependency on having hoisted buffers in the middle of parsing.
The RequantHelpers are a bonus that fixes the requantization mul and add hyperrectangles to keep the rank of the original tensors.
Added
Changed
Fixed
PR Merge Checklist
develcommit and pointing todevel.CHANGELOG.mdfile has been updated.