diff --git a/aidge_export_cpp/kernels/convolution.hpp b/aidge_export_cpp/kernels/convolution.hpp
index efc7ee7a192112018d6c582207560d93b4548add..6ea9f0579b84dd5a28a5ea66a778326fcd9c84ce 100644
--- a/aidge_export_cpp/kernels/convolution.hpp
+++ b/aidge_export_cpp/kernels/convolution.hpp
@@ -65,8 +65,8 @@ void convolution_forward(
int oOffset = NB_OUTPUTS * oPos;
// <--
-
- Bias_T weightedSum = biases[output];
+ // Check if the biases are defined
+ Bias_T weightedSum = biases ? biases[output] : 0;
for (int sy = 0; sy < KERNEL_HEIGHT; ++sy) {
if ((PADDING_Y != 0
@@ -116,4 +116,45 @@ void convolution_forward(
}
}
+// Template specialization when biases are not given to the convolution
+template<int NB_CHANNELS,
+ int CHANNELS_HEIGHT, int CHANNELS_WIDTH,
+ int NB_OUTPUTS,
+ int OUTPUTS_HEIGHT, int OUTPUTS_WIDTH,
+ int PADDING_Y, int PADDING_X,
+ int STRIDE_Y, int STRIDE_X,
+ int DILATION_Y, int DILATION_X,
+ int KERNEL_HEIGHT, int KERNEL_WIDTH,
+ ActivationFunction_T ACTIVATION,
+ typename Input_T, typename Output_T,
+ typename Weight_T,
+ typename Rescaling_T>
+__attribute__((always_inline)) inline
+void convolution_forward(
+ const Input_T* __restrict inputs,
+ Output_T* __restrict outputs,
+ const Weight_T* __restrict weights,
+ std::nullptr_t __restrict,
+ const Rescaling_T& __restrict rescaling)
+{
+ const float* b = nullptr;
+
+ convolution_forward<NB_CHANNELS,
+ CHANNELS_HEIGHT,
+ CHANNELS_WIDTH,
+ NB_OUTPUTS,
+ OUTPUTS_HEIGHT,
+ OUTPUTS_WIDTH,
+ PADDING_Y,
+ PADDING_X,
+ STRIDE_Y,
+ STRIDE_X,
+ DILATION_Y,
+ DILATION_X,
+ KERNEL_HEIGHT,
+ KERNEL_WIDTH,
+ ACTIVATION>
+ (inputs, outputs, weights, b, rescaling);
+}
+
#endif // __AIDGE_EXPORT_CPP_KERNELS_CONVOLUTION__
diff --git a/aidge_export_cpp/kernels/reshape.hpp b/aidge_export_cpp/kernels/reshape.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..a5828da39553222ef933f0355424e50956ca8490
--- /dev/null
+++ b/aidge_export_cpp/kernels/reshape.hpp
@@ -0,0 +1,27 @@
+#ifndef __AIDGE_EXPORT_CPP_KERNELS_RESHAPE__
+#define __AIDGE_EXPORT_CPP_KERNELS_RESHAPE__
+
+#include "network/typedefs.hpp"
+
+// Generic function for reshape and activation
+
+template<int M,
+ typename Input_T, typename Output_T>
+__attribute__((always_inline)) inline
+void reshape_forward (
+ const Input_T* __restrict, // First input is useless as it only dictate the resulting layout of the reshape
+ const Input_T* __restrict inputs2,
+ Output_T* __restrict outputs)
+{
+ // If inputs and outputs pointers are the same, the memory manager has already optimized this function so it is a no-op !
+ if (inputs2 == outputs)
+ return;
+
+ // A reshape in c++ world should equal to a Noop
+ // We only need to copy the input buffer to the output
+ for (int m = 0; m < M; ++m) {
+ outputs[m] = inputs2[m];
+ }
+}
+
+#endif // __AIDGE_EXPORT_CPP_KERNELS_RESHAPE__
\ No newline at end of file
diff --git a/aidge_export_cpp/operators.py b/aidge_export_cpp/operators.py
index 54c38055b2c622aa1796bc1e7ff2dd46e60afcba..346928f4a84c403df2172311cede8b99fd06eebe 100644
--- a/aidge_export_cpp/operators.py
+++ b/aidge_export_cpp/operators.py
@@ -94,27 +94,65 @@ class ReLUCPP(ExportNodeCpp):
str(ROOT / "kernels" / "rescaling.hpp")
]
-@ExportLibCpp.register("Conv2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
-class ConvCPP(ExportNodeCpp):
+@ExportLibCpp.register("Reshape", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
+class ReshapeCPP(ExportNodeCpp):
+ def __init__(self, node, mem_info):
+ super().__init__(node, mem_info)
+ self.config_template = str(
+ ROOT / "templates" / "configuration" / "reshape_config.jinja")
+ self.forward_template = str(
+ ROOT / "templates" / "kernel_forward" / "reshape_forward.jinja")
+ self.include_list = []
+ self.kernels_to_copy = [
+ str(ROOT / "kernels" / "reshape.hpp"),
+ ]
+
+@ExportLibCpp.register("MatMul", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
+class MatMulCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
super().__init__(node, mem_info)
- # No padding with Conv
- # Use PaddedConv to add padding attribute
- self.attributes["padding"] = [0, 0]
self.attributes["activation"] = "Linear"
self.attributes["rescaling"] = "NoScaling"
self.config_template = str(
- ROOT / "templates" / "configuration" / "convolution_config.jinja")
+ ROOT / "templates" / "configuration" / "matmul_config.jinja")
self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "convolution_forward.jinja")
+ ROOT / "templates" / "kernel_forward" / "matmul_forward.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "kernels" / "convolution.hpp"),
- str(ROOT / "kernels" / "macs.hpp"),
- str(ROOT / "kernels" / "activation.hpp"),
- str(ROOT / "kernels" / "rescaling.hpp")
+ str(ROOT / "kernels" / "matmul.hpp"),
]
+def _setup_conv2D(conv):
+ """Common setup code for convolutions: Conv2D and PaddedConv2D."""
+
+ # If biases are not provided we set it as nullptr instead of None
+ if (len(conv.attributes["in_name"]) > 2 and conv.attributes["in_name"][2] is None):
+ conv.attributes["in_name"][2] = "nullptr"
+
+ conv.attributes["activation"] = "Linear"
+ conv.attributes["rescaling"] = "NoScaling"
+ conv.config_template = str(
+ ROOT / "templates" / "configuration" / "convolution_config.jinja")
+ conv.forward_template = str(
+ ROOT / "templates" / "kernel_forward" / "convolution_forward.jinja")
+ conv.include_list = []
+ conv.kernels_to_copy = [
+ str(ROOT / "kernels" / "convolution.hpp"),
+ str(ROOT / "kernels" / "macs.hpp"),
+ str(ROOT / "kernels" / "activation.hpp"),
+ str(ROOT / "kernels" / "rescaling.hpp")
+ ]
+
+@ExportLibCpp.register("Conv2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
+class ConvCPP(ExportNodeCpp):
+ def __init__(self, node, mem_info):
+ super().__init__(node, mem_info)
+ # No padding with Conv
+ # Use PaddedConv to add padding attribute
+ self.attributes["padding"] = [0, 0]
+
+ _setup_conv2D(self)
+
@ExportLibCpp.register_metaop("PaddedConv2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class PaddedConvCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
@@ -131,74 +169,60 @@ class PaddedConvCPP(ExportNodeCpp):
).attr.stride_dims
self.attributes["dilation_dims"] = n.get_operator(
).attr.dilation_dims
- self.attributes["activation"] = "Linear"
- self.attributes["rescaling"] = "NoScaling"
- self.config_template = str(
- ROOT / "templates" / "configuration" / "convolution_config.jinja")
- self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "convolution_forward.jinja")
- self.include_list = []
- self.kernels_to_copy = [
- str(ROOT / "kernels" / "convolution.hpp"),
- str(ROOT / "kernels" / "macs.hpp"),
- str(ROOT / "kernels" / "activation.hpp"),
- str(ROOT / "kernels" / "rescaling.hpp")
- ]
+
+ _setup_conv2D(self)
+
+def _setup_elemwise_op(elemwise, op):
+ """Common code (template and kernel setup) shared across all the different elementWise operator (Add, Sub,...)."""
+
+ elemwise.attributes["elemwise_op"] = op
+ elemwise.attributes["activation"] = "Linear"
+ elemwise.attributes["rescaling"] = "NoScaling"
+ elemwise.config_template = str(
+ ROOT / "templates" / "configuration" / "elemwise_config.jinja")
+ elemwise.forward_template = str(
+ ROOT / "templates" / "kernel_forward" / "elemwise_forward.jinja")
+ elemwise.include_list = []
+ elemwise.kernels_to_copy = [
+ str(ROOT / "kernels" / "elemwise.hpp"),
+ str(ROOT / "kernels" / "activation.hpp"),
+ str(ROOT / "kernels" / "rescaling.hpp")
+ ]
@ExportLibCpp.register("Add", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class AddCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
super().__init__(node, mem_info)
- self.attributes["elemwise_op"] = "Add"
- self.attributes["activation"] = "Linear"
- self.attributes["rescaling"] = "NoScaling"
- self.config_template = str(
- ROOT / "templates" / "configuration" / "elemwise_config.jinja")
- self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "elemwise_forward.jinja")
- self.include_list = []
- self.kernels_to_copy = [
- str(ROOT / "kernels" / "elemwise.hpp"),
- str(ROOT / "kernels" / "activation.hpp"),
- str(ROOT / "kernels" / "rescaling.hpp")
- ]
+
+ _setup_elemwise_op(self, "Add")
@ExportLibCpp.register("Sub", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class SubCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
super().__init__(node, mem_info)
- self.attributes["elemwise_op"] = "Sub"
- self.attributes["activation"] = "Linear"
- self.attributes["rescaling"] = "NoScaling"
- self.config_template = str(
- ROOT / "templates" / "configuration" / "elemwise_config.jinja")
- self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "elemwise_forward.jinja")
- self.include_list = []
- self.kernels_to_copy = [
- str(ROOT / "kernels" / "elemwise.hpp"),
- str(ROOT / "kernels" / "activation.hpp"),
- str(ROOT / "kernels" / "rescaling.hpp")
- ]
+ _setup_elemwise_op(self, "Sub")
@ExportLibCpp.register("Mul", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class MulCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
super().__init__(node, mem_info)
- self.attributes["elemwise_op"] = "Mul"
- self.attributes["activation"] = "Linear"
- self.attributes["rescaling"] = "NoScaling"
- self.config_template = str(
- ROOT / "templates" / "configuration" / "elemwise_config.jinja")
- self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "elemwise_forward.jinja")
- self.include_list = []
- self.kernels_to_copy = [
- str(ROOT / "kernels" / "elemwise.hpp"),
- str(ROOT / "kernels" / "activation.hpp"),
- str(ROOT / "kernels" / "rescaling.hpp")
- ]
+
+ _setup_elemwise_op(self, "Mul")
+
+def _setup_pooling(pooling):
+ """Common code (template and kernel setup) shared across all the different pooling operator."""
+
+ pooling.config_template = str(
+ ROOT / "templates" / "configuration" / "pooling_config.jinja")
+ pooling.forward_template = str(
+ ROOT / "templates" / "kernel_forward" / "pooling_forward.jinja")
+ pooling.include_list = []
+ pooling.kernels_to_copy = [
+ str(ROOT / "kernels" / "pooling.hpp"),
+ str(ROOT / "kernels" / "activation.hpp"),
+ str(ROOT / "kernels" / "rescaling.hpp")
+ ]
@ExportLibCpp.register("MaxPooling2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class MaxPoolCPP(ExportNodeCpp):
@@ -211,17 +235,7 @@ class MaxPoolCPP(ExportNodeCpp):
self.attributes["pool_type"] = "Max"
self.attributes["activation"] = "Linear"
- self.config_template = str(
- ROOT / "templates" / "configuration" / "pooling_config.jinja")
- self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "pooling_forward.jinja")
- self.include_list = []
- self.kernels_to_copy = [
- str(ROOT / "kernels" / "pooling.hpp"),
- str(ROOT / "kernels" / "activation.hpp"),
- str(ROOT / "kernels" / "rescaling.hpp")
- ]
-
+ _setup_pooling(self)
@ExportLibCpp.register_metaop("PaddedMaxPooling2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class PaddedMaxPoolCPP(ExportNodeCpp):
@@ -239,17 +253,7 @@ class PaddedMaxPoolCPP(ExportNodeCpp):
self.attributes["pool_type"] = "Max"
self.attributes["activation"] = "Linear"
- self.config_template = str(
- ROOT / "templates" / "configuration" / "pooling_config.jinja")
- self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "pooling_forward.jinja")
- self.include_list = []
- self.kernels_to_copy = [
- str(ROOT / "kernels" / "pooling.hpp"),
- str(ROOT / "kernels" / "activation.hpp"),
- str(ROOT / "kernels" / "rescaling.hpp")
- ]
-
+ _setup_pooling(self)
@ExportLibCpp.register("GlobalAveragePooling", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class GlobalAveragePoolCPP(ExportNodeCpp):
@@ -267,16 +271,7 @@ class GlobalAveragePoolCPP(ExportNodeCpp):
self.attributes["pool_type"] = "Average"
self.attributes["activation"] = "Linear"
- self.config_template = str(
- ROOT / "templates" / "configuration" / "pooling_config.jinja")
- self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "pooling_forward.jinja")
- self.include_list = []
- self.kernels_to_copy = [
- str(ROOT / "kernels" / "pooling.hpp"),
- str(ROOT / "kernels" / "activation.hpp"),
- str(ROOT / "kernels" / "rescaling.hpp")
- ]
+ _setup_pooling(self)
@ExportLibCpp.register("FC", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class FcCPP(ExportNodeCpp):
diff --git a/aidge_export_cpp/templates/configuration/matmul_config.jinja b/aidge_export_cpp/templates/configuration/matmul_config.jinja
index fece988ac13b0136a8506abb39998114923817d6..38316f20947fa726085bf3577ead510e6c5096f3 100644
--- a/aidge_export_cpp/templates/configuration/matmul_config.jinja
+++ b/aidge_export_cpp/templates/configuration/matmul_config.jinja
@@ -2,10 +2,13 @@
#ifndef {{ name|upper }}_LAYER_H
#define {{ name|upper }}_LAYER_H
+{% include "./_def_io.jinja" %}
+{% include "./_meminfo.jinja" %}
+
{# For layer configuration -#}
-#define {{ name|upper }}_M {{ inputs_dims[0][0] }}
-#define {{ name|upper }}_K {{ inputs_dims[0][1] }}
-#define {{ name|upper }}_N {{ inputs_dims[1][1] }}
+#define {{ name|upper }}_M {{ in_dims[0][0] }}
+#define {{ name|upper }}_K {{ in_dims[0][1] }}
+#define {{ name|upper }}_N {{ in_dims[1][1] }}
#define {{ name|upper }}_ACTIVATION {{ activation }}
static const {{ rescaling }} {{ name|upper }}_RESCALING = {};
diff --git a/aidge_export_cpp/templates/configuration/reshape_config.jinja b/aidge_export_cpp/templates/configuration/reshape_config.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..041cf8ae898ed561e050354d137c1c676723a81f
--- /dev/null
+++ b/aidge_export_cpp/templates/configuration/reshape_config.jinja
@@ -0,0 +1,8 @@
+{#- For name header -#}
+#ifndef {{ name|upper }}_LAYER_H
+#define {{ name|upper }}_LAYER_H
+
+{% include "./_def_io.jinja" %}
+{% include "./_meminfo.jinja" %}
+{# For layer configuration -#}
+#define {{ name|upper }}_NB_ELTS {{ in_dims[0]|join('*') }}
diff --git a/aidge_export_cpp/templates/kernel_forward/matmul_forward.jinja b/aidge_export_cpp/templates/kernel_forward/matmul_forward.jinja
index ce80ffd2abc90ad611d3008c57aae36383691452..64b3df301794e1cb3d56170646a6b9524f18a6ab 100644
--- a/aidge_export_cpp/templates/kernel_forward/matmul_forward.jinja
+++ b/aidge_export_cpp/templates/kernel_forward/matmul_forward.jinja
@@ -1,5 +1,9 @@
+{% filter indent(width=4, first=False) %}
+{% include "./_mem_offset.jinja" %}
matmul_forward<{{name|upper}}_M,
{{name|upper}}_K,
{{name|upper}}_N,
{{name|upper}}_ACTIVATION>
- ({{inputs1_name}}, {{inputs2_name}}, {{outputs_name}}, {{name|upper}}_RESCALING);
\ No newline at end of file
+ ({{in_name[0]}}, {{in_name[1]}}, {{out_name[0]}}, {{name|upper}}_RESCALING);
+{% include "./_save_outputs.jinja" %}
+{% endfilter %}
diff --git a/aidge_export_cpp/templates/kernel_forward/reshape_forward.jinja b/aidge_export_cpp/templates/kernel_forward/reshape_forward.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..f9752bcc85255ba321082fbf5cf599f45b3ab4c4
--- /dev/null
+++ b/aidge_export_cpp/templates/kernel_forward/reshape_forward.jinja
@@ -0,0 +1,6 @@
+{% filter indent(width=4, first=False) %}
+{% include "./_mem_offset.jinja" %}
+reshape_forward<{{name|upper}}_NB_ELTS>
+ ({{in_name[0]}}, {{in_name[1]}}, {{out_name[0]}});
+{% include "./_save_outputs.jinja" %}
+{% endfilter %}