Skip to content

GitLab

Unverified Commit 3c31dc6c authored by PanZezhong1725's avatar PanZezhong1725 Committed by GitHub
Browse files

Merge pull request #45 from YdrMaster/main 

issue/52 代码格式化：机制和效果
parents 16dad776 e5ed9fa1
Hide whitespace changes
Inline Side-by-side
Showing with 542 additions and 297 deletions
.clang-format
View file @ 3c31dc6c
# Generated from CLion C/C++ Code Style settings
---
BasedOnStyle: LLVM
AccessModifierOffset: -4
AlignAfterOpenBracket: Align
# AlignConsecutiveAssignments: None
AlignOperands: Align
AllowAllArgumentsOnNextLine: false
AllowAllConstructorInitializersOnNextLine: false
AllowAllParametersOfDeclarationOnNextLine: false
AllowShortBlocksOnASingleLine: Always
AllowShortCaseLabelsOnASingleLine: false
AllowShortFunctionsOnASingleLine: All
AllowShortIfStatementsOnASingleLine: Always
AllowShortLambdasOnASingleLine: All
AllowShortLoopsOnASingleLine: true
AlwaysBreakAfterReturnType: None
AlwaysBreakTemplateDeclarations: No
BreakBeforeBraces: Custom
IndentWidth: 4 # 缩进宽度，LLVM 默认值为 2，改为 4
AccessModifierOffset: -4 # public/protected/private 访问控制符相对成员的偏移，与 IndentWidth 配合，LLVM 默认值为 -2
AlignOperands: AlignAfterOperator # 双目运算符的行间对齐，LLVM 默认值为 Align，改为带符号一起换行
BreakBeforeBinaryOperators: All # 在双目运算符之前换行，LLVM 默认值为 None，改为换行时总是把双目运算符放在行首，包括赋值（=）
ColumnLimit: 0 # 列宽限制，LLVM 默认值为 80，改为不限制
AllowShortBlocksOnASingleLine: Always # 是否允许短块（单个语句的块）不换行，LLVM 默认值为 Never，改为允许
AllowShortLoopsOnASingleLine: true # 是否允许短循环不换行，LLVM 默认值为 false，改为允许
InsertBraces: true # 是否在 if/for/while/switch 等语句后插入大括号，LLVM 默认值为 false，改为允许
BreakBeforeBraces: Custom # 大括号换行配置，LLVM 默认值为 LLVM，改为自定义以使 BraceWrapping 生效
BraceWrapping:
AfterCaseLabel: false
AfterClass: false
......@@ -23,44 +16,15 @@ BraceWrapping:
AfterEnum: false
AfterFunction: false
AfterNamespace: false
AfterObjCDeclaration: false
AfterStruct: false
AfterUnion: false
AfterExternBlock: false
BeforeCatch: false
BeforeElse: false
BeforeLambdaBody: false
BeforeWhile: false
IndentBraces: false
SplitEmptyFunction: false
SplitEmptyFunction: true
SplitEmptyRecord: true
BreakBeforeBinaryOperators: None
BreakBeforeTernaryOperators: true
BreakConstructorInitializers: BeforeColon
BreakInheritanceList: BeforeColon
ColumnLimit: 0
CompactNamespaces: true
ContinuationIndentWidth: 4
IndentCaseLabels: true
IndentPPDirectives: None
IndentWidth: 4
KeepEmptyLinesAtTheStartOfBlocks: true
MaxEmptyLinesToKeep: 2
NamespaceIndentation: All
ObjCSpaceAfterProperty: false
ObjCSpaceBeforeProtocolList: true
PointerAlignment: Right
ReflowComments: false
SpaceAfterCStyleCast: true
SpaceAfterLogicalNot: false
SpaceAfterTemplateKeyword: false
SpaceBeforeAssignmentOperators: true
SpaceBeforeCpp11BracedList: false
SpaceBeforeCtorInitializerColon: true
SpaceBeforeInheritanceColon: true
SpaceBeforeParens: ControlStatements
SpaceBeforeRangeBasedForLoopColon: true
SpaceInEmptyParentheses: false
SpacesBeforeTrailingComments: 0
SpacesInAngles: false
SpacesInCStyleCastParentheses: false
SpacesInContainerLiterals: false
SpacesInParentheses: false
SpacesInSquareBrackets: false
TabWidth: 4
UseTab: Never
SplitEmptyNamespace: true
.github/workflows/build.yml 0 → 100644
View file @ 3c31dc6c
name: Build and test
on:
pull_request:
push:
paths-ignore:
- '**.md'
- 'LICENSE'
jobs:
build:
name: Build
runs-on: ubuntu-latest
strategy:
fail-fast: false
matrix:
type: [debug, release]
steps:
- name: checkout code
uses: actions/checkout@v4
- name: install black
run: pip install black
- name: check format
run: python3 scripts/format.py --path src --check
- name: install xmake
uses: xmake-io/github-action-setup-xmake@v1
with:
xmake-version: latest
- name: configure xmake
run: xmake f -cv
- name: build with xmake
run: xmake build && xmake install
README.md
View file @ 3c31dc6c
# InfiniCore
InfiniCore是一个跨平台统一编程工具集，为不同芯片平台的功能（包括计算、运行时、通信等）提供统一 C 语言接口。目前支持的芯片包括CPU、英伟达GPU、华为昇腾NPU、寒武纪MLU、摩尔线程GPU、天数智芯GPU、沐曦GPU、曙光DCU、昆仑芯。
InfiniCore 是一个跨平台统一编程工具集，为不同芯片平台的功能（包括计算、运行时、通信等）提供统一 C 语言接口。目前支持的硬件和后端包括：
## 一、使用说明
- CPU；
- CUDA
- 英伟达 GPU；
- 摩尔线程 GPU；
- 天数智芯 GPU；
- 沐曦 GPU；
- 曙光 DCU；
- 华为昇腾 NPU；
- 寒武纪 MLU；
- 昆仑芯 XPU；
### 1. 配置
## 配置和使用
#### 查看当前配置
1. 项目配置
```xmake
xmake f -v
```
- 查看当前配置
#### 配置 CPU （默认配置）
```shell
xmake f -v
```
```xmake
xmake f -cv
```
- 配置 CPU（默认配置）
#### 配置加速卡
```shell
xmake f -cv
```
```xmake
# 英伟达
# 可以指定 CUDA 路径， 一般环境变量为 `CUDA_HOME` 或者 `CUDA_ROOT`
xmake f --nv-gpu=true --cuda=$CUDA_HOME -cv
- 配置加速卡
# 寒武纪
xmake f --cambricon-mlu=true -cv
```shell
# 英伟达
# 可以指定 CUDA 路径， 一般环境变量为 `CUDA_HOME` 或者 `CUDA_ROOT`
xmake f --nv-gpu=true --cuda=$CUDA_HOME -cv
# 华为昇腾
xmake f --ascend-npu=true -cv
```
# 寒武纪
xmake f --cambricon-mlu=true -cv
### 2. 编译安装
# 华为昇腾
xmake f --ascend-npu=true -cv
```
```xmake
xmake build && xmake install
# 默认安装路径为 $HOME/.infini
```
2. 编译安装
默认安装路径为 `$HOME/.infini`。
```shell
xmake build && xmake install
```
3. 设置环境变量
按输出提示设置 `INFINI_ROOT` 和 `LD_LIBRARY_PATH` 环境变量。
4. 运行算子测试
### 3. 设置环境变量
```shell
python test/infiniop/[operator].py [--cpu | --nvidia | --cambricon | --ascend]
```
按输出提示设置 `INFINI_ROOT` 和 `LD_LIBRARY_PATH` 环境变量。
## 开发指南
### 4. 运行算子测试
### 代码格式化
```bash
python test/infiniop/[operator].py [--cpu | --nvidia | --cambricon | --ascend]
本项目使用 [`scripts/format.py`](/scripts/format.py) 脚本实现代码格式化检查和操作。
使用
```shell
python scripts/format.py -h
```
查看脚本帮助信息：
```plaintext
usage: format.py [-h] [--ref REF] [--path [PATH ...]] [--check] [--c C] [--py PY]
options:
-h, --help show this help message and exit
--ref REF Git reference (commit hash) to compare against.
--path [PATH ...] Files to format or check.
--check Check files without modifying them.
--c C C formatter (default: clang-format-16)
--py PY Python formatter (default: black)
```
参数中：
- `ref` 和 `path` 控制格式化的文件范围
- 若 `ref` 和 `path` 都为空，格式化当前暂存（git added）的文件；
- 否则
- 若 `ref` 非空，将比较指定 commit 和当前代码的差异，只格式化修改过的文件；
- 若 `path` 非空，可传入多个路径（`--path p0 p1 p2`），只格式化指定路径及其子目录中的文件；
- 若设置 `--check`，将检查代码是否需要修改格式，不修改文件内容；
- 通过 `--c` 指定 c/c++ 格式化器，默认为 `clang-format-16`；
- 通过 `--python` 指定 python 格式化器 `black`；
include/infinicore.h
View file @ 3c31dc6c
......@@ -6,8 +6,7 @@
#define __INFINICORE_EXPORT_C__
#if defined(_WIN32)
#define __export __declspec(dllexport)
#elif defined(__GNUC__) && \
((__GNUC__ >= 4) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 3))
#elif defined(__GNUC__) && ((__GNUC__ >= 4) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 3))
#define __export __attribute__((visibility("default")))
#else
#define __export
......
include/infiniop/ops/causal_softmax.h
View file @ 3c31dc6c
......@@ -19,5 +19,4 @@ __C __export infiniopStatus_t infiniopCausalSoftmax(infiniopCausalSoftmaxDescrip
__C __export infiniopStatus_t infiniopDestroyCausalSoftmaxDescriptor(infiniopCausalSoftmaxDescriptor_t desc);
#endif
include/infiniop/ops/conv.h
View file @ 3c31dc6c
......@@ -21,5 +21,4 @@ __C __export infiniopStatus_t infiniopConv(infiniopConvDescriptor_t desc, void *
__C __export infiniopStatus_t infiniopDestroyConvDescriptor(infiniopConvDescriptor_t desc);
#endif
include/infiniop/ops/random_sample.h
View file @ 3c31dc6c
......@@ -22,5 +22,4 @@ __C __export infiniopStatus_t infiniopRandomSample(infiniopRandomSampleDescripto
__C __export infiniopStatus_t infiniopDestroyRandomSampleDescriptor(infiniopRandomSampleDescriptor_t desc);
#endif
include/infiniop/tensor_descriptor.h
View file @ 3c31dc6c
......@@ -21,4 +21,4 @@ __C __export infiniopStatus_t infiniopCreateTensorDescriptor(infiniopTensorDescr
__C __export infiniopStatus_t infiniopDestroyTensorDescriptor(infiniopTensorDescriptor_t desc);
#endif// __INFINIOP_TENSOR_DESCRIPTOR__
#endif // __INFINIOP_TENSOR_DESCRIPTOR__
scripts/format.py 0 → 100644
View file @ 3c31dc6c
import argparse
import subprocess
import os
from pathlib import Path
from colorama import Fore, Style
# 支持的文件类型
SUPPORTED_FILES = {
".h": "c",
".hh": "c",
".hpp": "c",
".c": "c",
".cc": "c",
".cpp": "c",
".cxx": "c",
".cu": "c",
".cuh": "c",
".mlu": "c",
".cl": "c",
".py": "py",
}
def format_file(file: Path, check: bool, formatter) -> bool:
formatter = formatter.get(SUPPORTED_FILES.get(file.suffix, None), None)
if not formatter:
return True # 文件类型不支持，跳过
try:
cmd = []
if formatter.startswith("clang-format"):
cmd = [formatter, "-style=file", "-i", file]
if check:
cmd.insert(2, "-dry-run")
process = subprocess.run(
cmd,
capture_output=True,
text=True,
check=True,
)
if process.stderr:
print(f"{Fore.YELLOW}{file} is not formatted.{Style.RESET_ALL}")
print(
f"Use {Fore.CYAN}{formatter} -style=file -i {file}{Style.RESET_ALL} to format it."
)
return False
else:
subprocess.run(
cmd,
capture_output=True,
text=True,
check=True,
)
print(f"{Fore.CYAN}Formatted: {file}{Style.RESET_ALL}")
elif formatter == "black":
cmd = [formatter, file]
if check:
cmd.insert(1, "--check")
process = subprocess.run(
cmd,
capture_output=True,
text=True,
check=True,
)
if process.stderr:
print(f"{Fore.YELLOW}{file} is not formatted.{Style.RESET_ALL}")
print(
f"Use {Fore.CYAN}{formatter} {file}{Style.RESET_ALL} to format it."
)
return False
else:
subprocess.run(
cmd,
capture_output=True,
text=True,
check=True,
)
print(f"{Fore.CYAN}Formatted: {file}{Style.RESET_ALL}")
except FileNotFoundError:
print(
f"{Fore.RED}Formatter {formatter} not found, {file} skipped.{Style.RESET_ALL}"
)
except subprocess.CalledProcessError as e:
print(f"{Fore.RED}Formatter {formatter} failed: {e}{Style.RESET_ALL}")
return True
def git_added_files():
"""获取所有已暂存更改的文件"""
try:
# 使用 git diff --cached --name-only 获取所有已添加到暂存区的文件
result = subprocess.run(
["git", "diff", "--cached", "--name-only"],
capture_output=True,
text=True,
check=True,
)
for file in result.stdout.splitlines():
yield Path(file.strip())
except subprocess.CalledProcessError as e:
print(f"{Fore.RED}Git diff failed: {e}{Style.RESET_ALL}")
def git_modified_since_ref(ref):
"""获取从指定的 Git 引用到当前状态的修改文件列表"""
try:
result = subprocess.run(
["git", "diff", f"{ref}..", "--diff-filter=AMR", "--name-only"],
capture_output=True,
text=True,
check=True,
)
for file in result.stdout.splitlines():
yield Path(file.strip())
except subprocess.CalledProcessError as e:
print(f"{Fore.RED}Git diff failed: {e}{Style.RESET_ALL}")
def list_files(paths):
"""递归获取指定路径下的所有文件"""
files = []
for path in paths:
if path.is_file():
yield path
elif path.is_dir():
for dirpath, _, filenames in os.walk(path):
for name in filenames:
yield Path(dirpath) / name
else:
print(
f"{Fore.RED}Error: {path} is not a file or directory.{Style.RESET_ALL}"
)
def filter_in_path(file: Path, path) -> bool:
"""判断文件是否在指定路径下"""
for p in path:
if file.is_relative_to(p):
return True
return False
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--ref", type=str, help="Git reference (commit hash) to compare against."
)
parser.add_argument(
"--path", nargs="*", type=Path, help="Files to format or check."
)
parser.add_argument(
"--check", action="store_true", help="Check files without modifying them."
)
parser.add_argument(
"--c", default="clang-format-16", help="C formatter (default: clang-format-16)"
)
parser.add_argument(
"--py", default="black", help="Python formatter (default: black)"
)
args = parser.parse_args()
if args.ref is None and args.path is None:
# Last commit.
print("{Fore.GREEN}Formating git added files.{Style.RESET_ALL}")
files = git_added_files()
else:
if args.ref is None:
print(f"{Fore.GREEN}Formating files in {args.path}.{Style.RESET_ALL}")
files = list_files(args.path)
elif args.path is None:
print(
f"{Fore.GREEN}Formating git modified files from {args.ref}.{Style.RESET_ALL}"
)
files = git_modified_since_ref(args.ref)
else:
print(
f"{Fore.GREEN}Formating git modified files from {args.ref} in {args.path}.{Style.RESET_ALL}"
)
files = (
file
for file in git_modified_since_ref(args.ref)
if filter_in_path(file, args.path)
)
formatted = True
for file in files:
if not format_file(
file,
args.check,
{
"c": args.c,
"py": args.py,
},
):
formatted = False
if not formatted:
exit(1)
if __name__ == "__main__":
main()
src/infiniop/devices/ascend/common_ascend.cc
View file @ 3c31dc6c
......@@ -31,115 +31,115 @@ infiniopStatus_t freeWorkspace(void *workspaceAddr) {
}
aclDataType toAclDataType(infiniDtype_t dt) {
if (dt == INFINI_DTYPE_I8)
if (dt == INFINI_DTYPE_I8) {
return aclDataType::ACL_INT8;
else if (dt == INFINI_DTYPE_I16)
} else if (dt == INFINI_DTYPE_I16) {
return aclDataType::ACL_INT16;
else if (dt == INFINI_DTYPE_I32)
} else if (dt == INFINI_DTYPE_I32) {
return aclDataType::ACL_INT32;
else if (dt == INFINI_DTYPE_I64)
} else if (dt == INFINI_DTYPE_I64) {
return aclDataType::ACL_INT64;
else if (dt == INFINI_DTYPE_U8)
} else if (dt == INFINI_DTYPE_U8) {
return aclDataType::ACL_UINT8;
else if (dt == INFINI_DTYPE_U16)
} else if (dt == INFINI_DTYPE_U16) {
return aclDataType::ACL_UINT16;
else if (dt == INFINI_DTYPE_U32)
} else if (dt == INFINI_DTYPE_U32) {
return aclDataType::ACL_UINT32;
else if (dt == INFINI_DTYPE_U64)
} else if (dt == INFINI_DTYPE_U64) {
return aclDataType::ACL_UINT64;
else if (dt == INFINI_DTYPE_F16)
} else if (dt == INFINI_DTYPE_F16) {
return aclDataType::ACL_FLOAT16;
else if (dt == INFINI_DTYPE_BF16)
} else if (dt == INFINI_DTYPE_BF16) {
return aclDataType::ACL_BF16;
else if (dt == INFINI_DTYPE_F32)
} else if (dt == INFINI_DTYPE_F32) {
return aclDataType::ACL_FLOAT;
else if (dt == INFINI_DTYPE_F64)
} else if (dt == INFINI_DTYPE_F64) {
return aclDataType::ACL_DOUBLE;
else
} else {
return aclDataType::ACL_DT_UNDEFINED;
}
}
const char *dataTypeToString(aclDataType dtype) {
switch (dtype) {
case ACL_DT_UNDEFINED:
return "ACL_DT_UNDEFINED";
case ACL_FLOAT:
return "ACL_FLOAT";
case ACL_FLOAT16:
return "ACL_FLOAT16";
case ACL_INT8:
return "ACL_INT8";
case ACL_INT32:
return "ACL_INT32";
case ACL_UINT8:
return "ACL_UINT8";
case ACL_INT16:
return "ACL_INT16";
case ACL_UINT16:
return "ACL_UINT16";
case ACL_UINT32:
return "ACL_UINT32";
case ACL_INT64:
return "ACL_INT64";
case ACL_UINT64:
return "ACL_UINT64";
case ACL_DOUBLE:
return "ACL_DOUBLE";
case ACL_BOOL:
return "ACL_BOOL";
case ACL_STRING:
return "ACL_STRING";
case ACL_COMPLEX64:
return "ACL_COMPLEX64";
case ACL_COMPLEX128:
return "ACL_COMPLEX128";
case ACL_BF16:
return "ACL_BF16";
case ACL_INT4:
return "ACL_INT4";
case ACL_UINT1:
return "ACL_UINT1";
case ACL_COMPLEX32:
return "ACL_COMPLEX32";
default:
return "UNKNOWN";
case ACL_DT_UNDEFINED:
return "ACL_DT_UNDEFINED";
case ACL_FLOAT:
return "ACL_FLOAT";
case ACL_FLOAT16:
return "ACL_FLOAT16";
case ACL_INT8:
return "ACL_INT8";
case ACL_INT32:
return "ACL_INT32";
case ACL_UINT8:
return "ACL_UINT8";
case ACL_INT16:
return "ACL_INT16";
case ACL_UINT16:
return "ACL_UINT16";
case ACL_UINT32:
return "ACL_UINT32";
case ACL_INT64:
return "ACL_INT64";
case ACL_UINT64:
return "ACL_UINT64";
case ACL_DOUBLE:
return "ACL_DOUBLE";
case ACL_BOOL:
return "ACL_BOOL";
case ACL_STRING:
return "ACL_STRING";
case ACL_COMPLEX64:
return "ACL_COMPLEX64";
case ACL_COMPLEX128:
return "ACL_COMPLEX128";
case ACL_BF16:
return "ACL_BF16";
case ACL_INT4:
return "ACL_INT4";
case ACL_UINT1:
return "ACL_UINT1";
case ACL_COMPLEX32:
return "ACL_COMPLEX32";
default:
return "UNKNOWN";
}
}
const char *formatToString(aclFormat format) {
switch (format) {
case ACL_FORMAT_UNDEFINED:
return "ACL_FORMAT_UNDEFINED";
case ACL_FORMAT_NCHW:
return "ACL_FORMAT_NCHW";
case ACL_FORMAT_NHWC:
return "ACL_FORMAT_NHWC";
case ACL_FORMAT_ND:
return "ACL_FORMAT_ND";
case ACL_FORMAT_NC1HWC0:
return "ACL_FORMAT_NC1HWC0";
case ACL_FORMAT_FRACTAL_Z:
return "ACL_FORMAT_FRACTAL_Z";
case ACL_FORMAT_NC1HWC0_C04:
return "ACL_FORMAT_NC1HWC0_C04";
case ACL_FORMAT_HWCN:
return "ACL_FORMAT_HWCN";
case ACL_FORMAT_NDHWC:
return "ACL_FORMAT_NDHWC";
case ACL_FORMAT_FRACTAL_NZ:
return "ACL_FORMAT_FRACTAL_NZ";
case ACL_FORMAT_NCDHW:
return "ACL_FORMAT_NCDHW";
case ACL_FORMAT_NDC1HWC0:
return "ACL_FORMAT_NDC1HWC0";
case ACL_FRACTAL_Z_3D:
return "ACL_FRACTAL_Z_3D";
case ACL_FORMAT_NC:
return "ACL_FORMAT_NC";
case ACL_FORMAT_NCL:
return "ACL_FORMAT_NCL";
default:
return "UNKNOWN";
case ACL_FORMAT_UNDEFINED:
return "ACL_FORMAT_UNDEFINED";
case ACL_FORMAT_NCHW:
return "ACL_FORMAT_NCHW";
case ACL_FORMAT_NHWC:
return "ACL_FORMAT_NHWC";
case ACL_FORMAT_ND:
return "ACL_FORMAT_ND";
case ACL_FORMAT_NC1HWC0:
return "ACL_FORMAT_NC1HWC0";
case ACL_FORMAT_FRACTAL_Z:
return "ACL_FORMAT_FRACTAL_Z";
case ACL_FORMAT_NC1HWC0_C04:
return "ACL_FORMAT_NC1HWC0_C04";
case ACL_FORMAT_HWCN:
return "ACL_FORMAT_HWCN";
case ACL_FORMAT_NDHWC:
return "ACL_FORMAT_NDHWC";
case ACL_FORMAT_FRACTAL_NZ:
return "ACL_FORMAT_FRACTAL_NZ";
case ACL_FORMAT_NCDHW:
return "ACL_FORMAT_NCDHW";
case ACL_FORMAT_NDC1HWC0:
return "ACL_FORMAT_NDC1HWC0";
case ACL_FRACTAL_Z_3D:
return "ACL_FRACTAL_Z_3D";
case ACL_FORMAT_NC:
return "ACL_FORMAT_NC";
case ACL_FORMAT_NCL:
return "ACL_FORMAT_NCL";
default:
return "UNKNOWN";
}
}
src/infiniop/devices/ascend/common_ascend.h
View file @ 3c31dc6c
......@@ -34,7 +34,6 @@ extern "C" {
return INFINIOP_STATUS_INTERNAL_ERROR; \
} while (0)
#ifdef __cplusplus
};
#endif
......
src/infiniop/devices/ascend/tensor_aclnn.cc
View file @ 3c31dc6c
......@@ -21,7 +21,6 @@ infiniopStatus_t aclnnTensorDescriptor::setDescriptor(aclDataType dtype, const s
return INFINIOP_STATUS_SUCCESS;
}
/// @brief Infer storage shape. For now this ruturns a 1D shape of the total tensor storage size.
/// We don't see why higher dimensional storage shape is ever needed. To change if necesary.
infiniopStatus_t aclnnTensorDescriptor::inferStorageShape() {
......@@ -93,8 +92,10 @@ char *aclnnTensorDescriptor::toString() {
// Assume bufferSize
size_t bufferSize = 1024 + this->ndim * 40 + this->storageNdim * 40;
char *buffer = (char *) malloc(bufferSize);
if (!buffer) return NULL;
char *buffer = (char *)malloc(bufferSize);
if (!buffer) {
return NULL;
}
// Write info into buffer
char *ptr = buffer;
......
src/infiniop/devices/cpu/common_cpu.cc
View file @ 3c31dc6c
......@@ -35,11 +35,10 @@ float f16_to_f32(uint16_t h) {
uint16_t f32_to_f16(float val) {
uint32_t f32;
memcpy(&f32, &val, sizeof(f32)); // Read the bits of the float32
uint16_t sign = (f32 >> 16) & 0x8000; // Extract the sign bit
int32_t exponent =
((f32 >> 23) & 0xFF) - 127; // Extract and de-bias the exponent
uint32_t mantissa = f32 & 0x7FFFFF; // Extract the mantissa (fraction part)
memcpy(&f32, &val, sizeof(f32)); // Read the bits of the float32
uint16_t sign = (f32 >> 16) & 0x8000; // Extract the sign bit
int32_t exponent = ((f32 >> 23) & 0xFF) - 127; // Extract and de-bias the exponent
uint32_t mantissa = f32 & 0x7FFFFF; // Extract the mantissa (fraction part)
if (exponent >= 31) { // Special cases for Inf and NaN
// NaN
......
src/infiniop/devices/cpu/common_cpu.h
View file @ 3c31dc6c
......@@ -19,7 +19,7 @@ size_t indexToReducedOffset(size_t flat_index, size_t ndim, int64_t const *broad
size_t indexToOffset(size_t flat_index, size_t ndim, size_t const *shape, int64_t const *strides);
/**
* get the total array size (element count) after applying padding for a
* get the total array size (element count) after applying padding for a
* ndim-ary tensor with the given shape
*/
size_t getPaddedSize(size_t ndim, size_t *shape, size_t const *pads);
......@@ -27,4 +27,4 @@ size_t getPaddedSize(size_t ndim, size_t *shape, size_t const *pads);
// calculate the padded shape and store the result in padded_shape
std::vector<size_t> getPaddedShape(size_t ndim, size_t const *shape, size_t const *pads);
#endif// __INFINIOP__COMMON_CPU_H__
#endif // __INFINIOP__COMMON_CPU_H__
src/infiniop/devices/cuda/common_cuda.cuh
View file @ 3c31dc6c
......@@ -47,18 +47,18 @@ struct InfiniopCudaHandle {
int compute_capability_minor;
};
template<typename T>
template <typename T>
void use_cublas(std::shared_ptr<Pool<cublasHandle_t>> cublas_handle_pool, int device_id, cudaStream_t stream, T const &f) {
auto handle = cublas_handle_pool->pop();
if (!handle) {
cublasCreate(&(*handle));
}
cublasSetStream(*handle, (cudaStream_t) stream);
cublasSetStream(*handle, (cudaStream_t)stream);
f(*handle);
cublas_handle_pool->push(std::move(*handle));
}
template<typename T>
template <typename T>
cudnnStatus_t use_cudnn(std::shared_ptr<Pool<cudnnHandle_t>> cudnn_handle_pool, int device_id, cudaStream_t stream, T const &f) {
auto handle = cudnn_handle_pool->pop();
if (!handle) {
......@@ -72,24 +72,24 @@ cudnnStatus_t use_cudnn(std::shared_ptr<Pool<cudnnHandle_t>> cudnn_handle_pool,
inline cudnnDataType_t getCudnnDtype(infiniDtype_t dt) {
switch (dt) {
case INFINI_DTYPE_F16:
return CUDNN_DATA_HALF;
case INFINI_DTYPE_F32:
return CUDNN_DATA_FLOAT;
case INFINI_DTYPE_F64:
return CUDNN_DATA_DOUBLE;
case INFINI_DTYPE_BF16:
return CUDNN_DATA_BFLOAT16;
case INFINI_DTYPE_I8:
return CUDNN_DATA_INT8;
case INFINI_DTYPE_I32:
return CUDNN_DATA_INT32;
case INFINI_DTYPE_I64:
return CUDNN_DATA_INT64;
case INFINI_DTYPE_U8:
return CUDNN_DATA_UINT8;
default:
return CUDNN_DATA_FLOAT;
case INFINI_DTYPE_F16:
return CUDNN_DATA_HALF;
case INFINI_DTYPE_F32:
return CUDNN_DATA_FLOAT;
case INFINI_DTYPE_F64:
return CUDNN_DATA_DOUBLE;
case INFINI_DTYPE_BF16:
return CUDNN_DATA_BFLOAT16;
case INFINI_DTYPE_I8:
return CUDNN_DATA_INT8;
case INFINI_DTYPE_I32:
return CUDNN_DATA_INT32;
case INFINI_DTYPE_I64:
return CUDNN_DATA_INT64;
case INFINI_DTYPE_U8:
return CUDNN_DATA_UINT8;
default:
return CUDNN_DATA_FLOAT;
}
}
......@@ -118,4 +118,4 @@ inline __device__ __host__ size_t indexToOffset(size_t flat_index, size_t ndim,
return res;
}
#endif// __INFINIOP_COMMON_CUDA_H__
#endif // __INFINIOP_COMMON_CUDA_H__
src/infiniop/devices/pool.h
View file @ 3c31dc6c
......@@ -5,7 +5,7 @@
#include <mutex>
#include <optional>
template<class T>
template <class T>
class Pool {
public:
Pool() : _head(nullptr) {}
......@@ -21,7 +21,7 @@ public:
void push(T &&val) const {
Node<T> *new_node = new Node<T>(std::move(val));
new_node->next = _head.load();
while (!_head.compare_exchange_weak(new_node->next, new_node));
while (!_head.compare_exchange_weak(new_node->next, new_node)) {}
}
std::optional<T> pop() const {
......@@ -37,7 +37,7 @@ public:
}
private:
template<class U>
template <class U>
struct Node {
U data;
Node<U> *next;
......
src/infiniop/ops/causal_softmax/operator.cc
View file @ 3c31dc6c
......@@ -6,35 +6,35 @@ __C infiniopStatus_t infiniopCreateCausalSoftmaxDescriptor(
infiniopTensorDescriptor_t y_desc) {
switch (handle->device) {
#ifdef ENABLE_CPU
case DevCpu:
return cpuCreateCausalSoftmaxDescriptor(handle, (CausalSoftmaxCpuDescriptor_t *) desc_ptr, y_desc);
case DevCpu:
return cpuCreateCausalSoftmaxDescriptor(handle, (CausalSoftmaxCpuDescriptor_t *)desc_ptr, y_desc);
#endif
#ifdef ENABLE_NV_GPU
case DevNvGpu: {
return cudaCreateCausalSoftmaxDescriptor((CudaHandle_t)handle, (CausalSoftmaxCudaDescriptor_t *) desc_ptr, y_desc);
}
case DevNvGpu: {
return cudaCreateCausalSoftmaxDescriptor((CudaHandle_t)handle, (CausalSoftmaxCudaDescriptor_t *)desc_ptr, y_desc);
}
#endif
#ifdef ENABLE_CAMBRICON_MLU
case DevCambriconMlu: {
return bangCreateCausalSoftmaxDescriptor((BangHandle_t) handle, (CausalSoftmaxBangDescriptor_t *) desc_ptr, y_desc);
// return cnnlCreateCausalSoftmaxDescriptor((BangHandle_t) handle, (CausalSoftmaxCnnlDescriptor_t *) desc_ptr, y_desc);
}
case DevCambriconMlu: {
return bangCreateCausalSoftmaxDescriptor((BangHandle_t)handle, (CausalSoftmaxBangDescriptor_t *)desc_ptr, y_desc);
// return cnnlCreateCausalSoftmaxDescriptor((BangHandle_t) handle, (CausalSoftmaxCnnlDescriptor_t *) desc_ptr, y_desc);
}
#endif
#ifdef ENABLE_ASCEND_NPU
case DevAscendNpu: {
return aclnnCreateCausalSoftmaxDescriptor((AscendHandle_t) handle, (CausalSoftmaxAclnnDescriptor_t *) desc_ptr, y_desc);
}
case DevAscendNpu: {
return aclnnCreateCausalSoftmaxDescriptor((AscendHandle_t)handle, (CausalSoftmaxAclnnDescriptor_t *)desc_ptr, y_desc);
}
#endif
#ifdef ENABLE_METAX_GPU
case DevMetaxGpu: {
return macaCreateCausalSoftmaxDescriptor((MacaHandle_t) handle, (CausalSoftmaxMacaDescriptor_t *) desc_ptr, y_desc);
}
case DevMetaxGpu: {
return macaCreateCausalSoftmaxDescriptor((MacaHandle_t)handle, (CausalSoftmaxMacaDescriptor_t *)desc_ptr, y_desc);
}
#endif
#ifdef ENABLE_MTHREADS_GPU
case DevMthreadsGpu: {
return musaCreateCausalSoftmaxDescriptor((MusaHandle_t) handle, (CausalSoftmaxMusaDescriptor_t *) desc_ptr, y_desc);
}
case DevMthreadsGpu: {
return musaCreateCausalSoftmaxDescriptor((MusaHandle_t)handle, (CausalSoftmaxMusaDescriptor_t *)desc_ptr, y_desc);
}
#endif
}
return INFINIOP_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
......@@ -43,36 +43,36 @@ __C infiniopStatus_t infiniopCreateCausalSoftmaxDescriptor(
__C infiniopStatus_t infiniopGetCausalSoftmaxWorkspaceSize(infiniopCausalSoftmaxDescriptor_t desc, uint64_t *size) {
switch (desc->device) {
#ifdef ENABLE_CPU
case DevCpu:
return cpuGetCausalSoftmaxWorkspaceSize((CausalSoftmaxCpuDescriptor_t) desc, size);
case DevCpu:
return cpuGetCausalSoftmaxWorkspaceSize((CausalSoftmaxCpuDescriptor_t)desc, size);
#endif
#ifdef ENABLE_NV_GPU
case DevNvGpu: {
return cudaGetCausalSoftmaxWorkspaceSize((CausalSoftmaxCudaDescriptor_t) desc, size);
}
case DevNvGpu: {
return cudaGetCausalSoftmaxWorkspaceSize((CausalSoftmaxCudaDescriptor_t)desc, size);
}
#endif
#ifdef ENABLE_CAMBRICON_MLU
case DevCambriconMlu: {
return bangGetCausalSoftmaxWorkspaceSize((CausalSoftmaxBangDescriptor_t) desc, size);
// return cnnlGetCausalSoftmaxWorkspaceSize((CausalSoftmaxCnnlDescriptor_t) desc, size);
}
case DevCambriconMlu: {
return bangGetCausalSoftmaxWorkspaceSize((CausalSoftmaxBangDescriptor_t)desc, size);
// return cnnlGetCausalSoftmaxWorkspaceSize((CausalSoftmaxCnnlDescriptor_t) desc, size);
}
#endif
#ifdef ENABLE_ASCEND_NPU
case DevAscendNpu: {
return aclnnGetCausalSoftmaxWorkspaceSize((CausalSoftmaxAclnnDescriptor_t) desc, size);
}
case DevAscendNpu: {
return aclnnGetCausalSoftmaxWorkspaceSize((CausalSoftmaxAclnnDescriptor_t)desc, size);
}
#endif
#ifdef ENABLE_METAX_GPU
case DevMetaxGpu: {
return macaGetCausalSoftmaxWorkspaceSize((CausalSoftmaxMacaDescriptor_t) desc, size);
}
case DevMetaxGpu: {
return macaGetCausalSoftmaxWorkspaceSize((CausalSoftmaxMacaDescriptor_t)desc, size);
}
#endif
#ifdef ENABLE_MTHREADS_GPU
case DevMthreadsGpu: {
return musaGetCausalSoftmaxWorkspaceSize((CausalSoftmaxMusaDescriptor_t) desc, size);
}
case DevMthreadsGpu: {
return musaGetCausalSoftmaxWorkspaceSize((CausalSoftmaxMusaDescriptor_t)desc, size);
}
#endif
}
return INFINIOP_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
......@@ -81,35 +81,35 @@ __C infiniopStatus_t infiniopGetCausalSoftmaxWorkspaceSize(infiniopCausalSoftmax
__C infiniopStatus_t infiniopCausalSoftmax(infiniopCausalSoftmaxDescriptor_t desc, void *workspace, uint64_t workspace_size, void *data, void *stream) {
switch (desc->device) {
#ifdef ENABLE_CPU
case DevCpu:
return cpuCausalSoftmax((CausalSoftmaxCpuDescriptor_t) desc, workspace, workspace_size, data, stream);
case DevCpu:
return cpuCausalSoftmax((CausalSoftmaxCpuDescriptor_t)desc, workspace, workspace_size, data, stream);
#endif
#ifdef ENABLE_NV_GPU
case DevNvGpu: {
return cudaCausalSoftmax((CausalSoftmaxCudaDescriptor_t) desc, workspace, workspace_size, data, stream);
}
case DevNvGpu: {
return cudaCausalSoftmax((CausalSoftmaxCudaDescriptor_t)desc, workspace, workspace_size, data, stream);
}
#endif
#ifdef ENABLE_CAMBRICON_MLU
case DevCambriconMlu: {
return bangCausalSoftmax((CausalSoftmaxBangDescriptor_t) desc, workspace, workspace_size, data, stream);
// return cnnlCausalSoftmax((CausalSoftmaxCnnlDescriptor_t) desc, workspace, workspace_size, data, stream);
}
case DevCambriconMlu: {
return bangCausalSoftmax((CausalSoftmaxBangDescriptor_t)desc, workspace, workspace_size, data, stream);
// return cnnlCausalSoftmax((CausalSoftmaxCnnlDescriptor_t) desc, workspace, workspace_size, data, stream);
}
#endif
#ifdef ENABLE_ASCEND_NPU
case DevAscendNpu: {
return aclnnCausalSoftmax((CausalSoftmaxAclnnDescriptor_t) desc, workspace, workspace_size, data, stream);
}
case DevAscendNpu: {
return aclnnCausalSoftmax((CausalSoftmaxAclnnDescriptor_t)desc, workspace, workspace_size, data, stream);
}
#endif
#ifdef ENABLE_METAX_GPU
case DevMetaxGpu: {
return macaCausalSoftmax((CausalSoftmaxMacaDescriptor_t) desc, workspace, workspace_size, data, stream);
}
case DevMetaxGpu: {
return macaCausalSoftmax((CausalSoftmaxMacaDescriptor_t)desc, workspace, workspace_size, data, stream);
}
#endif
#ifdef ENABLE_MTHREADS_GPU
case DevMthreadsGpu: {
return musaCausalSoftmax((CausalSoftmaxMusaDescriptor_t) desc, workspace, workspace_size, data, stream);
}
case DevMthreadsGpu: {
return musaCausalSoftmax((CausalSoftmaxMusaDescriptor_t)desc, workspace, workspace_size, data, stream);
}
#endif
}
return INFINIOP_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
......@@ -118,34 +118,34 @@ __C infiniopStatus_t infiniopCausalSoftmax(infiniopCausalSoftmaxDescriptor_t des
__C infiniopStatus_t infiniopDestroyCausalSoftmaxDescriptor(infiniopCausalSoftmaxDescriptor_t desc) {
switch (desc->device) {
#ifdef ENABLE_CPU
case DevCpu:
return cpuDestroyCausalSoftmaxDescriptor((CausalSoftmaxCpuDescriptor_t) desc);
case DevCpu:
return cpuDestroyCausalSoftmaxDescriptor((CausalSoftmaxCpuDescriptor_t)desc);
#endif
#ifdef ENABLE_NV_GPU
case DevNvGpu: {
return cudaDestroyCausalSoftmaxDescriptor((CausalSoftmaxCudaDescriptor_t) desc);
}
case DevNvGpu: {
return cudaDestroyCausalSoftmaxDescriptor((CausalSoftmaxCudaDescriptor_t)desc);
}
#endif
#ifdef ENABLE_CAMBRICON_MLU
case DevCambriconMlu: {
return bangDestroyCausalSoftmaxDescriptor((CausalSoftmaxBangDescriptor_t) desc);
// return cnnlDestroyCausalSoftmaxDescriptor((CausalSoftmaxCnnlDescriptor_t) desc);
}
case DevCambriconMlu: {
return bangDestroyCausalSoftmaxDescriptor((CausalSoftmaxBangDescriptor_t)desc);
// return cnnlDestroyCausalSoftmaxDescriptor((CausalSoftmaxCnnlDescriptor_t) desc);
}
#endif
#ifdef ENABLE_ASCEND_NPU
case DevAscendNpu: {
return aclnnDestroyCausalSoftmaxDescriptor((CausalSoftmaxAclnnDescriptor_t) desc);
}
case DevAscendNpu: {
return aclnnDestroyCausalSoftmaxDescriptor((CausalSoftmaxAclnnDescriptor_t)desc);
}
#endif
#ifdef ENABLE_METAX_GPU
case DevMetaxGpu: {
return macaDestroyCausalSoftmaxDescriptor((CausalSoftmaxMacaDescriptor_t) desc);
}
case DevMetaxGpu: {
return macaDestroyCausalSoftmaxDescriptor((CausalSoftmaxMacaDescriptor_t)desc);
}
#endif
#ifdef ENABLE_MTHREADS_GPU
case DevMthreadsGpu:
return musaDestroyCausalSoftmaxDescriptor((CausalSoftmaxMusaDescriptor_t) desc);
case DevMthreadsGpu:
return musaDestroyCausalSoftmaxDescriptor((CausalSoftmaxMusaDescriptor_t)desc);
#endif
}
return INFINIOP_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
......
src/infiniop/ops/matmul/ascend/matmul_aclnn.cc
View file @ 3c31dc6c
......@@ -123,17 +123,13 @@ infiniopStatus_t aclnnMatmul(MatmulAclnnDescriptor_t desc, void *workspace,
for (size_t i = 0; i < batch; i++) {
AclSetTensorAddr(desc->executor, 0, ta,
(char *)(a) + i * desc->info->a_matrix.stride *
infiniSizeof(desc->dtype));
(char *)(a) + i * desc->info->a_matrix.stride * infiniSizeof(desc->dtype));
AclSetTensorAddr(desc->executor, 1, tb,
(char *)(b) + i * desc->info->b_matrix.stride *
infiniSizeof(desc->dtype));
(char *)(b) + i * desc->info->b_matrix.stride * infiniSizeof(desc->dtype));
AclSetTensorAddr(desc->executor, 2, tc,
(char *)(c) + i * desc->info->c_matrix.stride *
infiniSizeof(desc->dtype));
(char *)(c) + i * desc->info->c_matrix.stride * infiniSizeof(desc->dtype));
AclSetTensorAddr(desc->executor, 3, tc,
(char *)(c) + i * desc->info->c_matrix.stride *
infiniSizeof(desc->dtype));
(char *)(c) + i * desc->info->c_matrix.stride * infiniSizeof(desc->dtype));
ret = aclnnGemm(workspace, workspaceSize, desc->executor, stream);
CHECK_RET(ret == ACL_SUCCESS,
LOG_PRINT("aclnnGemm failed. ERROR: %d\n", ret);
......
src/infiniop/ops/matmul/bang/matmul_cnnl.cc
View file @ 3c31dc6c
......@@ -73,8 +73,8 @@ bangDestroyMatmulDescriptor(infiniopMatmulBangDescriptor_t desc) {
}
void bangMatmulCnnl(infiniopMatmulBangDescriptor_t desc, void *workspace, void *c,
float beta, void const *a, void const *b, float alpha,
void *stream) {
float beta, void const *a, void const *b, float alpha,
void *stream) {
auto info = desc->info;
if (info.is_transed) {
std::swap(a, b);
......
src/infiniop/ops/matmul/blas.h
View file @ 3c31dc6c
......@@ -88,7 +88,7 @@ struct MatmulInfo {
return;
}
if (c_matrix.rows != a_matrix.rows || c_matrix.cols != b_matrix.cols || a_matrix.cols != b_matrix.rows){
if (c_matrix.rows != a_matrix.rows || c_matrix.cols != b_matrix.cols || a_matrix.cols != b_matrix.rows) {
*status = INFINIOP_STATUS_BAD_TENSOR_SHAPE;
return;
}
......@@ -113,4 +113,4 @@ struct MatmulInfo {
}
};
#endif// __BLAS_H__
#endif // __BLAS_H__
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment