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
PaddlePaddle
Paddle-Lite
提交
538ff20a
P
Paddle-Lite
项目概览
PaddlePaddle
/
Paddle-Lite
通知
331
Star
4
Fork
1
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
271
列表
看板
标记
里程碑
合并请求
78
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle-Lite
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
271
Issue
271
列表
看板
标记
里程碑
合并请求
78
合并请求
78
Pages
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
提交
Issue看板
提交
538ff20a
编写于
9月 01, 2018
作者:
L
liuruilong
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix conv add 1x5
上级
40642fe0
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
95 addition
and
38 deletion
+95
-38
metal/paddle-mobile/paddle-mobile/Executor.swift
metal/paddle-mobile/paddle-mobile/Executor.swift
+3
-2
metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvAddKernel.swift
...obile/paddle-mobile/Operators/Kernels/ConvAddKernel.swift
+6
-3
metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvAddMetal.metal
.../paddle-mobile/Operators/Kernels/metal/ConvAddMetal.metal
+86
-33
未找到文件。
metal/paddle-mobile/paddle-mobile/Executor.swift
浏览文件 @
538ff20a
...
...
@@ -14,7 +14,7 @@
import
Foundation
let
testTo
=
54
let
testTo
=
61
var
isTest
=
false
...
...
@@ -133,7 +133,8 @@ public class Executor<P: PrecisionType> {
print
(
" 第
\(
i
)
个 op: "
)
op
.
delogOutput
()
}
// self.ops[53].delogOutput()
// self.ops[59].delogOutput()
// self.ops[60].delogOutput()
return
...
...
metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvAddKernel.swift
浏览文件 @
538ff20a
...
...
@@ -17,6 +17,7 @@ import Foundation
class
ConvAddKernel
<
P
:
PrecisionType
>
:
Kernel
,
Computable
{
var
metalParam
:
MetalConvParam
!
required
init
(
device
:
MTLDevice
,
param
:
ConvAddParam
<
P
>
)
{
if
computePrecision
==
.
Float16
{
if
param
.
filter
.
width
==
1
&&
param
.
filter
.
height
==
1
{
super
.
init
(
device
:
device
,
inFunctionName
:
"conv_add_1x1_half"
)
...
...
@@ -30,6 +31,8 @@ class ConvAddKernel<P: PrecisionType>: Kernel, Computable {
super
.
init
(
device
:
device
,
inFunctionName
:
"conv_add_1x1"
)
}
else
if
param
.
filter
.
channel
==
1
{
super
.
init
(
device
:
device
,
inFunctionName
:
"depthwise_conv_add_3x3"
)
}
else
if
param
.
filter
.
width
==
1
&&
param
.
filter
.
height
==
5
{
super
.
init
(
device
:
device
,
inFunctionName
:
"conv_add_5x1"
)
}
else
{
super
.
init
(
device
:
device
,
inFunctionName
:
"conv_add_3x3"
)
}
...
...
@@ -37,11 +40,11 @@ class ConvAddKernel<P: PrecisionType>: Kernel, Computable {
fatalError
()
}
param
.
output
.
initTexture
(
device
:
device
,
inTranspose
:
[
0
,
2
,
3
,
1
],
computePrecision
:
computePrecision
)
let
offsetY
=
(
Int
(
param
.
dilations
[
1
])
*
(
param
.
filter
.
height
-
1
)
+
1
)
/
2
-
Int
(
param
.
paddings
[
1
]
)
let
offsetX
=
(
Int
(
param
.
dilations
[
0
])
*
(
param
.
filter
.
width
-
1
)
+
1
)
/
2
-
Int
(
param
.
paddings
[
0
])
let
offsetY
=
(
Int
(
param
.
dilations
[
1
])
*
(
param
.
filter
.
height
-
1
)
+
1
)
/
2
-
Int
(
param
.
paddings
[
1
]
)
param
.
output
.
initTexture
(
device
:
device
,
inTranspose
:
[
0
,
2
,
3
,
1
],
computePrecision
:
computePrecision
)
param
.
filter
.
initBuffer
(
device
:
device
,
precision
:
computePrecision
)
param
.
y
.
initBuffer
(
device
:
device
,
precision
:
computePrecision
)
...
...
metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvAddMetal.metal
浏览文件 @
538ff20a
...
...
@@ -93,15 +93,6 @@ kernel void conv_add_3x3(texture2d_array<float, access::sample> inTexture [[text
float4 input[9];
for (uint i = 0; i < input_arr_size; ++i) {
// input[0] = inTexture.sample(sample, float2(posInInput.x - 1, posInInput.y - 1), i);
// input[1] = inTexture.sample(sample, float2(posInInput.x, posInInput.y - 1), i);
// input[2] = inTexture.sample(sample, float2(posInInput.x + 1, posInInput.y - 1), i);
// input[3] = inTexture.sample(sample, float2(posInInput.x - 1, posInInput.y), i);
// input[4] = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
// input[5] = inTexture.sample(sample, float2(posInInput.x + 1, posInInput.y), i);
// input[6] = inTexture.sample(sample, float2(posInInput.x - 1, posInInput.y + 1), i);
// input[7] = inTexture.sample(sample, float2(posInInput.x, posInInput.y + 1), i);
// input[8] = inTexture.sample(sample, float2(posInInput.x + 1, posInInput.y + 1), i);
input[0] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y - dilation_y), i);
input[1] = inTexture.sample(sample, float2(posInInput.x, posInInput.y - dilation_y), i);
...
...
@@ -138,8 +129,7 @@ kernel void conv_add_3x3(texture2d_array<float, access::sample> inTexture [[text
outTexture.write(output, gid.xy, gid.z);
}
kernel void test_conv_add_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
kernel void conv_add_5x1(texture2d_array<float, access::sample> inTexture [[texture(0)]],
texture2d_array<float, access::write> outTexture [[texture(1)]],
constant MetalConvParam ¶m [[buffer(0)]],
const device float4 *weights [[buffer(1)]],
...
...
@@ -152,14 +142,12 @@ kernel void test_conv_add_3x3(texture2d_array<float, access::sample> inTexture [
return;
}
if (gid.x > 0 || gid.y > 0 || gid.z > 0) { return; }
ushort2 stride = ushort2(param.strideX, param.strideY);
const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
const uint kernelHXW =
9
;
const uint kernelHXW =
5
;
uint input_arr_size = inTexture.get_array_size();
...
...
@@ -167,32 +155,21 @@ kernel void test_conv_add_3x3(texture2d_array<float, access::sample> inTexture [
float4 output = float4(0.0);
ushort dilation_x = param.dilationX;
ushort dilation_y = param.dilationY;
float4 input[9];
float4 input[5];
for (uint i = 0; i < input_arr_size; ++i) {
input[0] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y - dilation_y), i);
input[0] = inTexture.sample(sample, float2(posInInput.x, posInInput.y - 2 * dilation_y), i);
input[1] = inTexture.sample(sample, float2(posInInput.x, posInInput.y - dilation_y), i);
input[2] = inTexture.sample(sample, float2(posInInput.x
+ dilation_x, posInInput.y - dilation_
y), i);
input[2] = inTexture.sample(sample, float2(posInInput.x
, posInInput.
y), i);
input[3] = inTexture.sample(sample, float2(posInInput.x
- dilation_x, posInInput.
y), i);
input[3] = inTexture.sample(sample, float2(posInInput.x
, posInInput.y + dilation_
y), i);
input[4] = inTexture.sample(sample, float2(posInInput.x,
posInInput.
y), i);
input[4] = inTexture.sample(sample, float2(posInInput.x,
posInInput.y + 2 * dilation_
y), i);
input[5] = inTexture.sample(sample, float2(posInInput.x + dilation_x, posInInput.y), i);
input[6] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y + dilation_y), i);
input[7] = inTexture.sample(sample, float2(posInInput.x, posInInput.y + dilation_y), i);
input[8] = inTexture.sample(sample, float2(posInInput.x + dilation_x, posInInput.y + dilation_y), i);
for (int j = 0; j < 9; ++j) {
for (int j = 0; j < 5; ++j) {
float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
output.x += dot(input[j], weight_x);
...
...
@@ -206,10 +183,11 @@ kernel void test_conv_add_3x3(texture2d_array<float, access::sample> inTexture [
output.w += dot(input[j], weight_w);
}
}
// output = output + biase[gid.z];
outTexture.write(output, gid.xy, gid.z);
}
kernel void depthwise_conv_add_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
texture2d_array<float, access::write> outTexture [[texture(1)]],
constant MetalConvParam ¶m [[buffer(0)]],
...
...
@@ -390,3 +368,78 @@ kernel void depthwise_conv_add_3x3_half(texture2d_array<half, access::sample> in
output = output + float4(biase[gid.z]);
outTexture.write(half4(output), gid.xy, gid.z);
}
kernel void test_conv_add_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
texture2d_array<float, access::write> outTexture [[texture(1)]],
constant MetalConvParam ¶m [[buffer(0)]],
const device float4 *weights [[buffer(1)]],
const device float4 *biase [[buffer(2)]],
uint3 gid [[thread_position_in_grid]]) {
if (gid.x >= outTexture.get_width() ||
gid.y >= outTexture.get_height() ||
gid.z >= outTexture.get_array_size()) {
return;
}
if (gid.x > 0 || gid.y > 0 || gid.z > 0) { return; }
ushort2 stride = ushort2(param.strideX, param.strideY);
const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
const uint kernelHXW = 9;
uint input_arr_size = inTexture.get_array_size();
uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
float4 output = float4(0.0);
ushort dilation_x = param.dilationX;
ushort dilation_y = param.dilationY;
float4 input[9];
for (uint i = 0; i < input_arr_size; ++i) {
input[0] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y - dilation_y), i);
input[1] = inTexture.sample(sample, float2(posInInput.x, posInInput.y - dilation_y), i);
input[2] = inTexture.sample(sample, float2(posInInput.x + dilation_x, posInInput.y - dilation_y), i);
input[3] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y), i);
input[4] = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
input[5] = inTexture.sample(sample, float2(posInInput.x + dilation_x, posInInput.y), i);
input[6] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y + dilation_y), i);
input[7] = inTexture.sample(sample, float2(posInInput.x, posInInput.y + dilation_y), i);
input[8] = inTexture.sample(sample, float2(posInInput.x + dilation_x, posInInput.y + dilation_y), i);
for (int j = 0; j < 9; ++j) {
float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
output.x += dot(input[j], weight_x);
float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
output.y += dot(input[j], weight_y);
float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
output.z += dot(input[j], weight_z);
float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
output.w += dot(input[j], weight_w);
}
}
// output = output + biase[gid.z];
outTexture.write(output, gid.xy, gid.z);
}
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