Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
Xiaomi
Mace
提交
422360c1
Mace
项目概览
Xiaomi
/
Mace
通知
106
Star
40
Fork
27
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
列表
看板
标记
里程碑
合并请求
0
DevOps
流水线
流水线任务
计划
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
Mace
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
0
Issue
0
列表
看板
标记
里程碑
合并请求
0
合并请求
0
Pages
DevOps
DevOps
流水线
流水线任务
计划
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
流水线任务
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
提交
422360c1
编写于
3月 05, 2018
作者:
李
李寅
浏览文件
操作
浏览文件
下载
差异文件
Merge branch 'benchmark-dilation' into 'master'
Add dilation benchmark See merge request !261
上级
fe8b202b
c6170b0f
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
79 addition
and
83 deletion
+79
-83
mace/ops/conv_2d_benchmark.cc
mace/ops/conv_2d_benchmark.cc
+79
-68
mace/third_party/ncnn.BUILD
mace/third_party/ncnn.BUILD
+0
-15
未找到文件。
mace/ops/conv_2d_benchmark.cc
浏览文件 @
422360c1
...
...
@@ -20,6 +20,7 @@ static void Conv2d(int iters,
int
kernel_h
,
int
kernel_w
,
int
stride
,
int
dilation
,
Padding
padding
,
int
output_channels
)
{
mace
::
testing
::
StopTiming
();
...
...
@@ -46,7 +47,7 @@ static void Conv2d(int iters,
.
Output
(
"Output"
)
.
AddIntsArg
(
"strides"
,
{
stride
,
stride
})
.
AddIntArg
(
"padding"
,
padding
)
.
AddIntsArg
(
"dilations"
,
{
1
,
1
})
.
AddIntsArg
(
"dilations"
,
{
dilation
,
dilation
})
.
AddIntArg
(
"T"
,
static_cast
<
int
>
(
DataTypeToEnum
<
T
>::
value
))
.
Finalize
(
net
.
NewOperatorDef
());
}
else
{
...
...
@@ -57,7 +58,7 @@ static void Conv2d(int iters,
.
Output
(
"Output"
)
.
AddIntsArg
(
"strides"
,
{
stride
,
stride
})
.
AddIntArg
(
"padding"
,
padding
)
.
AddIntsArg
(
"dilations"
,
{
1
,
1
})
.
AddIntsArg
(
"dilations"
,
{
dilation
,
dilation
})
.
AddIntArg
(
"T"
,
static_cast
<
int
>
(
DataTypeToEnum
<
T
>::
value
))
.
Finalize
(
net
.
NewOperatorDef
());
}
...
...
@@ -79,71 +80,81 @@ static void Conv2d(int iters,
// approximate the amortized latency. The OpenCL runtime for Mali/Adreno is
// in-order.
#define BM_CONV_2D_MACRO(N, C, H, W, KH, KW, STRIDE, P, OC, TYPE, DEVICE) \
static void \
BM_CONV_2D_##N##_##C##_##H##_##W##_K##KH##x##KW##S##STRIDE##_##P##_##OC##_##TYPE##_##DEVICE( \
int iters) { \
const int64_t dilation = 1; \
const int64_t tot = static_cast<int64_t>(iters) * N * C * H * W; \
int64_t pad_h = 0, pad_w = 0; \
if (P == SAME) { \
pad_h = KH / 2; \
pad_w = KW / 2; \
} \
int64_t oh = \
(H + 2 * pad_h - KH - (KH - 1) * (dilation - 1)) / STRIDE + 1; \
int64_t ow = \
(W + 2 * pad_w - KW - (KW - 1) * (dilation - 1)) / STRIDE + 1; \
const int64_t macc = \
static_cast<int64_t>(iters) * N * OC * oh * ow * (KH * KW * C + 1); \
mace::testing::MaccProcessed(macc); \
mace::testing::BytesProcessed(tot *(sizeof(TYPE))); \
Conv2d<DEVICE, TYPE>(iters, N, C, H, W, KH, KW, STRIDE, mace::Padding::P, \
OC); \
} \
BENCHMARK( \
BM_CONV_2D_##N##_##C##_##H##_##W##_K##KH##x##KW##S##STRIDE##_##P##_##OC##_##TYPE##_##DEVICE)
#define BM_CONV_2D(N, C, H, W, KH, KW, S, P, OC) \
BM_CONV_2D_MACRO(N, C, H, W, KH, KW, S, P, OC, float, CPU); \
BM_CONV_2D_MACRO(N, C, H, W, KH, KW, S, P, OC, float, OPENCL); \
BM_CONV_2D_MACRO(N, C, H, W, KH, KW, S, P, OC, half, OPENCL);
BM_CONV_2D
(
1
,
512
,
15
,
15
,
1
,
1
,
1
,
VALID
,
1024
);
BM_CONV_2D
(
1
,
64
,
60
,
60
,
1
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
32
,
60
,
60
,
1
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
128
,
60
,
60
,
3
,
3
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
32
,
60
,
60
,
3
,
3
,
1
,
SAME
,
32
);
BM_CONV_2D
(
1
,
3
,
512
,
512
,
7
,
7
,
2
,
SAME
,
64
);
BM_CONV_2D
(
1
,
512
,
64
,
64
,
1
,
1
,
1
,
SAME
,
256
);
BM_CONV_2D
(
1
,
128
,
16
,
16
,
3
,
3
,
1
,
VALID
,
32
);
BM_CONV_2D
(
1
,
128
,
64
,
64
,
3
,
3
,
1
,
VALID
,
32
);
BM_CONV_2D
(
1
,
128
,
128
,
128
,
3
,
3
,
1
,
VALID
,
32
);
BM_CONV_2D
(
1
,
3
,
480
,
480
,
1
,
1
,
1
,
VALID
,
3
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
1
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
64
,
33
,
31
,
1
,
1
,
1
,
VALID
,
128
);
// Test bad alignments
BM_CONV_2D
(
1
,
3
,
512
,
512
,
1
,
1
,
1
,
VALID
,
3
);
BM_CONV_2D
(
1
,
32
,
112
,
112
,
1
,
1
,
1
,
VALID
,
64
);
BM_CONV_2D
(
1
,
64
,
56
,
56
,
1
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
256
,
28
,
28
,
1
,
1
,
1
,
VALID
,
256
);
BM_CONV_2D
(
1
,
1024
,
7
,
7
,
1
,
1
,
1
,
VALID
,
1024
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
3
,
3
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
64
,
33
,
31
,
3
,
3
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
3
,
512
,
512
,
3
,
3
,
1
,
VALID
,
3
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
3
,
3
,
1
,
SAME
,
128
);
BM_CONV_2D
(
1
,
64
,
33
,
31
,
3
,
3
,
1
,
SAME
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
3
,
3
,
2
,
VALID
,
128
);
BM_CONV_2D
(
1
,
3
,
512
,
512
,
3
,
3
,
2
,
VALID
,
3
);
BM_CONV_2D
(
1
,
64
,
33
,
31
,
3
,
3
,
2
,
VALID
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
3
,
3
,
2
,
SAME
,
128
);
BM_CONV_2D
(
1
,
64
,
33
,
31
,
3
,
3
,
2
,
SAME
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
5
,
5
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
31
,
5
,
5
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
5
,
5
,
1
,
SAME
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
31
,
5
,
5
,
1
,
SAME
,
128
);
#define BM_CONV_2D_MACRO(N, C, H, W, KH, KW, STRIDE, DILATION, P, OC, TYPE, \
DEVICE) \
static void \
BM_CONV_2D_##N##_##C##_##H##_##W##_K##KH##x##KW##S##STRIDE##D##DILATION##_##P##_##OC##_##TYPE##_##DEVICE( \
int iters) { \
const int64_t tot = static_cast<int64_t>(iters) * N * C * H * W; \
int64_t pad_h = 0, pad_w = 0; \
if (P == SAME) { \
pad_h = KH / 2; \
pad_w = KW / 2; \
} \
int64_t oh = \
(H + 2 * pad_h - KH - (KH - 1) * (DILATION - 1)) / STRIDE + 1; \
int64_t ow = \
(W + 2 * pad_w - KW - (KW - 1) * (DILATION - 1)) / STRIDE + 1; \
const int64_t macc = \
static_cast<int64_t>(iters) * N * OC * oh * ow * (KH * KW * C + 1); \
mace::testing::MaccProcessed(macc); \
mace::testing::BytesProcessed(tot *(sizeof(TYPE))); \
Conv2d<DEVICE, TYPE>(iters, N, C, H, W, KH, KW, STRIDE, DILATION, \
mace::Padding::P, OC); \
} \
BENCHMARK( \
BM_CONV_2D_##N##_##C##_##H##_##W##_K##KH##x##KW##S##STRIDE##D##DILATION##_##P##_##OC##_##TYPE##_##DEVICE)
#define BM_CONV_2D(N, C, H, W, KH, KW, S, D, P, OC) \
BM_CONV_2D_MACRO(N, C, H, W, KH, KW, S, D, P, OC, float, CPU); \
BM_CONV_2D_MACRO(N, C, H, W, KH, KW, S, D, P, OC, float, OPENCL); \
BM_CONV_2D_MACRO(N, C, H, W, KH, KW, S, D, P, OC, half, OPENCL);
BM_CONV_2D
(
1
,
256
,
64
,
64
,
3
,
3
,
1
,
1
,
VALID
,
256
);
BM_CONV_2D
(
1
,
512
,
15
,
15
,
1
,
1
,
1
,
1
,
VALID
,
1024
);
BM_CONV_2D
(
1
,
64
,
60
,
60
,
1
,
1
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
32
,
60
,
60
,
1
,
1
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
128
,
60
,
60
,
3
,
3
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
32
,
60
,
60
,
3
,
3
,
1
,
1
,
SAME
,
32
);
BM_CONV_2D
(
1
,
3
,
512
,
512
,
7
,
7
,
2
,
1
,
SAME
,
64
);
BM_CONV_2D
(
1
,
512
,
64
,
64
,
1
,
1
,
1
,
1
,
SAME
,
256
);
BM_CONV_2D
(
1
,
128
,
16
,
16
,
3
,
3
,
1
,
1
,
VALID
,
32
);
BM_CONV_2D
(
1
,
128
,
64
,
64
,
3
,
3
,
1
,
1
,
VALID
,
32
);
BM_CONV_2D
(
1
,
128
,
128
,
128
,
3
,
3
,
1
,
1
,
VALID
,
32
);
BM_CONV_2D
(
1
,
3
,
480
,
480
,
1
,
1
,
1
,
1
,
VALID
,
3
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
1
,
1
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
64
,
33
,
31
,
1
,
1
,
1
,
1
,
VALID
,
128
);
// Test bad alignments
BM_CONV_2D
(
1
,
3
,
512
,
512
,
1
,
1
,
1
,
1
,
VALID
,
3
);
BM_CONV_2D
(
1
,
32
,
112
,
112
,
1
,
1
,
1
,
1
,
VALID
,
64
);
BM_CONV_2D
(
1
,
64
,
56
,
56
,
1
,
1
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
256
,
28
,
28
,
1
,
1
,
1
,
1
,
VALID
,
256
);
BM_CONV_2D
(
1
,
1024
,
7
,
7
,
1
,
1
,
1
,
1
,
VALID
,
1024
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
3
,
3
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
64
,
33
,
31
,
3
,
3
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
3
,
512
,
512
,
3
,
3
,
1
,
1
,
VALID
,
3
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
3
,
3
,
1
,
1
,
SAME
,
128
);
BM_CONV_2D
(
1
,
64
,
33
,
31
,
3
,
3
,
1
,
1
,
SAME
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
3
,
3
,
2
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
3
,
512
,
512
,
3
,
3
,
2
,
1
,
VALID
,
3
);
BM_CONV_2D
(
1
,
64
,
33
,
31
,
3
,
3
,
2
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
3
,
3
,
2
,
1
,
SAME
,
128
);
BM_CONV_2D
(
1
,
64
,
33
,
31
,
3
,
3
,
2
,
1
,
SAME
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
5
,
5
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
31
,
5
,
5
,
1
,
1
,
VALID
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
32
,
5
,
5
,
1
,
1
,
SAME
,
128
);
BM_CONV_2D
(
1
,
64
,
32
,
31
,
5
,
5
,
1
,
1
,
SAME
,
128
);
// Dilation
BM_CONV_2D
(
1
,
32
,
256
,
256
,
3
,
3
,
1
,
2
,
VALID
,
32
);
BM_CONV_2D
(
1
,
32
,
256
,
256
,
3
,
3
,
1
,
4
,
VALID
,
32
);
// MobileNet
BM_CONV_2D
(
1
,
128
,
56
,
56
,
1
,
1
,
1
,
1
,
SAME
,
128
);
BM_CONV_2D
(
1
,
1024
,
7
,
7
,
1
,
1
,
1
,
1
,
SAME
,
1024
);
}
// namespace mace
mace/third_party/ncnn.BUILD
已删除
100644 → 0
浏览文件 @
fe8b202b
genrule(
name = "cmake",
outs = [
"build-android-aarch64/install/lib/libncnn.a",
"build-android-aarch64/install/include/blob.h",
"build-android-aarch64/install/include/cpu.h",
"build-android-aarch64/install/include/layer.h",
"build-android-aarch64/install/include/mat.h",
"build-android-aarch64/install/include/net.h",
"build-android-aarch64/install/include/opencv.h",
"build-android-aarch64/install/include/layer_type_enum.h",
"build-android-aarch64/install/include/platform.h",
],
cmd = "pwd; echo $$(dirname $(location cpu.h)); mkdir -p build-android-aarch64; pushd build-android-aarch64; cmake -DCMAKE_TOOLCHAIN_FILE=../external/ncnn/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_NATIVE_API_LEVEL=android-21 -DANDROID_FORCE_ARM_BUILD=OFF -DANDROID_STL_FORCE_FEATURES=OFF ../external/ncnn && make && make install; popd",
)
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录