Chiebot-Mirror
/
opencv
8
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Merge pull request #10672 from pengli:dnn

Browse Source
pull/10692/head
Alexander Alekhin 7 years ago
parent c401168bea 2493083935
commit 8c2563be41
8 changed files with 110 additions and 18 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 12
      modules/dnn/src/dnn.cpp
  2. 9
      modules/dnn/src/layers/batch_norm_layer.cpp
  3. 10
      modules/dnn/src/layers/convolution_layer.cpp
  4. 58
      modules/dnn/src/layers/mvn_layer.cpp
  5. 4
      modules/dnn/src/ocl4dnn/include/ocl4dnn.hpp
  6. 14
      modules/dnn/src/ocl4dnn/src/ocl4dnn_conv_spatial.cpp
  7. 3
      modules/dnn/src/opencl/conv_layer_spatial.cl
  8. 18
      modules/dnn/src/opencl/mvn.cl

12
modules/dnn/src/dnn.cpp
Unescape View File

@ -1190,7 +1190,8 @@ struct Net::Impl
// TODO: OpenCL target support more fusion styles.
if ( preferableTarget == DNN_TARGET_OPENCL &&
(!cv::ocl::useOpenCL() || ld.layerInstance->type.compare("Convolution")) )
(!cv::ocl::useOpenCL() || (ld.layerInstance->type != "Convolution" &&
ld.layerInstance->type != "MVN")) )
continue;
Ptr<Layer>& currLayer = ld.layerInstance;
@ -1239,13 +1240,14 @@ struct Net::Impl
}
}
// For now, OpenCL target only support fusion with activation of ReLU/ChannelsPReLU/Power
// For now, OpenCL target support fusion with activation of ReLU/ChannelsPReLU/Power/Tanh
if ( preferableTarget != DNN_TARGET_OPENCL ||
(preferableTarget == DNN_TARGET_OPENCL &&
nextData &&
(!nextData->type.compare("ReLU") ||
!nextData->type.compare("ChannelsPReLU") ||
!nextData->type.compare("Power"))) )
((nextData->type == "ReLU") ||
(nextData->type == "ChannelsPReLU") ||
(nextData->type == "TanH") ||
(nextData->type == "Power"))) )
{
Ptr<ActivationLayer> nextActivLayer;

9
modules/dnn/src/layers/batch_norm_layer.cpp
Unescape View File

@ -81,9 +81,6 @@ public:
dstWeightsData[i] = w;
dstBiasData[i] = (hasBias ? biasData[i] : 0.0f) - w * meanData[i] * varMeanScale;
}
umat_weight = weights_.getUMat(ACCESS_READ);
umat_bias = bias_.getUMat(ACCESS_READ);
}
void getScaleShift(Mat& scale, Mat& shift) const
@ -119,6 +116,12 @@ public:
CV_Assert(blobs.size() >= 2);
CV_Assert(inputs.size() == 1);
if (umat_weight.empty())
{
umat_weight = weights_.getUMat(ACCESS_READ);
umat_bias = bias_.getUMat(ACCESS_READ);
}
UMat &inpBlob = inputs[0];
CV_Assert(inpBlob.dims == 2 || inpBlob.dims == 4);
int groups = inpBlob.size[0];

10
modules/dnn/src/layers/convolution_layer.cpp
Unescape View File

@ -246,6 +246,11 @@ public:
power = activ_power->power;
activType = OCL4DNN_CONV_FUSED_ACTIV_POWER;
}
Ptr<TanHLayer> activ_tanh = activ.dynamicCast<TanHLayer>();
if (!activ_tanh.empty())
{
activType = OCL4DNN_CONV_FUSED_ACTIV_TANH;
}
}
#endif
return !activ.empty();
@ -877,11 +882,16 @@ public:
{
convolutionOp->setActivPower(true, power);
}
else if ( activType == OCL4DNN_CONV_FUSED_ACTIV_TANH)
{
convolutionOp->setActivTanh(true);
}
else
{
convolutionOp->setActivReLU(false, 0);
convolutionOp->setActivPReLU(false, reluslope);
convolutionOp->setActivPower(false, 1.f);
convolutionOp->setActivTanh(false);
}
newActiv = false;
}

58
modules/dnn/src/layers/mvn_layer.cpp
Unescape View File

@ -60,6 +60,36 @@ public:
normVariance = params.get<bool>("normalize_variance", true);
acrossChannels = params.get<bool>("across_channels", false);
eps = params.get<double>("eps", 1e-9);
fuse_batch_norm = false;
fuse_relu = false;
relu_slope = 0.f;
}
Ptr<BatchNormLayer> bnorm;
Mat scale, shift;
UMat bnorm_weight, bnorm_bias;
bool fuse_batch_norm;
bool setBatchNorm(const Ptr<BatchNormLayer>& layer )
{
bnorm = layer;
fuse_batch_norm = !bnorm.empty() && (preferableTarget == DNN_TARGET_OPENCL);
return fuse_batch_norm;
}
Ptr<ReLULayer> activ_relu;
float relu_slope;
bool fuse_relu;
bool setActivation(const Ptr<ActivationLayer>& layer)
{
if (!layer.empty() && preferableTarget == DNN_TARGET_OPENCL)
{
activ_relu = layer.dynamicCast<ReLULayer>();
if( !activ_relu.empty() )
relu_slope = activ_relu->negativeSlope;
}
fuse_relu = !activ_relu.empty();
return fuse_relu;
}
#ifdef HAVE_OPENCL
@ -71,19 +101,24 @@ public:
inputs_.getUMatVector(inputs);
outputs_.getUMatVector(outputs);
if( fuse_batch_norm && scale.empty())
{
bnorm->getScaleShift(scale, shift);
bnorm_weight = scale.getUMat(ACCESS_READ);
bnorm_bias = shift.getUMat(ACCESS_READ);
}
for (size_t inpIdx = 0; inpIdx < inputs.size(); inpIdx++)
{
UMat &inpBlob = inputs[inpIdx];
UMat &outBlob = outputs[inpIdx];
UMat &inpMat = inputs[inpIdx];
UMat &outMat = outputs[inpIdx];
int splitDim = (acrossChannels) ? 1 : 2;
int i, newRows = 1;
for( i = 0; i < splitDim; i++ )
newRows *= inpBlob.size[i];
newRows *= inpMat.size[i];
MatShape s = shape(newRows, inpBlob.total() / newRows);
UMat& inpMat = inpBlob;
UMat& outMat = outBlob;
MatShape s = shape(newRows, inpMat.total() / newRows);
UMat oneMat = UMat::ones(s[1], 1, CV_32F);
UMat meanMat = UMat(s[0], 1, CV_32F);
UMat devMat = UMat(s[0], 1, CV_32F);
@ -121,8 +156,9 @@ public:
}
String kname = format("mvn%d", number);
if (normVariance)
buildopt += "-DNORM_VARIANCE";
buildopt += format("%s %s %s ", (normVariance) ? "-DNORM_VARIANCE" : "",
(fuse_batch_norm) ? "-DFUSE_BATCH_NORM" : "",
(fuse_relu) ? "-DFUSE_RELU" : "");
ocl::Kernel kernel1(kname.c_str(), ocl::dnn::mvn_oclsrc, buildopt);
if (kernel1.empty())
return false;
@ -132,7 +168,11 @@ public:
kernel1.set(3, (float)eps);
kernel1.set(4, ocl::KernelArg::PtrReadOnly(meanMat));
kernel1.set(5, ocl::KernelArg::PtrReadOnly(devMat));
kernel1.set(6, ocl::KernelArg::PtrWriteOnly(outMat));
kernel1.set(6, ocl::KernelArg::PtrReadOnly(bnorm_weight));
kernel1.set(7, ocl::KernelArg::PtrReadOnly(bnorm_bias));
kernel1.set(8, (int)inpMat.size[1]);
kernel1.set(9, (float)relu_slope);
kernel1.set(10, ocl::KernelArg::PtrWriteOnly(outMat));
ret = kernel1.run(2, global, NULL, false);
if (!ret)
return false;

4
modules/dnn/src/ocl4dnn/include/ocl4dnn.hpp
Unescape View File

@ -77,7 +77,8 @@ typedef enum {
OCL4DNN_CONV_FUSED_ACTIV_NONE = 0,
OCL4DNN_CONV_FUSED_ACTIV_RELU = 1,
OCL4DNN_CONV_FUSED_ACTIV_PRELU = 2,
OCL4DNN_CONV_FUSED_ACTIV_POWER = 3
OCL4DNN_CONV_FUSED_ACTIV_POWER = 3,
OCL4DNN_CONV_FUSED_ACTIV_TANH = 4
} ocl4dnnFusedActiv_t;
template<typename Dtype>
@ -94,6 +95,7 @@ class OCL4DNNConvSpatial
void setActivReLU(bool fuse_activ, float slope);
void setActivPReLU(bool fuse_activ, std::vector<float> &slope);
void setActivPower(bool fuse_activ, float power);
void setActivTanh(bool fuse_activ);
void setBias(bool bias_term);
private:

14
modules/dnn/src/ocl4dnn/src/ocl4dnn_conv_spatial.cpp
Unescape View File

@ -159,6 +159,9 @@ void OCL4DNNConvSpatial<Dtype>::setFusionDefine(ocl4dnnFusedActiv_t fused_activ,
case OCL4DNN_CONV_FUSED_ACTIV_POWER:
addDef("FUSED_CONV_POWER", 1);
break;
case OCL4DNN_CONV_FUSED_ACTIV_TANH:
addDef("FUSED_CONV_TANH", 1);
break;
default:
;
}
@ -415,6 +418,17 @@ void OCL4DNNConvSpatial<Dtype>::setActivPower(bool fuse_activ, float power)
fused_activ_ = OCL4DNN_CONV_FUSED_ACTIV_NONE;
}
template<typename Dtype>
void OCL4DNNConvSpatial<Dtype>::setActivTanh(bool fuse_activ)
{
if ( fuse_activ )
{
fused_activ_ = OCL4DNN_CONV_FUSED_ACTIV_TANH;
}
else
fused_activ_ = OCL4DNN_CONV_FUSED_ACTIV_NONE;
}
template<typename Dtype>
bool OCL4DNNConvSpatial<Dtype>::Forward(const UMat& bottom,
const UMat& bottom2,

3
modules/dnn/src/opencl/conv_layer_spatial.cl
Unescape View File

@ -55,6 +55,9 @@
#elif defined(FUSED_CONV_POWER)
#define ACTIVATION_RELU_FUNCTION(x, c) pow(x, power)
#define NEGATIVE_SLOPE_ARG Dtype power,
#elif defined(FUSED_CONV_TANH)
#define ACTIVATION_RELU_FUNCTION(x, c) tanh(x)
#define NEGATIVE_SLOPE_ARG
#else
#define ACTIVATION_RELU_FUNCTION(x, c) (x)
#define NEGATIVE_SLOPE_ARG

18
modules/dnn/src/opencl/mvn.cl
Unescape View File

@ -89,6 +89,10 @@ __kernel void MVN(__global const Dtype* src,
const Dtype eps,
__global const Dtype* mean,
__global const Dtype* dev,
__global const Dtype* bnorm_weight,
__global const Dtype* bnorm_bias,
const int channels,
const float relu_slope,
__global Dtype* dst)
{
int x = get_global_id(0);
@ -106,7 +110,21 @@ __kernel void MVN(__global const Dtype* src,
#else
alpha = 1;
#endif
Dtype w = 1.f, b = 0.f;
#ifdef FUSE_BATCH_NORM
w = bnorm_weight[x % channels];
b = bnorm_bias[x % channels];
#endif
vec_type src_vec = load(src, index) - (vec_type)mean_val;
vec_type dst_vec = src_vec * alpha;
dst_vec = dst_vec * w + (vec_type)b;
#ifdef FUSE_RELU
vec_type new_val = dst_vec * relu_slope;
dst_vec = select(new_val, dst_vec, dst_vec > (vec_type)0.f);
#endif
store(dst_vec, dst, index);
}

Write Preview
Loading…
Cancel
Save