From 1d453e58386ca95e5c3d54141dcb40e327af5db8 Mon Sep 17 00:00:00 2001
From: NAUD Maxence <maxence.naud@cea.fr>
Date: Mon, 15 Jan 2024 15:53:00 +0000
Subject: [PATCH] Slight optimization of TensorImpl_cpu
---
include/aidge/backend/cpu/data/TensorImpl.hpp | 49 +++++++++++--------
1 file changed, 29 insertions(+), 20 deletions(-)
diff --git a/include/aidge/backend/cpu/data/TensorImpl.hpp b/include/aidge/backend/cpu/data/TensorImpl.hpp
index 3fc67c33..c451b4a5 100644
--- a/include/aidge/backend/cpu/data/TensorImpl.hpp
+++ b/include/aidge/backend/cpu/data/TensorImpl.hpp
@@ -1,3 +1,14 @@
+/********************************************************************************
+ * Copyright (c) 2023 CEA-List
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ *
+ ********************************************************************************/
+
#ifndef AIDGE_CPU_DATA_TENSORIMPL_H_
#define AIDGE_CPU_DATA_TENSORIMPL_H_
@@ -10,9 +21,10 @@
#include "aidge/utils/future_std/span.hpp"
namespace Aidge {
+
template <class T>
class TensorImpl_cpu : public TensorImpl {
- private:
+private:
const Tensor &mTensor; // Impl needs to access Tensor information, but is not
// supposed to change it!
/// Pointer to the data and its capacity
@@ -20,7 +32,7 @@ class TensorImpl_cpu : public TensorImpl {
/// If this instance own the data, std::unique_ptr manages it
std::unique_ptr<T[]> mDataOwner;
- public:
+public:
static constexpr const char *Backend = "cpu";
TensorImpl_cpu(const Tensor &tensor) : TensorImpl(Backend), mTensor(tensor) {}
@@ -31,8 +43,8 @@ class TensorImpl_cpu : public TensorImpl {
std::size_t i = 0;
for (; i < mTensor.size() &&
- mData[i] == typedOtherImpl.data()[i];
- ++i) {
+ *(mData.data()+i) == *static_cast<const T*>(typedOtherImpl.rawPtr(i));
+ ++i) {
}
return i == mTensor.size();
}
@@ -41,23 +53,20 @@ class TensorImpl_cpu : public TensorImpl {
return std::make_unique<TensorImpl_cpu<T>>(tensor);
}
- // native interface
- auto data() const -> decltype(mData.data()) { return mData.data(); }
-
- std::size_t size() const override { return mData.size(); }
- std::size_t scalarSize() const override { return sizeof(T); }
+ inline std::size_t size() const noexcept override final { return mData.size(); }
+ inline std::size_t scalarSize() const noexcept override final { return sizeof(T); }
- void setDevice(DeviceIdx_t device) override {
+ void setDevice(DeviceIdx_t device) override final {
AIDGE_ASSERT(device == 0, "device cannot be != 0 for CPU backend");
}
- void copy(const void *src, NbElts_t length, NbElts_t offset = 0) override {
+ void copy(const void *src, NbElts_t length, NbElts_t offset = 0) override final {
AIDGE_ASSERT(length <= mData.size() || length <= mTensor.size(), "copy length is above capacity");
std::copy(static_cast<const T *>(src), static_cast<const T *>(src) + length,
static_cast<T *>(rawPtr()) + offset);
}
- void copyCast(const void *src, NbElts_t length, const DataType srcDt) override {
+ void copyCast(const void *src, NbElts_t length, const DataType srcDt) override final {
if (length == 0) {
return;
}
@@ -101,7 +110,7 @@ class TensorImpl_cpu : public TensorImpl {
std::copy(static_cast<const uint16_t*>(src), static_cast<const uint16_t*>(src) + length,
static_cast<T *>(rawPtr()));
break;
- case ataType::Int8:
+ case DataType::Int8:
std::copy(static_cast<const int8_t*>(src), static_cast<const int8_t*>(src) + length,
static_cast<T *>(rawPtr()));
break;
@@ -115,39 +124,39 @@ class TensorImpl_cpu : public TensorImpl {
}
}
- void copyFromDevice(const void *src, NbElts_t length, const std::pair<std::string, DeviceIdx_t>& device) override {
+ void copyFromDevice(const void *src, NbElts_t length, const std::pair<std::string, DeviceIdx_t>& device) override final {
AIDGE_ASSERT(device.first == Backend, "backend must match");
AIDGE_ASSERT(device.second == 0, "device cannot be != 0 for CPU backend");
copy(src, length);
}
- void copyFromHost(const void *src, NbElts_t length) override {
+ inline void copyFromHost(const void *src, NbElts_t length) override final {
copy(src, length);
}
- void copyToHost(void *dst, NbElts_t length) const override {
+ void copyToHost(void *dst, NbElts_t length) const override final {
AIDGE_ASSERT(length <= mData.size() || length <= mTensor.size(), "copy length is above capacity");
const T* src = static_cast<const T*>(rawPtr());
std::copy(static_cast<const T *>(src), static_cast<const T *>(src) + length,
static_cast<T *>(dst));
}
- void *rawPtr(NbElts_t offset = 0) override {
+ void *rawPtr(NbElts_t offset = 0) override final {
lazyInit();
return (mData.data() + offset);
};
- const void *rawPtr(NbElts_t offset = 0) const override {
+ const void *rawPtr(NbElts_t offset = 0) const override final {
AIDGE_ASSERT(mData.size() >= mTensor.size(), "accessing uninitialized const rawPtr");
return (mData.data() + offset);
};
- void *hostPtr(NbElts_t offset = 0) override {
+ void *hostPtr(NbElts_t offset = 0) override final {
lazyInit();
return (mData.data() + offset);
};
- const void *hostPtr(NbElts_t offset = 0) const override {
+ const void *hostPtr(NbElts_t offset = 0) const override final {
AIDGE_ASSERT(mData.size() >= mTensor.size(), "accessing uninitialized const hostPtr");
return (mData.data() + offset);
};
--
GitLab