From 321a39dfc2bc63798581d598990d4609f2050cd2 Mon Sep 17 00:00:00 2001
From: Wissam Boussella <wissam.boussella@cea.fr>
Date: Fri, 6 Dec 2024 14:18:42 +0100
Subject: [PATCH] [Fix] Fix for conv depthwise and softmax, mbnet-v1 now
working
---
.../_Aidge_Arm/kernels/Convolution/ConvDW.hpp | 4 +-
.../Softmax/aidge_softmax_chw_float32.h | 8 ++--
.../configuration/conv_dw_cinfig.jinja | 40 +++++++++++++++++++
.../templates/configuration/softmax.jinja | 7 +++-
.../forward_call/conv_dw_kernel.jinja | 28 +++++++++++++
aidge_export_arm_cortexm/operators.py | 8 ++--
6 files changed, 83 insertions(+), 12 deletions(-)
create mode 100644 aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/conv_dw_cinfig.jinja
create mode 100644 aidge_export_arm_cortexm/_Aidge_Arm/templates/forward_call/conv_dw_kernel.jinja
diff --git a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Convolution/ConvDW.hpp b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Convolution/ConvDW.hpp
index cefe933..e1cca08 100644
--- a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Convolution/ConvDW.hpp
+++ b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Convolution/ConvDW.hpp
@@ -95,8 +95,8 @@ __attribute__((always_inline)) inline void convcellDWPropagate(
for (int output = 0; output < NB_OUTPUTS; ++output) {
const int channel = (output * NB_CHANNELS) / NB_OUTPUTS;
-
- SUM_T weightedSum = biasses[output];
+ using Sum_T = typename std::conditional<std::is_floating_point<Input_T>::value, float, int32_t>::type;
+ Sum_T weightedSum = biasses[output];
for (int sy = 0; sy < KERNEL_HEIGHT; ++sy) {
if ((PADDING_Y != 0
diff --git a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Softmax/aidge_softmax_chw_float32.h b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Softmax/aidge_softmax_chw_float32.h
index 93d710e..6ec9122 100644
--- a/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Softmax/aidge_softmax_chw_float32.h
+++ b/aidge_export_arm_cortexm/_Aidge_Arm/kernels/Softmax/aidge_softmax_chw_float32.h
@@ -2,16 +2,16 @@
void aidge_softmax_chw_float32 (float* inputs,
float* outputs,
- int inputDims[],
+ const int inputDims[],
int axis,
- unsigned int size_inputDim,
- unsigned int size)
+ const unsigned int size_inputDim,
+ const unsigned int size)
{
axis += (axis >= 0 ) ? 0 : size_inputDim;
int postAxisElems = 1;
- for (int i = axis+1; i < size_inputDim; ++i) {
+ for (unsigned int i = axis+1; i < size_inputDim; ++i) {
postAxisElems *= inputDims[i];
}
int preAxisElems = 1;
diff --git a/aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/conv_dw_cinfig.jinja b/aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/conv_dw_cinfig.jinja
new file mode 100644
index 0000000..71c8b59
--- /dev/null
+++ b/aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/conv_dw_cinfig.jinja
@@ -0,0 +1,40 @@
+#define {{ name|upper }}_LAYER_H
+
+#include "typedefs.h"
+#include "nn_scaling_functions.hpp"
+
+{% include "./_def_io.jinja" %}
+{% include "./_meminfo.jinja" %}
+
+// Attributes
+#define {{ name|upper }}_KERNEL_HEIGHT {{ kernel_dims[1] }}
+#define {{ name|upper }}_KERNEL_WIDTH {{ kernel_dims[0] }}
+#define {{ name|upper }}_PADDING_Y {{ padding[1] }}
+#define {{ name|upper }}_PADDING_X {{ padding[0] }}
+#define {{ name|upper }}_STRIDE_Y {{ stride_dims[1] }}
+#define {{ name|upper }}_STRIDE_X {{ stride_dims[0] }}
+#define {{ name|upper }}_DILATION_Y {{ dilation_dims[1] }}
+#define {{ name|upper }}_DILATION_X {{ dilation_dims[0] }}
+{# #define {{ name|upper }}_GROUP {{ group }} #}
+
+// Activation/Scaling
+#define {{ name|upper }}_ACTIVATION {{ activation }}
+
+{%- if scaling_type == "floating_point" %}
+static const N2D2_Export::FloatingPointScaling {{ name|upper }}_SCALING = { {{scaling_value}} };
+{%- elif scaling_type == "fixed_point" %}
+static const N2D2_Export::FixedPointScaling<{{scaling_value}}, {{fractional_bits}}> {{ name|upper }}_SCALING;
+{%- elif scaling_type == "single_shift" %}
+static const N2D2_Export::SingleShiftScaling<{{shift_value}}> {{ name|upper }}_SCALING;
+{%- else %}
+static const N2D2_Export::NoScaling {{ name|upper }}_SCALING;
+{%- endif %}
+
+// Sizes
+#define {{ name|upper }}_WEIGHTS_SIZE {{ out_chan[0] * in_chan[0] * kernel_dims[1] * kernel_dims[0] }}
+#define {{ name|upper }}_BIASES_SIZE {{ out_chan[0] }}
+#define {{ name|upper }}_OUTPUTS_SIZE {{ out_chan[0] * out_height[0] * out_width[0] }}
+#define {{ name|upper }}_CHANNELS_SIZE {{ in_chan[0] * in_height[0] * in_width[0] }}
+
+
+#endif /* {{ name|upper }}_LAYER_H */
diff --git a/aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/softmax.jinja b/aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/softmax.jinja
index 9f1eb50..b48a02c 100644
--- a/aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/softmax.jinja
+++ b/aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/softmax.jinja
@@ -1,3 +1,7 @@
+
+{% include "./_def_io.jinja" %}
+{% include "./_meminfo.jinja" %}
+
{#- For name header -#}
#ifndef {{ name|upper }}_LAYER_H
#define {{ name|upper }}_LAYER_H
@@ -5,9 +9,8 @@
{# For layer configuration -#}
#define {{ name|upper }}_INPUTS_SIZE {{ in_size[0] }}
#define {{ name|upper }}_OUTPUTS_SIZE {{ out_size[0] }}
-#define {{ name|upper }}_DIMS {{ in_dims[0] }}
#define {{ name|upper }}_AXIS {{ axis }}
#define {{ name|upper }}_INPUT_DIMS_SIZE {{ in_dims[0]|length}}
-
+static const int {{ name|upper }}_DIMS[] = { {{ in_dims[0] | join(', ') }} };
#endif /* {{ name|upper }}_LAYER_H */
diff --git a/aidge_export_arm_cortexm/_Aidge_Arm/templates/forward_call/conv_dw_kernel.jinja b/aidge_export_arm_cortexm/_Aidge_Arm/templates/forward_call/conv_dw_kernel.jinja
new file mode 100644
index 0000000..af00c26
--- /dev/null
+++ b/aidge_export_arm_cortexm/_Aidge_Arm/templates/forward_call/conv_dw_kernel.jinja
@@ -0,0 +1,28 @@
+{% filter indent(width=4, first=False) %}
+{% include "./_mem_offset.jinja" %}
+N2D2_Export::convcellDWPropagate<{{ in_name[0]|upper }}_NB_CHANNELS,
+ {{ in_name[0]|upper }}_IN_HEIGHT,
+ {{ in_name[0]|upper }}_IN_WIDTH,
+ {{ out_name[0]|upper }}_NB_OUTPUTS,
+ {{ out_name[0]|upper }}_OUT_HEIGHT,
+ {{ out_name[0]|upper }}_OUT_WIDTH,
+ {{ name|upper }}_PADDING_Y,
+ {{ name|upper }}_PADDING_X,
+ {{ name|upper }}_STRIDE_Y,
+ {{ name|upper }}_STRIDE_X,
+ {{ name|upper }}_KERNEL_HEIGHT,
+ {{ name|upper }}_KERNEL_WIDTH,
+ {{ name|upper }}_ACTIVATION,
+ {{ in_name[0]|upper }}_CONT_OFFSET,
+ {{ in_name[0]|upper }}_CONT_SIZE,
+ {{ in_name[0]|upper }}_WRAP_OFFSET,
+ {{ in_name[0]|upper }}_WRAP_SIZE,
+ {{ in_name[0]|upper }}_STRIDE,
+ {{ out_name[0]|upper }}_CONT_OFFSET,
+ {{ out_name[0]|upper }}_CONT_SIZE,
+ {{ out_name[0]|upper }}_WRAP_OFFSET,
+ {{ out_name[0]|upper }}_WRAP_SIZE,
+ {{ out_name[0]|upper }}_STRIDE>
+ ({{in_name[0]}}, {{out_name[0]}}, {{in_name[2]}}, {{in_name[1]}}, {{ name|upper }}_SCALING);
+{% endfilter %}
+
diff --git a/aidge_export_arm_cortexm/operators.py b/aidge_export_arm_cortexm/operators.py
index 9f00632..b27efa7 100644
--- a/aidge_export_arm_cortexm/operators.py
+++ b/aidge_export_arm_cortexm/operators.py
@@ -250,8 +250,8 @@ class ConvDW_ARMCortexM(ExportNodeCpp):
# Use PaddedConv to add padding attribute
self.attributes["padding"] = [0, 0]
- self.config_template = str(ROOT / "_Aidge_Arm" / "templates" / "configuration" / "conv_config.jinja")
- self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "conv_kernel.jinja")
+ self.config_template = str(ROOT / "_Aidge_Arm" / "templates" / "configuration" / "conv_dw_config.jinja")
+ self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "conv_dw_kernel.jinja")
self.include_list = []
self.kernels_to_copy = [
str(ROOT / "_Aidge_Arm" / "kernels" / "Convolution" / "ConvDW.hpp")
@@ -275,8 +275,8 @@ class PaddedConvDW_ARMCortexM(ExportNodeCpp):
self.attributes["dilation_dims"] = n.get_operator(
).attr.dilation_dims
- self.config_template = str(ROOT / "_Aidge_Arm" / "templates" / "configuration" / "conv_config.jinja")
- self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "conv_kernel.jinja")
+ self.config_template = str(ROOT / "_Aidge_Arm" / "templates" / "configuration" / "conv_dw_config.jinja")
+ self.forward_template = str(ROOT / "_Aidge_Arm" / "templates" / "forward_call" / "conv_dw_kernel.jinja")
self.include_list = []
self.kernels_to_copy = [
str(ROOT / "_Aidge_Arm" / "kernels" / "Convolution" / "ConvDW.hpp")
--
GitLab