From 29a0fec1af61ea5fb77cf63e5ebf36b0d6273e82 Mon Sep 17 00:00:00 2001
From: cmoineau <cyril.moineau@cea.fr>
Date: Thu, 12 Dec 2024 09:38:50 +0000
Subject: [PATCH 01/22] Update export_cpp with
https://gitlab.eclipse.org/eclipse/aidge/aidge_core/-/merge_requests/277
---
aidge_export_cpp/__init__.py | 2 +-
aidge_export_cpp/utils/__init__.py | 25 ++++++-------------------
pyproject.toml | 17 +++++++++++------
setup.cfg | 4 ++++
4 files changed, 22 insertions(+), 26 deletions(-)
create mode 100644 setup.cfg
diff --git a/aidge_export_cpp/__init__.py b/aidge_export_cpp/__init__.py
index 99df130..16c9be6 100644
--- a/aidge_export_cpp/__init__.py
+++ b/aidge_export_cpp/__init__.py
@@ -10,7 +10,7 @@ import aidge_core
from aidge_export_cpp.utils import ROOT
-from ._version import *
+# from ._version import *
from .export import *
diff --git a/aidge_export_cpp/utils/__init__.py b/aidge_export_cpp/utils/__init__.py
index 0728388..5b15131 100644
--- a/aidge_export_cpp/utils/__init__.py
+++ b/aidge_export_cpp/utils/__init__.py
@@ -1,27 +1,14 @@
from pathlib import Path
-import os
+from importlib.metadata import version
# Constants
FILE = Path(__file__).resolve()
ROOT = FILE.parents[1]
-OPERATORS_REGISTRY = {}
+def show_version():
+ version_aidge_export_cpp = version("aidge_export_cpp")
+ print(f"Aidge Export CPP: {version_aidge_export_cpp}")
-def operator_register(*args):
-
- key_list = [arg for arg in args]
-
- def decorator(operator):
- class Wrapper(operator):
- def __init__(self, *args, **kwargs):
- return operator(*args, **kwargs)
-
- for key in key_list:
- OPERATORS_REGISTRY[key] = operator
-
- return Wrapper
- return decorator
-
-def supported_operators():
- return list(OPERATORS_REGISTRY.keys())
+def get_project_version()->str:
+ return version("aidge_export_cpp")
diff --git a/pyproject.toml b/pyproject.toml
index 870f193..b3d85aa 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -9,7 +9,7 @@ dependencies = [
requires-python = ">= 3.7"
readme = "README.md"
license = { file = "LICENSE" }
-classifiers = [
+classifiers = [
"Development Status :: 2 - Pre-Alpha",
"Intended Audience :: Developers",
"Intended Audience :: Education",
@@ -26,7 +26,14 @@ classifiers = [
"Topic :: Scientific/Engineering :: Artificial Intelligence",
"Topic :: Software Development"
]
-dynamic = ["version"] # defined in tool.setuptools_scm
+dynamic = ["version"] # defined by pbr
+
+[project.urls]
+Homepage = "https://www.deepgreen.ai/en/platform"
+Documentation = "https://eclipse-aidge.readthedocs.io/en/latest/"
+Repository = "https://gitlab.eclipse.org/eclipse/aidge/aidge_export_cpp"
+Issues = "https://gitlab.eclipse.org/eclipse/aidge/aidge_export_cpp/-/issues/"
+Changelog = "https://gitlab.eclipse.org/eclipse/aidge/aidge_export_cpp/-/releases"
[project.optional-dependencies]
test = ["pytest"]
@@ -34,7 +41,7 @@ test = ["pytest"]
[build-system]
requires = [
"setuptools>=64",
- "setuptools_scm[toml]==7.1.0"
+ "pbr"
]
build-backend = "setuptools.build_meta"
@@ -47,6 +54,4 @@ include = ["aidge_export_cpp"] # package names should match these glob patterns
namespaces = false # to disable scanning PEP 420 namespaces (true by default)
[tool.setuptools.exclude-package-data]
aidge_export_cpp = ["unit_tests*"] # exclude unit_tests which may be included as data
-# SETUPTOOLS_SCM
-[tool.setuptools_scm]
-write_to = "aidge_export_cpp/_version.py"
+
diff --git a/setup.cfg b/setup.cfg
new file mode 100644
index 0000000..bb5e124
--- /dev/null
+++ b/setup.cfg
@@ -0,0 +1,4 @@
+
+# pbr file
+[metadata]
+version = file: version.txt
--
GitLab
From d7d6330bdfadb2e09fe541ddff1f8c9315e679c3 Mon Sep 17 00:00:00 2001
From: cmoineau <cyril.moineau@cea.fr>
Date: Thu, 2 Jan 2025 14:07:54 +0000
Subject: [PATCH 02/22] Fix import path + remove deprecated functions.
---
aidge_export_cpp/__init__.py | 8 +-
aidge_export_cpp/export.py | 128 +-----------------
aidge_export_cpp/export_registry.py | 2 +-
aidge_export_cpp/operators.py | 17 ++-
.../{utils/__init__.py => utils.py} | 2 +-
aidge_export_cpp/utils/converter.py | 18 ---
aidge_export_cpp/utils/generation.py | 51 -------
7 files changed, 25 insertions(+), 201 deletions(-)
rename aidge_export_cpp/{utils/__init__.py => utils.py} (93%)
delete mode 100644 aidge_export_cpp/utils/converter.py
delete mode 100644 aidge_export_cpp/utils/generation.py
diff --git a/aidge_export_cpp/__init__.py b/aidge_export_cpp/__init__.py
index 16c9be6..4eff598 100644
--- a/aidge_export_cpp/__init__.py
+++ b/aidge_export_cpp/__init__.py
@@ -2,15 +2,9 @@ r"""
Aidge Export for CPP standalone projects
"""
+from .utils import ROOT
from .export_registry import ExportLibCpp
-
from .operators import *
from collections import defaultdict
-import aidge_core
-
-from aidge_export_cpp.utils import ROOT
-
-# from ._version import *
-
from .export import *
diff --git a/aidge_export_cpp/export.py b/aidge_export_cpp/export.py
index ebac7a8..42bf90f 100644
--- a/aidge_export_cpp/export.py
+++ b/aidge_export_cpp/export.py
@@ -1,131 +1,15 @@
-import re
-import os
-import numpy as np
-
import aidge_core
-
-from aidge_core.export_utils.code_generation import *
-from aidge_core.mem_info import compute_default_mem_info
-
-from aidge_export_cpp.utils import ROOT
-from aidge_export_cpp.utils.converter import numpy_dtype2ctype
from aidge_export_cpp import ExportLibCpp
-from aidge_export_cpp.utils.generation import *
-# from aidge_export_cpp.memory import *
-
-
-def generate_input_file(export_folder:str,
- array_name:str,
- array: np.ndarray):
-
- # If directory doesn't exist, create it
- if not os.path.exists(export_folder):
- os.makedirs(export_folder)
-
- generate_file(
- file_path=f"{export_folder}/{array_name}.h",
- template_path=str(ROOT / "templates" / "data" / "inputs.jinja"),
- dims = array.shape,
- data_t = numpy_dtype2ctype(array.dtype),
- name = array_name,
- values = array.tolist()
- )
-
def export(export_folder_name, graphview, scheduler, mem_wrapping=False):
+ print("Warning: This function is deprecated, check tutorial https://eclipse.dev/aidge/source/Tutorial/export_cpp.html to find the new way to generate a C++ export.")
aidge_core.export_utils.scheduler_export(
scheduler,
export_folder_name,
ExportLibCpp,
- memory_manager=compute_default_mem_info
+ memory_manager=aidge_core.mem_info.generate_optimized_memory_info,
+ memory_manager_args={
+ "stats_folder": f"{export_folder_name}/stats",
+ "wrapping": mem_wrapping
+ }
)
-
- # export_folder = Path().absolute() / export_folder_name
-
- # os.makedirs(str(export_folder), exist_ok=True)
-
- # dnn_folder = export_folder / "dnn"
- # os.makedirs(str(dnn_folder), exist_ok=True)
-
- # list_actions = []
- # list_configs = []
- # peak_mem, mem_info = compute_default_mem_info(scheduler)
- # list_forward_nodes = scheduler.get_static_scheduling()
-
- # for node in list_forward_nodes:
- # if ExportLibCpp.exportable(node):
- # op = ExportLibCpp.get_export_node(node)(node, mem_info[node])
- # # For configuration files
- # list_configs = op.export(dnn_folder, list_configs)
-
- # # For forward file
- # list_actions = op.forward(list_actions)
- # else:
- # raise RuntimeError(f"Operator not supported: {node.type()} !")
-
- # # Memory management
- # # stats_folder = export_folder / "statistics"
- # # os.makedirs(str(stats_folder), exist_ok=True)
- # # mem_size, mem_info = generate_optimized_memory_info(stats_folder, scheduler, mem_wrapping)
- # # peak_mem, mem_info = compute_default_mem_info(scheduler)
-
- # # Generate the memory file
- # # generate_file(
- # # str(dnn_folder / "memory" / "mem_info.h"),
- # # str(ROOT / "templates" / "memory" / "mem_info.jinja"),
- # # mem_size = mem_size,
- # # mem_info_legends = MEMORY_INFO_TEMPLATE,
- # # mem_info = mem_info
- # # )
- # # list_configs.append("memory/mem_info.h")
-
- # # Get entry nodes
- # # Store the datatype & name
- # list_inputs_name = []
- # for node in graphview.get_input_nodes():
- # for idx, node_input_tuple in enumerate(node.inputs()):
- # node_input, _ = node_input_tuple
- # if node_input is None:
- # export_type = aidge2c(node.get_operator().get_output(0).dtype())
- # list_inputs_name.append((export_type, f"{node.name()}_input_{idx}"))
- # elif node_input not in graphview.get_nodes():
- # export_type = aidge2c(node_input.get_operator().get_output(0).dtype())
- # list_inputs_name.append((export_type, node_input.name()))
-
-
- # # Get output nodes
- # # Store the datatype & name, like entry nodes
- # list_outputs_name = []
- # for node in graphview.get_nodes():
- # if len(node.get_children()) == 0:
- # export_type = aidge2c(node.get_operator().get_output(0).dtype())
- # list_outputs_name.append((export_type, f"{node.name()}_output_0"))
-
- # # Generate forward file
- # # TODO: for now the mem type is bound for all intermediate results, should change.
- # # Note that we may have all inputs constants, hence select output type
- # assert len(list_outputs_name) >= 1, f"TODO: requires some output to determine mem type"
- # mem_ctype = list_outputs_name[0][0]
- # generate_file(
- # str(dnn_folder / "src" / "forward.cpp"),
- # str(ROOT / "templates" / "network" / "network_forward.jinja"),
- # headers=set(list_configs),
- # actions=list_actions,
- # inputs= list_inputs_name,
- # outputs=list_outputs_name,
- # mem_ctype=mem_ctype,
- # peak_mem=peak_mem
- # )
-
- # # Generate dnn API
- # generate_file(
- # str(dnn_folder / "include" / "dnn.hpp"),
- # str(ROOT / "templates" / "network" / "dnn_header.jinja"),
- # libraries=[],
- # functions=get_functions_from_c_file(str(dnn_folder / "src" / "forward.cpp")),
- # )
-
- # # Copy all static files in the export
- # shutil.copy(str(ROOT / "static" / "main.cpp"), str(export_folder))
- # shutil.copy(str(ROOT / "static" / "Makefile"), str(export_folder))
- # shutil.copytree(str(ROOT / "static" / "include"), str(dnn_folder / "include"), dirs_exist_ok=True)
diff --git a/aidge_export_cpp/export_registry.py b/aidge_export_cpp/export_registry.py
index f1aa83b..876e4ff 100644
--- a/aidge_export_cpp/export_registry.py
+++ b/aidge_export_cpp/export_registry.py
@@ -1,5 +1,5 @@
from aidge_core.export_utils import ExportLib
-from aidge_export_cpp.utils import ROOT
+from aidge_export_cpp import ROOT
class ExportLibCpp(ExportLib):
_name="export_cpp"
diff --git a/aidge_export_cpp/operators.py b/aidge_export_cpp/operators.py
index 9654a20..f04dbb3 100644
--- a/aidge_export_cpp/operators.py
+++ b/aidge_export_cpp/operators.py
@@ -4,12 +4,27 @@ from pathlib import Path
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 import ExportLibCpp
##############################################
############## Export functions ##############
##############################################
+def numpy_dtype2ctype(dtype):
+ if dtype == np.int8:
+ return "int8_t"
+ elif dtype == np.int16:
+ return "int16_t"
+ elif dtype == np.int32:
+ return "int32_t"
+ elif dtype == np.int64:
+ return "int64_t"
+ elif dtype == np.float32:
+ return "float"
+ elif dtype == np.float64:
+ return "double"
+ # Add more dtype mappings as needed
+ else:
+ raise ValueError(f"Unsupported {dtype} dtype")
def export_params(name: str,
array: np.ndarray,
diff --git a/aidge_export_cpp/utils/__init__.py b/aidge_export_cpp/utils.py
similarity index 93%
rename from aidge_export_cpp/utils/__init__.py
rename to aidge_export_cpp/utils.py
index 5b15131..915c2c6 100644
--- a/aidge_export_cpp/utils/__init__.py
+++ b/aidge_export_cpp/utils.py
@@ -3,7 +3,7 @@ from importlib.metadata import version
# Constants
FILE = Path(__file__).resolve()
-ROOT = FILE.parents[1]
+ROOT = FILE.parents[0]
def show_version():
diff --git a/aidge_export_cpp/utils/converter.py b/aidge_export_cpp/utils/converter.py
deleted file mode 100644
index d4af124..0000000
--- a/aidge_export_cpp/utils/converter.py
+++ /dev/null
@@ -1,18 +0,0 @@
-import numpy as np
-
-def numpy_dtype2ctype(dtype):
- if dtype == np.int8:
- return "int8_t"
- elif dtype == np.int16:
- return "int16_t"
- elif dtype == np.int32:
- return "int32_t"
- elif dtype == np.int64:
- return "int64_t"
- elif dtype == np.float32:
- return "float"
- elif dtype == np.float64:
- return "double"
- # Add more dtype mappings as needed
- else:
- raise ValueError(f"Unsupported {dtype} dtype")
diff --git a/aidge_export_cpp/utils/generation.py b/aidge_export_cpp/utils/generation.py
deleted file mode 100644
index 4478ef7..0000000
--- a/aidge_export_cpp/utils/generation.py
+++ /dev/null
@@ -1,51 +0,0 @@
-import re
-import os
-import shutil
-from jinja2 import Environment, FileSystemLoader
-
-
-def get_functions_from_c_file(file_path):
- functions = []
- pattern = r'\w+\s+(\w+)\s*\(([^)]*)\)\s*{'
-
- with open(file_path, 'r') as file:
- file_content = file.read()
-
- matches = re.findall(pattern, file_content)
- for match in matches:
- function_name = match[0]
- arguments = match[1].split(',')
- arguments = [arg.strip() for arg in arguments]
-
- return_type = get_return_type(file_content, function_name)
-
- function_string = f"{return_type} {function_name}({', '.join(arguments)});"
- functions.append(function_string)
-
- return functions
-
-
-def get_return_type(file_content, function_name):
- pattern = rf'\w+\s+{function_name}\s*\([^)]*\)\s*{{'
- return_type = re.search(pattern, file_content).group()
- return_type = return_type.split()[0].strip()
- return return_type
-
-
-def get_functions_from_c_folder(folder_path):
- functions = []
-
- for _, _, files in os.walk(folder_path):
- for file in files:
- functions += get_functions_from_c_file(os.path.join(folder_path, file))
-
- return functions
-
-
-def copyfile(filename, dst_folder):
-
- # If directory doesn't exist, create it
- if not os.path.exists(dst_folder):
- os.makedirs(dst_folder)
-
- shutil.copy(filename, dst_folder)
--
GitLab
From 847a27b15d16bf6b62964c070573f6e4761194ea Mon Sep 17 00:00:00 2001
From: Olivier BICHLER <olivier.bichler@cea.fr>
Date: Sun, 19 Jan 2025 15:48:57 +0100
Subject: [PATCH 03/22] Hotfix: source files were not included anymore
---
pyproject.toml | 2 ++
1 file changed, 2 insertions(+)
diff --git a/pyproject.toml b/pyproject.toml
index b3d85aa..25fb9d3 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -52,6 +52,8 @@ build-backend = "setuptools.build_meta"
where = ["."] # list of folders that contain the packages (["."] by default)
include = ["aidge_export_cpp"] # package names should match these glob patterns (["*"] by default)
namespaces = false # to disable scanning PEP 420 namespaces (true by default)
+[tool.setuptools.package-data]
+'aidge_export_cpp' = ['**/*']
[tool.setuptools.exclude-package-data]
aidge_export_cpp = ["unit_tests*"] # exclude unit_tests which may be included as data
--
GitLab
From 4b92baa187381698f8cee9619b19cebf31a40a63 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Wed, 12 Feb 2025 14:17:37 +0000
Subject: [PATCH 04/22] Upload New File
---
aidge_export_cpp/kernels/erf.hpp | 40 ++++++++++++++++++++++++++++++++
1 file changed, 40 insertions(+)
create mode 100644 aidge_export_cpp/kernels/erf.hpp
diff --git a/aidge_export_cpp/kernels/erf.hpp b/aidge_export_cpp/kernels/erf.hpp
new file mode 100644
index 0000000..b509133
--- /dev/null
+++ b/aidge_export_cpp/kernels/erf.hpp
@@ -0,0 +1,40 @@
+#ifndef __AIDGE_EXPORT_CPP_KERNELS_ERP__
+#define __AIDGE_EXPORT_CPP_KERNELS_ERP__
+
+#include "network/typedefs.hpp"
+#include <cmath>
+#include <math.h>
+#include <iostream>
+
+template<int _NB_ELTS,
+ typename Input_T, typename Output_T>
+__attribute__((always_inline)) inline
+void erf_forward (
+ const Input_T* __restrict inputs,
+ Output_T* __restrict outputs)
+{
+ double a1 = 0.254829592;
+ double a2 = -0.284496736;
+ double a3 = 1.421413741;
+ double a4 = -1.453152027;
+ double a5 = 1.061405429;
+ double p = 0.3275911;
+
+
+#pragma omp parallel for
+ for (int i = 0; i < _NB_ELTS; ++i) {
+ int sign = 1;
+ if (inputs[i] < 0)
+ sign = -1;
+ double abs_value = abs(inputs[i]);
+
+ // A&S formula 7.1.26
+ double t = 1.0/(1.0 + p*abs_value);
+ double y = 1.0 - (((((a5*t + a4)*t) + a3)*t + a2)*t + a1)*t*exp(-abs_value*abs_value);
+ outputs[i] = sign*y;
+
+ }
+}
+
+
+#endif // __AIDGE_EXPORT_CPP_KERNELS_ERP_
\ No newline at end of file
--
GitLab
From 9b6fb714ff531a47da042fce67ef28cd864418ae Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Wed, 12 Feb 2025 14:21:31 +0000
Subject: [PATCH 05/22] Upload New File
---
aidge_export_cpp/kernels/transpose.hpp | 56 ++++++++++++++++++++++++++
1 file changed, 56 insertions(+)
create mode 100644 aidge_export_cpp/kernels/transpose.hpp
diff --git a/aidge_export_cpp/kernels/transpose.hpp b/aidge_export_cpp/kernels/transpose.hpp
new file mode 100644
index 0000000..a6ddce2
--- /dev/null
+++ b/aidge_export_cpp/kernels/transpose.hpp
@@ -0,0 +1,56 @@
+#ifndef __AIDGE_EXPORT_CPP_KERNELS_TRANSPOSE__
+#define __AIDGE_EXPORT_CPP_KERNELS_TRANSPOSE__
+
+#include "network/typedefs.hpp"
+#include <cstring>
+#include <cstdio>
+#include <iostream>
+
+using namespace std;
+
+template< int INPUT_DIMS[], int PERM[], int OUTPUT_DIMS[],
+ int SIZE_OUTPUT_DIMS, int SIZE,
+ typename Input_T, typename Output_T>
+__attribute__((always_inline)) inline
+void transpose_forward (
+ const Input_T* __restrict inputs,
+ Output_T* __restrict outputs)
+ {
+
+ int newStrides[SIZE_OUTPUT_DIMS];
+ for (int i = 0; i<SIZE_OUTPUT_DIMS;++i){newStrides[i] = 1;}
+ for (int i = 0; i < SIZE_OUTPUT_DIMS; ++i) {
+ for (int j = i + 1; j < SIZE_OUTPUT_DIMS; ++j) {
+ newStrides[i] *= OUTPUT_DIMS[j];
+ }
+ }
+
+ int indices[SIZE_OUTPUT_DIMS];
+ for (int i = 0; i<SIZE_OUTPUT_DIMS;++i){indices[i] = 0;}
+
+ for (int i = 0; i < SIZE; ++i) {
+ int idx = 0;
+ for (int j = SIZE_OUTPUT_DIMS -1; j >=0; --j) {
+ idx += indices[PERM[j]] * newStrides[j];
+ }
+
+ outputs[idx] = inputs[i];
+
+
+ for (int j = SIZE_OUTPUT_DIMS - 1; j >= 0; --j) {
+ if (indices[j] < INPUT_DIMS[j] - 1) {
+ indices[j]++;
+ break;
+ }
+ else {
+ indices[j] = 0;
+ }
+ }
+ }
+
+
+}
+
+
+
+#endif // __AIDGE_EXPORT_CPP_KERNELS_TRANSPOSE__
\ No newline at end of file
--
GitLab
From 1fd2f8da29e2b797bbc9e036cfff5e869b6fa52c Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Tue, 25 Feb 2025 12:56:48 +0000
Subject: [PATCH 06/22] Upload Erf jinja forward
---
aidge_export_cpp/templates/kernel_forward/erf_forward.jinja | 6 ++++++
1 file changed, 6 insertions(+)
create mode 100644 aidge_export_cpp/templates/kernel_forward/erf_forward.jinja
diff --git a/aidge_export_cpp/templates/kernel_forward/erf_forward.jinja b/aidge_export_cpp/templates/kernel_forward/erf_forward.jinja
new file mode 100644
index 0000000..9f3fbf3
--- /dev/null
+++ b/aidge_export_cpp/templates/kernel_forward/erf_forward.jinja
@@ -0,0 +1,6 @@
+{% filter indent(width=4, first=False) %}
+{% include "./_mem_offset.jinja" %}
+erf_forward<{{name|upper}}_NB_ELTS>
+ ({{in_name[0]}}, {{out_name[0]}});
+{% include "./_save_outputs.jinja" %}
+{% endfilter %}
\ No newline at end of file
--
GitLab
From 0264ce4b97ff330aea589f650a10b30c07bccf18 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Tue, 25 Feb 2025 12:57:58 +0000
Subject: [PATCH 07/22] Upload Transpose Forward File
---
.../templates/kernel_forward/transpose_forward.jinja | 11 +++++++++++
1 file changed, 11 insertions(+)
create mode 100644 aidge_export_cpp/templates/kernel_forward/transpose_forward.jinja
diff --git a/aidge_export_cpp/templates/kernel_forward/transpose_forward.jinja b/aidge_export_cpp/templates/kernel_forward/transpose_forward.jinja
new file mode 100644
index 0000000..2f8d939
--- /dev/null
+++ b/aidge_export_cpp/templates/kernel_forward/transpose_forward.jinja
@@ -0,0 +1,11 @@
+{% filter indent(width=4, first=False) %}
+{% include "./_mem_offset.jinja" %}
+transpose_forward<{{ name|upper }}_INPUT_DIMS,
+ {{ name|upper }}_PERM,
+ {{ name|upper}}_OUTPUT_DIMS,
+ {{ name|upper}}_SIZE_OUTPUT_DIMS,
+ {{name|upper}}_SIZE>
+ ({{in_name[0]}}, {{out_name[0]}});
+{% include "./_save_outputs.jinja" %}
+{% endfilter %}
+
--
GitLab
From bb72026146e36ecdf51188f2f9d047dcfd6c4329 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Wed, 26 Feb 2025 14:27:14 +0000
Subject: [PATCH 08/22] Adding erf and transpose config
---
.../templates/configuration/erf_config.jinja | 11 +++++++++++
.../configuration/transpose_config.jinja | 17 +++++++++++++++++
2 files changed, 28 insertions(+)
create mode 100644 aidge_export_cpp/templates/configuration/erf_config.jinja
create mode 100644 aidge_export_cpp/templates/configuration/transpose_config.jinja
diff --git a/aidge_export_cpp/templates/configuration/erf_config.jinja b/aidge_export_cpp/templates/configuration/erf_config.jinja
new file mode 100644
index 0000000..b273472
--- /dev/null
+++ b/aidge_export_cpp/templates/configuration/erf_config.jinja
@@ -0,0 +1,11 @@
+{#- For name header -#}
+#ifndef {{ name|upper }}_LAYER_H
+#define {{ name|upper }}_LAYER_H
+
+{# For layer configuration -#}
+{# For layer configuration -#}
+{% include "./_def_io.jinja" %}
+{% include "./_meminfo.jinja" %}
+#define {{ name|upper }}_NB_ELTS {{ in_dims[0]|join('*') }}
+
+#endif /* {{ name|upper }}_LAYER_H */
\ No newline at end of file
diff --git a/aidge_export_cpp/templates/configuration/transpose_config.jinja b/aidge_export_cpp/templates/configuration/transpose_config.jinja
new file mode 100644
index 0000000..c3eabc5
--- /dev/null
+++ b/aidge_export_cpp/templates/configuration/transpose_config.jinja
@@ -0,0 +1,17 @@
+{#- For name header -#}
+#ifndef {{ name|upper }}_LAYER_H
+#define {{ name|upper }}_LAYER_H
+
+{# For layer configuration -#}
+{% include "./_def_io.jinja" %}
+{% include "./_meminfo.jinja" %}
+#define {{ name|upper }}_SIZE {{out_size[0]}}
+#define {{name|upper }}_SIZE_OUTPUT_DIMS {{out_dims[0]|length}}
+
+int {{name|upper}}_OUTPUT_DIMS[] = { {{ out_dims[0]|join(", ") }} };
+int {{name|upper}}_INPUT_DIMS[] = { {{ in_dims[0]|join(", ") }} };
+int {{name|upper}}_PERM[] = { {{ output_dims_order|join(", ") }} };
+
+
+
+#endif /* {{ name|upper }}_LAYER_H */
--
GitLab
From 659a8f4bb519f6b7f1863bb9de994608b4b5ff57 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Wed, 26 Feb 2025 15:19:25 +0000
Subject: [PATCH 09/22] Adding new files for add, div, mul and sub to make it
easier to read and manipulated
---
aidge_export_cpp/kernels/add.hpp | 104 ++++++++++++++++++
aidge_export_cpp/kernels/div.hpp | 103 +++++++++++++++++
aidge_export_cpp/kernels/mul.hpp | 104 ++++++++++++++++++
aidge_export_cpp/kernels/sub.hpp | 103 +++++++++++++++++
.../templates/configuration/add_config.jinja | 25 +++++
.../templates/configuration/div_config.jinja | 25 +++++
.../templates/configuration/mul_config.jinja | 25 +++++
.../templates/configuration/sub_config.jinja | 25 +++++
.../kernel_forward/add_forward.jinja | 14 +++
.../kernel_forward/div_forward.jinja | 14 +++
.../kernel_forward/mul_forward.jinja | 14 +++
.../kernel_forward/sub_forward.jinja | 14 +++
12 files changed, 570 insertions(+)
create mode 100644 aidge_export_cpp/kernels/add.hpp
create mode 100644 aidge_export_cpp/kernels/div.hpp
create mode 100644 aidge_export_cpp/kernels/mul.hpp
create mode 100644 aidge_export_cpp/kernels/sub.hpp
create mode 100644 aidge_export_cpp/templates/configuration/add_config.jinja
create mode 100644 aidge_export_cpp/templates/configuration/div_config.jinja
create mode 100644 aidge_export_cpp/templates/configuration/mul_config.jinja
create mode 100644 aidge_export_cpp/templates/configuration/sub_config.jinja
create mode 100644 aidge_export_cpp/templates/kernel_forward/add_forward.jinja
create mode 100644 aidge_export_cpp/templates/kernel_forward/div_forward.jinja
create mode 100644 aidge_export_cpp/templates/kernel_forward/mul_forward.jinja
create mode 100644 aidge_export_cpp/templates/kernel_forward/sub_forward.jinja
diff --git a/aidge_export_cpp/kernels/add.hpp b/aidge_export_cpp/kernels/add.hpp
new file mode 100644
index 0000000..03ba2c5
--- /dev/null
+++ b/aidge_export_cpp/kernels/add.hpp
@@ -0,0 +1,104 @@
+#ifndef __AIDGE_EXPORT_CPP_KERNELS_ADD__
+#define __AIDGE_EXPORT_CPP_KERNELS_ADD__
+
+#include "network/typedefs.hpp"
+#include "kernels/activation.hpp"
+#include <iostream>
+#include <cassert>
+
+
+
+template<int NB_ELTS,
+ int INPUT_A_DIMS[], int INPUT_B_DIMS[], int OUTPUT_DIMS[],
+ int SIZE_DIM_IN_A, int SIZE_DIM_IN_B, int SIZE_DIM_OUT, int OUT_SIZE,
+ ActivationFunction_T ACTIVATION,
+ typename Input_T, typename Output_T>
+__attribute__((always_inline)) inline
+void add_forward (
+ Output_T* __restrict outputs,
+ const Input_T* __restrict inputs1,
+ const Input_T* __restrict inputs2)
+{
+ int ndim_a[SIZE_DIM_OUT];
+ int ndim_b[SIZE_DIM_OUT];
+ for (int i= 0; i<SIZE_DIM_OUT; i++){
+ int idx = SIZE_DIM_OUT-SIZE_DIM_IN_A;
+ ndim_a[i] = (i< idx) ? 1 : INPUT_A_DIMS[i-idx];
+ }
+ for (int i= 0; i<SIZE_DIM_OUT; i++){
+ int idx = SIZE_DIM_OUT-SIZE_DIM_IN_B;
+ ndim_b[i] = (i< idx) ? 1 : INPUT_B_DIMS[i-idx];
+ }
+
+ // Find the highest equal dimension
+ int contiguousidx = SIZE_DIM_OUT -1 ;
+
+ for (int i = contiguousidx ; ndim_a[i] == ndim_b[i]; i--) {
+ contiguousidx = i;
+ }
+
+ // Compute the highest number of contiguous data for each Tensor
+ int input0_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ input0_contiguous_size *= ndim_a[i];
+ }
+ int input1_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ input1_contiguous_size *= ndim_b[i];
+ }
+ int output_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ output_contiguous_size *= OUTPUT_DIMS[i];
+ }
+ // initialize strides to iterate through data because of broadcasting
+ int stride_post0[contiguousidx ] ;
+ int stride_post1[contiguousidx ] ;
+ int stride_step0[contiguousidx ] ;
+ int stride_step1[contiguousidx ] ;
+ if (contiguousidx > 0) {
+ stride_post0[contiguousidx - 1] = 1;
+ stride_post1[contiguousidx - 1] = 1;
+ #pragma omp parallel for
+ for (int i = contiguousidx -2; i != -1; --i) {
+ stride_post0[i] = stride_post0[i+1]*ndim_a[i+1];
+ stride_post1[i] = stride_post1[i+1]*ndim_b[i+1];
+ }
+ #pragma omp parallel for
+ for (int i = 0; i < contiguousidx ; ++i) {
+ stride_step0[i] = (ndim_a[i] == 1) ? 1 - stride_post0[i] : 1;
+ stride_step1[i] = (ndim_b[i] == 1) ? 1 - stride_post1[i] : 1;
+ }
+ }
+ int offsetIn0 = 0;
+ int offsetIn1 = 0;
+ int offsetOut = 0;
+ int nbMatrices = 1;
+ for(int i = 0; i<contiguousidx ; ++i){
+ nbMatrices *= OUTPUT_DIMS[i];
+
+ }
+ int dim = contiguousidx - 1;
+ for(int stack = 0; stack < nbMatrices;){
+ for(int i = 0; i < output_contiguous_size; ++i){
+ int in0_id = (input0_contiguous_size != 1) ? i : 0;
+ int in1_id = (input1_contiguous_size != 1) ? i : 0;
+ outputs[i + offsetOut*output_contiguous_size] = inputs1[in0_id + offsetIn0*input0_contiguous_size] + inputs2[in1_id + offsetIn1*input1_contiguous_size];
+ }
+ if (++stack < nbMatrices) {
+ int tmp_stack = stack;
+ while(tmp_stack % OUTPUT_DIMS[dim] == 0) {
+ tmp_stack /= OUTPUT_DIMS[dim];
+ dim--;
+ }
+ offsetIn0 += stride_step0[dim];
+ offsetIn1 += stride_step1[dim];
+ ++offsetOut;
+ dim = contiguousidx - 1;
+ }
+ }
+}
+
+
+
+
+#endif // __AIDGE_EXPORT_CPP_KERNELS_ADD__
\ No newline at end of file
diff --git a/aidge_export_cpp/kernels/div.hpp b/aidge_export_cpp/kernels/div.hpp
new file mode 100644
index 0000000..f1ff7d0
--- /dev/null
+++ b/aidge_export_cpp/kernels/div.hpp
@@ -0,0 +1,103 @@
+#ifndef __AIDGE_EXPORT_CPP_KERNELS_DIV__
+#define __AIDGE_EXPORT_CPP_KERNELS_DIV__
+
+#include "network/typedefs.hpp"
+#include "kernels/activation.hpp"
+#include <iostream>
+#include <cassert>
+
+
+
+template<int NB_ELTS,
+ int INPUT_A_DIMS[], int INPUT_B_DIMS[], int OUTPUT_DIMS[],
+ int SIZE_DIM_IN_A, int SIZE_DIM_IN_B, int SIZE_DIM_OUT, int OUT_SIZE,
+ ActivationFunction_T ACTIVATION,
+ typename Input_T, typename Output_T>
+__attribute__((always_inline)) inline
+void div_forward (
+ Output_T* __restrict outputs,
+ const Input_T* __restrict inputs1,
+ const Input_T* __restrict inputs2)
+{
+
+ int ndim_a[SIZE_DIM_OUT];
+ int ndim_b[SIZE_DIM_OUT];
+ for (int i= 0; i<SIZE_DIM_OUT; i++){
+ int idx = SIZE_DIM_OUT-SIZE_DIM_IN_A;
+ ndim_a[i] = (i< idx) ? 1 : INPUT_A_DIMS[i-idx];
+ }
+ for (int i= 0; i<SIZE_DIM_OUT; i++){
+ int idx = SIZE_DIM_OUT-SIZE_DIM_IN_B;
+ ndim_b[i] = (i< idx) ? 1 : INPUT_B_DIMS[i-idx];
+ }
+
+ // Find the highest equal dimension
+ int contiguousidx = SIZE_DIM_OUT -1 ;
+
+ for (int i = contiguousidx ; ndim_a[i] == ndim_b[i]; i--) {
+ contiguousidx = i;
+ }
+
+ // Compute the highest number of contiguous data for each Tensor
+ int input0_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ input0_contiguous_size *= ndim_a[i];
+ }
+ int input1_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ input1_contiguous_size *= ndim_b[i];
+ }
+ int output_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ output_contiguous_size *= OUTPUT_DIMS[i];
+ }
+ // initialize strides to iterate through data because of broadcasting
+ int stride_post0[contiguousidx ] ;
+ int stride_post1[contiguousidx ] ;
+ int stride_step0[contiguousidx ] ;
+ int stride_step1[contiguousidx ] ;
+ if (contiguousidx > 0) {
+ stride_post0[contiguousidx - 1] = 1;
+ stride_post1[contiguousidx - 1] = 1;
+ for (int i = contiguousidx -2; i != -1; --i) {
+ stride_post0[i] = stride_post0[i+1]*ndim_a[i+1];
+ stride_post1[i] = stride_post1[i+1]*ndim_b[i+1];
+ }
+ for (int i = 0; i < contiguousidx ; ++i) {
+ stride_step0[i] = (ndim_a[i] == 1) ? 1 - stride_post0[i] : 1;
+ stride_step1[i] = (ndim_b[i] == 1) ? 1 - stride_post1[i] : 1;
+ }
+ }
+ int offsetIn0 = 0;
+ int offsetIn1 = 0;
+ int offsetOut = 0;
+ int nbMatrices = 1;
+ for(int i = 0; i<contiguousidx ; ++i){
+ nbMatrices *= OUTPUT_DIMS[i];
+
+ }
+ int dim = contiguousidx - 1;
+ for(int stack = 0; stack < nbMatrices;){
+ for(int i = 0; i < output_contiguous_size; ++i){
+ int in0_id = (input0_contiguous_size != 1) ? i : 0;
+ int in1_id = (input1_contiguous_size != 1) ? i : 0;
+ outputs[i + offsetOut*output_contiguous_size] = inputs1[in0_id + offsetIn0*input0_contiguous_size] / inputs2[in1_id + offsetIn1*input1_contiguous_size];
+ }
+ if (++stack < nbMatrices) {
+ int tmp_stack = stack;
+ while(tmp_stack % OUTPUT_DIMS[dim] == 0) {
+ tmp_stack /= OUTPUT_DIMS[dim];
+ dim--;
+ }
+ offsetIn0 += stride_step0[dim];
+ offsetIn1 += stride_step1[dim];
+ ++offsetOut;
+ dim = contiguousidx - 1;
+ }
+ }
+}
+
+
+
+
+#endif // __AIDGE_EXPORT_CPP_KERNELS_DIV__
\ No newline at end of file
diff --git a/aidge_export_cpp/kernels/mul.hpp b/aidge_export_cpp/kernels/mul.hpp
new file mode 100644
index 0000000..cbed0f6
--- /dev/null
+++ b/aidge_export_cpp/kernels/mul.hpp
@@ -0,0 +1,104 @@
+#ifndef __AIDGE_EXPORT_CPP_KERNELS_MUL__
+#define __AIDGE_EXPORT_CPP_KERNELS_MUL__
+
+#include "network/typedefs.hpp"
+#include "kernels/activation.hpp"
+#include <iostream>
+#include <cassert>
+
+
+
+template<int NB_ELTS,
+ int INPUT_A_DIMS[], int INPUT_B_DIMS[], int OUTPUT_DIMS[],
+ int SIZE_DIM_IN_A, int SIZE_DIM_IN_B, int SIZE_DIM_OUT, int OUT_SIZE,
+ ActivationFunction_T ACTIVATION,
+ typename Input_T, typename Output_T>
+__attribute__((always_inline)) inline
+void mul_forward (
+ Output_T* __restrict outputs,
+ const Input_T* __restrict inputs1,
+ const Input_T* __restrict inputs2)
+{
+ int ndim_a[SIZE_DIM_OUT];
+ int ndim_b[SIZE_DIM_OUT];
+ for (int i= 0; i<SIZE_DIM_OUT; i++){
+ int idx = SIZE_DIM_OUT-SIZE_DIM_IN_A;
+ ndim_a[i] = (i< idx) ? 1 : INPUT_A_DIMS[i-idx];
+ }
+ for (int i= 0; i<SIZE_DIM_OUT; i++){
+ int idx = SIZE_DIM_OUT-SIZE_DIM_IN_B;
+ ndim_b[i] = (i< idx) ? 1 : INPUT_B_DIMS[i-idx];
+ }
+
+ // Find the highest equal dimension
+ int contiguousidx = SIZE_DIM_OUT -1 ;
+
+ for (int i = contiguousidx ; ndim_a[i] == ndim_b[i]; i--) {
+ contiguousidx = i;
+ }
+
+ // Compute the highest number of contiguous data for each Tensor
+ int input0_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ input0_contiguous_size *= ndim_a[i];
+ }
+ int input1_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ input1_contiguous_size *= ndim_b[i];
+ }
+ int output_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ output_contiguous_size *= OUTPUT_DIMS[i];
+ }
+ // initialize strides to iterate through data because of broadcasting
+ int stride_post0[contiguousidx ] ;
+ int stride_post1[contiguousidx ] ;
+ int stride_step0[contiguousidx ] ;
+ int stride_step1[contiguousidx ] ;
+ if (contiguousidx > 0) {
+ stride_post0[contiguousidx - 1] = 1;
+ stride_post1[contiguousidx - 1] = 1;
+ #pragma omp parallel for
+ for (int i = contiguousidx -2; i != -1; --i) {
+ stride_post0[i] = stride_post0[i+1]*ndim_a[i+1];
+ stride_post1[i] = stride_post1[i+1]*ndim_b[i+1];
+ }
+ #pragma omp parallel for
+ for (int i = 0; i < contiguousidx ; ++i) {
+ stride_step0[i] = (ndim_a[i] == 1) ? 1 - stride_post0[i] : 1;
+ stride_step1[i] = (ndim_b[i] == 1) ? 1 - stride_post1[i] : 1;
+ }
+ }
+ int offsetIn0 = 0;
+ int offsetIn1 = 0;
+ int offsetOut = 0;
+ int nbMatrices = 1;
+ for(int i = 0; i<contiguousidx ; ++i){
+ nbMatrices *= OUTPUT_DIMS[i];
+
+ }
+ int dim = contiguousidx - 1;
+ for(int stack = 0; stack < nbMatrices;){
+ for(int i = 0; i < output_contiguous_size; ++i){
+ int in0_id = (input0_contiguous_size != 1) ? i : 0;
+ int in1_id = (input1_contiguous_size != 1) ? i : 0;
+ outputs[i + offsetOut*output_contiguous_size] = inputs1[in0_id + offsetIn0*input0_contiguous_size] * inputs2[in1_id + offsetIn1*input1_contiguous_size];
+ }
+ if (++stack < nbMatrices) {
+ int tmp_stack = stack;
+ while(tmp_stack % OUTPUT_DIMS[dim] == 0) {
+ tmp_stack /= OUTPUT_DIMS[dim];
+ dim--;
+ }
+ offsetIn0 += stride_step0[dim];
+ offsetIn1 += stride_step1[dim];
+ ++offsetOut;
+ dim = contiguousidx - 1;
+ }
+ }
+}
+
+
+
+
+#endif // __AIDGE_EXPORT_CPP_KERNELS_MUL__
\ No newline at end of file
diff --git a/aidge_export_cpp/kernels/sub.hpp b/aidge_export_cpp/kernels/sub.hpp
new file mode 100644
index 0000000..07637cd
--- /dev/null
+++ b/aidge_export_cpp/kernels/sub.hpp
@@ -0,0 +1,103 @@
+#ifndef __AIDGE_EXPORT_CPP_KERNELS_SUB__
+#define __AIDGE_EXPORT_CPP_KERNELS_SUB__
+
+#include "network/typedefs.hpp"
+#include "kernels/activation.hpp"
+#include <iostream>
+#include <cassert>
+
+
+
+template<int NB_ELTS,
+ int INPUT_A_DIMS[], int INPUT_B_DIMS[], int OUTPUT_DIMS[],
+ int SIZE_DIM_IN_A, int SIZE_DIM_IN_B, int SIZE_DIM_OUT, int OUT_SIZE,
+ ActivationFunction_T ACTIVATION,
+ typename Input_T, typename Output_T>
+__attribute__((always_inline)) inline
+void sub_forward (
+ Output_T* __restrict outputs,
+ const Input_T* __restrict inputs1,
+ const Input_T* __restrict inputs2)
+{
+
+ int ndim_a[SIZE_DIM_OUT];
+ int ndim_b[SIZE_DIM_OUT];
+ for (int i= 0; i<SIZE_DIM_OUT; i++){
+ int idx = SIZE_DIM_OUT-SIZE_DIM_IN_A;
+ ndim_a[i] = (i< idx) ? 1 : INPUT_A_DIMS[i-idx];
+ }
+ for (int i= 0; i<SIZE_DIM_OUT; i++){
+ int idx = SIZE_DIM_OUT-SIZE_DIM_IN_B;
+ ndim_b[i] = (i< idx) ? 1 : INPUT_B_DIMS[i-idx];
+ }
+
+ // Find the highest equal dimension
+ int contiguousidx = SIZE_DIM_OUT -1 ;
+
+ for (int i = contiguousidx ; ndim_a[i] == ndim_b[i]; i--) {
+ contiguousidx = i;
+ }
+
+ // Compute the highest number of contiguous data for each Tensor
+ int input0_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ input0_contiguous_size *= ndim_a[i];
+ }
+ int input1_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ input1_contiguous_size *= ndim_b[i];
+ }
+ int output_contiguous_size = 1;
+ for(int i = contiguousidx ; i<SIZE_DIM_OUT; ++i){
+ output_contiguous_size *= OUTPUT_DIMS[i];
+ }
+ // initialize strides to iterate through data because of broadcasting
+ int stride_post0[contiguousidx ] ;
+ int stride_post1[contiguousidx ] ;
+ int stride_step0[contiguousidx ] ;
+ int stride_step1[contiguousidx ] ;
+ if (contiguousidx > 0) {
+ stride_post0[contiguousidx - 1] = 1;
+ stride_post1[contiguousidx - 1] = 1;
+ for (int i = contiguousidx -2; i != -1; --i) {
+ stride_post0[i] = stride_post0[i+1]*ndim_a[i+1];
+ stride_post1[i] = stride_post1[i+1]*ndim_b[i+1];
+ }
+ for (int i = 0; i < contiguousidx ; ++i) {
+ stride_step0[i] = (ndim_a[i] == 1) ? 1 - stride_post0[i] : 1;
+ stride_step1[i] = (ndim_b[i] == 1) ? 1 - stride_post1[i] : 1;
+ }
+ }
+ int offsetIn0 = 0;
+ int offsetIn1 = 0;
+ int offsetOut = 0;
+ int nbMatrices = 1;
+ for(int i = 0; i<contiguousidx ; ++i){
+ nbMatrices *= OUTPUT_DIMS[i];
+
+ }
+ int dim = contiguousidx - 1;
+ for(int stack = 0; stack < nbMatrices;){
+ for(int i = 0; i < output_contiguous_size; ++i){
+ int in0_id = (input0_contiguous_size != 1) ? i : 0;
+ int in1_id = (input1_contiguous_size != 1) ? i : 0;
+ outputs[i + offsetOut*output_contiguous_size] = inputs1[in0_id + offsetIn0*input0_contiguous_size] - inputs2[in1_id + offsetIn1*input1_contiguous_size];
+ }
+ if (++stack < nbMatrices) {
+ int tmp_stack = stack;
+ while(tmp_stack % OUTPUT_DIMS[dim] == 0) {
+ tmp_stack /= OUTPUT_DIMS[dim];
+ dim--;
+ }
+ offsetIn0 += stride_step0[dim];
+ offsetIn1 += stride_step1[dim];
+ ++offsetOut;
+ dim = contiguousidx - 1;
+ }
+ }
+}
+
+
+
+
+#endif // __AIDGE_EXPORT_CPP_KERNELS_SUB__
\ No newline at end of file
diff --git a/aidge_export_cpp/templates/configuration/add_config.jinja b/aidge_export_cpp/templates/configuration/add_config.jinja
new file mode 100644
index 0000000..143d004
--- /dev/null
+++ b/aidge_export_cpp/templates/configuration/add_config.jinja
@@ -0,0 +1,25 @@
+{#- For name header -#}
+#ifndef {{ name|upper }}_LAYER_H
+#define {{ name|upper }}_LAYER_H
+#include "kernels/rescaling.hpp"
+
+{% include "./_def_io.jinja" %}
+{% include "./_meminfo.jinja" %}
+{# For layer configuration -#}
+#define {{ name|upper }}_NB_ELTS {{ in_dims[0]|join('*') }}
+#define {{ name|upper }}_NB_ELTS_B {{ in_dims[1]|join('*')}}
+
+int {{name|upper}}_OUTPUT_DIMS[] = { {{ out_dims[0]|join(", ") }} };
+int {{name|upper}}_INPUT_A_DIMS[] = { {{ in_dims[0]|join(", ") }} };
+int {{name|upper}}_INPUT_B_DIMS[] = { {{ in_dims[1]|join(", ") }} };
+
+#define {{name|upper}}_SIZE_DIM_IN_A {{in_dims[0]|length}}
+#define {{name|upper}}_SIZE_DIM_IN_B {{in_dims[1]|length}}
+#define {{name|upper}}_SIZE_DIM_OUT {{out_dims[0]|length}}
+
+#define {{ name|upper }}_OUT_SIZE {{out_size[0]}}
+#define {{name|upper }}_SIZE_DIM_OUT {{out_dims[0]|length}}
+
+#define {{ name|upper }}_ACTIVATION {{ activation }}
+static const {{ rescaling }} {{ name|upper }}_RESCALING = {};
+#endif /* {{ name|upper }}_LAYER_H */
diff --git a/aidge_export_cpp/templates/configuration/div_config.jinja b/aidge_export_cpp/templates/configuration/div_config.jinja
new file mode 100644
index 0000000..143d004
--- /dev/null
+++ b/aidge_export_cpp/templates/configuration/div_config.jinja
@@ -0,0 +1,25 @@
+{#- For name header -#}
+#ifndef {{ name|upper }}_LAYER_H
+#define {{ name|upper }}_LAYER_H
+#include "kernels/rescaling.hpp"
+
+{% include "./_def_io.jinja" %}
+{% include "./_meminfo.jinja" %}
+{# For layer configuration -#}
+#define {{ name|upper }}_NB_ELTS {{ in_dims[0]|join('*') }}
+#define {{ name|upper }}_NB_ELTS_B {{ in_dims[1]|join('*')}}
+
+int {{name|upper}}_OUTPUT_DIMS[] = { {{ out_dims[0]|join(", ") }} };
+int {{name|upper}}_INPUT_A_DIMS[] = { {{ in_dims[0]|join(", ") }} };
+int {{name|upper}}_INPUT_B_DIMS[] = { {{ in_dims[1]|join(", ") }} };
+
+#define {{name|upper}}_SIZE_DIM_IN_A {{in_dims[0]|length}}
+#define {{name|upper}}_SIZE_DIM_IN_B {{in_dims[1]|length}}
+#define {{name|upper}}_SIZE_DIM_OUT {{out_dims[0]|length}}
+
+#define {{ name|upper }}_OUT_SIZE {{out_size[0]}}
+#define {{name|upper }}_SIZE_DIM_OUT {{out_dims[0]|length}}
+
+#define {{ name|upper }}_ACTIVATION {{ activation }}
+static const {{ rescaling }} {{ name|upper }}_RESCALING = {};
+#endif /* {{ name|upper }}_LAYER_H */
diff --git a/aidge_export_cpp/templates/configuration/mul_config.jinja b/aidge_export_cpp/templates/configuration/mul_config.jinja
new file mode 100644
index 0000000..143d004
--- /dev/null
+++ b/aidge_export_cpp/templates/configuration/mul_config.jinja
@@ -0,0 +1,25 @@
+{#- For name header -#}
+#ifndef {{ name|upper }}_LAYER_H
+#define {{ name|upper }}_LAYER_H
+#include "kernels/rescaling.hpp"
+
+{% include "./_def_io.jinja" %}
+{% include "./_meminfo.jinja" %}
+{# For layer configuration -#}
+#define {{ name|upper }}_NB_ELTS {{ in_dims[0]|join('*') }}
+#define {{ name|upper }}_NB_ELTS_B {{ in_dims[1]|join('*')}}
+
+int {{name|upper}}_OUTPUT_DIMS[] = { {{ out_dims[0]|join(", ") }} };
+int {{name|upper}}_INPUT_A_DIMS[] = { {{ in_dims[0]|join(", ") }} };
+int {{name|upper}}_INPUT_B_DIMS[] = { {{ in_dims[1]|join(", ") }} };
+
+#define {{name|upper}}_SIZE_DIM_IN_A {{in_dims[0]|length}}
+#define {{name|upper}}_SIZE_DIM_IN_B {{in_dims[1]|length}}
+#define {{name|upper}}_SIZE_DIM_OUT {{out_dims[0]|length}}
+
+#define {{ name|upper }}_OUT_SIZE {{out_size[0]}}
+#define {{name|upper }}_SIZE_DIM_OUT {{out_dims[0]|length}}
+
+#define {{ name|upper }}_ACTIVATION {{ activation }}
+static const {{ rescaling }} {{ name|upper }}_RESCALING = {};
+#endif /* {{ name|upper }}_LAYER_H */
diff --git a/aidge_export_cpp/templates/configuration/sub_config.jinja b/aidge_export_cpp/templates/configuration/sub_config.jinja
new file mode 100644
index 0000000..143d004
--- /dev/null
+++ b/aidge_export_cpp/templates/configuration/sub_config.jinja
@@ -0,0 +1,25 @@
+{#- For name header -#}
+#ifndef {{ name|upper }}_LAYER_H
+#define {{ name|upper }}_LAYER_H
+#include "kernels/rescaling.hpp"
+
+{% include "./_def_io.jinja" %}
+{% include "./_meminfo.jinja" %}
+{# For layer configuration -#}
+#define {{ name|upper }}_NB_ELTS {{ in_dims[0]|join('*') }}
+#define {{ name|upper }}_NB_ELTS_B {{ in_dims[1]|join('*')}}
+
+int {{name|upper}}_OUTPUT_DIMS[] = { {{ out_dims[0]|join(", ") }} };
+int {{name|upper}}_INPUT_A_DIMS[] = { {{ in_dims[0]|join(", ") }} };
+int {{name|upper}}_INPUT_B_DIMS[] = { {{ in_dims[1]|join(", ") }} };
+
+#define {{name|upper}}_SIZE_DIM_IN_A {{in_dims[0]|length}}
+#define {{name|upper}}_SIZE_DIM_IN_B {{in_dims[1]|length}}
+#define {{name|upper}}_SIZE_DIM_OUT {{out_dims[0]|length}}
+
+#define {{ name|upper }}_OUT_SIZE {{out_size[0]}}
+#define {{name|upper }}_SIZE_DIM_OUT {{out_dims[0]|length}}
+
+#define {{ name|upper }}_ACTIVATION {{ activation }}
+static const {{ rescaling }} {{ name|upper }}_RESCALING = {};
+#endif /* {{ name|upper }}_LAYER_H */
diff --git a/aidge_export_cpp/templates/kernel_forward/add_forward.jinja b/aidge_export_cpp/templates/kernel_forward/add_forward.jinja
new file mode 100644
index 0000000..3176ced
--- /dev/null
+++ b/aidge_export_cpp/templates/kernel_forward/add_forward.jinja
@@ -0,0 +1,14 @@
+{% filter indent(width=4, first=False) %}
+{% include "./_mem_offset.jinja" %}
+add_forward<{{name|upper}}_NB_ELTS,
+ {{name|upper}}_INPUT_A_DIMS,
+ {{name|upper}}_INPUT_B_DIMS,
+ {{name|upper}}_OUTPUT_DIMS,
+ {{name|upper}}_SIZE_DIM_IN_A,
+ {{name|upper}}_SIZE_DIM_IN_B,
+ {{name|upper}}_SIZE_DIM_OUT,
+ {{name|upper}}_OUT_SIZE,
+ {{name|upper}}_ACTIVATION>
+ ({{out_name[0]}}, {{in_name[0]}}, {{in_name[1]}});
+{% include "./_save_outputs.jinja" %}
+{% endfilter %}
diff --git a/aidge_export_cpp/templates/kernel_forward/div_forward.jinja b/aidge_export_cpp/templates/kernel_forward/div_forward.jinja
new file mode 100644
index 0000000..4b79357
--- /dev/null
+++ b/aidge_export_cpp/templates/kernel_forward/div_forward.jinja
@@ -0,0 +1,14 @@
+{% filter indent(width=4, first=False) %}
+{% include "./_mem_offset.jinja" %}
+div_forward<{{name|upper}}_NB_ELTS,
+ {{name|upper}}_INPUT_A_DIMS,
+ {{name|upper}}_INPUT_B_DIMS,
+ {{name|upper}}_OUTPUT_DIMS,
+ {{name|upper}}_SIZE_DIM_IN_A,
+ {{name|upper}}_SIZE_DIM_IN_B,
+ {{name|upper}}_SIZE_DIM_OUT,
+ {{name|upper}}_OUT_SIZE,
+ {{name|upper}}_ACTIVATION>
+ ({{out_name[0]}}, {{in_name[0]}}, {{in_name[1]}});
+{% include "./_save_outputs.jinja" %}
+{% endfilter %}
diff --git a/aidge_export_cpp/templates/kernel_forward/mul_forward.jinja b/aidge_export_cpp/templates/kernel_forward/mul_forward.jinja
new file mode 100644
index 0000000..9a7170b
--- /dev/null
+++ b/aidge_export_cpp/templates/kernel_forward/mul_forward.jinja
@@ -0,0 +1,14 @@
+{% filter indent(width=4, first=False) %}
+{% include "./_mem_offset.jinja" %}
+mul_forward<{{name|upper}}_NB_ELTS,
+ {{name|upper}}_INPUT_A_DIMS,
+ {{name|upper}}_INPUT_B_DIMS,
+ {{name|upper}}_OUTPUT_DIMS,
+ {{name|upper}}_SIZE_DIM_IN_A,
+ {{name|upper}}_SIZE_DIM_IN_B,
+ {{name|upper}}_SIZE_DIM_OUT,
+ {{name|upper}}_OUT_SIZE,
+ {{name|upper}}_ACTIVATION>
+ ({{out_name[0]}}, {{in_name[0]}}, {{in_name[1]}});
+{% include "./_save_outputs.jinja" %}
+{% endfilter %}
diff --git a/aidge_export_cpp/templates/kernel_forward/sub_forward.jinja b/aidge_export_cpp/templates/kernel_forward/sub_forward.jinja
new file mode 100644
index 0000000..51b47a8
--- /dev/null
+++ b/aidge_export_cpp/templates/kernel_forward/sub_forward.jinja
@@ -0,0 +1,14 @@
+{% filter indent(width=4, first=False) %}
+{% include "./_mem_offset.jinja" %}
+sub_forward<{{name|upper}}_NB_ELTS,
+ {{name|upper}}_INPUT_A_DIMS,
+ {{name|upper}}_INPUT_B_DIMS,
+ {{name|upper}}_OUTPUT_DIMS,
+ {{name|upper}}_SIZE_DIM_IN_A,
+ {{name|upper}}_SIZE_DIM_IN_B,
+ {{name|upper}}_SIZE_DIM_OUT,
+ {{name|upper}}_OUT_SIZE,
+ {{name|upper}}_ACTIVATION>
+ ({{out_name[0]}}, {{in_name[0]}}, {{in_name[1]}});
+{% include "./_save_outputs.jinja" %}
+{% endfilter %}
--
GitLab
From a4dd75a361f3d98565287160a44426ebe1fa8515 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Tue, 4 Mar 2025 12:29:22 +0000
Subject: [PATCH 10/22] Debug for batchnorm and for matmul and transpose able
to broadcast
---
aidge_export_cpp/kernels/batchnorm.hpp | 32 +++++---
aidge_export_cpp/kernels/matmul.hpp | 108 +++++++++++++++++++++----
aidge_export_cpp/kernels/transpose.hpp | 3 +-
3 files changed, 114 insertions(+), 29 deletions(-)
diff --git a/aidge_export_cpp/kernels/batchnorm.hpp b/aidge_export_cpp/kernels/batchnorm.hpp
index 740ea21..0ed5080 100644
--- a/aidge_export_cpp/kernels/batchnorm.hpp
+++ b/aidge_export_cpp/kernels/batchnorm.hpp
@@ -3,7 +3,9 @@
#include "network/typedefs.hpp"
#include "kernels/rescaling.hpp"
+#include "kernels/activation.hpp"
#include <math.h>
+#include <iostream>
// WARNING: this kernel only works for 32-bits floating point values
@@ -12,30 +14,34 @@ template<int NB_OUTPUTS,
ActivationFunction_T ACTIVATION,
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,
+ const Param_T* __restrict scales,
const Param_T* __restrict biases,
- const Param_T* __restrict variances,
const Param_T* __restrict means,
- const Param_T* __restrict scales,
+ const Param_T* __restrict variances,
const double epsilon)
{
- for (unsigned int output = 0; output < NB_OUTPUTS; ++output) {
- const Output_T var = sqrt(variances[output] + epsilon);
- for (int oy = 0; oy < OUTPUTS_HEIGHT; ++oy) {
- for (int ox = 0; ox < OUTPUTS_WIDTH; ++ox) {
- const int outputOffset = OUTPUTS_HEIGHT * oy + ox;
-
- const Output_T normalized = (inputs[outputOffset + output] - means[output]) / var;
- const Output_T sAs = scales[output] * normalized + biases[output];
- outputs[outputOffset + output] = sat<Output_T>(sAs, output, ACTIVATION, NoScaling);
- }
+ int featureMapSize = OUTPUTS_HEIGHT * OUTPUTS_WIDTH;
+ #pragma omp parallel for
+ for (int ch = 0; ch < NB_OUTPUTS; ++ch) {
+ int ioIndex = ch * featureMapSize;
+ #pragma omp parallel for
+ for (int i = ioIndex; i < ioIndex + featureMapSize; i++) {
+ outputs[i] = biases[ch];
+ }
+ float var = sqrt(variances[ch] + epsilon);
+ #pragma omp parallel for
+ for (int feature = 0; feature < featureMapSize; ++feature) {
+ outputs[ioIndex + feature] += (scales[ch] * (inputs[ioIndex + feature] - means[ch]) / var);
}
}
+
}
+
#endif // __AIDGE_EXPORT_CPP_KERNELS_BATCHNORM__
diff --git a/aidge_export_cpp/kernels/matmul.hpp b/aidge_export_cpp/kernels/matmul.hpp
index 4500993..b284214 100644
--- a/aidge_export_cpp/kernels/matmul.hpp
+++ b/aidge_export_cpp/kernels/matmul.hpp
@@ -3,31 +3,109 @@
#include "network/typedefs.hpp"
#include "kernels/activation.hpp"
-
+#include <iostream>
// Generic function for matmul and activation
-template<int M,
- int K,
- int N,
+template<int INPUT_A_DIMS[], int INPUT_B_DIMS[], int OUTPUT_DIMS[],
+ int _SIZE_DIM_IN_A, int _SIZE_DIM_IN_B, int SIZE_DIM_OUT,
ActivationFunction_T ACTIVATION,
- typename Input_T, typename Output_T,
- typename Rescaling_T>
+ typename Input_T, typename Output_T>
__attribute__((always_inline)) inline
void matmul_forward (
const Input_T* __restrict inputs1,
const Input_T* __restrict inputs2,
- Output_T* __restrict outputs,
- const Rescaling_T& __restrict rescaling)
+ Output_T* __restrict outputs)
{
- for (int m = 0; m < M; ++m) {
- for (int n = 0; n < N; ++n) {
- Output_T sum = Output_T(0);
- for (int k = 0; k < K; ++k) {
- sum += inputs1[K*m + k] * inputs2[N*k + n];
+
+ //initialize arrays storing broadcasted(or not) dims
+ int ndim_a[SIZE_DIM_OUT];
+ int ndim_b[SIZE_DIM_OUT];
+ if ( _SIZE_DIM_IN_A == 1){
+ ndim_a[0] = 1;
+ ndim_a[1] =INPUT_A_DIMS[0];
+ }
+ if ( _SIZE_DIM_IN_B == 1){
+ ndim_b[0] =INPUT_B_DIMS[0];
+ ndim_b[1] = 1;
+ }
+
+ for (int i= 0; i<SIZE_DIM_OUT; i++){
+ int idx = SIZE_DIM_OUT-_SIZE_DIM_IN_A;
+ ndim_a[i] = (i< idx) ? 1 :INPUT_A_DIMS[i-idx];
+ }
+ for (int i= 0; i<SIZE_DIM_OUT; i++){
+ int idx = SIZE_DIM_OUT-_SIZE_DIM_IN_B;
+ ndim_b[i] = (i< idx) ? 1 :INPUT_B_DIMS[i-idx];
+ }
+
+ // initialize strides to iterate through data because of broadcasting
+ int stride_post0[SIZE_DIM_OUT-2] ;
+ int stride_post1[SIZE_DIM_OUT-2] ;
+ int stride_step0[SIZE_DIM_OUT-2] ;
+ int stride_step1[SIZE_DIM_OUT-2] ;
+ if (SIZE_DIM_OUT > 2){
+ stride_post0[SIZE_DIM_OUT - 3] = 1;
+ stride_post1[SIZE_DIM_OUT - 3] = 1;
+ for (int i = SIZE_DIM_OUT-4; i != -1; --i) {
+ stride_post0[i] = stride_post0[i+1]*ndim_a[i+1];
+ stride_post1[i] = stride_post1[i+1]*ndim_b[i+1];
+ }
+ for (int i = 0; i < SIZE_DIM_OUT-2; ++i) {
+ stride_step0[i] = (ndim_a[i] == 1) ? 1 - stride_post0[i] : 1;
+ stride_step1[i] = (ndim_b[i] == 1) ? 1 - stride_post1[i] : 1;
+ }
+
+ }
+
+
+ // if _SIZE_DIM_IN_B == _SIZE_DIM_IN_A, then _SIZE_DIM_IN_A == SIZE_DIM_OUT == _SIZE_DIM_IN_B;
+ // else it will be broadcasted to the correct dims
+
+ int nbMatrices = 1;
+ for(int i = SIZE_DIM_OUT -3; i>=0; --i){
+ nbMatrices *= OUTPUT_DIMS[i];
+ }
+ int dim = SIZE_DIM_OUT -3;
+
+
+ int offsetIn0 = 0;
+ int offsetIn1 = 0;
+ int offsetOut = 0;
+ const int n = ndim_a[SIZE_DIM_OUT - 2];
+ const int k = ndim_a[SIZE_DIM_OUT - 1];
+ const int m = ndim_b[SIZE_DIM_OUT - 1];
+ const int matrix0Size = n*k;
+ const int matrix1Size = k*m;
+ const int matrixOutSize = n*m;
+
+ for(int stack = 0; stack < nbMatrices;){
+
+ for (int i = 0; i < n; ++i) {
+
+ for (int j = 0; j < m; ++j) {
+ float sum = 0;
+
+ for (int l = 0; l < k; ++l) {
+ sum += (inputs1[ offsetIn0*matrix0Size + i*k + l] * inputs2[offsetIn1*matrix1Size + l*m + j]);
+ }
+ outputs[ offsetOut*matrixOutSize + i*m + j] = sum;
+ }
+ }
+
+ if (++stack < nbMatrices) {
+ int tmp_stack = stack;
+ while(tmp_stack % OUTPUT_DIMS[dim] == 0) {
+ tmp_stack /= OUTPUT_DIMS[dim];
+ dim--;
}
- outputs[N*m + n] = activation_forward_value<Output_T>(sum, 0/*not applicable*/, ACTIVATION, rescaling);
+ offsetIn0 += stride_step0[dim];
+ offsetIn1 += stride_step1[dim];
+ ++offsetOut;
+ dim = SIZE_DIM_OUT -3;
}
+
}
+
}
-#endif // __AIDGE_EXPORT_CPP_KERNELS_MATMUL__
+#endif // __AIDGE_EXPORT_CPP_KERNELS_MATMUL__
\ No newline at end of file
diff --git a/aidge_export_cpp/kernels/transpose.hpp b/aidge_export_cpp/kernels/transpose.hpp
index a6ddce2..2a89e3c 100644
--- a/aidge_export_cpp/kernels/transpose.hpp
+++ b/aidge_export_cpp/kernels/transpose.hpp
@@ -27,9 +27,10 @@ void transpose_forward (
int indices[SIZE_OUTPUT_DIMS];
for (int i = 0; i<SIZE_OUTPUT_DIMS;++i){indices[i] = 0;}
-
+ #pragma omp parallel for
for (int i = 0; i < SIZE; ++i) {
int idx = 0;
+ #pragma omp parallel for
for (int j = SIZE_OUTPUT_DIMS -1; j >=0; --j) {
idx += indices[PERM[j]] * newStrides[j];
}
--
GitLab
From ace17ef9620f1ba1843f795c349be513ac47d21c Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Wed, 5 Mar 2025 14:45:09 +0000
Subject: [PATCH 11/22] Jinja files matmul modified to accomodate for
broadcasting modificatio of matmul
---
.../configuration/matmul_config.jinja | 21 +++++++++++++++----
.../kernel_forward/matmul_forward.jinja | 15 +++++++++----
2 files changed, 28 insertions(+), 8 deletions(-)
diff --git a/aidge_export_cpp/templates/configuration/matmul_config.jinja b/aidge_export_cpp/templates/configuration/matmul_config.jinja
index fece988..0c28e06 100644
--- a/aidge_export_cpp/templates/configuration/matmul_config.jinja
+++ b/aidge_export_cpp/templates/configuration/matmul_config.jinja
@@ -1,13 +1,26 @@
{#- For name header -#}
#ifndef {{ name|upper }}_LAYER_H
#define {{ name|upper }}_LAYER_H
+#include "kernels/rescaling.hpp"
{# 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] }}
+{% include "./_def_io.jinja" %}
+{% include "./_meminfo.jinja" %}
+#define {{ name|upper }}_B {{ in_dims[0][0]}}
+#define {{ name|upper }}_C {{ in_chan[0]}}
+#define {{ name|upper }}_M {{ in_height[0]}}
+#define {{ name|upper }}_K {{ in_width[0] }}
+#define {{ name|upper }}_N {{ out_width[0] }}
+
+#define {{name|upper}}_SIZE_DIM_IN_A {{in_dims[0]|length}}
+#define {{name|upper}}_SIZE_DIM_IN_B {{in_dims[1]|length}}
+#define {{name|upper}}_SIZE_DIM_OUT {{out_dims[0]|length}}
+
+int {{name|upper}}_OUTPUT_DIMS[] = { {{ out_dims[0]|join(", ") }} };
+int {{name|upper}}_INPUT_A_DIMS[] = { {{ in_dims[0]|join(", ") }} };
+int {{name|upper}}_INPUT_B_DIMS[] = { {{ in_dims[1]|join(", ") }} };
+
#define {{ name|upper }}_ACTIVATION {{ activation }}
-static const {{ rescaling }} {{ name|upper }}_RESCALING = {};
{#- Calculate sizes #}
diff --git a/aidge_export_cpp/templates/kernel_forward/matmul_forward.jinja b/aidge_export_cpp/templates/kernel_forward/matmul_forward.jinja
index ce80ffd..4ed0264 100644
--- a/aidge_export_cpp/templates/kernel_forward/matmul_forward.jinja
+++ b/aidge_export_cpp/templates/kernel_forward/matmul_forward.jinja
@@ -1,5 +1,12 @@
-matmul_forward<{{name|upper}}_M,
- {{name|upper}}_K,
- {{name|upper}}_N,
+{% filter indent(width=4, first=False) %}
+{% include "./_mem_offset.jinja" %}
+matmul_forward<{{name|upper}}_INPUT_A_DIMS,
+ {{name|upper}}_INPUT_B_DIMS,
+ {{name|upper}}_OUTPUT_DIMS,
+ {{name|upper}}_SIZE_DIM_IN_A,
+ {{name|upper}}_SIZE_DIM_IN_B,
+ {{name|upper}}_SIZE_DIM_OUT,
{{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]}});
+{% include "./_save_outputs.jinja" %}
+{% endfilter %}
\ No newline at end of file
--
GitLab
From 8a315efbf4986f9903d163216bdac188ee2faeec Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Tue, 25 Mar 2025 08:37:11 +0000
Subject: [PATCH 12/22] Add operator helpers for transpose, erf, matmul and
batchnorm
---
aidge_export_cpp/operators.py | 166 ++++++++++++++++++++++++++++------
1 file changed, 138 insertions(+), 28 deletions(-)
diff --git a/aidge_export_cpp/operators.py b/aidge_export_cpp/operators.py
index f04dbb3..16890e1 100644
--- a/aidge_export_cpp/operators.py
+++ b/aidge_export_cpp/operators.py
@@ -4,27 +4,14 @@ from pathlib import Path
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 import ExportLibCpp
##############################################
############## Export functions ##############
##############################################
-def numpy_dtype2ctype(dtype):
- if dtype == np.int8:
- return "int8_t"
- elif dtype == np.int16:
- return "int16_t"
- elif dtype == np.int32:
- return "int32_t"
- elif dtype == np.int64:
- return "int64_t"
- elif dtype == np.float32:
- return "float"
- elif dtype == np.float64:
- return "double"
- # Add more dtype mappings as needed
- else:
- raise ValueError(f"Unsupported {dtype} dtype")
+
+
def export_params(name: str,
array: np.ndarray,
@@ -103,6 +90,7 @@ class ConvCPP(ExportNodeCpp):
self.attributes["padding"] = [0, 0]
self.attributes["activation"] = "Linear"
self.attributes["rescaling"] = "NoScaling"
+ self.attributes["groups"] = 1
self.config_template = str(
ROOT / "templates" / "configuration" / "convolution_config.jinja")
self.forward_template = str(
@@ -144,21 +132,53 @@ class PaddedConvCPP(ExportNodeCpp):
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
+ self.attributes["groups"] = 1
+
+@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)
+ # 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
+ 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")
+ ]
+ self.attributes["groups"] = self.attributes["out_chan"][0]
@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["add_op"] = "Add"
self.attributes["activation"] = "Linear"
self.attributes["rescaling"] = "NoScaling"
self.config_template = str(
- ROOT / "templates" / "configuration" / "elemwise_config.jinja")
+ ROOT / "templates" / "configuration" / "add_config.jinja")
self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "elemwise_forward.jinja")
+ ROOT / "templates" / "kernel_forward" / "add_forward.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "kernels" / "elemwise.hpp"),
+ str(ROOT / "kernels" / "add.hpp"),
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
@@ -167,16 +187,16 @@ class AddCPP(ExportNodeCpp):
class SubCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
super().__init__(node, mem_info)
- self.attributes["elemwise_op"] = "Sub"
+ self.attributes["sub_op"] = "Sub"
self.attributes["activation"] = "Linear"
self.attributes["rescaling"] = "NoScaling"
self.config_template = str(
- ROOT / "templates" / "configuration" / "elemwise_config.jinja")
+ ROOT / "templates" / "configuration" / "sub_config.jinja")
self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "elemwise_forward.jinja")
+ ROOT / "templates" / "kernel_forward" / "sub_forward.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "kernels" / "elemwise.hpp"),
+ str(ROOT / "kernels" / "sub.hpp"),
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
]
@@ -186,20 +206,39 @@ class SubCPP(ExportNodeCpp):
class MulCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
super().__init__(node, mem_info)
- self.attributes["elemwise_op"] = "Mul"
+ self.attributes["mul_op"] = "Mul"
+ self.attributes["activation"] = "Linear"
+ self.attributes["rescaling"] = "NoScaling"
+ self.config_template = str(
+ ROOT / "templates" / "configuration" / "mul_config.jinja")
+ self.forward_template = str(
+ ROOT / "templates" / "kernel_forward" / "mul_forward.jinja")
+ self.include_list = []
+ self.kernels_to_copy = [
+ str(ROOT / "kernels" / "mul.hpp"),
+ str(ROOT / "kernels" / "activation.hpp"),
+ str(ROOT / "kernels" / "rescaling.hpp")
+ ]
+
+@ExportLibCpp.register("Div", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
+class DivCPP(ExportNodeCpp):
+ def __init__(self, node, mem_info):
+ super().__init__(node, mem_info)
+ self.attributes["div_op"] = "Div"
self.attributes["activation"] = "Linear"
self.attributes["rescaling"] = "NoScaling"
self.config_template = str(
- ROOT / "templates" / "configuration" / "elemwise_config.jinja")
+ ROOT / "templates" / "configuration" / "div_config.jinja")
self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "elemwise_forward.jinja")
+ ROOT / "templates" / "kernel_forward" / "div_forward.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "kernels" / "elemwise.hpp"),
+ str(ROOT / "kernels" / "div.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):
def __init__(self, node, mem_info):
@@ -295,3 +334,74 @@ class FcCPP(ExportNodeCpp):
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.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)
+ self.attributes["activation"] = "Linear"
+ self.attributes["rescaling"] = "NoScaling"
+ self.config_template = str(
+ ROOT / "templates" / "configuration" / "matmul_config.jinja")
+ self.forward_template = str(
+ ROOT / "templates" / "kernel_forward" / "matmul_forward.jinja")
+ self.include_list = []
+ self.kernels_to_copy = [
+ str(ROOT / "kernels" / "matmul.hpp"),
+ str(ROOT / "kernels" / "activation.hpp"),
+ str(ROOT / "kernels" / "rescaling.hpp")
+ ]
+
+@ExportLibCpp.register("Erf", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
+class ErfCPP(ExportNodeCpp):
+ def __init__(self, node, mem_info):
+ super().__init__(node, mem_info)
+ self.attributes["activation"] = "Linear"
+ self.attributes["rescaling"] = "NoScaling"
+ self.config_template = str(
+ ROOT / "templates" / "configuration" / "erf_config.jinja")
+ self.forward_template = str(
+ ROOT / "templates" / "kernel_forward" / "erf_forward.jinja")
+ self.include_list = []
+ self.kernels_to_copy = [
+ str(ROOT / "kernels" / "erf.hpp"),
+ str(ROOT / "kernels" / "activation.hpp"),
+ str(ROOT / "kernels" / "rescaling.hpp")
+ ]
+
+@ExportLibCpp.register("Transpose", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
+class TransposeCPP(ExportNodeCpp):
+ def __init__(self, node, mem_info):
+ super().__init__(node, mem_info)
+ # Get parameter permutation from transpose
+ self.attributes["output_dims_order"] = self.operator.attr.get_attr("output_dims_order")
+
+ self.attributes["activation"] = "Linear"
+ self.attributes["rescaling"] = "NoScaling"
+ self.config_template = str(
+ ROOT / "templates" / "configuration" / "transpose_config.jinja")
+ self.forward_template = str(
+ ROOT / "templates" / "kernel_forward" / "transpose_forward.jinja")
+ self.include_list = []
+ self.kernels_to_copy = [
+ str(ROOT / "kernels" / "transpose.hpp"),
+ str(ROOT / "kernels" / "activation.hpp"),
+ str(ROOT / "kernels" / "rescaling.hpp")
+ ]
+
+@ExportLibCpp.register("BatchNorm2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
+class BatchNorm2DCPP(ExportNodeCpp):
+ def __init__(self, node, mem_info):
+ super().__init__(node, mem_info)
+ self.attributes["activation"] = "Linear"
+ self.attributes["rescaling"] = "NoScaling"
+ self.config_template = str(
+ ROOT / "templates" / "configuration" / "batchnorm_config.jinja")
+ self.forward_template = str(
+ ROOT / "templates" / "kernel_forward" / "batchnorm_forward.jinja")
+ self.include_list = []
+ self.kernels_to_copy = [
+ str(ROOT / "kernels" / "batchnorm.hpp"),
+ str(ROOT / "kernels" / "activation.hpp"),
+ str(ROOT / "kernels" / "rescaling.hpp")
+ ]
\ No newline at end of file
--
GitLab
From f64c927a746af341cd70a75937ced5992cb4827a Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Thu, 27 Mar 2025 10:26:36 +0000
Subject: [PATCH 13/22] Added new parameter groups
---
aidge_export_cpp/kernels/convolution.hpp | 149 +++++++++--------------
1 file changed, 56 insertions(+), 93 deletions(-)
diff --git a/aidge_export_cpp/kernels/convolution.hpp b/aidge_export_cpp/kernels/convolution.hpp
index efc7ee7..b623369 100644
--- a/aidge_export_cpp/kernels/convolution.hpp
+++ b/aidge_export_cpp/kernels/convolution.hpp
@@ -6,114 +6,77 @@
#include "network/utils.hpp"
#include "kernels/macs.hpp"
#include "kernels/activation.hpp"
+#include <omp.h>
+#include <iostream>
+
+// Weights index en NHWC
+constexpr int inds_pos(int n, int c, int h, int w, int N, int C, int H, int W) {
+ return n * (H * W * C) +
+ h * (W * C) +
+ w * C +
+ c;
+}
+// Image index in CHW
+constexpr int inds_pos(int c, int h, int w, int C, int H, int W) {
+ return c * (H * W) +
+ h * W +
+ w;
+}
-template<int NB_CHANNELS,
- int CHANNELS_HEIGHT, int CHANNELS_WIDTH,
- int NB_OUTPUTS,
- int OUTPUTS_HEIGHT, int OUTPUTS_WIDTH,
+template<int NB_CHANNELS,
+ int IN_HEIGHT, int IN_WIDTH,
+ int NB_OUTPUTS, int GROUPS,
+ int OUT_HEIGHT, int OUT_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 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,
+ const Input_T* __restrict inputs,
Output_T* __restrict outputs,
const Weight_T* __restrict weights,
const Bias_T* __restrict biases,
const Rescaling_T& __restrict rescaling)
{
- 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;
-
-#pragma omp parallel for collapse(2)
- 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);
- int oOffset = NB_OUTPUTS * oPos;
-
- // <--
-
- Bias_T weightedSum = biases[output];
- 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);
- int iOffset = NB_CHANNELS * iPos;
-
- const int wOffset = (output*KERNEL_HEIGHT + sy) * KERNEL_WIDTH * NB_CHANNELS;
-
- if (DILATION_X == 1 && ((PADDING_X == 0 && OUTPUTS_WIDTH == OUTPUTS_WIDTH_NOPAD)
- || sxMax - sxMin == KERNEL_WIDTH))
- {
- macsOnRange<KERNEL_WIDTH * NB_CHANNELS>(
- inputs + iOffset,
- 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;
- }
-
- 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,
- weightedSum);
- }
- }
- }
-
- outputs[oOffset + output] = activation_forward_value<Output_T>(weightedSum, output, ACTIVATION, rescaling);
- }
- }
+ if (NB_CHANNELS % GROUPS != 0 || NB_OUTPUTS % GROUPS != 0) {
+ throw std::invalid_argument("Groups must be a divisor of both NB_CHANNELS and NB_OUTPUTS!");
+ }
+
+ int c_in_g = NB_CHANNELS / GROUPS;
+ int c_out_g = NB_OUTPUTS / GROUPS;
+ #pragma omp parallel for
+ for (int oc = 0; oc < NB_OUTPUTS; oc++) {
+ int g_oc = oc / c_out_g;
+ #pragma omp parallel for
+ for (int i = 0; i < OUT_HEIGHT; ++i) {
+ #pragma omp parallel for
+ for (int j = 0; j < OUT_WIDTH; ++j) {
+ Output_T value = biases[oc];
+ #pragma omp parallel for
+ for (int ic = g_oc * c_in_g; ic < (g_oc + 1) * c_in_g; ++ic) {
+ #pragma omp parallel for
+ for (int m = 0; m < KERNEL_HEIGHT; ++m) {
+ #pragma omp parallel for
+ for (int n = 0; n < KERNEL_WIDTH; ++n) {
+ int i_p = i * STRIDE_X - PADDING_X + m * DILATION_X;
+ int j_p = j * STRIDE_Y - PADDING_Y + n * DILATION_Y;
+ if (i_p >= 0 && i_p < IN_HEIGHT && j_p >= 0 && j_p < IN_WIDTH) {
+ value += weights[inds_pos(oc, ic % c_in_g, m, n, NB_OUTPUTS, c_in_g, KERNEL_HEIGHT, KERNEL_WIDTH)] *
+ inputs[inds_pos(ic, i_p, j_p, NB_CHANNELS, IN_HEIGHT, IN_WIDTH)];
+ }
+ }
+ }
+ }
+ outputs[inds_pos(oc, i, j, NB_OUTPUTS, OUT_HEIGHT, OUT_WIDTH)] = activation_forward_value<Output_T>(value, oc, ACTIVATION, rescaling);
+ }
+ }
}
}
-
-#endif // __AIDGE_EXPORT_CPP_KERNELS_CONVOLUTION__
+#endif // __AIDGE_EXPORT_CPP_KERNELS_CONVOLUTION__
\ No newline at end of file
--
GitLab
From 05c867bef62d9508b7ef1eae0c33edec1718b887 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Fri, 28 Mar 2025 13:12:16 +0000
Subject: [PATCH 14/22] Cleaned file
---
aidge_export_cpp/kernels/elemwise.hpp | 4 ----
1 file changed, 4 deletions(-)
diff --git a/aidge_export_cpp/kernels/elemwise.hpp b/aidge_export_cpp/kernels/elemwise.hpp
index 6f73bc0..1b45c59 100644
--- a/aidge_export_cpp/kernels/elemwise.hpp
+++ b/aidge_export_cpp/kernels/elemwise.hpp
@@ -26,7 +26,6 @@ void elemwise_forward (
switch (ELEM_OP) {
case Add: {
- // std::cout <<"Add " << std::endl;
int ndim_a[SIZE_DIM_OUT];
int ndim_b[SIZE_DIM_OUT];
for (int i= 0; i<SIZE_DIM_OUT; i++){
@@ -182,7 +181,6 @@ void elemwise_forward (
break;
}
case Mul: {
- // std::cout<< "MUL " << std::endl;
int ndim_a[SIZE_DIM_OUT];
int ndim_b[SIZE_DIM_OUT];
@@ -258,8 +256,6 @@ void elemwise_forward (
break;
}
case Div: {
-
- std::cout<< "DIV " << std::endl;
int ndim_a[SIZE_DIM_OUT];
int ndim_b[SIZE_DIM_OUT];
for (int i= 0; i<SIZE_DIM_OUT; i++){
--
GitLab
From 81d728acbb0b50c07b5c86dff61c87d250b97113 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Fri, 28 Mar 2025 13:48:54 +0000
Subject: [PATCH 15/22] New transpose method used for exporting ConvNeXt onnx
---
aidge_export_cpp/kernels/transpose_diff.hpp | 57 +++++++++++++++++++
.../configuration/transpose_config.jinja | 4 +-
.../kernel_forward/transpose_forward.jinja | 3 +-
3 files changed, 59 insertions(+), 5 deletions(-)
create mode 100644 aidge_export_cpp/kernels/transpose_diff.hpp
diff --git a/aidge_export_cpp/kernels/transpose_diff.hpp b/aidge_export_cpp/kernels/transpose_diff.hpp
new file mode 100644
index 0000000..298e5d9
--- /dev/null
+++ b/aidge_export_cpp/kernels/transpose_diff.hpp
@@ -0,0 +1,57 @@
+#ifndef __AIDGE_EXPORT_CPP_KERNELS_TRANSPOSE__
+#define __AIDGE_EXPORT_CPP_KERNELS_TRANSPOSE__
+
+#include "network/typedefs.hpp"
+#include <cstring>
+#include <cstdio>
+#include <iostream>
+
+using namespace std;
+
+template< int INPUT_DIMS[], int PERM[], int OUTPUT_DIMS[],
+ int SIZE_OUTPUT_DIMS, int SIZE,
+ typename Input_T, typename Output_T>
+__attribute__((always_inline)) inline
+void transpose_forward (
+ const Input_T* __restrict inputs,
+ Output_T* __restrict outputs)
+ {
+
+ int newStrides[SIZE_OUTPUT_DIMS];
+ for (int i = 0; i<SIZE_OUTPUT_DIMS;++i){newStrides[i] = 1;}
+ for (int i = 0; i < SIZE_OUTPUT_DIMS; ++i) {
+ for (int j = i + 1; j < SIZE_OUTPUT_DIMS; ++j) {
+ newStrides[i] *= OUTPUT_DIMS[j];
+ }
+ }
+
+ int indices[SIZE_OUTPUT_DIMS];
+ for (int i = 0; i<SIZE_OUTPUT_DIMS;++i){indices[i] = 0;}
+// #pragma omp parallel for
+ for (int i = 0; i < SIZE; ++i) {
+ int idx = 0;
+// #pragma omp parallel for
+ for (int j = SIZE_OUTPUT_DIMS -1; j >=0; --j) {
+ idx += indices[PERM[j]] * newStrides[j];
+ }
+
+ outputs[idx] = inputs[i];
+
+
+ for (int j = SIZE_OUTPUT_DIMS - 1; j >= 0; --j) {
+ if (indices[j] < INPUT_DIMS[j] - 1) {
+ indices[j]++;
+ break;
+ }
+ else {
+ indices[j] = 0;
+ }
+ }
+ }
+
+
+}
+
+
+
+#endif // __AIDGE_EXPORT_CPP_KERNELS_TRANSPOSE__
diff --git a/aidge_export_cpp/templates/configuration/transpose_config.jinja b/aidge_export_cpp/templates/configuration/transpose_config.jinja
index c3eabc5..01e57b8 100644
--- a/aidge_export_cpp/templates/configuration/transpose_config.jinja
+++ b/aidge_export_cpp/templates/configuration/transpose_config.jinja
@@ -12,6 +12,4 @@ int {{name|upper}}_OUTPUT_DIMS[] = { {{ out_dims[0]|join(", ") }} };
int {{name|upper}}_INPUT_DIMS[] = { {{ in_dims[0]|join(", ") }} };
int {{name|upper}}_PERM[] = { {{ output_dims_order|join(", ") }} };
-
-
-#endif /* {{ name|upper }}_LAYER_H */
+#endif /* {{ name|upper }}_LAYER_H */
\ No newline at end of file
diff --git a/aidge_export_cpp/templates/kernel_forward/transpose_forward.jinja b/aidge_export_cpp/templates/kernel_forward/transpose_forward.jinja
index 2f8d939..2a5433c 100644
--- a/aidge_export_cpp/templates/kernel_forward/transpose_forward.jinja
+++ b/aidge_export_cpp/templates/kernel_forward/transpose_forward.jinja
@@ -7,5 +7,4 @@ transpose_forward<{{ name|upper }}_INPUT_DIMS,
{{name|upper}}_SIZE>
({{in_name[0]}}, {{out_name[0]}});
{% include "./_save_outputs.jinja" %}
-{% endfilter %}
-
+{% endfilter %}
\ No newline at end of file
--
GitLab
From 2a06f059865e9581daf58fece9f128c06a333ee2 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Fri, 28 Mar 2025 14:07:16 +0000
Subject: [PATCH 16/22] Add new parameter groups for convolution
---
.../kernels/convolution_groups.hpp | 84 ++++++++++
aidge_export_cpp/operators.py | 150 ++++++++----------
.../configuration/convolution_config.jinja | 1 -
.../kernel_forward/convolution_forward.jinja | 2 +-
4 files changed, 154 insertions(+), 83 deletions(-)
create mode 100644 aidge_export_cpp/kernels/convolution_groups.hpp
diff --git a/aidge_export_cpp/kernels/convolution_groups.hpp b/aidge_export_cpp/kernels/convolution_groups.hpp
new file mode 100644
index 0000000..7d73f79
--- /dev/null
+++ b/aidge_export_cpp/kernels/convolution_groups.hpp
@@ -0,0 +1,84 @@
+#ifndef __AIDGE_EXPORT_CPP_KERNELS_CONVOLUTION__
+#define __AIDGE_EXPORT_CPP_KERNELS_CONVOLUTION__
+
+#include "network/typedefs.hpp"
+#include "kernels/rescaling.hpp"
+#include "network/utils.hpp"
+#include "kernels/macs.hpp"
+#include "kernels/activation.hpp"
+#include <omp.h>
+#include <iostream>
+
+// Weights index en NHWC
+constexpr int inds_pos(int n, int c, int h, int w, int N, int C, int H, int W) {
+ return n * (H * W * C) +
+ h * (W * C) +
+ w * C +
+ c;
+}
+
+// Image index in CHW
+constexpr int inds_pos(int c, int h, int w, int C, int H, int W) {
+ return c * (H * W) +
+ h * W +
+ w;
+}
+
+
+
+template<int NB_CHANNELS,
+ int IN_HEIGHT, int IN_WIDTH,
+ int NB_OUTPUTS, int GROUPS,
+ int OUT_HEIGHT, int OUT_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_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)
+{
+
+ if (NB_CHANNELS % GROUPS != 0 || NB_OUTPUTS % GROUPS != 0) {
+ throw std::invalid_argument("Groups must be a divisor of both NB_CHANNELS and NB_OUTPUTS!");
+ }
+
+ int c_in_g = NB_CHANNELS / GROUPS;
+ int c_out_g = NB_OUTPUTS / GROUPS;
+ #pragma omp parallel for
+ for (int oc = 0; oc < NB_OUTPUTS; oc++) {
+ int g_oc = oc / c_out_g;
+ #pragma omp parallel for
+ for (int i = 0; i < OUT_HEIGHT; ++i) {
+ #pragma omp parallel for
+ for (int j = 0; j < OUT_WIDTH; ++j) {
+ Output_T value = biases[oc];
+ #pragma omp parallel for
+ for (int ic = g_oc * c_in_g; ic < (g_oc + 1) * c_in_g; ++ic) {
+ #pragma omp parallel for
+ for (int m = 0; m < KERNEL_HEIGHT; ++m) {
+ #pragma omp parallel for
+ for (int n = 0; n < KERNEL_WIDTH; ++n) {
+ int i_p = i * STRIDE_X - PADDING_X + m * DILATION_X;
+ int j_p = j * STRIDE_Y - PADDING_Y + n * DILATION_Y;
+ if (i_p >= 0 && i_p < IN_HEIGHT && j_p >= 0 && j_p < IN_WIDTH) {
+ value += weights[inds_pos(oc, ic % c_in_g, m, n, NB_OUTPUTS, c_in_g, KERNEL_HEIGHT, KERNEL_WIDTH)] *
+ inputs[inds_pos(ic, i_p, j_p, NB_CHANNELS, IN_HEIGHT, IN_WIDTH)];
+ }
+ }
+ }
+ }
+ outputs[inds_pos(oc, i, j, NB_OUTPUTS, OUT_HEIGHT, OUT_WIDTH)] = activation_forward_value<Output_T>(value, oc, ACTIVATION, rescaling);
+ }
+ }
+ }
+}
+#endif // __AIDGE_EXPORT_CPP_KERNELS_CONVOLUTION__
\ No newline at end of file
diff --git a/aidge_export_cpp/operators.py b/aidge_export_cpp/operators.py
index a511637..5ee9992 100644
--- a/aidge_export_cpp/operators.py
+++ b/aidge_export_cpp/operators.py
@@ -4,14 +4,27 @@ from pathlib import Path
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 import ExportLibCpp
##############################################
############## Export functions ##############
##############################################
-
-
+def numpy_dtype2ctype(dtype):
+ if dtype == np.int8:
+ return "int8_t"
+ elif dtype == np.int16:
+ return "int16_t"
+ elif dtype == np.int32:
+ return "int32_t"
+ elif dtype == np.int64:
+ return "int64_t"
+ elif dtype == np.float32:
+ return "float"
+ elif dtype == np.float64:
+ return "double"
+ # Add more dtype mappings as needed
+ else:
+ raise ValueError(f"Unsupported {dtype} dtype")
def export_params(name: str,
array: np.ndarray,
@@ -43,7 +56,7 @@ class ProducerCPP(ExportNode):
super().__init__(node, mem_info)
self.values = np.array(self.operator.get_output(0))
- if len(self.values.shape) == 4: # Note: export in HWC
+ if len(self.values.shape) == 4: # Note: export in HWC
self.values = np.transpose(self.values, (0, 2, 3, 1))
def export(self, export_folder: Path):
@@ -130,6 +143,24 @@ def _setup_conv2D(conv):
str(ROOT / "kernels" / "rescaling.hpp")
]
+
+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("Conv2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class ConvCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
@@ -137,7 +168,7 @@ class ConvCPP(ExportNodeCpp):
# No padding with Conv
# Use PaddedConv to add padding attribute
self.attributes["padding"] = [0, 0]
-
+ self.attributes["groups"] = 1
_setup_conv2D(self)
@ExportLibCpp.register_metaop("PaddedConv2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
@@ -156,25 +187,28 @@ class PaddedConvCPP(ExportNodeCpp):
).attr.stride_dims
self.attributes["dilation_dims"] = n.get_operator(
).attr.dilation_dims
-
+ self.attributes["groups"] = 1
_setup_conv2D(self)
-def _setup_elemwise_op(elemwise, op):
- """Common code (template and kernel setup) shared across all the different elementWise operator (Add, Sub,...)."""
+@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)
+ # 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
- 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")
- ]
+ self.attributes["groups"] = self.attributes["out_chan"][0]
+ _setup_conv2D(self)
@ExportLibCpp.register("Add", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class AddCPP(ExportNodeCpp):
@@ -197,6 +231,14 @@ class MulCPP(ExportNodeCpp):
_setup_elemwise_op(self, "Mul")
+@ExportLibCpp.register("Div", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
+class MulCPP(ExportNodeCpp):
+ def __init__(self, node, mem_info):
+ super().__init__(node, mem_info)
+
+ _setup_elemwise_op(self, "Div")
+
+
def _setup_pooling(pooling):
"""Common code (template and kernel setup) shared across all the different pooling operator."""
@@ -211,25 +253,6 @@ def _setup_pooling(pooling):
str(ROOT / "kernels" / "rescaling.hpp")
]
-@ExportLibCpp.register("Div", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
-class DivCPP(ExportNodeCpp):
- def __init__(self, node, mem_info):
- super().__init__(node, mem_info)
- self.attributes["div_op"] = "Div"
- self.attributes["activation"] = "Linear"
- self.attributes["rescaling"] = "NoScaling"
- self.config_template = str(
- ROOT / "templates" / "configuration" / "div_config.jinja")
- self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "div_forward.jinja")
- self.include_list = []
- self.kernels_to_copy = [
- str(ROOT / "kernels" / "div.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):
def __init__(self, node, mem_info):
@@ -297,23 +320,20 @@ class FcCPP(ExportNodeCpp):
str(ROOT / "kernels" / "rescaling.hpp")
]
-@ExportLibCpp.register("MatMul", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
-class MatMulCPP(ExportNodeCpp):
+@ExportLibCpp.register("Transpose", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.any)))
+class TransposeCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
super().__init__(node, mem_info)
- self.attributes["activation"] = "Linear"
- self.attributes["rescaling"] = "NoScaling"
self.config_template = str(
- ROOT / "templates" / "configuration" / "matmul_config.jinja")
+ ROOT / "templates" / "configuration" / "transpose_ND_config.jinja")
self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "matmul_forward.jinja")
+ ROOT / "templates" / "kernel_forward" / "transpose_ND_forward.jinja")
self.include_list = []
self.kernels_to_copy = [
- str(ROOT / "kernels" / "matmul.hpp"),
- str(ROOT / "kernels" / "activation.hpp"),
- str(ROOT / "kernels" / "rescaling.hpp")
+ str(ROOT / "kernels" / "transpose.hpp")
]
+
@ExportLibCpp.register("Erf", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class ErfCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
@@ -331,26 +351,6 @@ class ErfCPP(ExportNodeCpp):
str(ROOT / "kernels" / "rescaling.hpp")
]
-@ExportLibCpp.register("Transpose", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
-class TransposeCPP(ExportNodeCpp):
- def __init__(self, node, mem_info):
- super().__init__(node, mem_info)
- # Get parameter permutation from transpose
- self.attributes["output_dims_order"] = self.operator.attr.get_attr("output_dims_order")
-
- self.attributes["activation"] = "Linear"
- self.attributes["rescaling"] = "NoScaling"
- self.config_template = str(
- ROOT / "templates" / "configuration" / "transpose_config.jinja")
- self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "transpose_forward.jinja")
- self.include_list = []
- self.kernels_to_copy = [
- str(ROOT / "kernels" / "transpose.hpp"),
- str(ROOT / "kernels" / "activation.hpp"),
- str(ROOT / "kernels" / "rescaling.hpp")
- ]
-
@ExportLibCpp.register("BatchNorm2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
class BatchNorm2DCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
@@ -366,16 +366,4 @@ class BatchNorm2DCPP(ExportNodeCpp):
str(ROOT / "kernels" / "batchnorm.hpp"),
str(ROOT / "kernels" / "activation.hpp"),
str(ROOT / "kernels" / "rescaling.hpp")
- ]
-@ExportLibCpp.register("Transpose", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.any)))
-class TransposeCPP(ExportNodeCpp):
- def __init__(self, node, mem_info):
- super().__init__(node, mem_info)
- self.config_template = str(
- ROOT / "templates" / "configuration" / "transpose_ND_config.jinja")
- self.forward_template = str(
- ROOT / "templates" / "kernel_forward" / "transpose_ND_forward.jinja")
- self.include_list = []
- self.kernels_to_copy = [
- str(ROOT / "kernels" / "transpose.hpp")
]
\ No newline at end of file
diff --git a/aidge_export_cpp/templates/configuration/convolution_config.jinja b/aidge_export_cpp/templates/configuration/convolution_config.jinja
index 041e5b5..417f240 100644
--- a/aidge_export_cpp/templates/configuration/convolution_config.jinja
+++ b/aidge_export_cpp/templates/configuration/convolution_config.jinja
@@ -23,5 +23,4 @@ static const {{ rescaling }} {{ name|upper }}_RESCALING = {};
#define {{ name|upper }}_WEIGHTS_SIZE {{ weights_size }}
#define {{ name|upper }}_BIASES_SIZE {{ out_chan[0] }}
-
#endif /* {{ name|upper }}_LAYER_H */
diff --git a/aidge_export_cpp/templates/kernel_forward/convolution_forward.jinja b/aidge_export_cpp/templates/kernel_forward/convolution_forward.jinja
index 1760aa3..98b9e03 100644
--- a/aidge_export_cpp/templates/kernel_forward/convolution_forward.jinja
+++ b/aidge_export_cpp/templates/kernel_forward/convolution_forward.jinja
@@ -19,4 +19,4 @@ convolution_forward<{{ in_name[0]|upper }}_NB_CHANNELS,
({{in_name[0]}}, {{out_name[0]}}, {{in_name[1]}}, {{in_name[2]}}, {{name|upper}}_RESCALING);
{% include "./_save_outputs.jinja" %}
-{% endfilter %}
+{% endfilter %}
\ No newline at end of file
--
GitLab
From f06fd92b9b665e781d1fb0c9370d880a6751dce6 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Fri, 28 Mar 2025 14:09:16 +0000
Subject: [PATCH 17/22] Add for div operator
---
aidge_export_cpp/static/include/network/typedefs.hpp | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/aidge_export_cpp/static/include/network/typedefs.hpp b/aidge_export_cpp/static/include/network/typedefs.hpp
index acece91..9b83602 100644
--- a/aidge_export_cpp/static/include/network/typedefs.hpp
+++ b/aidge_export_cpp/static/include/network/typedefs.hpp
@@ -19,7 +19,8 @@ typedef enum {
typedef enum {
Add,
Sub,
- Mul
+ Mul,
+ Div
} ElemWise_T;
typedef enum {
--
GitLab
From 9fc76615f2efcb283fd5d2c3ceb1a2170b5ab665 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Tue, 1 Apr 2025 07:44:24 +0000
Subject: [PATCH 18/22] Delete debug includes
---
aidge_export_cpp/kernels/add.hpp | 3 ---
aidge_export_cpp/kernels/batchnorm.hpp | 1 -
aidge_export_cpp/kernels/convolution_groups.hpp | 1 -
aidge_export_cpp/kernels/erf.hpp | 7 +++----
aidge_export_cpp/kernels/matmul.hpp | 2 +-
5 files changed, 4 insertions(+), 10 deletions(-)
diff --git a/aidge_export_cpp/kernels/add.hpp b/aidge_export_cpp/kernels/add.hpp
index 03ba2c5..eb8a93b 100644
--- a/aidge_export_cpp/kernels/add.hpp
+++ b/aidge_export_cpp/kernels/add.hpp
@@ -3,9 +3,6 @@
#include "network/typedefs.hpp"
#include "kernels/activation.hpp"
-#include <iostream>
-#include <cassert>
-
template<int NB_ELTS,
diff --git a/aidge_export_cpp/kernels/batchnorm.hpp b/aidge_export_cpp/kernels/batchnorm.hpp
index 0ed5080..201ef16 100644
--- a/aidge_export_cpp/kernels/batchnorm.hpp
+++ b/aidge_export_cpp/kernels/batchnorm.hpp
@@ -5,7 +5,6 @@
#include "kernels/rescaling.hpp"
#include "kernels/activation.hpp"
#include <math.h>
-#include <iostream>
// WARNING: this kernel only works for 32-bits floating point values
diff --git a/aidge_export_cpp/kernels/convolution_groups.hpp b/aidge_export_cpp/kernels/convolution_groups.hpp
index 7d73f79..321ffc7 100644
--- a/aidge_export_cpp/kernels/convolution_groups.hpp
+++ b/aidge_export_cpp/kernels/convolution_groups.hpp
@@ -7,7 +7,6 @@
#include "kernels/macs.hpp"
#include "kernels/activation.hpp"
#include <omp.h>
-#include <iostream>
// Weights index en NHWC
constexpr int inds_pos(int n, int c, int h, int w, int N, int C, int H, int W) {
diff --git a/aidge_export_cpp/kernels/erf.hpp b/aidge_export_cpp/kernels/erf.hpp
index b509133..768f3b9 100644
--- a/aidge_export_cpp/kernels/erf.hpp
+++ b/aidge_export_cpp/kernels/erf.hpp
@@ -1,10 +1,9 @@
-#ifndef __AIDGE_EXPORT_CPP_KERNELS_ERP__
-#define __AIDGE_EXPORT_CPP_KERNELS_ERP__
+#ifndef __AIDGE_EXPORT_CPP_KERNELS_ERF__
+#define __AIDGE_EXPORT_CPP_KERNELS_ERF__
#include "network/typedefs.hpp"
#include <cmath>
#include <math.h>
-#include <iostream>
template<int _NB_ELTS,
typename Input_T, typename Output_T>
@@ -37,4 +36,4 @@ void erf_forward (
}
-#endif // __AIDGE_EXPORT_CPP_KERNELS_ERP_
\ No newline at end of file
+#endif // __AIDGE_EXPORT_CPP_KERNELS_ERF_
\ No newline at end of file
diff --git a/aidge_export_cpp/kernels/matmul.hpp b/aidge_export_cpp/kernels/matmul.hpp
index b284214..1403a01 100644
--- a/aidge_export_cpp/kernels/matmul.hpp
+++ b/aidge_export_cpp/kernels/matmul.hpp
@@ -3,7 +3,7 @@
#include "network/typedefs.hpp"
#include "kernels/activation.hpp"
-#include <iostream>
+
// Generic function for matmul and activation
template<int INPUT_A_DIMS[], int INPUT_B_DIMS[], int OUTPUT_DIMS[],
--
GitLab
From 4a4517821e3d9c70f6844f74b581e0c34304920a Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Tue, 1 Apr 2025 07:52:41 +0000
Subject: [PATCH 19/22] Add missing files
---
.../templates/configuration/batchnorm_config.jinja | 2 +-
.../templates/configuration/matmul_config.jinja | 6 ------
.../templates/kernel_forward/batchnorm_forward.jinja | 2 +-
3 files changed, 2 insertions(+), 8 deletions(-)
diff --git a/aidge_export_cpp/templates/configuration/batchnorm_config.jinja b/aidge_export_cpp/templates/configuration/batchnorm_config.jinja
index 701ba7c..3706ee6 100644
--- a/aidge_export_cpp/templates/configuration/batchnorm_config.jinja
+++ b/aidge_export_cpp/templates/configuration/batchnorm_config.jinja
@@ -8,4 +8,4 @@
#define {{ name|upper }}_ACTIVATION {{ activation }}
#define {{ name|upper }}_EPSILON {{ epsilon }}
-#endif /* {{ name|upper }}_LAYER_H */
+#endif /* {{ name|upper }}_LAYER_H */
\ No newline at end of file
diff --git a/aidge_export_cpp/templates/configuration/matmul_config.jinja b/aidge_export_cpp/templates/configuration/matmul_config.jinja
index 4e380e5..6ef27fe 100644
--- a/aidge_export_cpp/templates/configuration/matmul_config.jinja
+++ b/aidge_export_cpp/templates/configuration/matmul_config.jinja
@@ -7,7 +7,6 @@
{% include "./_meminfo.jinja" %}
{# For layer configuration -#}
-<<<<<<< HEAD
{% include "./_def_io.jinja" %}
{% include "./_meminfo.jinja" %}
#define {{ name|upper }}_B {{ in_dims[0][0]}}
@@ -24,11 +23,6 @@ int {{name|upper}}_OUTPUT_DIMS[] = { {{ out_dims[0]|join(", ") }} };
int {{name|upper}}_INPUT_A_DIMS[] = { {{ in_dims[0]|join(", ") }} };
int {{name|upper}}_INPUT_B_DIMS[] = { {{ in_dims[1]|join(", ") }} };
-=======
-#define {{ name|upper }}_M {{ in_dims[0][0] }}
-#define {{ name|upper }}_K {{ in_dims[0][1] }}
-#define {{ name|upper }}_N {{ in_dims[1][1] }}
->>>>>>> origin/dev
#define {{ name|upper }}_ACTIVATION {{ activation }}
{#- Calculate sizes #}
diff --git a/aidge_export_cpp/templates/kernel_forward/batchnorm_forward.jinja b/aidge_export_cpp/templates/kernel_forward/batchnorm_forward.jinja
index 5a759b8..a18e3a7 100644
--- a/aidge_export_cpp/templates/kernel_forward/batchnorm_forward.jinja
+++ b/aidge_export_cpp/templates/kernel_forward/batchnorm_forward.jinja
@@ -6,4 +6,4 @@ batchnorm_forward<{{ out_name[0]|upper }}_NB_OUTPUTS,
{{name|upper}}_ACTIVATION>
({{in_name[0]}}, {{out_name[0]}}, {{in_name[1]}}, {{in_name[2]}}, {{in_name[3]}}, {{in_name[4]}}, {{name|upper}}_EPSILON);
{% include "./_save_outputs.jinja" %}
-{% endfilter %}
+{% endfilter %}
\ No newline at end of file
--
GitLab
From b40da6dbf9c19b2bffb8aafb8e533dc59ffc1118 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Tue, 1 Apr 2025 07:53:16 +0000
Subject: [PATCH 20/22] Corrected typo
---
aidge_export_cpp/operators.py | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/aidge_export_cpp/operators.py b/aidge_export_cpp/operators.py
index 5ee9992..18089fd 100644
--- a/aidge_export_cpp/operators.py
+++ b/aidge_export_cpp/operators.py
@@ -232,7 +232,7 @@ class MulCPP(ExportNodeCpp):
_setup_elemwise_op(self, "Mul")
@ExportLibCpp.register("Div", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.float32)))
-class MulCPP(ExportNodeCpp):
+class DivCPP(ExportNodeCpp):
def __init__(self, node, mem_info):
super().__init__(node, mem_info)
--
GitLab
From dc0225cb2a571a9b995bcc31c151c3820f889f05 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Wed, 2 Apr 2025 14:25:23 +0000
Subject: [PATCH 21/22] Delete debug includes
---
aidge_export_cpp/kernels/div.hpp | 4 ----
aidge_export_cpp/kernels/elemwise.hpp | 3 ---
aidge_export_cpp/kernels/mul.hpp | 4 ----
aidge_export_cpp/kernels/pooling.hpp | 2 --
aidge_export_cpp/kernels/sub.hpp | 3 ---
aidge_export_cpp/kernels/transpose_diff.hpp | 7 ++-----
6 files changed, 2 insertions(+), 21 deletions(-)
diff --git a/aidge_export_cpp/kernels/div.hpp b/aidge_export_cpp/kernels/div.hpp
index f1ff7d0..44640aa 100644
--- a/aidge_export_cpp/kernels/div.hpp
+++ b/aidge_export_cpp/kernels/div.hpp
@@ -3,10 +3,6 @@
#include "network/typedefs.hpp"
#include "kernels/activation.hpp"
-#include <iostream>
-#include <cassert>
-
-
template<int NB_ELTS,
int INPUT_A_DIMS[], int INPUT_B_DIMS[], int OUTPUT_DIMS[],
diff --git a/aidge_export_cpp/kernels/elemwise.hpp b/aidge_export_cpp/kernels/elemwise.hpp
index 1b45c59..9b97959 100644
--- a/aidge_export_cpp/kernels/elemwise.hpp
+++ b/aidge_export_cpp/kernels/elemwise.hpp
@@ -3,9 +3,6 @@
#include "network/typedefs.hpp"
#include "kernels/activation.hpp"
-#include <iostream>
-#include <cassert>
-
template<int NB_ELTS, ElemWise_T ELEM_OP,
diff --git a/aidge_export_cpp/kernels/mul.hpp b/aidge_export_cpp/kernels/mul.hpp
index cbed0f6..b3ff9e1 100644
--- a/aidge_export_cpp/kernels/mul.hpp
+++ b/aidge_export_cpp/kernels/mul.hpp
@@ -3,10 +3,6 @@
#include "network/typedefs.hpp"
#include "kernels/activation.hpp"
-#include <iostream>
-#include <cassert>
-
-
template<int NB_ELTS,
int INPUT_A_DIMS[], int INPUT_B_DIMS[], int OUTPUT_DIMS[],
diff --git a/aidge_export_cpp/kernels/pooling.hpp b/aidge_export_cpp/kernels/pooling.hpp
index 14a2473..8f6de40 100644
--- a/aidge_export_cpp/kernels/pooling.hpp
+++ b/aidge_export_cpp/kernels/pooling.hpp
@@ -5,8 +5,6 @@
#include "network/utils.hpp"
#include <limits>
#include <stdexcept>
-#include <iostream>
-
void reorder_NCHW_NHWC_pool(const float* input, float* output, int N, int C, int H, int W, bool direct = true) {
diff --git a/aidge_export_cpp/kernels/sub.hpp b/aidge_export_cpp/kernels/sub.hpp
index 07637cd..2576edc 100644
--- a/aidge_export_cpp/kernels/sub.hpp
+++ b/aidge_export_cpp/kernels/sub.hpp
@@ -3,9 +3,6 @@
#include "network/typedefs.hpp"
#include "kernels/activation.hpp"
-#include <iostream>
-#include <cassert>
-
template<int NB_ELTS,
diff --git a/aidge_export_cpp/kernels/transpose_diff.hpp b/aidge_export_cpp/kernels/transpose_diff.hpp
index 298e5d9..712d9b4 100644
--- a/aidge_export_cpp/kernels/transpose_diff.hpp
+++ b/aidge_export_cpp/kernels/transpose_diff.hpp
@@ -2,9 +2,7 @@
#define __AIDGE_EXPORT_CPP_KERNELS_TRANSPOSE__
#include "network/typedefs.hpp"
-#include <cstring>
-#include <cstdio>
-#include <iostream>
+
using namespace std;
@@ -27,10 +25,9 @@ void transpose_forward (
int indices[SIZE_OUTPUT_DIMS];
for (int i = 0; i<SIZE_OUTPUT_DIMS;++i){indices[i] = 0;}
-// #pragma omp parallel for
+
for (int i = 0; i < SIZE; ++i) {
int idx = 0;
-// #pragma omp parallel for
for (int j = SIZE_OUTPUT_DIMS -1; j >=0; --j) {
idx += indices[PERM[j]] * newStrides[j];
}
--
GitLab
From f74ec2c3a86cfb5783036d9abee25c7baa9289c7 Mon Sep 17 00:00:00 2001
From: Matthew Newson <matthew.newson@cea.fr>
Date: Thu, 3 Apr 2025 12:10:41 +0000
Subject: [PATCH 22/22] Add ifdef pragma or delete unneeded pragma
---
aidge_export_cpp/kernels/add.hpp | 2 --
aidge_export_cpp/kernels/batchnorm.hpp | 6 ++++++
aidge_export_cpp/kernels/convolution_groups.hpp | 14 ++++++++------
aidge_export_cpp/kernels/erf.hpp | 3 ++-
aidge_export_cpp/kernels/mul.hpp | 2 --
5 files changed, 16 insertions(+), 11 deletions(-)
diff --git a/aidge_export_cpp/kernels/add.hpp b/aidge_export_cpp/kernels/add.hpp
index eb8a93b..52b58f5 100644
--- a/aidge_export_cpp/kernels/add.hpp
+++ b/aidge_export_cpp/kernels/add.hpp
@@ -55,12 +55,10 @@ void add_forward (
if (contiguousidx > 0) {
stride_post0[contiguousidx - 1] = 1;
stride_post1[contiguousidx - 1] = 1;
- #pragma omp parallel for
for (int i = contiguousidx -2; i != -1; --i) {
stride_post0[i] = stride_post0[i+1]*ndim_a[i+1];
stride_post1[i] = stride_post1[i+1]*ndim_b[i+1];
}
- #pragma omp parallel for
for (int i = 0; i < contiguousidx ; ++i) {
stride_step0[i] = (ndim_a[i] == 1) ? 1 - stride_post0[i] : 1;
stride_step1[i] = (ndim_b[i] == 1) ? 1 - stride_post1[i] : 1;
diff --git a/aidge_export_cpp/kernels/batchnorm.hpp b/aidge_export_cpp/kernels/batchnorm.hpp
index 201ef16..4100e6d 100644
--- a/aidge_export_cpp/kernels/batchnorm.hpp
+++ b/aidge_export_cpp/kernels/batchnorm.hpp
@@ -25,15 +25,21 @@ void batchnorm_forward (
{
int featureMapSize = OUTPUTS_HEIGHT * OUTPUTS_WIDTH;
+#ifdef _OPENMP
#pragma omp parallel for
+#endif
for (int ch = 0; ch < NB_OUTPUTS; ++ch) {
int ioIndex = ch * featureMapSize;
+#ifdef _OPENMP
#pragma omp parallel for
+#endif
for (int i = ioIndex; i < ioIndex + featureMapSize; i++) {
outputs[i] = biases[ch];
}
float var = sqrt(variances[ch] + epsilon);
+#ifdef _OPENMP
#pragma omp parallel for
+#endif
for (int feature = 0; feature < featureMapSize; ++feature) {
outputs[ioIndex + feature] += (scales[ch] * (inputs[ioIndex + feature] - means[ch]) / var);
}
diff --git a/aidge_export_cpp/kernels/convolution_groups.hpp b/aidge_export_cpp/kernels/convolution_groups.hpp
index 321ffc7..17cb1bf 100644
--- a/aidge_export_cpp/kernels/convolution_groups.hpp
+++ b/aidge_export_cpp/kernels/convolution_groups.hpp
@@ -6,7 +6,6 @@
#include "network/utils.hpp"
#include "kernels/macs.hpp"
#include "kernels/activation.hpp"
-#include <omp.h>
// Weights index en NHWC
constexpr int inds_pos(int n, int c, int h, int w, int N, int C, int H, int W) {
@@ -52,19 +51,22 @@ void convolution_forward(
int c_in_g = NB_CHANNELS / GROUPS;
int c_out_g = NB_OUTPUTS / GROUPS;
- #pragma omp parallel for
+#ifdef _OPENMP
+ #pragma omp parallel for collapse(3)
+#endif
for (int oc = 0; oc < NB_OUTPUTS; oc++) {
- int g_oc = oc / c_out_g;
- #pragma omp parallel for
for (int i = 0; i < OUT_HEIGHT; ++i) {
- #pragma omp parallel for
for (int j = 0; j < OUT_WIDTH; ++j) {
+ int g_oc = oc / c_out_g;
Output_T value = biases[oc];
- #pragma omp parallel for
for (int ic = g_oc * c_in_g; ic < (g_oc + 1) * c_in_g; ++ic) {
+#ifdef _OPENMP
#pragma omp parallel for
+#endif
for (int m = 0; m < KERNEL_HEIGHT; ++m) {
+#ifdef _OPENMP
#pragma omp parallel for
+#endif
for (int n = 0; n < KERNEL_WIDTH; ++n) {
int i_p = i * STRIDE_X - PADDING_X + m * DILATION_X;
int j_p = j * STRIDE_Y - PADDING_Y + n * DILATION_Y;
diff --git a/aidge_export_cpp/kernels/erf.hpp b/aidge_export_cpp/kernels/erf.hpp
index 768f3b9..88aafe2 100644
--- a/aidge_export_cpp/kernels/erf.hpp
+++ b/aidge_export_cpp/kernels/erf.hpp
@@ -19,8 +19,9 @@ void erf_forward (
double a5 = 1.061405429;
double p = 0.3275911;
-
+#ifdef _OPENMP
#pragma omp parallel for
+#endif
for (int i = 0; i < _NB_ELTS; ++i) {
int sign = 1;
if (inputs[i] < 0)
diff --git a/aidge_export_cpp/kernels/mul.hpp b/aidge_export_cpp/kernels/mul.hpp
index b3ff9e1..5c1ba62 100644
--- a/aidge_export_cpp/kernels/mul.hpp
+++ b/aidge_export_cpp/kernels/mul.hpp
@@ -54,12 +54,10 @@ void mul_forward (
if (contiguousidx > 0) {
stride_post0[contiguousidx - 1] = 1;
stride_post1[contiguousidx - 1] = 1;
- #pragma omp parallel for
for (int i = contiguousidx -2; i != -1; --i) {
stride_post0[i] = stride_post0[i+1]*ndim_a[i+1];
stride_post1[i] = stride_post1[i+1]*ndim_b[i+1];
}
- #pragma omp parallel for
for (int i = 0; i < contiguousidx ; ++i) {
stride_step0[i] = (ndim_a[i] == 1) ? 1 - stride_post0[i] : 1;
stride_step1[i] = (ndim_b[i] == 1) ? 1 - stride_post1[i] : 1;
--
GitLab