diff --git a/aidge_export_cpp/kernels/convolution.hpp b/aidge_export_cpp/kernels/convolution.hpp
index 6ea9f0579b84dd5a28a5ea66a778326fcd9c84ce..5855654b39d5d7faf09e81735fbe80fa248ace94 100644
--- a/aidge_export_cpp/kernels/convolution.hpp
+++ b/aidge_export_cpp/kernels/convolution.hpp
@@ -48,7 +48,9 @@ void convolution_forward(
0, DILATED_KERNEL_HEIGHT);
const int iy = (oy * STRIDE_Y) - PADDING_Y;
+#ifdef _OPENMP
#pragma omp parallel for collapse(2)
+#endif
for (int ox = 0; ox < OUTPUTS_WIDTH; ++ox) {
for (int output = 0; output < NB_OUTPUTS; ++output) {
// moved to inner loop for collapsing -->
@@ -62,7 +64,7 @@ void convolution_forward(
const int ix = (ox * STRIDE_X) - PADDING_X;
const int oPos = (ox + OUTPUTS_WIDTH * oy);
- int oOffset = NB_OUTPUTS * oPos;
+ const int oOffset = NB_OUTPUTS * oPos;
// <--
// Check if the biases are defined
@@ -77,7 +79,7 @@ void convolution_forward(
}
const int iPos = ix + CHANNELS_WIDTH * (iy + sy*DILATION_Y);
- int iOffset = NB_CHANNELS * iPos;
+ const int iOffset = NB_CHANNELS * iPos;
const int wOffset = (output*KERNEL_HEIGHT + sy) * KERNEL_WIDTH * NB_CHANNELS;
@@ -98,7 +100,7 @@ void convolution_forward(
continue;
}
- int iOffsetInRange = iOffset
+ const int iOffsetInRange = iOffset
+ sx * DILATION_X * NB_CHANNELS;
macsOnRange<NB_CHANNELS>(
@@ -157,4 +159,158 @@ void convolution_forward(
(inputs, outputs, weights, b, rescaling);
}
+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 Bias_T,
+ typename Rescaling_T>
+__attribute__((always_inline)) inline
+void convolution_depthwise_forward(
+ const Input_T* __restrict inputs,
+ Output_T* __restrict outputs,
+ const Weight_T* __restrict weights,
+ const Bias_T* __restrict biases,
+ const Rescaling_T& __restrict rescaling)
+{
+ static_assert(NB_OUTPUTS % NB_CHANNELS == 0,
+ "NB_OUTPUTS should be a multiple of NB_CHANNELS.");
+
+ constexpr int DILATED_KERNEL_HEIGHT
+ = KERNEL_HEIGHT + (DILATION_Y - 1) * (KERNEL_HEIGHT - 1);
+
+ constexpr int DILATED_KERNEL_WIDTH
+ = KERNEL_WIDTH + (DILATION_X - 1) * (KERNEL_WIDTH - 1);
+
+ constexpr int OUTPUTS_HEIGHT_NOPAD
+ = (CHANNELS_HEIGHT - DILATION_Y * (KERNEL_HEIGHT - 1) - 1 + STRIDE_Y) / STRIDE_Y;
+ constexpr int OUTPUTS_WIDTH_NOPAD
+ = (CHANNELS_WIDTH - DILATION_X * (KERNEL_WIDTH - 1) - 1 + STRIDE_X) / STRIDE_X;
+
+ for (int oy = 0; oy < OUTPUTS_HEIGHT; ++oy) {
+ const int syMin = (PADDING_Y == 0) ? 0
+ : max(PADDING_Y - (oy * STRIDE_Y), 0);
+ const int syMax = (PADDING_Y == 0
+ && OUTPUTS_HEIGHT == OUTPUTS_HEIGHT_NOPAD) ? DILATED_KERNEL_HEIGHT
+ : clamp(CHANNELS_HEIGHT + PADDING_Y - (oy * STRIDE_Y),
+ 0, DILATED_KERNEL_HEIGHT);
+ const int iy = (oy * STRIDE_Y) - PADDING_Y;
+
+#ifdef _OPENMP
+#pragma omp parallel for collapse(2)
+#endif
+ for (int ox = 0; ox < OUTPUTS_WIDTH; ++ox) {
+ for (int output = 0; output < NB_OUTPUTS; ++output) {
+ // moved to inner loop for collapsing -->
+ const int sxMin = (PADDING_X == 0) ? 0
+ : max(PADDING_X - (ox * STRIDE_X), 0);
+ const int sxMax = (PADDING_X == 0
+ && OUTPUTS_WIDTH == OUTPUTS_WIDTH_NOPAD)
+ ? DILATED_KERNEL_WIDTH
+ : clamp(CHANNELS_WIDTH + PADDING_X - (ox * STRIDE_X),
+ 0, DILATED_KERNEL_WIDTH);
+ const int ix = (ox * STRIDE_X) - PADDING_X;
+
+ const int oPos = (ox + OUTPUTS_WIDTH * oy);
+ const int oOffset = NB_OUTPUTS * oPos;
+ // <--
+
+ const int channel = (output * NB_CHANNELS) / NB_OUTPUTS;
+
+ Bias_T weightedSum = biases ? biases[output] : 0;
+
+ for (int sy = 0; sy < KERNEL_HEIGHT; ++sy) {
+ if ((PADDING_Y != 0
+ || OUTPUTS_HEIGHT != OUTPUTS_HEIGHT_NOPAD)
+ && ((sy*DILATION_Y < syMin) || (sy*DILATION_Y >= syMax)))
+ {
+ continue;
+ }
+
+ const int iPos = ix + CHANNELS_WIDTH * (iy + sy*DILATION_Y);
+ const int iOffset = NB_CHANNELS * iPos;
+
+ const int wOffset = (output*KERNEL_HEIGHT + sy)
+ * KERNEL_WIDTH;
+
+ if (DILATION_X == 1 && ((PADDING_X == 0
+ && OUTPUTS_WIDTH == OUTPUTS_WIDTH_NOPAD)
+ || sxMax - sxMin == KERNEL_WIDTH))
+ {
+ macsOnRange<KERNEL_WIDTH, NB_CHANNELS>(
+ inputs + iOffset + channel,
+ weights + wOffset,
+ weightedSum);
+ }
+ else {
+ for (int sx = 0; sx < KERNEL_WIDTH; ++sx) {
+ if ((PADDING_X != 0
+ || OUTPUTS_WIDTH != OUTPUTS_WIDTH_NOPAD)
+ && ((sx*DILATION_X < sxMin) || (sx*DILATION_X >= sxMax)))
+ {
+ continue;
+ }
+
+ const int iOffsetInRange = iOffset
+ + sx * DILATION_X * NB_CHANNELS;
+
+ weightedSum += inputs[iOffsetInRange + channel]
+ * weights[wOffset + sx];
+ }
+ }
+ }
+
+ outputs[oOffset + output] = activation_forward_value<Output_T>(weightedSum, output, ACTIVATION, rescaling);
+ }
+ }
+ }
+}
+
+// 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_depthwise_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_depthwise_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/fullyconnected.hpp b/aidge_export_cpp/kernels/fullyconnected.hpp
index 2780de2deadd9c519597627e64eb5bedeee948f8..60805e7b90fa29ba00c6736bb8771985aeca19b4 100644
--- a/aidge_export_cpp/kernels/fullyconnected.hpp
+++ b/aidge_export_cpp/kernels/fullyconnected.hpp
@@ -28,6 +28,9 @@ void fullyconnected_forward (
// It is only an issue if the FC was after a flatten layer.
// Otherwise it is not an issue for the other FC because CHANNELS_WIDTH = CHANNELS_HEIGHT = 1
// Solution: Add a system to check dataformat
+#ifdef _OPENMP
+#pragma omp parallel for
+#endif
for (int och = 0; och < NB_OUTPUTS; och++) {
Bias_T weightedSum = (biases) ? biases[och] : Bias_T(0);
@@ -45,7 +48,9 @@ void fullyconnected_forward (
}
/*
Here the kernel to use with inputs in NHWC and weights in NHWC
+#ifdef _OPENMP
#pragma omp parallel for
+#endif
for (int och = 0; och < NB_OUTPUTS; och++) {
Bias_T weightedSum = (biases) ? biases[och] : Bias_T(0);
diff --git a/aidge_export_cpp/kernels/leakyrelu.hpp b/aidge_export_cpp/kernels/leakyrelu.hpp
index 07352cd2bf8d73eb1bb3afdcca381fcec4729bbd..5e6598d8fe5d43d9ae9320498289577ab7695e97 100644
--- a/aidge_export_cpp/kernels/leakyrelu.hpp
+++ b/aidge_export_cpp/kernels/leakyrelu.hpp
@@ -11,7 +11,9 @@ void leakyrelu_forward (
Output_T* __restrict outputs,
const float negative_slope)
{
+#ifdef _OPENMP
#pragma omp parallel for
+#endif
for (int i = 0; i < NB_DATA; ++i) {
if (inputs[i] >= 0) {
outputs[i] = inputs[i];
diff --git a/aidge_export_cpp/kernels/pooling.hpp b/aidge_export_cpp/kernels/pooling.hpp
index a86fd4196a9f6e19f45dbdc4f1035c1e94e7d285..30fa766abbeded7eb55caf01902c216d95a2ed17 100644
--- a/aidge_export_cpp/kernels/pooling.hpp
+++ b/aidge_export_cpp/kernels/pooling.hpp
@@ -36,7 +36,9 @@ void pooling_forward(
0, POOL_HEIGHT);
const int iy = (oy * STRIDE_Y) - PADDING_Y;
+#ifdef _OPENMP
#pragma omp parallel for collapse(2)
+#endif
for (int ox = 0; ox < OUTPUTS_WIDTH; ++ox) {
for (int output = 0; output < NB_OUTPUTS; ++output) {
// moved to inner loop for collapsing -->
diff --git a/aidge_export_cpp/operators.py b/aidge_export_cpp/operators.py
index 26ca62155401707573d9625ad91a9b63cb1b4d2b..cb7a09c9fa01eee433f2cbcdb09cc66dc0bb17a6 100644
--- a/aidge_export_cpp/operators.py
+++ b/aidge_export_cpp/operators.py
@@ -187,6 +187,39 @@ class PaddedConvCPP(ExportNodeCpp):
_setup_conv2D(self)
+@ExportLibCpp.register("ConvDepthWise2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
+class ConvDepthWiseCPP(ExportNodeCpp):
+ def __init__(self, node, mem_info):
+ super().__init__(node, mem_info)
+ self.attributes["depthwise"] = True
+
+ # No padding with Conv
+ # Use PaddedConv to add padding attribute
+ self.attributes["padding"] = [0, 0]
+
+ _setup_conv2D(self)
+
+@ExportLibCpp.register_metaop("PaddedConvDepthWise2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
+class PaddedConvDepthWiseCPP(ExportNodeCpp):
+ def __init__(self, node, mem_info):
+ super().__init__(node, mem_info)
+ self.attributes["depthwise"] = True
+
+ # TODO find a way to retrive attr for meta op
+ for n in self.operator.get_micro_graph().get_nodes():
+ if n.type() == "Pad2D":
+ self.attributes["padding"] = n.get_operator(
+ ).attr.begin_end_borders
+ if n.type() == "ConvDepthWise2D":
+ self.attributes["kernel_dims"] = n.get_operator(
+ ).attr.kernel_dims
+ self.attributes["stride_dims"] = n.get_operator(
+ ).attr.stride_dims
+ self.attributes["dilation_dims"] = n.get_operator(
+ ).attr.dilation_dims
+
+ _setup_conv2D(self)
+
def _setup_elemwise_op(elemwise, op):
"""Common code (template and kernel setup) shared across all the different elementWise operator (Add, Sub,...)."""
diff --git a/aidge_export_cpp/templates/kernel_forward/convolution_forward.jinja b/aidge_export_cpp/templates/kernel_forward/convolution_forward.jinja
index 421013b9590dabe6ee0ac12f969494913414a530..7d0af8c6f75df47825e67a8b47258c3f8469fc6a 100644
--- a/aidge_export_cpp/templates/kernel_forward/convolution_forward.jinja
+++ b/aidge_export_cpp/templates/kernel_forward/convolution_forward.jinja
@@ -1,6 +1,6 @@
{% filter indent(width=4, first=False) %}
{% include "./_mem_offset.jinja" %}
-convolution_forward<{{ in_name[0]|upper }}_NB_CHANNELS,
+convolution{{ "_depthwise" if depthwise is defined else "" }}_forward<{{ in_name[0]|upper }}_NB_CHANNELS,
{{ in_name[0]|upper }}_IN_HEIGHT,
{{ in_name[0]|upper }}_IN_WIDTH,
{{ out_name[0]|upper }}_NB_OUTPUTS,
diff --git a/aidge_export_cpp/unit_tests/test_export.py b/aidge_export_cpp/unit_tests/test_export.py
index 607778d23deda862db73f5908fd1caa6ccc1d95b..65db07e33c68e4900476f97d92879ad7ca3ed4e2 100644
--- a/aidge_export_cpp/unit_tests/test_export.py
+++ b/aidge_export_cpp/unit_tests/test_export.py
@@ -410,6 +410,14 @@ class test_operator_export(unittest.TestCase):
self.unit_test_export(model, "Conv2D", [[1, 3, 12, 12]], False, False)
+ def test_export_convDepthWise2D(self):
+ print("ConvDepthWise2D")
+ model = aidge_core.sequential([
+ aidge_core.ConvDepthWise2D(nb_channels=3, kernel_dims=(3, 3), name="conv")
+ ])
+
+ self.unit_test_export(model, "ConvDepthWise2D", [[1, 3, 12, 12]], False, False)
+
def test_export_max_pooling(self):
print("MaxPooling2D")
model = aidge_core.sequential([