Skip to content

Fix VisualWakeWords on GAP9 #187

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
Xeratec wants to merge 5 commits into
base: devel
Choose a base branch
from
Draft

Fix VisualWakeWords on GAP9 #187

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
65eda17
Fix DWConv im2col scratch buffer sizing on PULPOpen
Xeratec Apr 30, 2026
0e0d596
Fix pulp-nn linear leftover-neuron RQS parameter aliasing
Xeratec Apr 30, 2026
06e2250
Enable VisualWakeWord Test on GAP9
Xeratec Apr 30, 2026
2519c50
Fix transient-buffer hoisting to dispatch through subclass overrides
Xeratec Apr 30, 2026
8c769c7
Add depthwise im2col override for PULP1DDWConvTemplate
Xeratec Apr 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 17 additions & 0 deletions Deeploy/Targets/PULPOpen/Templates/ConvTemplate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,23 @@ def alignToContext(self, ctxt: NetworkContext,

return ctxt, operatorRepresentation, []

@staticmethod
def computeTransientBuffersSize(
ctxt: NetworkContext,
operatorRepresentation: OperatorRepresentation) -> List[Tuple[str, Union[int, IntVar]]]:
# The depthwise pulp-nn kernel reuses one column-shaped im2col scratch
# per core. Per core it needs `dim_kernel_x * (dim_in_y + pad_top +
# pad_bot) + dim_kernel_x` bytes (the trailing `+ dim_kernel_x` is the
# safety zone for the last v4u write).
pad_top = operatorRepresentation['padding_y_top']
pad_bot = operatorRepresentation['padding_y_bottom']
per_core = (operatorRepresentation['dim_kernel_x'] *
(operatorRepresentation['dim_im_in_y'] + pad_top + pad_bot) +
operatorRepresentation['dim_kernel_x'])
im2col_dim = 8 * per_core
im2col_name = operatorRepresentation['nodeName'] + "_buffer"
return [(im2col_name, im2col_dim)]
Comment thread
coderabbitai[bot] marked this conversation as resolved.
Outdated


class PULP1DConvTemplate(NodeTemplate):

Expand Down
Binary file added DeeployTest/Tests/Kernels/Integer/Conv/DW_2D_RQ_8x16x16/activations.npz
View file
Open in desktop
Binary file not shown.
Binary file added DeeployTest/Tests/Kernels/Integer/Conv/DW_2D_RQ_8x16x16/inputs.npz
View file
Open in desktop
Binary file not shown.
Binary file added DeeployTest/Tests/Kernels/Integer/Conv/DW_2D_RQ_8x16x16/network.onnx
View file
Open in desktop
Binary file not shown.
Binary file added DeeployTest/Tests/Kernels/Integer/Conv/DW_2D_RQ_8x16x16/outputs.npz
View file
Open in desktop
Binary file not shown.
6 changes: 5 additions & 1 deletion DeeployTest/test_gap9_tiled_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,8 @@
"Kernels/Integer/MatMul/Regular": [64000, 32000, 16000],
"Kernels/Integer/Conv/Regular_2D_RQ": [8000, 6000, 4000],
"Kernels/Integer/Conv/StriddedPadded_2D_RQ": [600],
"Kernels/Integer/Conv/DW_2D_RQ": [2561],
"Kernels/Integer/Conv/DW_2D_RQ": [64000, 2561],
"Kernels/Integer/Conv/DW_2D_RQ_8x16x16": [64000],
"Kernels/Integer/Softmax/Regular": [800, 500, 300],
"Kernels/Integer/Concat": [32000, 16000, 8000],
"Kernels/Integer/Hardswish/Regular": [750],
Expand All @@ -39,6 +40,7 @@
"Kernels/Integer/MatMul/Regular": [64000, 32000, 16000],
"Kernels/Integer/Conv/Regular_2D_RQ": [8000, 6000, 5000],
"Kernels/Integer/Conv/DW_2D_RQ": [5121],
"Kernels/Integer/Conv/DW_2D_RQ_8x16x16": [64000],
"Kernels/Integer/Softmax/Regular": [1600, 1000, 600],
"Kernels/Integer/Concat": [64000, 32000, 16000],
"Kernels/Integer/Hardswish/Regular": [750],
Expand All @@ -64,6 +66,7 @@
"Models/MLPerf/KeywordSpotting": [64000],
"Models/MLPerf/ImageClassification": [64000],
"Models/MLPerf/AnomalyDetection": [64000],
"Models/MLPerf/VisualWakeWords": [64000],
}

L2_DOUBLEBUFFER_MODELS = {
Expand All @@ -72,6 +75,7 @@
"Models/MLPerf/KeywordSpotting": [64000],
"Models/MLPerf/ImageClassification": [64000],
"Models/MLPerf/AnomalyDetection": [64000],
"Models/MLPerf/VisualWakeWords": [64000],
}

L3_SINGLEBUFFER_MODELS = {
Expand Down
2 changes: 2 additions & 0 deletions DeeployTest/test_siracusa_tiled_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -112,6 +112,7 @@
"Models/MLPerf/KeywordSpotting": [64000],
"Models/MLPerf/ImageClassification": [64000],
"Models/MLPerf/AnomalyDetection": [64000],
"Models/MLPerf/VisualWakeWords": [64000],
"Models/CCT/FP32/CCT_1_16_16_8": [64000],
"Models/TinyViT/Demo": [4000],
}
Expand All @@ -127,6 +128,7 @@
"Models/MLPerf/KeywordSpotting": [128000],
"Models/MLPerf/ImageClassification": [128000],
"Models/MLPerf/AnomalyDetection": [128000],
"Models/MLPerf/VisualWakeWords": [128000],
"Models/CCT/FP32/CCT_1_16_16_8": [128000],
"Models/TinyViT/Demo": [8000],
}
Expand Down
2 changes: 2 additions & 0 deletions TargetLibraries/GAP9/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,8 @@ target_compile_options(deeploygap9 PUBLIC
)

target_compile_options(deeploygap9 PRIVATE
-Wno-incompatible-pointer-types
-Wno-implicit-function-declaration
-Wno-sign-conversion
-Wno-sign-compare
-Wno-type-limits
Expand Down
Binary file modified TargetLibraries/GAP9/prebuilt/libpulp-nn-mixed.a
View file
Open in desktop
Binary file not shown.
3 changes: 2 additions & 1 deletion TargetLibraries/PULPOpen/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,7 +62,8 @@ if (platform IN_LIST PULP_NNX_PLATFORMS)
-Wno-sign-conversion
-Wno-typedef-redefinition
-Wno-unused-parameter
-Wno-incompatible-pointer-types-discards-qualifiers
-Wno-incompatible-pointer-types
-Wno-discards-qualifiers
)
target_link_libraries(deeploypulp INTERFACE pulp-nnx)
endif()
2 changes: 1 addition & 1 deletion TargetLibraries/third_party/pulp-nn-mixed
Open in desktop
Submodule pulp-nn-mixed updated 74 files
+2 −1 .gitignore
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i2_i8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i2_i8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i2_i8_i8.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i2_u8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i2_u8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i2_u8_i8.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i4_i8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i4_i8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i4_i8_i8.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i4_u8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i4_u8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i4_u8_i8.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i8_i8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i8_i8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i8_i8_i8.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i8_u8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i8_u8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_i8_u8_i8.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u2_i8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u2_i8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u2_i8_i8.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u2_u8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u2_u8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u2_u8_i8.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u4_i8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u4_i8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u4_i8_i8.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u4_u8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u4_u8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u4_u8_i8.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u8_i8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u8_i8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u8_i8_i8.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u8_u8_i2.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u8_u8_i4.c
+3 −3 XpulpV2/32bit/src/LinearQuant/pulp_nn_linear_u8_u8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i2_i8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i2_i8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i2_i8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i2_u8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i2_u8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i2_u8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i4_i8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i4_i8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i4_i8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i4_u8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i4_u8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i4_u8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i8_i8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i8_i8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i8_i8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i8_u8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i8_u8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_i8_u8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u2_i8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u2_i8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u2_i8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u2_u8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u2_u8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u2_u8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u4_i8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u4_i8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u4_i8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u4_u8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u4_u8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u4_u8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u8_i8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u8_i8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u8_i8_i8.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u8_u8_i2.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u8_u8_i4.c
+3 −3 XpulpV2/64bit/src/LinearQuant/pulp_nn_linear_u8_u8_i8.c
+3 −3 generators/templates/pulp_nn_linear_q_x_y_z.t
Loading