Skip to content
Snippets Groups Projects
Commit 838aa704 authored by Cyril Moineau's avatar Cyril Moineau
Browse files

Update export cpp with new registry methods.

parent 7fef0535
No related branches found
No related tags found
3 merge requests!27v0.2.0,!22v0.4.0,!15Export refactor
Hide whitespace changes
Inline Side-by-side
Showing
with 78 additions and 179 deletions
aidge_export_cpp/export_registry.py
+ 6
− 1
View file @ 838aa704
......@@ -2,9 +2,14 @@ from aidge_core.export_utils import ExportLib
from aidge_export_cpp.utils import ROOT
class ExportLibCpp(ExportLib):
name="export_cpp"
_name="export_cpp"
static_files={
str(ROOT / "static" / "Makefile"): "",
str(ROOT / "static" / "include" / "network" / "typedefs.hpp"): "dnn/include/network",
str(ROOT / "static" / "include" / "network" / "utils.hpp"): "dnn/include/network",
}
# TODO ugly fix for Tensor registration issue...
import aidge_core
aidge_core.register_Tensor(["export_cpp", aidge_core.dtype.float32],
aidge_core.get_key_value_Tensor(["cpu", aidge_core.dtype.float32]))
aidge_export_cpp/kernels/activation.hpp
+ 15
− 15
View file @ 838aa704
......@@ -4,36 +4,36 @@
#include <type_traits>
#include "network/typedefs.hpp"
#include "network/utils.hpp"
#include "network/rescaling.hpp"
#include "kernels/rescaling.hpp"
template<typename Output_T, typename T,
template<typename Output_T, typename T,
typename std::enable_if<std::is_floating_point<T>::value>::type* = nullptr>
__attribute__((always_inline)) inline
Output_T saturate (T value, int32_t /*sat*/)
Output_T saturate (T value, int32_t /*sat*/)
{
return value;
}
template<typename Output_T, typename T,
template<typename Output_T, typename T,
typename std::enable_if<!std::is_floating_point<T>::value>::type* = nullptr>
__attribute__((always_inline)) inline
Output_T saturate (T value, uint32_t sat)
Output_T saturate (T value, uint32_t sat)
{
if (std::is_unsigned<Output_T>::value) {
return clamp(value, T(0), (T(1) << sat) - 1);
} else {
return clamp(value, -(T(1) << (sat - 1)), (T(1) << (sat - 1)) - 1);
}
}
}
template<typename Output_T,
typename Sum_T,
template<typename Output_T,
typename Sum_T,
typename Rescaling_T>
__attribute__((always_inline)) inline
Output_T activation_forward_value (Sum_T weightedSum,
int output,
ActivationFunction_T func,
const Rescaling_T& __restrict rescaling)
__attribute__((always_inline)) inline
Output_T activation_forward_value (Sum_T weightedSum,
int output,
ActivationFunction_T func,
const Rescaling_T& __restrict rescaling)
{
switch(func) {
case Linear:
......@@ -49,7 +49,7 @@ Output_T activation_forward_value (Sum_T weightedSum,
break;
}
// Value fixed here for now but it should be generated by
// Value fixed here for now but it should be generated by
// the export module or determined by the type of Output_T
// For now only works for int8_t and uint8_t
const uint32_t NB_BITS = 8;
......@@ -60,7 +60,7 @@ Output_T activation_forward_value (Sum_T weightedSum,
template<int NB_DATA,
ActivationFunction_T ACTIVATION,
typename Input_T, typename Output_T, typename Rescaling_T>
__attribute__((always_inline)) inline
__attribute__((always_inline)) inline
void activation_forward (
const Input_T* __restrict inputs,
Output_T* __restrict outputs,
......
aidge_export_cpp/kernels/batchnorm.hpp
+ 4
− 4
View file @ 838aa704
......@@ -2,17 +2,17 @@
#define __AIDGE_EXPORT_CPP_KERNELS_BATCHNORM__
#include "network/typedefs.hpp"
#include "network/rescaling.hpp"
#include "kernels/rescaling.hpp"
#include <math.h>
// WARNING: this kernel only works for 32-bits floating point values
template<int NB_OUTPUTS,
template<int NB_OUTPUTS,
int OUTPUTS_HEIGHT, int OUTPUTS_WIDTH,
ActivationFunction_T ACTIVATION,
typename Input_T, typename Output_T,
typename Input_T, typename Output_T,
typename Param_T>
__attribute__((always_inline)) inline
__attribute__((always_inline)) inline
void batchnorm_forward (
const Input_T* __restrict inputs,
Output_T* __restrict outputs,
......
aidge_export_cpp/kernels/convolution.hpp
+ 13
− 13
View file @ 838aa704
......@@ -2,13 +2,13 @@
#define __AIDGE_EXPORT_CPP_KERNELS_CONVOLUTION__
#include "network/typedefs.hpp"
#include "network/rescaling.hpp"
#include "kernels/rescaling.hpp"
#include "network/utils.hpp"
#include "kernels/macs.hpp"
#include "kernels/activation.hpp"
template<int NB_CHANNELS,
template<int NB_CHANNELS,
int CHANNELS_HEIGHT, int CHANNELS_WIDTH,
int NB_OUTPUTS,
int OUTPUTS_HEIGHT, int OUTPUTS_WIDTH,
......@@ -17,10 +17,10 @@ template<int NB_CHANNELS,
int DILATION_Y, int DILATION_X,
int KERNEL_HEIGHT, int KERNEL_WIDTH,
ActivationFunction_T ACTIVATION,
typename Input_T, typename Output_T,
typename Input_T, typename Output_T,
typename Weight_T, typename Bias_T,
typename Rescaling_T>
__attribute__((always_inline)) inline
__attribute__((always_inline)) inline
void convolution_forward(
const Input_T* __restrict inputs,
Output_T* __restrict outputs,
......@@ -28,10 +28,10 @@ void convolution_forward(
const Bias_T* __restrict biases,
const Rescaling_T& __restrict rescaling)
{
constexpr int DILATED_KERNEL_HEIGHT
constexpr int DILATED_KERNEL_HEIGHT
= KERNEL_HEIGHT + (DILATION_Y - 1) * (KERNEL_HEIGHT - 1);
constexpr int DILATED_KERNEL_WIDTH
constexpr int DILATED_KERNEL_WIDTH
= KERNEL_WIDTH + (DILATION_X - 1) * (KERNEL_WIDTH - 1);
constexpr int OUTPUTS_HEIGHT_NOPAD
......@@ -44,7 +44,7 @@ void convolution_forward(
: 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),
: clamp(CHANNELS_HEIGHT + PADDING_Y - (oy * STRIDE_Y),
0, DILATED_KERNEL_HEIGHT);
const int iy = (oy * STRIDE_Y) - PADDING_Y;
......@@ -57,7 +57,7 @@ void convolution_forward(
const int sxMax = (PADDING_X == 0
&& OUTPUTS_WIDTH == OUTPUTS_WIDTH_NOPAD)
? DILATED_KERNEL_WIDTH
: clamp(CHANNELS_WIDTH + PADDING_X - (ox * STRIDE_X),
: clamp(CHANNELS_WIDTH + PADDING_X - (ox * STRIDE_X),
0, DILATED_KERNEL_WIDTH);
const int ix = (ox * STRIDE_X) - PADDING_X;
......@@ -85,8 +85,8 @@ void convolution_forward(
|| sxMax - sxMin == KERNEL_WIDTH))
{
macsOnRange<KERNEL_WIDTH * NB_CHANNELS>(
inputs + iOffset,
weights + wOffset,
inputs + iOffset,
weights + wOffset,
weightedSum);
}
else {
......@@ -100,11 +100,11 @@ void convolution_forward(
int iOffsetInRange = iOffset
+ sx * DILATION_X * NB_CHANNELS;
macsOnRange<NB_CHANNELS>(
// same input line so no wrapping can occur
inputs + iOffsetInRange,
weights + wOffset + sx * NB_CHANNELS,
inputs + iOffsetInRange,
weights + wOffset + sx * NB_CHANNELS,
weightedSum);
}
}
......
aidge_export_cpp/kernels/fullyconnected.hpp
+ 8
− 8
View file @ 838aa704
......@@ -2,20 +2,20 @@
#define __AIDGE_EXPORT_CPP_KERNELS_FULLYCONNECTED__
#include "network/typedefs.hpp"
#include "network/rescaling.hpp"
#include "kernels/rescaling.hpp"
#include "network/utils.hpp"
#include "kernels/macs.hpp"
#include "kernels/activation.hpp"
template<int NB_CHANNELS,
template<int NB_CHANNELS,
int CHANNELS_HEIGHT, int CHANNELS_WIDTH,
int NB_OUTPUTS,
int OUTPUTS_HEIGHT, int OUTPUTS_WIDTH,
ActivationFunction_T ACTIVATION,
typename Input_T, typename Output_T,
typename Input_T, typename Output_T,
typename Weight_T, typename Bias_T,
typename Rescaling_T>
__attribute__((always_inline)) inline
__attribute__((always_inline)) inline
void fullyconnected_forward (
const Input_T* __restrict inputs,
Output_T* __restrict outputs,
......@@ -35,7 +35,7 @@ void fullyconnected_forward (
for (int iy = 0; iy < CHANNELS_HEIGHT; ++iy) {
for (int ix = 0; ix < CHANNELS_WIDTH; ++ix) {
for (int ch = 0; ch < NB_CHANNELS; ++ch) {
weightedSum += inputs[CHANNELS_WIDTH*NB_CHANNELS*iy + NB_CHANNELS*ix + ch]
weightedSum += inputs[CHANNELS_WIDTH*NB_CHANNELS*iy + NB_CHANNELS*ix + ch]
* weights[CHANNELS_HEIGHT*CHANNELS_WIDTH*NB_CHANNELS*och + CHANNELS_HEIGHT*CHANNELS_WIDTH*ch + CHANNELS_HEIGHT*iy + ix];
}
}
......@@ -58,8 +58,8 @@ Here the kernel to use with inputs in NHWC and weights in NHWC
* (iy + CHANNELS_HEIGHT * och);
macsOnRange<NB_CHANNELS * CHANNELS_WIDTH>(
inputs + iOffset,
weights + wOffset,
inputs + iOffset,
weights + wOffset,
weightedSum);
}
......@@ -69,4 +69,4 @@ Here the kernel to use with inputs in NHWC and weights in NHWC
}
#endif // __AIDGE_EXPORT_CPP_KERNELS_FULLYCONNECTED__
\ No newline at end of file
#endif // __AIDGE_EXPORT_CPP_KERNELS_FULLYCONNECTED__
aidge_export_cpp/operators.py
+ 28
− 75
View file @ 838aa704
import os
import numpy as np
from pathlib import Path
from aidge_core.export_utils import ExportNode, ExportNodeCpp, operator_register, generate_file
import aidge_core
from aidge_core.export_utils import ExportNode, ExportNodeCpp, generate_file
from aidge_export_cpp.utils import ROOT
from aidge_export_cpp.utils.converter import numpy_dtype2ctype
from aidge_export_cpp.utils.generation import *
from aidge_export_cpp import ExportLibCpp
##############################################
################### Utils ####################
##############################################
# def get_node_parents(node):
# parents = []
# for parent in node.get_parents():
# if parent.type() != "Producer":
# parents.append(parent)
# return parents
# def get_producer_parents(node):
# parents = []
# for parent in node.get_parents():
# if parent.type() == "Producer":
# parents.append(parent)
# return parents
##############################################
############## Export functions ##############
......@@ -55,10 +35,8 @@ def export_params(name: str,
############## Operators helper ##############
##############################################
@operator_register(ExportLibCpp, "Producer")
@ExportLibCpp.register("Producer", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.any)))
class ProducerCPP(ExportNode):
def __init__(self, node, mem_info, is_input, is_output):
super().__init__(node, mem_info, is_input, is_output)
self.values = np.array(self.operator.get_output(0))
......@@ -77,11 +55,15 @@ class ProducerCPP(ExportNode):
def forward(self):
# A Producer does nothing during forward
return []
@classmethod
def exportable(cls, node):
return True # TODO add check i/o NCHW
@operator_register(ExportLibCpp, "ReLU")
# TODO : find a way to remove this dummy exportnode
@ExportLibCpp.register("Pad2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.any)))
class Pad_ARMCortexM(ExportNodeCpp):
def __init__(self, node, mem_info, is_input, is_output):
raise NotImplementedError("Pad2D nodes is not implemented")
@ExportLibCpp.register("ReLU", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class ReLUCPP(ExportNodeCpp):
def __init__(self, node, mem_info, is_input, is_output):
super().__init__(node, mem_info, is_input, is_output)
......@@ -97,11 +79,7 @@ class ReLUCPP(ExportNodeCpp):
str(ROOT / "kernels" / "rescaling.hpp")
]
@classmethod
def exportable(cls, node):
return True # TODO add check i/o NCHW
@operator_register(ExportLibCpp, "Conv")
@ExportLibCpp.register("Conv2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class ConvCPP(ExportNodeCpp):
def __init__(self, node, mem_info, is_input, is_output):
super().__init__(node, mem_info, is_input, is_output)
......@@ -121,22 +99,17 @@ class ConvCPP(ExportNodeCpp):
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
@classmethod
def exportable(cls, node):
return True # TODO add check i/o NCHW
@operator_register(ExportLibCpp, "PaddedConv")
@ExportLibCpp.register_metaop("PaddedConv2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class PaddedConvCPP(ExportNodeCpp):
def __init__(self, node, mem_info, is_input, is_output):
super().__init__(node, mem_info, is_input, is_output)
# TODO find a way to retrive attr for meta op
for n in self.operator.get_micro_graph().get_nodes():
if n.type() == "Pad":
if n.type() == "Pad2D":
self.attributes["padding"] = n.get_operator(
).attr.begin_end_borders
if n.type() == "Conv":
if n.type() == "Conv2D":
self.attributes["kernel_dims"] = n.get_operator(
).attr.kernel_dims
self.attributes["stride_dims"] = n.get_operator(
......@@ -156,11 +129,8 @@ class PaddedConvCPP(ExportNodeCpp):
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
@classmethod
def exportable(cls, node):
return True # TODO add check i/o NCHW
@operator_register(ExportLibCpp, "Add")
@ExportLibCpp.register("Add", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class AddCPP(ExportNodeCpp):
def __init__(self, node, mem_info, is_input, is_output):
super().__init__(node, mem_info, is_input, is_output)
......@@ -177,11 +147,8 @@ class AddCPP(ExportNodeCpp):
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
@classmethod
def exportable(cls, node):
return True # TODO add check i/o NCHW
@operator_register(ExportLibCpp, "Sub")
@ExportLibCpp.register("Sub", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class SubCPP(ExportNodeCpp):
def __init__(self, node, mem_info, is_input, is_output):
super().__init__(node, mem_info, is_input, is_output)
......@@ -198,12 +165,9 @@ class SubCPP(ExportNodeCpp):
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
@classmethod
def exportable(cls, node):
return True # TODO add check i/o NCHW
@operator_register(ExportLibCpp, "Mul")
@ExportLibCpp.register("Mul", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class MulCPP(ExportNodeCpp):
def __init__(self, node, mem_info, is_input, is_output):
super().__init__(node, mem_info, is_input, is_output)
......@@ -220,11 +184,8 @@ class MulCPP(ExportNodeCpp):
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
@classmethod
def exportable(cls, node):
return True # TODO add check i/o NCHW
@operator_register(ExportLibCpp, "MaxPooling")
@ExportLibCpp.register("MaxPooling2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class MaxPoolCPP(ExportNodeCpp):
def __init__(self, node, mem_info, is_input, is_output):
super().__init__(node, mem_info, is_input, is_output)
......@@ -243,21 +204,19 @@ class MaxPoolCPP(ExportNodeCpp):
self.kernels_to_copy = [
str(ROOT / "kernels" / "pooling.hpp"),
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
@classmethod
def exportable(cls, node):
return True # TODO add check i/o NCHW
@operator_register(ExportLibCpp, "PaddedMaxPooling")
@ExportLibCpp.register_metaop("PaddedMaxPooling2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class PaddedMaxPoolCPP(ExportNodeCpp):
def __init__(self, node, mem_info, is_input, is_output):
super().__init__(node, mem_info, is_input, is_output)
for n in self.operator.get_micro_graph().get_nodes():
if n.type() == "Pad":
if n.type() == "Pad2D":
self.attributes["padding"] = n.get_operator(
).attr.begin_end_borders
if n.type() == "MaxPooling":
if n.type() == "MaxPooling2D":
self.attributes["kernel_dims"] = n.get_operator(
).attr.kernel_dims
self.attributes["stride_dims"] = n.get_operator(
......@@ -273,12 +232,11 @@ class PaddedMaxPoolCPP(ExportNodeCpp):
self.kernels_to_copy = [
str(ROOT / "kernels" / "pooling.hpp"),
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
@classmethod
def exportable(cls, node):
return True # TODO add check i/o NCHW
@operator_register(ExportLibCpp, "GlobalAveragePooling")
@ExportLibCpp.register("GlobalAveragePooling", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class GlobalAveragePoolCPP(ExportNodeCpp):
def __init__(self, node, mem_info, is_input, is_output):
super().__init__(node, mem_info, is_input, is_output)
......@@ -302,12 +260,10 @@ class GlobalAveragePoolCPP(ExportNodeCpp):
self.kernels_to_copy = [
str(ROOT / "kernels" / "pooling.hpp"),
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
@classmethod
def exportable(cls, node):
return True # TODO add check i/o NCHW
@operator_register(ExportLibCpp, "FC")
@ExportLibCpp.register("FC", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class FcCPP(ExportNodeCpp):
def __init__(self, node, mem_info, is_input, is_output):
super().__init__(node, mem_info, is_input, is_output)
......@@ -324,6 +280,3 @@ class FcCPP(ExportNodeCpp):
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
@classmethod
def exportable(cls, node):
return True # TODO add check i/o NCHW
aidge_export_cpp/templates/configuration/activation_config.jinja
+ 1
− 0
View file @ 838aa704
{#- For name header -#}
#ifndef {{ name|upper }}_LAYER_H
#define {{ name|upper }}_LAYER_H
#include "kernels/rescaling.hpp"
{# For layer configuration -#}
{%- set nb_data = in_chan[0] * in_height[0] * in_width[0] %}
......
aidge_export_cpp/templates/configuration/convolution_config.jinja
+ 1
− 1
View file @ 838aa704
{#- For name header -#}
#ifndef {{ name|upper }}_LAYER_H
#define {{ name|upper }}_LAYER_H
#include "kernels/rescaling.hpp"
{# For layer configuration -#}
{% include "./_def_io.jinja" %}
{% include "./_meminfo.jinja" %}
......
aidge_export_cpp/templates/configuration/elemwise_config.jinja
+ 1
− 0
View file @ 838aa704
{#- For name header -#}
#ifndef {{ name|upper }}_LAYER_H
#define {{ name|upper }}_LAYER_H
#include "kernels/rescaling.hpp"
{% include "./_def_io.jinja" %}
{% include "./_meminfo.jinja" %}
......
aidge_export_cpp/templates/configuration/fullyconnected_config.jinja
+ 1
− 1
View file @ 838aa704
{#- For name header -#}
#ifndef {{ name|upper }}_LAYER_H
#define {{ name|upper }}_LAYER_H
#include "kernels/rescaling.hpp"
{# For layer configuration -#}
{% include "./_def_io.jinja" %}
{% include "./_meminfo.jinja" %}
......
aidge_export_cpp/templates/network/environment.jinja deleted 100644 → 0
+ 0
− 25
View file @ 7fef0535
#ifndef ENV_LAYER_H
#define ENV_LAYER_H
#include <stdint.h>
#define ENV_SIZE_X {{ size_x }}
#define ENV_SIZE_Y {{ size_y }}
#define ENV_NB_OUTPUTS {{ nb_outputs }}
#define ENV_DATA_UNSIGNED {{ is_unsigned }}
#define ENV_OUTPUTS_SIZE (ENV_NB_OUTPUTS*ENV_SIZE_X*ENV_SIZE_Y)
#define NETWORK_TARGETS 1
//Output targets network dimension definition:
static unsigned int OUTPUTS_HEIGHT[NETWORK_TARGETS] = {1};
static unsigned int OUTPUTS_WIDTH[NETWORK_TARGETS] = {1};
static unsigned int NB_OUTPUTS[NETWORK_TARGETS] = {1000};
static unsigned int NB_TARGET[NETWORK_TARGETS] = {1000};
static unsigned int OUTPUTS_SIZE[NETWORK_TARGETS] = {(OUTPUTS_WIDTH[0]*OUTPUTS_HEIGHT[0])};
typedef int32_t Target_0_T;
typedef Target_0_T Target_T;
#endif // ENV_LAYER_H
aidge_export_cpp/templates/network/network_forward.jinja deleted 100644 → 0
+ 0
− 36
View file @ 7fef0535
{#- For libraries #}
#include <stdint.h>
#ifdef SAVE_OUTPUTS
#include <sys/types.h>
#include <sys/stat.h>
#endif
#include "network/rescaling.hpp"
// Layer & memory configurations
{%- for header in headers %}
#include "{{ header }}"
{%- endfor %}
{# mem has the datatype of the firt output #}
{#- Change here to improve it -#}
static {{outputs[0][0]}} mem[{{peak_mem}}];
{# Forward function #}
{#- Support multiple inputs with different datatypes and multiple outputs with different datatypes -#}
void model_forward({% for inp in inputs %}const {{inp[0]}}* {{inp[1]}}, {% endfor %}{% for out in outputs %}{{out[0]}}* {{out[1]}}{{ ", " if not loop.last else "" }}{% endfor %})
{
#ifdef SAVE_OUTPUTS
// Creation of the outputs directory
struct stat st {};
if (stat("outputs", &st) == -1) {
mkdir("outputs", 0700);
}
#endif
{%- for action in actions %}
{{ action }}
{%- endfor %}
}
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment

Copyright © Eclipse Foundation, Inc. All Rights Reserved.     Privacy Policy | Terms of Use | Copyright Agent